KAI: An AI-powered Chatbot To

Support Therapy
Bachelor Thesis Project
Specilization in Computer Science

Mariama C. Djalo D.

Date: 23/01/2023
Director: Javier Béjar Alonso
Department: Computer Science
Degree: Bachelor of Computer Science
Center: FACULTAT D’INFORMÀTICA DE BARCELONA (FIB)
University: UNIVERSITAT POLITÈCNICA DE CATALUNYA (UPC) – BarcelonaTech
Contents
1 Context 8
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2 Cognitive Behavioral Therapy . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2.1 Self-Directed Therapy . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2.2 Cognitive Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2.3 Cognitive Distortions . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3 Artificial Intelligence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.3.1 NLP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.3.2 Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.3.3 Text Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.4 Justification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.5 Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.5.1 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.6 Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.7 Risks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.8 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.9 Stakeholders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2 Project Planning 18
2.1 Description of tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.1.1 Project Management . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.1.2 Project Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.1.3 Project Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.1.4 Data Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.1.5 Project Experimentation . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.1.6 Project Development . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.1.7 Project Documentation . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.1.8 Thesis Defense Preparation . . . . . . . . . . . . . . . . . . . . . . . 22
2.2 Risk management: alternative plans . . . . . . . . . . . . . . . . . . . . . . . 22
2.3 Budget . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.3.1 Costs Per Activity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.3.2 Generic Costs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.3.3 Contingency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.3.4 Incidental Costs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.3.5 Management control . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.4 Deviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.4.1 Deviations in the project development . . . . . . . . . . . . . . . . . 29
2.4.2 Deviations in the project documentation . . . . . . . . . . . . . . . . 30
2.4.3 Deviations in the budget . . . . . . . . . . . . . . . . . . . . . . . . . 30

1
3 Identification of Laws and Regulations 31
3.1 Academic Regulations for the Degree Final Project . . . . . . . . . . . . . . 31
3.2 GDPR Privacy Policy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2.1 Company’s contact details . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2.2 The Purposes and Legal Basis for Processing . . . . . . . . . . . . . . 32
3.2.3 Sharing of user’s personal data . . . . . . . . . . . . . . . . . . . . . 32
3.2.4 Sharing of user’s personal data to a third country . . . . . . . . . . . 32
3.2.5 Period of time storage of user’s personal data . . . . . . . . . . . . . 33
3.2.6 User’s Rights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.3 The EU Regulatory Environment of Medical Device Software Development . 33

4 Sustainability report 35
4.1 Self assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.2 Environmental dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.3 Economic dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.4 Social dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

5 Technical Competences 37

6 Dialogue Flow 38
6.1 Identifying Potential Cognitive Distortions . . . . . . . . . . . . . . . . . . . 38
6.2 Challenging Potential Cognitive Distortions . . . . . . . . . . . . . . . . . . 38
6.3 Chatbot’s Feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
6.3.1 Relaxation Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

7 Dataset 40
7.1 Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
7.1.1 CountVectorizer and TfidfTransformer . . . . . . . . . . . . . . . . . 40
7.1.2 Balancing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

8 Model Experimentation 42
8.1 Multinomial Naive Bayes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
8.1.1 alpha Hyperparameter . . . . . . . . . . . . . . . . . . . . . . . . . . 42
8.2 Multinomial Logistic Regression . . . . . . . . . . . . . . . . . . . . . . . . . 43
8.2.1 penalty Hyperparameter . . . . . . . . . . . . . . . . . . . . . . . . . 44
8.2.2 C Hyperparameter . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
8.2.3 solver Hyperparameter . . . . . . . . . . . . . . . . . . . . . . . . . . 44
8.3 Support Vector Machine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
8.3.1 decision function shape Hyperparameter . . . . . . . . . . . . . . . . 45
8.4 K-Nearest Neighbour . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
8.4.1 n neighbors Hyperparameter . . . . . . . . . . . . . . . . . . . . . . . 46
8.4.2 weights Hyperparameter . . . . . . . . . . . . . . . . . . . . . . . . . 46
8.4.3 p Hyperparameter . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
8.5 Random Forest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
8.6 Model Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

2
 8.6.1 Hyperparameters Tuning . . . . . . . . . . . . . . . . . . . . . . . . . 47
8.6.2 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

9 Implementation 50
9.1 Model Experimentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
9.2 Chatbot Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
9.2.1 Rule-Based Chatbot . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
9.3 Dialogue example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
9.3.1 Dialogue Flow Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
9.3.2 Rule-Based Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

10 Conclusions 54
10.1 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
10.2 Reflexions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

A 15 Major Cognitive Distortions 56

B Dialogue Flow 57

C EU Regulatory Environment for MDSW Development 58

D Standard and guidance documents useful to demonstrate MDSW compli-

ance with MDR 59

E Number of examples grouped by cognitive distortions 60

F Support Vector Machine Hyperplanes 60

List of Figures
1 A drawing that depicts how a situation causes feelings, from the book A
Therapist’s Guide to Brief Cognitive Behavioral Therapy [1]. . . . . . . . . . 10
2 Diagram illustrating the structure of the Cognitive Model, from the book A
Therapist’s Guide to Brief Cognitive Behavioral Therapy [1]. . . . . . . . . . 10
3 A diagram showing the different subfields of Artificial Intelligence, from Eu-
ropeanValley [2], a website. . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
4 ML model workflow overview, from Google Machine Learning Education [3],
a website. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
5 Gantt Chart illustrating the project’s schedule following a Waterfall model.
[Own Creation] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
6 Results of the Hyperparameters Tuning [Own Creation]. . . . . . . . . . . . 48
7 Model Experimentation’s Results [Own Creation] . . . . . . . . . . . . . . . 49
8 Screenshot showing the first part of the Dialogue flow [Own Creation] . . . . 52
9 Screenshot showing the second part of the Dialogue flow [Own Creation] . . 52
10 Screenshot showing how the chatbot correctly detects the the keywords and
responds accordingly [Own Creation] . . . . . . . . . . . . . . . . . . . . . . 53

3
 11 15 major cognitive distortions by PositivePyschology, a website [4] . . . . . . 56
12 Chatbot’s Dialogue Flow [Own Creation] . . . . . . . . . . . . . . . . . . . . 57
13 EU Regulatory Environment for MDSW Development. [5] . . . . . . . . . . 58
14 Number of examples grouped by cognitive distortions [Own Creation] . . . . 60
15 There are numerous alternative hyperplanes that might be used to divide the
two groups of data points (left image). Finding a plane with the greatest mar-
gin—that is, the greatest separation between data points from both classes—is
our goal (right image) in SVM. Images obtained from Towards Data Science
[6], a website. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

List of Tables
1 Task Table containing a summary of all task information. T and GEPT means
Tutor and GEP Tutor respectively. [Own Creation] . . . . . . . . . . . . . . 24
2 Salary of the different roles extracted from PayScale, a compensation software
company [7] multiplied by 1.35 to include the cost of social security [Own
creation]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3 Budget Structure of the project [Own creation] . . . . . . . . . . . . . . . . . 28
4 Final version of the task table [Own creation] . . . . . . . . . . . . . . . . . 29
5 Final version of the budget structure [Own creation] . . . . . . . . . . . . . . 30
6 Standard and guidance documents useful to demonstrate MDSW compliance
with MDR. [5] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

4
Acknowledgement
First of all, I want to thank my family for supporting me in my worst times and for sharing
with me my best moments. Also, I want to thank my friends for accompanying me through
this journey and for forming part of my good memories from College. Finally, I must express
my gratitude to my director for guiding me during the development of the project, I really
appreciate his support.

5
 Abstract
This project attempts to bridge the huge gap between people who struggle with
mental health and people who actually get treated with the design and implementation
of Kai, an AI-powered Chatbot that supports Cognitive Behavioral Therapy (CBT).
CBT is based on identifying cognitive distortions (negative thoughts) and challeng-
ing them to improve mood and overall mental health. This was done by using Text
Classification, a Natural Language Processing (NLP) technique to identify potential
cognitive distortions in text.
During the project, a model experimentation was done to compare different super-
vised machine learning models in order to choose the best one for the text classification.
The dataset needed to train the models was generated by manually giving labelled ex-
amples of the 15 major cognitive distortions. The model experimentation was entirely
done in python.
Furthermore, the design of the dialogue flow of the chatbot was done following the
CBT’s guidelines and the implementation of the chatbot was done in Python using the
TKinter in Framework for the interface. Finally, two test were made to check for the
correct functioning of the chatbot.

KEYWORDS: Chatbot, CBT, Cognitive Distortions, Text Classification, NLP, Su-

pervised Machine Learning.

Resum

Aquest projecte intenta reduir l’enorme bretxa existent entre les persones que
pateixen problemes de salut mental i les que realment reben tractament amb el disseny
i la implementació de Kai, un xatbot impulsat per IA que dóna suport a la teràpia
cognitivoconductual (TCC).
La TCC es basa en la identificació de les distorsions cognitives (pensaments negatius)
per qüestionar-les amb l’objectiu de millorar l’estat d’ànim i la salut mental. Per això
es va utilitzar la Classificació de Text, una tècnica de Processament del Llenguatge
Natural (PLN) per identificar potencials distorsions cognitives en el text.
Durant el projecte es va fer una experimentació de models per comparar diferents
models supervisats d’aprenentatge automàtic per triar el millor per a la classificació
de textos. El conjunt de dades necessari per entrenar els models es va generar pro-
porcionant manualment exemples etiquetats de les 15 distorsions cognitives principals.
L’experimentació del model es va fer ı́ntegrament en python.
A més, el disseny del flux de diàleg del chatbot es va fer seguint les directrius
del TCC i la implementació del chatbot es va fer a Python utilitzant TKinter, un
Framework per a la interfı́cie. Finalment, es van realitzar dos tests per comprovar el
funcionament correcte del chatbot.

PARAULES CLAU: Chatbot, CBT, Distorsions Cognitives, Classificació de Text,

NLP, Aprenentatge Automàtic Supervisat.

6
 Resumen

Este proyecto intenta reducir la enorme brecha existente entre las personas que
sufren problemas de salud mental y las que realmente reciben tratamiento con el diseño
y la implementación de Kai, un chatbot impulsado por IA que da soporte a la terapia
cognitivo-conductual (TCC).
La TCC se basa en la identificación de las distorsiones cognitivas (pensamientos neg-
ativos) para luego cuestionarlas con el objetivo de mejorar el estado de ánimo y la salud
mental. Para ello se utilizó la Clasificación de Texto, una técnica de Procesamiento
del Lenguaje Natural (PLN) para identificar potenciales distorsiones cognitivas en el
texto.
Durante el proyecto se realizó una experimentación de modelos para comparar dis-
tintos modelos supervisados de aprendizaje automático con el fin de elegir el mejor para
la clasificación de textos. El conjunto de datos necesario para entrenar los modelos se
generó proporcionando manualmente ejemplos etiquetados de las 15 distorsiones cog-
nitivas principales. La experimentación del modelo se realizó ı́ntegramente en python.
Además, el diseño del flujo de diálogo del chatbot se hizo siguiendo las directrices
del TCC y la implementación del chatbot se hizo en Python utilizando TKinter, un
Framework para la interfaz. Por último, se realizaron dos tests para comprobar el
correcto funcionamiento del chatbot.

PALABRAS CLAVE: Chatbot, CBT, Distorsiones Cognitivas, Clasificación de Texto,

NLP, Aprendizaje Automático Supervisado.

7
1 Context
This is a Bachelor Project Thesis done at the Barcelona School of Informatics (FIB) under
the supervision of Javier Béjar Alonso for the Computer Science Degree with a specialization
in Computing.

1.1 Introduction
According to WHO [8], depression is a leading cause of disability and among people aged
15 to 29, suicide is the fourth most common cause of death. There is still a substantial
disparity between those who require care and those who have access to care, despite the fact
that many mental health illnesses may be adequately treated at low cost [9].
Therapeutic and mental health Chatbots are an intriguing way to bridge this gap. A
chatbot is a computer system tool that is designed to simulate human communication.
On the other hand, it is widely known that a few sessions of cognitive-behavioral therapy
(CBT) can be extremely beneficial in the treatment of anxiety and depression. However,
many people do not have access to a CBT therapist because they cannot afford it, it is not
covered by their insurance, or there are none nearby. It can also be difficult to go to therapy
due to lack of time because of work or having to take care of kids for instance. However, in
some cases, a therapist may not be required. There are numerous options for doing CBT
without a therapist, such as self-help books and internet-based treatment, or by conducting
your own research on the material. Self-directed CBT has been shown in numerous studies
to be very effective [10]. This particular type of therapy promotes independence and self-
therapy. The methodology of CBT is based on identifying cognitive distortions, known
as negative thoughts and challenging them by replacing them for alternative and more
positive thoughts in order to improve mood and the overall mental health.
The goal of this project is to combine the two previously mentioned methodologies to help
people dealing with mental health issues. KAI, a Chatbot that provides therapy support in
self-directed CBT, will be designed and implemented as part of the project. KAI will detect
Cognitive Distortions in text using Text Classification, an NLP technique that will assist in
the application of CBT.
This project aims to address the problem of a significant gap that exists between people
who are struggling with mental health issues and need assistance and those who have access
to mental health care. In this project, the design and development of an AI-powered chat-
bot that supports therapy by automatically detecting cognitive distortions is done in order
to achieve this goal. The goal of this project is to train various supervised machine learn-
ing techniques for text classification in order to identify cognitive distortions from chatbot
conversations and use them to aid in self-directed CBT.
The project is aimed at people who meet all of the following conditions:

• People who have mild to moderate symptoms of mental health issues and can function
normally. 1
1
Those who are severely depressed or have severe mental health problems will most likely require one-on-
one therapy with a professional

8
 • People who do not have access to or prefer not to speak with a therapist due to privacy
concerns and/or a fear of being judged as a result of the stigma associated with mental
health issues.

• People who prefer to be autonomous and prefer to learn and treat themselves with
self-directed therapy.

Bear in mind that the chatbot is not a therapist and thus does not provide
mental health diagnosis; rather, the chatbot is a tool that assists with the automatic
detection of potential cognitive distortions and guides you through the process of identifying
and challenging them. In other words, KAI is a self-directed AI-based CBT tool that assists
people in learning how to deal with cognitive distortions.

1.2 Cognitive Behavioral Therapy

One of the core concepts of the project is cognitive-behavioral therapy (CBT), thus, in this
section is going ot be explained.
The goal of cognitive-behavioral therapy (CBT), a type of psychotherapy, is to help
patients understand how their thoughts, beliefs, and attitudes affect how they feel and
behave. The foundation of cognitive behavioral therapy (CBT) is the premise that our
thoughts, feelings, and behaviors are connected, and that by altering our thoughts and
beliefs, we can alter our behaviors and emotions.

1.2.1 Self-Directed Therapy

With little to no supervision from a therapist, patients can work independently during self-
directed cognitive-behavioral therapy (CBT) to identify and change the harmful thoughts
(cognitive distortions) and behaviors that contribute to their emotional and behavioral prob-
lems. Self-directed CBT may be a helpful treatment option for a patient who chooses to
work independently or does not have access to a therapist.
Self-help tools like books, websites, or apps are widely used in self-directed CBT to guide
the patient through the process of identifying and altering harmful thinking and behavior
patterns.

1.2.2 Cognitive Model

According to Cully et al. [1] ”The cognitive model is a theoretical paradigm for ex-
plaining how thoughts, feelings, and behaviors are associated . Most individuals
believe that situations give rise to their emotions” as shown in Figure 1. ”The cog-
nitive model challenges this subjective experience and suggests, instead, that it
is the thoughts we have about situations that give rise to emotions” as illustrated
in Figure 2.

9
Figure 1: A drawing that depicts how a situation causes feelings, from the book A Therapist’s
Guide to Brief Cognitive Behavioral Therapy [1].

Figure 2: Diagram illustrating the structure of the Cognitive Model, from the book A Ther-
apist’s Guide to Brief Cognitive Behavioral Therapy [1].

1.2.3 Cognitive Distortions

Cognitive distortions are inaccurate, incorrect, or distorted ways of thinking that can have
a negative impact on feelings and actions. These distortions alter how we perceive and
understand events, which may lead us to form inaccurate or erroneous assumptions that
have harmful consequences.
Typical examples of cognitive distortions are as follows:
• Jumping to conclusions: Drawing judgments without enough evidence. For exam-
ple, ”She’s not responding so she must be ignoring me or even worse, she must be mad
at me”
• Mental filter: Focusing only on the negative part of everything. For example, ”I
did well on most of the exam, but I got one question wrong, so I must be a complete
failure.”
• Emotional reasoning: Thinking that because you feel something it must be real.
For example, ”I feel like a failure, so I must be a failure.”

10
1.3 Artificial Intelligence
Artificial intelligence (AI) is another important concept of the project and will be explained in
this section. Artificial intelligence (AI) are computer systems that mimic human intelligence,
for instance, learning, problem-solving, decision-making and understanding languages. AI
applications include face recognition, chatbots, and self-driving cars.

1.3.1 NLP
Natural language processing (NLP), as illustrated in Figure 3, is a subfield of artificial
intelligence. NLP focuses on teaching computers to comprehend text and spoken language
in the same way that humans do.

Figure 3: A diagram showing the different subfields of Artificial Intelligence, from European-
Valley [2], a website.

1.3.2 Machine Learning

Machine Learning, as shown in Figure 3, is another subfield of AI based on the idea that
computers are capable of learning from data, recognizing patterns and making decisions with
little to no human intervention.
Machine learning algorithms are used to build models that can make decisions without
being coded to perform an specific task, in other words, autonomous systems. These al-
gorithms are able to learn from data, and improve over time, by identifying patterns and
relationships within the data.
Supervised learning and unsupervised learning are the two main types of machine learn-
ing. In supervised learning, the model is trained on labeled data, where the correct output
is provided for each example in the training set. The goal is to learn a function that can
predict the output for a new input.

11
 In unsupervised learning, the model is not given labeled training examples; instead, it
must use methods like clustering to determine the underlying structure of the data.

Features and Feature Selection

In machine learning, features refer to the input variables or characteristics that are used
to describe and predict the output or target variable. Feature selection is the process of
selecting a subset of the most relevant and informative features for building a model.

Overfitting in Machine learning

Overfitting occurs when a model learns the information and noise in the training data
to the extent that it negatively affects the model’s performance on new data. This indicates
that the machine learns concepts from the noise or random fluctuations in the training data.
These concepts don’t apply to new data, which poses a difficulty for the models’ capacity to
generalize.
Nonparametric and nonlinear models, which have more flexibility while learning a target
function, are more susceptible to overfitting. As a result, a lot of nonparametric machine
learning algorithms additionally incorporate parameters or methods to restrict and limit the
amount of information the model may learn.

High Variance

In machine learning, ”high variance” refers to a model’s propensity to have a significant

difference between its training error and its testing error. This can occur when the model is
overly sensitive to the specific details of the training data and is not able to generalize well
to unseen data.
High variance might be problematic since it indicates that the model will probably per-
form badly when applied to new, unseen data. Additionally, it can be a sign of overfitting,
in which case the model learns the random fluctuations and noise seen in the training data
rather than the underlying pattern.

Bias

In machine learning, bias is the systematic discrimination or inaccuracy that appears in

a model’s predictions. A number of factors, such as the use of biased training data or the
model’s construction itself, can lead to bias.
Because bias can result in unfair or erroneous predictions that might reinforce or magnify
already-existing societal inequalities, bias can have major negative effects. For instance, a
biased employment model might disproportionately reject specific minority groups, while a
biased lending model might unfairly refuse loans to particular borrowers.
To make sure that the model’s predictions are accurate and fair, bias in machine learning
must be carefully considered and addressed. This may require using a varied and represen-
tative training dataset, minimizing bias in the model using strategies like regularization, and
regularly evaluating the model’s performance on different subgroups.

12
 Data Crowdsourcing

Data crowdsourcing is the process of gathering data from a big group of people, fre-
quently via an internet platform. This may be a useful method for quickly and inexpensively
gathering a lot of data.

1.3.3 Text Classification

Text Classification is a supervised machine learning technique in NLP used to classify text
into different categories. Common types of Text Classification are Sentiment Analysis and
Spam Email detector.
As shown in Figure 4, the Text Classification Workflow begins with gathering data (ob-
taining a dataset) and continues with data exploration using descriptive analytics. The next
step is to prepare the data. Since I’ll be working with texts and ML models don’t under-
stand them, I’ll need to do Text Vectorisation, which is the process of converting text into
numerical representation.
Following that, the model is built, trained with the dataset, and evaluated to see how
well it performs. In my case, I’ll train several models before selecting the best one.
Finally, after Hyperparameters Tuning for improved performance, the model is deployed.
Hyperparameters are the parameters that define the model architecture, and Hyperparameter
Tuning is the process of searching for the best model architecture. The maximum depth is
an example of a Hyperparamter in decision trees (a ML Model).

Figure 4: ML model workflow overview, from Google Machine Learning Education [3], a
website.

1.4 Justification
The use of AI in mental healthcare and psychiatry, particularly therapeutic chatbots, is still
in its early stages, with limited research data and datasets available to fully explore the
field’s true potential.
Shickel et al. [11] published the most relevant research in 2019. The authors used su-
pervised machine learning techniques such as SVM, XGBoost, and RNN to find cognitive
distortions in text automatically. The authors obtained a weighted F1 scores of 0.68.
There are currently no datasets containing labeled text sections with distortions that
are available to the general public because the detection of cognitive distortions is a novel

13
machine learning task. They gathered information using a real-world online therapy program
and crowdsourcing in order to resolve this difficult problem.
Toledo et al. [12], on the other hand, took a different and very innovative approach:
they developed cognitive distortion responses to support CBT interactions. Which is ex-
tremely important because Cognitive Behavioural Therapy’s (CBT’s) core idea is the ability
to change distorted or negative beliefs (cognitive distortions) into more realistic alternatives
(positive thoughts). The authors used Transformers learners to generate the responses.
It’s also worth mentioning that there are self-help apps and chatbots for mental health,
many of which use CBT, on the market, including in Spain2 . Wysa and Woebot are the
most well-known.
As mentioned before, there are currently no publicly available datasets. Crouwdsourcing
and data generation are two options for resolving this challenging issue. Since Crowdsourcing
is a paid service, I will opt for the second option, in other words: I will create the dataset
myself. The process will be completed by providing sufficient examples of the 15 major
cognitive distortions. To ensure that the data is accurate, examples will be drawn or be
inspired from psychological books, websites, and articles specializing in psychology.
Unfortunately there is no public information available on how to design a CBT chatbot.
Therefore, the design will be done from scratch with the aid of a very helpful and introductory
manual book to CBT [1] that will help have a better understanding about the structure of
cognitive behavioural therapy session.

1.5 Scope
The project’s objectives, functional requirements, and potential risks and obstacles are all
discussed in this section.

1.5.1 Objectives
The main goal of this project is to design and implement an AI-powered Chatbot that
supports therapy to assist people suffering from mental health issues such as anxiety or de-
pression. Kai will use CBT technique and with the help of AI he will automatically detect
cognitive distortions. To achieve this goal, the project has been divided into several sub-
objective:

Theoretical part

1. Investigate CBT to gain a thorough understanding of how it works.

2. Design the chatbot’s conversation structure and dialogue flow.

3. Investigate the best supervised machine learning techniques for text classification.

2
This fact confirms and clarifies that it is perfectly legal to make support therapy chatbots

14
 Practical Part

1. Generate a suitable dataset to train the machines.

2. Train different supervised machine learning models.

3. With the help of metrics, choose the best model.

4. Implement the chatbot.

1.6 Requirements
There are some requirements that must be met in order to ensure the final project’s quality:

• Data pre-processing. Data pre-processing is essential for ensuring that the models
perform optimally.

• Hyperparameter Tuning. Understanding how different algorithms work is essential

for knowing how to make good Hyperparameter Tuning decisions to achieve the best
results.

• Optimization of the code. Optimize the code for all imputation methods as well to
improve efficiency.

• Avoid bias. Avoiding bias is key to ensure that the models are accurate.

• Good programming practices. Using good programming practices such as a read-

able style, good comments, and as little complexity as possible.

1.7 Risks
There may be some risks that prevent the project from progressing smoothly:

• Not being able to generate an appropriate or a representative dataset to

train models. The main risk would be not being able to find/generate a suitable
(without bias) dataset to train the models, which would prevent the project from
progressing.

• Deadline of the project. The project also has a deadline for completion that must
be met. This forces to make difficult decisions. As a result, strong organizational skills
and the ability to meet deadlines are essential for finishing the project on time. Some
libraries, on the other hand, have flaws.

• Bugs in libraries. Some libraries may have bugs in certain functions, resulting in
incorrect code.

15
1.8 Methodology
In order to have more flexibility, I will use a hybrid of Waterfall and Agile workflow method-
ology for this project. Waterfall is a project management methodology that is based on
a sequential design process. Agile is a methodology that prioritizes development through
evolution. This method enables sprint work and the resolution of issues that arise during
iterations. The Agile-Waterfall hybrid method combines the best features of both methods:
Agile allows you to check for bugs, test the code, and correct it in a progressive way without
having to wait until the entire implementation is completed. Waterfall, on the other hand,
allows you to keep track of all the dependencies between tasks to better organize the project.
The Kanban framework which falls under the Agile methodology will be used. Since the
1950s, the Japanese phrase ”kanban,” which means ”visual board” or a ”sign,” has been
used to refer to a process definition. Toyota invented it and used it as the first just-in-time
factory scheduling system. The ”Kanban Method,” which was initially defined in 2007, is
known and connected with the capitalized term ”Kanban,” on the other hand.
Kanban boards are used to efficiently show and control the workflows. The essential
elements are:

- Kanban Cards are used to express tasks visually. Each card contains details on the
work and its progress, including the due date, the person assigned to it, the description,
etc.
- Kanban Columns: On the board, each column corresponds to a distinct step of your
operation. The workflow is applied to the cards till they are finished completely.

In my case, I’ll have four columns.

1. To Do: composed by all of the tasks that haven’t been started yet.
2. In progress: composed by all of the tasks that are still in progress.
3. Tested: composed by all of the tasks that have already been completed but need to
be tested to ensure they work properly.
4. Completed: composed by all completed and tested tasks.

There are many project management tools that follow the kanban methodology, but I’m
going to use Trello because I believe it’s the better option for small teams or one-person
teams, such as freelancers (which is my case).
On the other hand, a Gantt chart with a Waterfall workflow will also be used to keep
track of the dependencies and the required time for each task.
I’ll use a Github repository as a version control tool to make sure I can restore earlier
versions in the event of serious errors because it’s securely kept in the cloud.
I’m planning to use the GitHub Flow as my Git branching technique. Its branches
are organized into a main branch where the code that is ready for production is kept and
additional branches, referred to as feature branches, where work on new features and bug
fixes is done and then merged back into the main branch. Smaller teams, like mine, as I’ve
already mentioned, benefit most from this approach.

16
 For the practical part, the cross-validation method will be used to choose the optimum
Hyper-parameters for the models and also to check their performance and if there is
bias. I’ll also set the random state to an integer to prevent having different outcomes each
time I run the model.
Last but not least, whenever I have queries or run into problems, I shall email or meet
with my tutor online. Extraordinary in-person meetings will be scheduled if I encounter any
serious issues with my project or if I believe that communicating with my tutor in person
will be more convenient.

1.9 Stakeholders
In this section the stakeholders who will benefits from the completion of this project will be
enumerated.
The project’s completion is important not only for people suffering from mental illnesses,
but also for hospitals, clinics, educational institutions, and communities/groups in general.
They could use the chatbot to treat and improve the mental health of patients/community
members, thereby improving people’s overall well-being.

- The project’s completion is specially important for people struggling with mental
health.

- Hospitals, clinics, educational institutions, and communities/groups in general also

benefit from this project. They could use the chatbot to treat and improve the mental
health of patients/community members, thereby improving people’s overall well-being

17
2 Project Planning
The project will last approximately 579 hours spread over 126 days, beginning on September
20th, 2022 and ending on January 23rd, 2022. Since the date for the project defense has
not yet been determined, the previous deadline is the earliest we can have. It is planned to
work an average of 5 hours per day, but some flexibility may be required due to exams or
personal issues.
This section begins with a task description, followed by the resources required for the
project’s development, and finally by an explanation of risk management. Furthermore,
Table 1 summarizes all of the defined tasks, as well as their dependencies and required
resources, and Figure 5 captures the project schedule.

2.1 Description of tasks

The identification and description of the tasks that will be completed during the course of
the project are presented in this section. It is given an estimate of the time needed for each
task in hours, as well as a description of the logical sequence and dependencies between
them. The tasks are divided into the following groups: Project Management, Project
Research, Project Theory, Data Generation, Project Experimentation, Project
Development, Project Documentation and Thesis Defense Preparation.

2.1.1 Project Management

Project management is most likely one of the project’s pillars. It defines the scope of the
project, the tasks and their planning, as well as the budget and sustainability.

• PM1 - ICT tools for project and team management.

- Description: To support the project’s development, the best technology, devices,

and concepts that fit the nature of our project are required. To accomplish this,
it is necessary to conduct research on various types of software for various tasks.
- Resources: PC with internet acces.
- Approximate duration: 1 hour.

• PM2 - Context and Scope.

- Description: The project’s scope, as well as its contextualization, are defined.

The general goal of the project, its justification, developments, and tools are all
discussed in this section.
- Resources: This part requires a PC with internet connection, Overleaf to docu-
ment and both The GEP 3 Tutor and the Tutor of the project for the feedback.
- Approximate duration 35 hours.
3
GEP is a course that must be passed before submitting the final version of the thesis. It stands for
”Gestió de Projectes” which translates to Project Management.

18
 • PM3 - Time planning.

- Description: It is critical to plan ahead of time to ensure that the deadline is

met. A good plan can help us identify which tasks require more attention than
others and which are critical.
- Resources: Making the Time Planning requires a PC, Overleaf, The GEP Tutor
and the Tutor and TeamGantt to make the Gantt chart.
- Approximate duration: 30 hours.

• PM4 - Budget and sustainability.

- Description: The main objectives of this assignment are to develop a budget

and assess the project’s sustainability. This is important to know in order to
calculate the project’s overall cost and its development impact.
- Resources: This part needs a PC, Overleaf, The GEP Tutor and the Tutor as
resources.
- Approximate duration: 30 hours.

• PM5 - Meetings.

- Description: Meetings with the project’s tutor will be scheduled as needed (E.g.
when doubts or critical problems that impede the proper development of the
project arise. I have added a ”reserved time” for the meetings to make a better
estimation of the total hours required for the project.
- Resources: Tutor.
- Approximate duration: 1 hour a week (in total, 18 hours).

2.1.2 Project Research

This part has been divided into the following tasks :

• PR1 - Psychology research.

- Description: This project’s methodology is based on psychology, specifically

Cognitive Behavioral Therapy. Before embarking on the implementation phase,
thorough research in CBT is essential.
- Resources: PC, Books, Research Papers, Articles.
- Approximate duration: 10 hours.

• PR2 - ML research.

- Description: Documentation on various types of supervised machine learning

models, as well as the statistics behind them, is also highly needed.
- Resources: PC, Books, Research Papers, Articles.
- Approximate duration: 10 hours.

19
2.1.3 Project Theory
In the theoretical part, I will study the structure of a CBT Therapy session in order to
properly design the dialogue flow. Furthermore, I will consider what are the best ML models
for both text classification and small datasets 4 .
This part is divided into the following tasks:

• PT1 - Design.

- Description: Design of the structure and dialogue flow of the chatbot.

- Resources: PC, Books, Research Papers, Articles.
- Approximate duration: 10 hours.

• PT2 - Choose.

- Description: Choose between the top 3-5 supervised ML models for text classi-
fication and small datasets.
- Resources: PC, Books, Research Papers, Articles.
- Approximate duration: 10 hours.

• PT3 - Select the Hyperparameters.

- Description: Select the Hyperparameters of every model to optimize.

- Resources: PC, Books, Research Papers, Articles.
- Approximate duration: 10 hours.

2.1.4 Data Generation

- Description: It is necessary to generate the dataset before beginning the experi-
mentation with the models to determine which is the best. As previously stated, the
generation will be created by using examples or finding inspiration on the internet,
books, and articles specialized in psychology. The dataset will include examples of the
15 major cognitive distortions shown in Figure 11.

- Resources: PC, Books, Research Papers, Articles.

- Approximate duration: 14 hours.

4
Since I will be creating the data, the dataset will most likely be quite small (less than 1000 examples of
cognitive distortions).

20
2.1.5 Project Experimentation
In the experimentation section, the ml model workflow is completed as described in the
previous assignment, and the best model is selected after analyzing the metrics.
In summary, the tasks are divided into the following:

• PE1 - Apply the workflow. This part will require 16 hours.

• PE2 - Hyperparameters tuning. Experiment with every model optimizing them

with the aid Hyperparameters tuning selected in the Project Theory part (thus there
is a dependency). This part will require 8 hours.

• PE3 - Choose the best ml model. Analyse the performance of every model and
choose the best one for the text classification. This part will require 2 hours.

All these tasks will be done in Colaboratory, best known as ”Colab”. This tool is a product
from Google Research. Colab is the best tool for machine learning, data analysis, and
education since it enables anyone to create and execute arbitrary Python code in the browser.
Google Drive is where Colab notebooks are kept, making it a secure cloud storage option.
Furthermore, a PC, books, research papers, articles, programming languages and Github
will be needed.

2.1.6 Project Development

The Project Development takes place after designing the structure and dialogue flow of the
chatbot and consists on the implementation of a telegram chatbot.
The tasks are organized as follows:

• PDEV1 - Implementation.

- Description: Implement the telegram chatbot.

- Resources: PC, Github, Colab and programming languages.
- Approximate duration: 200 hours.

• PDEV2 - Testing.

- Description: Test the correct functioning of the telegram bot. This will be done
during and after the implementation of the apps in order to make sure all parts
work correctly on time.
- Resources: PC, Github, Colab and programming languages.
- Approximate duration: 60 hours.

21
2.1.7 Project Documentation
To avoid having to do everything at the end, the project documentation will be completed
concurrently with the project’s development (after the research part is done). For these
task we need a PC, Overleaf/Texifier and Trello to keep track of the progress. The Project
Documentation has been broken down into the following tasks:

• PDOC1 - Annotation of events: Annotation of all the events that are done during
the project development. This task will be done intermittently and will aproximately
require 10 hours.

• PDOC2 - Revision of the events: Once the project development is done is time
to check all the documentation done during the project to better organize the ideas,
correct changes and structure the final document correctly. This task will require 20
hours.

• PDOC3 - Write Final Document: After PDOC2 is done the writing of the final
documentation begins. This task will require approximately 60 hours.

2.1.8 Thesis Defense Preparation

Finally, after the project documentation is completed, the oral defense preparation begins.
To do so, it is necessary to practice and prepare for potential tribunal questions. This will
require 25 hours.

2.2 Risk management: alternative plans

During the course of the project, difficulties that could jeopardize the project’s proper
progress may arise. All of the previously introduced potential problems in section 1.7 will
be addressed in this section by introducing new tasks and accommodating the planning. A
level of risk is also added to each of them.

• Not being able to generate an appropriate or a representative dataset to

train the models [Extreme Risk]. Bias may be introduced into the data by pro-
viding examples by myself because I am forcing the examples to fit the criteria for
cognitive distortions, which may or may not reflect how cognitive distortions occur in
the real world. CD ”in the wild” may be more subtle or more than one can appear in
a thought/phrase. If that were to occur these are the steps that would be followed:

– Scrapping the internet or obtaining a dataset from a public mental

health/therapy forum. In my case, I would use a public dataset obtained from
Reddit (specifically, from mental health subreddits) [13]. This method would yield
”real-world” examples. Unfortunately, Manual labeling of the dataset into the 15
major cognitive distortions would be required.
– Crowdsourcing. People will give examples that fit the CD criteria, so there
may still be some bias. However, it would be significantly more varied than giving
examples by myself in this case because people from different parts of the world

22
 would give examples (that may or may not apply to their real-life situation) that
would most likely differ from mine due to differences in backgrounds, for example.
– Resources to reuse: PC, programming languages, and TeamGantt.
– Estimated delay: between 1-2 weeks.

• Deadline of the project [High Risk]. It could be caused by an accurate preliminary

estimation of the tasks and their duration, which is completely normal because we do
it before we begin. It is important to plan ahead of time, but it is also important to
be flexible and adapt to changing circumstances. If that were to occur these are the
steps that would be followed:

– Replanning in a more advanced of the project. We can easily solve this

problem by replanning in a more advanced part of the project, allowing us to do
it more accurately.
– Increasing the hours dedicated to the project. If, despite planning, it is
still difficult to meet the deadline, it may be resolved by increasing the number
of hours dedicated to the project as a last resort.
– Resources to reuse: PC and TeamGantt.
– Estimated delay: 1-2 weeks.

• Bugs in libraries [Medium Risk]. Third-party libraries will be used during project
development, and they may contain bugs. Waiting until the library is updated, which
should hopefully fix the bug, is one possible solution. However, due to the tight
deadline, this option is out of the question. As a result, coding the function from scratch
and testing its correct operation would be required, increasing the overall duration of
the project.

– Resorces to reuse: PC, TeamGantt and programming languages.

– Estimated delay: 1-2 weeks.

23
24

Table 1: Task Table containing a summary of all task information. T and GEPT means Tutor and GEP Tutor respectively.
[Own Creation]
25

Figure 5: Gantt Chart illustrating the project’s schedule following a Waterfall model. [Own Creation]
2.3 Budget
In this section the economic cost of the project is discussed. First, the staff cost is described
and analysed, then the generic and indirect costs are also calculated. Furthermore, the
mechanism for controlling potential budget deviations is also explained. Finally, in Figure 3
we can see that the budget estimation is 17066,08€.

2.3.1 Costs Per Activity

To accurately estimate the project’s costs and create a budget, we must consider all of the
resources required. Human resources is one of them. Even though the project is going to
be done only by my with the guidance of my tutors, 7 roles are created to better simulate
the required human resources to develop the project in order to better estimate the cost by
task. The following is a description of the various responsibilities of each role:

• Project Manager. The project manager is in charge of the project’s planning and
development; in other words, the project manager oversees the project’s progress.

• Software engineer. The software engineers implements the chatbot.

• Tester. The tester is in charge of verifying that the implementation is correct.

• Research ML. The researcher is responsible for investigating the best supervised
machine learning models for the project and selecting the best hyperparameters for
optimization.

• Research psychologist is responsible of designing the proper dialogue flow structure

following the cbt method.

• Technical writer is in charge of documenting the project and presenting.

• ML engineer is responsible of implementing, tuning and analysing the models to

choose the best one.

The Project Managements roles are going to be played by the tutors and the rest of roles
by me.
In this section it is computed the Total Personnel Cost Per Activity (CPA). Each task
or activity (previously defined in 2.1) is associated with the staff cost who are involved in
that task. In this project there are 7 roles, each one with a different hourly salary which
translates into cost per hour shown in Figure 2.

26
 Role Gross Annual Salary (€) Price per hour (€)
Project Manager 52899,6 25,425
Software Engineer 46556,9 22,375
Tester 39533 19
Research ML 47239,4 22,7
Research psychologist 71436,3 34,35
Technical Writer 41600 20
ML engineer 47239,4 22,7

Table 2: Salary of the different roles extracted from PayScale, a compensation software
company [7] multiplied by 1.35 to include the cost of social security [Own creation].

The computation of CPA is done by multiplying the hours required per task/activity
with the cost per hour of the role that is involved in the activity. The total CPA is the
sum of the CPA of every task of the Gantt Chart. As shown in Figure 3, the total cost of
recruitment (CPA) is 13366,05€.

2.3.2 Generic Costs

There are many resources that aren’t directly tied to a task: the generic costs. To calculate
the generic costs we need to take into account the amortisation of the resources used. In
this project, all the software products are free so we are going to focus on the calculation
of the hardware costs. I will be working 5 hours a day on average during 126 days. The
computation of the amortisation is done with formula 1.

1 1 1
Amortisation(e ) = Resource Price· · · ·Hours Used
Years of Use Days of Work Hours per Day
(1)
The indirect costs are identified to make the budget more realistic. Since I’ll be working
from home for the project (unless an extraordinary in-person meeting with the tutor is
required), the transportation cost is zero. On the other hand, internet costs around 70€ per
month, and electricity costs 100€ per month. The total Generic Cost, as shown in Figure 3,
is 1223,11€.

2.3.3 Contingency
Unexpected events are common during the development of a project, and one must plan
ahead to account for them. As a result, a contingency plan is created in order to avoid
potential delays during the planning process. Since contingency margins in the IT sector
typically range from 10% to 20%, I decided to have a 15% contingency margin for this
project, which amounts to 2188,37€.

27
2.3.4 Incidental Costs
Incidental costs define all potential risks that could cause project delays. The most extreme
risk of the project in this case is detecting bias in the machine learning models and thus
having to generate more data alternatively, as explained in previous sections. The project is
delayed as a result of this risk. Total Incidental Costs are 288,54€.

2.3.5 Management control

The budget control mechanisms are discussed in this section. Additionally, the control
indicators that help monitor cost variances throughout the project’s development are defined.
While doing each planned task, the deviation from the estimated cost is calculated with
the Formula 2.

Deviation(e ) = CostEstimated − CostReal (2)

If the deviation is negative, part of the contingency fund must reallocated in order to
cover the deviation. In the positive case, it means there has been a overestimation of cost
and reallocation of the extra money to incidents would be more productive.

Activity Amount (€) Observations

PM1 - ICT tools for project and team management 25,43 Project Manager , 1 hour
PM2 - Context and Scope 889,88 Project Manager , 35 hours
PM3 - Time planning 762,75 Project Manager , 30 hours
PM4 - Budget and sustainability 762,75 Project Manager , 30 hours
PM5 - Meetings 457,65 Project Manager , 18 hours
PR1 - Psychology Research 343,50 Research psychologist, 10 hours
PR2 - ML Research 340,50 Research ML, 10 hours
PT1 - Design Chatbot 343,50 Research psychologist, 10 hours
PT2 - Choose supervised ML models 227,00 Research ML, 10 hours
PT3 - Select the Hyperparameters to optimize 227,00 Research ML, 10 hours
DG - Data Generation 480,90 Research psychologist, 14 hours
PE1 - Apply the Workflow 363,20 ML engineer, 16 hours
PE2 - Hyperparameters Tuning 181,60 ML engineer, 8 hours
PE3 - Performance analysis of every model 45,40 ML engineer, 2 hours
PDEV1 - Telegram bot Implementation 4475,00 Software Engineer, 200 hours
PDEV2 - Testing 1140,00 Tester, 60 hours
PDOC1 - Annotation of events 200,00 Technical Writer, 10 hours
PDOC2 - Revision of the events 400,00 Technical Writer, 20 hours
PDOC3 - Write final documentation 1200,00 Technical Writer, 60 hours
TDP - Thesis Defense Preparation 500,00 Technical Writer, 25 hours
Total CPA (Cost Per Activity) 13366,05 Total personnel costs by activty (Gantt actvities)
Hardware
Laptop 220,57 Mackbook Air 2017, Purchase Price: 1200e
Peripheral devices 122,54 Display + mouse + keyboard, Purchase Price: 400e
Software
Overleaf 0,00 Free to use
Google sheets 0,00 Free to use
TeamGantt 0,00 Free to use
Colab 0,00 Free to use
GitHub 0,00 Free to use
Space
Electricity 400,00 100€/month x 4 months (duration of project)
Furniture 200,00 Table + Chair
Internet 280,00 70€/month x 4 months (duration of project)
Transport 0,00 Work from home
Total GC (Cost computed Generically) 1223,11
Total Cost (Total CPA + Total GC) 14589,16
Contingency 2188,37 Contingency margin = 15%
Total DC (direct cost) + IC (indirect cost) + Contingency 16777,54
Data Generation Delay (1 Week) 240,45 Cost: Research psychologist, 14 hours. Risk: 50%
Data Generation Delay (2 Week) 48,09 Cost: Research psychologist, 14 hours. Risk: 10%
Total incidentals (or unforseen costs) 288,54
TOTAL 17066,08

Table 3: Budget Structure of the project [Own creation]

28
2.4 Deviations
The project’s methodology hasn’t changed; the hybrid mode between waterfall and agile
technique is well suited for the project, and it’s because of this that the previously mentioned
deviations haven’t had a significant impact on the project’s proper development.
The gantt chart, which employs a waterfall methodology, was used throughout the
project’s development to determine the dependencies and the order in which to strategize the
tasks. On the other hand, the project’s coding and testing phases used an agile methodology
with kanban boards to keep track of all the tasks.
As can be seen in Table 4, there were two significant changes: one that affected the
project development and the other that affected the project documentation. Both changes
had an impact on the budget. The changes will be described in the section that follows.

Table 4: Final version of the task table [Own creation]

2.4.1 Deviations in the project development

After revising the documentation of the telegram app API to develop the chatbot, I opted to
use another alternative: BotUI a javascript framework. The reason to this was that I found
the Telegram API documentation rather unclear and not easy to use for the functionalities
needed for the chatbot.
Unfortunately, there was a problem with the API connecting the front end and the model,
and I had to make another change because of the time constraints. The implementation was
done in a Jupyter notebook and TKinter, a python framework to develop the interface.

29
 Additionally, the time required to do the project development was reduced from 200
hours to 146 hours.

2.4.2 Deviations in the project documentation

Writing the monitoring report wasn’t factored into the initial project planning. As a result,
the monitoring report has been included in the project’s final planning. This modification
affected both the computation of the budget, as shown in Table 5, and the projected number
of hours required to complete the project, as shown in Table 4.

2.4.3 Deviations in the budget

Using the Formula 2, the total deviation cost is 17066,08 - 15995,01 = 1071.07e . Since
the deviation is positive, this amount is the saved amount of budget.

Activity Amount (€) Observations

PM1 - ICT tools for project and team management 25,43 Project Manager , 1 hour
PM2 - Context and Scope 889,88 Project Manager , 35 hours
PM3 - Time planning 762,75 Project Manager , 30 hours
PM4 - Budget and sustainability 762,75 Project Manager , 30 hours
PM5 - Meetings 457,65 Project Manager , 18 hours
PR1 - Psychology Research 343,50 Research psychologist, 10 hours
PR2 - ML Research 340,50 Research ML, 10 hours
PT1 - Design Chatbot 343,50 Research psychologist, 10 hours
PT2 - Choose supervised ML models 227,00 Research ML, 10 hours
PT3 - Select the Hyperparameters to optimize 227,00 Research ML, 10 hours
DG - Data Generation 480,90 Research psychologist, 14 hours
PE1 - Apply the Workflow 363,20 ML engineer, 16 hours
PE2 - Hyperparameters Tuning 181,60 ML engineer, 8 hours
PE3 - Performance analysis of every model 45,40 ML engineer, 2 hours
PDEV1 - Chatbot Implementation 3266,75 Software Engineer, 146 hours
PDEV2 - Testing 1140,00 Tester, 60 hours
PDOC1 - Monitoring Report 300,00 Technical Writer, 15 hours
PDOC2 - Annotation of events 200,00 Technical Writer, 10 hours
PDOC3 - Revision of the events 400,00 Technical Writer, 20 hours
PDOC4 - Write final documentation 1200,00 Technical Writer, 60 hours
TDP - Thesis Defense Preparation 500,00 Technical Writer, 25 hours
Total CPA (Cost Per Activity) 12457,80 Total personnel costs by activty (Gantt actvities)
Hardware
Laptop 205,71 Mackbook Air 2017, Purchase Price: 1200e
Peripheral devices 114,29 Display + mouse + keyboard, Purchase Price: 400e
Software
Overleaf 0,00 Free to use
Google sheets 0,00 Free to use
TeamGantt 0,00 Free to use
Colab 0,00 Free to use
GitHub 0,00 Free to use
Space
Electricity 400,00 100€/month x 4 months (duration of project)
Furniture 200,00 Table + Chair
Internet 280,00 70€/month x 4 months (duration of project)
Transport 0,00 Work from home
Total GC (Cost computed Generically) 1200,00
Total Cost (Total CPA + Total GC) 13657,80
Contingency 2048,67 Contingency margin = 15%
Total DC (direct cost) + IC (indirect cost) + Contingency 15706,47
Data Generation Delay (1 Week) 240,45 Cost: Research psychologist, 14 hours. Risk: 50%
Data Generation Delay (2 Week) 48,09 Cost: Research psychologist, 14 hours. Risk: 10%
Total incidentals (or unforseen costs) 288,54
TOTAL 15995,01

Table 5: Final version of the budget structure [Own creation]

30
3 Identification of Laws and Regulations
Understanding the laws and regulations that have an impact on the design and development
of the chatbot is one of the most crucial components of the project thesis.

3.1 Academic Regulations for the Degree Final Project

The UPC has a documentation of the regulations for the Degree Final Project available
online [14]. This documents defines and describes the characteristics of the final project.
The document also explains all the process needed to do the project. This document is of
course very important and must be followed to ensure the correct development of the project.

3.2 GDPR Privacy Policy

A first step toward granting EU people and residents more control over how their data are
used in organizations is the EU General Data Protection Regulation (GDPR). No matter
where they are in the globe, businesses must adhere to the GDPR if they handle the personal
data of people who reside in the EU.
A key requirement for businesses subject to the GDPR is that they make transparent
and easily accessible information about the personal data they are processing available to
the public. A clear and thorough privacy policy will help one to achieve this.
A privacy notice is a public statement from a company outlining how it manages customer
information and adheres to data protection laws. A GDPR privacy notice is a crucial tool
for assisting customers and users in making informed choices regarding the data you gather
and use.
According to the GDPR [15], organizations are required to give customers a privacy
disclosure that is:

• In a clear, visible, understandable, and readily available format

• Written in a straightforward manner, especially for any information aimed exclusively

towards children

• Delivered on schedule

• Provided free of charge

In the following sections, all the information that must be included in a privacy notice is
explained.

3.2.1 Company’s contact details

Article 13(1)(a) [16] of the GDPR requires providing to users with: ”the identity and
the contact details of the controller and, where applicable, of the controller’s
representative”. An individual who determines how and why personal data is handled is
referred to as ”the controller” or a ”data controller.”

31
 Article 13(1)(b) [16] of the GDPR also requires providing: ”the contact details of the
data protection officer, where applicable”. A data protection officer is required for
some firms of a specific size or those that consistently handle sensitive personal data (DPO).

3.2.2 The Purposes and Legal Basis for Processing

Article 13 (1)(c) [16] of the GDPR requires providing information about: ”the purposes
of the processing for which the personal data are intended as well as the legal
basis for the processing”. To put it another way, is not allowed to process personal data
unless there is a purpose for doing so. Additionally, there must be a legal justification for
every form of data processing is carried out.
The GDPR sets out six legal bases at Article 6.
A person’s personal data may only be processed if at least one of the following conditions
is met [16]:

• You have their consent.

• To carry out or enter into a contract with them, you must process their personal data.

• It’s required by law that you handle their personal information.

• Failure to process their personal data could endanger their lives or the life of another
person.

• Processing their personal data is something you’re doing in the public interest.

• You have a legitimate interest in processing their personal data.

The app falls under the category ”You have a legitimate interest in processing
their personal data” [16] since it collects user data in order to identify potential cognitive
distortions based on user input.

3.2.3 Sharing of user’s personal data

Article 13 (1)(e) [16] requires to provide information about: ”the recipients or categories
of recipients of the personal data, if any”. In the app’s case the data is never shared
with third party companies.

3.2.4 Sharing of user’s personal data to a third country

Article 13(1)(f) [16] of the GDPR requires providing information about: ”the fact that
the controller intends to transfer personal data to a third country or interna-
tional organization and the existence or absence of an adequacy decision by the
commission”. A ”third country” refers to a country outside of the EU.
The list of nations with ”sufficient” data protection rules is maintained by the European
Commission. You must indicate if a country is on the list if you are sending data to a third
country. In the app’s case the data is never shared with third countries.

32
3.2.5 Period of time storage of user’s personal data
Article 13(2)(a) [16] of the GDPR requires informing users: ”the period for which the
personal data will be stored, or if that is not possible, the criteria used to de-
termine that period”. It’s crucial to comply with the GDPR’s prohibition on keeping
personal data longer than necessary. In the case of the app, user data is never stored.

3.2.6 User’s Rights

Chapter 3 of the GDPR [17] sets out the rights that people have over their data. The GDPR
not only requires you to not only make it easier for your users to access these rights, but
also to inform them of those rights in your Privacy Policy. Additionally, you must
let the user know how to file a complaint with their local data protection authority.

3.3 The EU Regulatory Environment of Medical Device Software

Development
The International Medical Device Regulators Forum (IMDRF) defines SaMD (Software as a
Medical Device) as “Software intended to be used for one or more medical purposes
that perform these purposes without being part of a hardware medical device”.
Taking into account this definition, the chatbot falls into this category.
The relevant General Safety and Performance Requirements (GSPRs) will have to be
complied with by all software that falls within the Medical Device category. MDSW lawful
manufacturers must put out a dossier or technical document (TD) for their product in order
to prove conformity with GSPR. The details and explanations of the documents’ contents
are provided in the section that follows.
Applicable GSPRs mainly refer to one of the following general fields:
• Quality Management System (QMS) requirements. MDSW developers must
work following a QMS methdology.
• Risk Management System (RMS) requirements. The basic objective of an RMS
is to make sure that any potential dangers are recognized, categorised, and minimized
without negatively influencing the device’s risk-benefit ratio. To do this, the man-
ufacturer must develop and implement a risk management strategy that accurately
identifies all hazards related to the devices, establishes the necessary risk mitigation
measures, and evaluates the effectiveness of each strategy.
• Clinical Evaluation and Post-market surveillance requirements. The clinical
evaluation of a MDSW must be carried out following MDCG 2020-1, guidance on
Clinical Evaluation of MDSW, and a Clinical Evaluation Report drafted providing the
following information:
– A valid clinical association of the software with the targeted clinical condition or
physiological state, usually by means of literature references.
– An analytical evaluation of the software to show that it is capable of processing
data appropriately.

33
 – The software’s output is then validated clinically to guarantee that it is accurate
and dependable in the context of the clinical setting.

• Usability requirements. SW developers must make sure that as many user errors
as possible are prevented via the user interface. IEC 62366 Medical devices — Part 1:
Application of usability engineering to medical devices must be followed when planning
and conducting usability tests for this. From a cybersecurity and safety standpoint,
each and every one of the discovered user errors must be taken into account in the
risk analysis and contributed to the risk management strategy and report. As with
any other risk, preventive steps must be taken if the possibility of user errors cannot
be entirely removed. Some of them include increasing training or adding particular
warnings to the user handbook.

In addition, theses requirements are specifically applicable to MDSW:

• Software lifecycle requirements. The software lifecyle is described in IEC 62304

Medical device software — Software life cycle processes. It provides a series of steps
that should be taken by SW developers. In Figure 13 we can see the Software Lifecycle.

• Cybersecurity requirements. This mainly regard patient data protection and pro-
tection from other cyber threats.

As with any other MD, it is advised that MDSW developers use approved techniques and
standardized procedures like the ones listed below in order to adhere to the relevant GSPRs:

• International standards (mainly ISO 5 , IEC 6

and ANSI/AAMI 7
standards)

• MDCG or IMDRF guidance documents

Table 6 provides an exhaustive list of the standards and guidelines now in existence that
are advised to be followed by MDSW developers in order to accomplish with applicable
GSPRs, together with their most recent updates.

5
International Organization for Standardization (ISO)
6
International Electrotechnical Commission (IEC)
7
American National Standards Institute (ANSI)/ Association for the Advancement of Medical Instru-
mentation (AAMI)

34
4 Sustainability report
It is well known about climate change and the consequences that we are going to suffer or that
we are even suffering now. Thus is really important for humans and companies to stop being
selfish and think about the future of the world and cooperate in order to reduce pollution
urgently. Thus, it is important to check the footprint of a project to see how does it impact
in the environment. Assessing for the economical impact is also important, it helps us to
optimize cost and savings. Finally keeping track of the social impact of a company/project
is also very important. New technologies in particular have changed millions of people life
including minorities and in developing countries.

4.1 Self assessment

Students were asked to complete a survey for their bachelor thesis. This survey asks respon-
dents about their knowledge of sustainability in various fields. Some of these fields include
economic, environmental, and social sustainability. After doing the survey, I realize that the
Environmental field is my weak spot. In particular, I don’t know which indicators to use to
measure the impact in this aspect. Regarding the economical field, I have some intuition on
how to measure and control the economical impact since I previously did a a budget. Finally
in the social field, I think I know how new technologies and in particular my project impacts
society. So in conclusion, I have a below average level of knowledge about sustainability,
especially in the environmental field.

4.2 Environmental dimension

Regardind the PPP, As of 2023-01-10 19:00 according to Nowtricity a website that offers
real time information of the emissions of every country [18], the current emissions in Spain is
137 grams C02 / kWh. The computer used for the project consumes an average of 0.2 kWh.
Taking into account the previous information, the environmental impact can be calculated
as shown in Formula 3.

gr CO2
137 · 0.2 kWh · 540 h (total duration of the project) = 14796 gr of CO2) (3)
kW h
One approach to reduce the the impact would be to execute in parallel to reduce training
and evaluation time. This can be accomplished by increasing the njobs parameters above 0:
the sklearn python library includes a parameter for determining the number of jobs to run
in parallel for cross-validation.
Regarding the exploitation, as mentioned before, most people go to therapy in the tra-
ditional way. Since people can have access to KAI without the need to move from home,
the project helps in reducing the pollution because they no longer need to take a mains of
transport to go to a session. So we can conclude that KAI is more environmentally-friendly.
Regarding the risks, the project doesn’t pose any, in fact the projects helps in reducing
people’s ecological footprint as mentioned previously.

35
4.3 Economic dimension
Regarding PPP, in section 2.3 the estimated costs of the project are identified and calculated
and the budget is also shown. The hours required for the project have been revised and finally
reduced resulting in saving costs.
Regarding the exploitation, nowadays most people go to in-person therapy with a pro-
fessional therapist which is really expensive: in Spain the average price for one session is
50e and taking into account that on average a person needs between 8-20 sessions, the final
costs amounts to 400-1000e . Online therapy sessions are getting popular giving people the
flexibility to receive support and help with the need to transport. This option is usually
more economical than traditional therapy (in person). Self-help therapy chatbots currently
available in the market work on a free basis but if you want access to more content you need
to subscribe, furthermore if you want to have access to a therapist: rates vary depending on
the therapist or works on a subscription basis. Since KAI is free, it will help people embark
in their self-help journey in therapy with the guide of the chatbot in a more affordable way.
In the future, the project will have an almost inevitable cost: human resources. This cost
could be reduce with the automation of tasks and with the availability of datasets.
Regarding the risks, the project is very dependant on data, if the quality of data is not
good enough it could lead to very inaccurate predictions.

4.4 Social dimension

Regarding the PPP, it has aided me in learning more about psychology, a field in which I
have always been interested, and how technology can help people with their psychological
needs. Furthermore, it has helped me in becoming more familiar with the machine learning
field, particularly in the healthcare/medical sector, in which I am very interested. It also
made me realize that Python is a powerful programming language, particularly for AI, due
to the existence of an extensive library. What’s more, the project experience has assisted
me in determining whether I truly want to pursue a career in AI in healthcare or medicine.
Finally, the most significant and special contribution of the project to me has been the op-
portunity to provide a mental health support tool like KAI to my sister, who unfortunately
suffers from a severe mental illness.

Regarding the exploitation, the project will help to close the gap between those who need
and those who receive mental health care. They will also have 24/7 access to support. More
importantly, it will allow people to have access to mental health support in a more
affordable, autonomous, and time-efficient way. This project is aimed at people who
have mild to low symptoms of depression or anxiety. People who suffer from severe mental
health or struggle with severe depression and anxiety are advised against using the chatbot.
Regarding the risks, as mentioned before, the project is not aimed at people who suffer
from severe mental issues, instead is a tool for people with mild symptoms of anxiety and/or
depression to do a self-guided CBT therapy.

36
5 Technical Competences
During the development of the project thesis, the following technical competences from com-
puting specialization were followed:

CCO2.1
To demonstrate knowledge about the fundamentals, paradigms and the own
techniques of intelligent systems, and analyse, design and build computer sys-
tems, services and applications which use these techniques in any applicable
field. [Quite]

During the project design and development of an application (a chatbot) using machine
learning has been done.

CCO2.2
Capacity to acquire, obtain, formalize and represent human knowledge in a com-
putable way to solve problems through a computer system in any applicable field,
in particular in the fields related to computation, perception and operation in
intelligent environments. [Quite]

This competence was achieved with the acquisition of human knowledge and the repre-
sentation through machine learning to detect possible cognitive distortions.

CCO2.3
To develop and evaluate interactive systems and systems that show complex in-
formation, and its application to solve person-computer interaction problems.
[A little]

This competence was achieved with the design and development of the chatbot that
extracts and shows complex information (possible cognitive distortions) from user input, a
form of human-computer interaction (conversational user interface).

CCO2.4
To demonstrate knowledge and develop techniques about computational learn-
ing; to design and implement applications and system that use them, including
these ones dedicated to the automatic extraction of information and knowledge
from large data volumes. [In depth]

The study included extensive research on the top machine learning methods for topic
classification. In addition, the design and development of the chatbot that automatically
detects potential cognitive distortions using a machine learning technique.

37
6 Dialogue Flow
The design of the chatbot’s dialogue flow is described in this section. The dialogue flow
is divided in three main parts as shown in Figure 12 which are the following: identifying
potential cognitive distortions, challenging potential cognitive distortions, and
chatbot’s feedback. The design of the dialogue flow is based on Module 9 and Module 10
of A Therapist’s Guide to Brief Cognitive Behavioral Therpy [1].

6.1 Identifying Potential Cognitive Distortions

Finding cognitive distortions is the initial stage in the therapy’s cognitive component. To
accomplish this, one must first identify the user’s automatic thoughts. An automatic thought
is a thought that comes to you without conscious thought. Automatic thoughts are typically
associated with negative emotions and might be triggered by certain events or circumstances.
The red elements in Figure 12 are a part of the first step, which is to identify potential
cognitive distortions. The steps to accomplishing this are as follows:

• Asking how the user has been feeling. If the user’s answer is positive then the support
to therapy ends and the dialogue starts over. Otherwise the dialogue flow continues.

• Asking the user what made them feel this negative feeling.

• Asking the user what where they thinking when they were in that situation that made
them feel bad in order to detect an automatic thought. In this part is where the ml
model to detect potential cognitive distortions will be used.

• Asking the user how would they rate their negative feeling/mood that they previously
mention in the first step.

6.2 Challenging Potential Cognitive Distortions

After identifying the potential cognitive distortions, the next step is challenging them. This
will be done by using Dysfunctional Thought Record (DTR).
In cognitive-behavioral therapy (CBT), a Dysfunctional Thought Record (DTR) is a
tool used to identify and address negative ideas and beliefs that contribute to emotional and
behavioral problems. The DTR is a systematic form that assists people in recognizing their
negative behaviors and beliefs, weighing the evidence supporting and refuting them, and
coming up with more reasonable alternatives.
Figure 12 shows that the second step—challenging potential cognitive distortions—includes
the green components. This is done by asking the the user for evidence that their thought
is true and not true and finally asking the user to think of an alternative way to see the
situation
The last step consist in asking the user to rate their mood after providing their alternative
and more objective way of seeing the situation to see if there has been an improvement and
the user feels better.

38
6.3 Chatbot’s Feedback
The chatbot provides feedback to the user as the final stage of the conversation flow. This
is accomplished by providing the user with a definition of the potential cognitive distortion
that has been identified during the conversation, along with some helpful advice. If the
chatbot notices that the user’s mood hasn’t improved, there is also a relaxing exercise (a
link to a video).

6.3.1 Relaxation Exercise

Deep breathing is used in this part of the session. Slowing down the shallow, irregular
breathing that usually occurs when people are agitated, worried, or anxious is the aim of
deep breathing. The patient may have symptoms like hyperventilation and dizziness as a
result of rapid and shallow breathing, which can cause their blood oxygen levels to fall and
affect their ability to concentrate. A deep, complete breath may instead increase the amount
of oxygen-rich blood flow, which may result in a sense of calm or slowness.

39
7 Dataset
The dataset of the project recollects phrases examples of 15 cognitive distortions (see Figure
11 to see the lists of cognitive distortions with their definitions). The dataset is composed
of 595 rows and 2 columns. In the first columns there are example phrases of cognitive
distortions and the second column indicates the type of cognitive distortion. Since not all
automatic thoughts are negative or cognitive distortions, non-cognitive distortions examples
have been added too.
As mentioned in section 2.1.3, the data has been generated by recollecting examples of
cognitive distortions from trustful sources (books, articles, official psychology pages, etc).

7.1 Preprocessing
In this section the preprocessing done to the dataset is explained.

7.1.1 CountVectorizer and TfidfTransformer

Since machine learning models can’t work with texts, it necessary to transform it into numer-
ical representation. This is done by using two functions: CountVectorizer and TfidfTrans-
former. According to the official scikit library [19], CountVectorizer is used to ”Convert a
collection of text documents to a matrix of token counts”, while TfidfTransformer
is used to ”Transform a count matrix to a normalized tf or tf-idf representation”.
TF-IDF (Term Frequency-Inverse Document Frequency) is used to determine the impor-
tance of a word in a document or group of documents. By employing TF-IDF, it is intended
to provide less weight to terms that are frequently used and more weight to words that are
uncommon or unique to the documents under consideration.
The number of times a word appears in a document, normalized by the number of words
in the document, is known as its frequency (TF). The logarithm of the total number of
documents in the corpus divided by the number of documents where the word appears gives
the inverse document frequency (IDF). By multiplying a word’s TF and IDF values, one can
determine the overall weight of the word in a document.

7.1.2 Balancing
When there are significantly less samples from one or more classes than there are from other
classes, the dataset is said to be unbalanced. This might happen if the data were gathered
from a real-world situation where there might not be an equal distribution of examples among
the various classes. The dataset is clearly unbalanced, as can be seen in Figure 14 (there are
almost 50 examples of Emotional Reasoning but only a little more of 20 examples of Always
Being Right).
Machine learning algorithms may encounter difficulties when given unbalanced datasets
because they may be biased in favor of the dominant class and may not adequately repre-
sent the minority class. This may result in models that are not generalizable to real-world
scenarios and poor performance on the minority class.
The SMOTE methodology is used for balancing in this project. SMOTE (Synthetic
Minority Oversampling Technique) is an oversampling technique used in machine learning to

40
overcome the problem of unbalanced datasets. Instead of just duplicating existing examples,
it generates synthetic examples of the minority class to balance the distribution of classes.
To generate synthetic examples, SMOTE first selects a minority class example and finds
its K nearest minority class neighbors. It then interpolates a new synthetic example between
the selected example and one of its neighbors, by sampling from a line between the two
examples. This process is repeated until the desired amount of oversampling is achieved.

41
8 Model Experimentation
Each machine learning model that was used in the model experimentation is introduced
in the subsections that follow, along with an explanation of how each hyperparameter was
tuned. Additionally, the model experimentation results are presented and analyzed.

8.1 Multinomial Naive Bayes

The Multinomial Naive Bayes algorithm is a classification method that is based on the Naive
Bayes algorithm and is specifically designed for text classification tasks with multiple classes.
It estimates the probability of each class label occurring and the probabily of each feature
occurring given a specific class label, and uses the Bayes theorem seen in Formula 4 to classify
new data points.
The Bayes theorem can be expressed as follows:

P (B|A) · P (A)
P (A|B) = (4)
P (B)
where:

• P (A|B) is the probability of event A occurring given that condition B is true.

• P (B|A) is the probability of condition B being true given that event A has occurred.

• P(A) is the probability of event A occurring.

• P(B) is the probability of condition B being true.

8.1.1 alpha Hyperparameter

The alpha parameter is an Additive (Laplace/Lidstone) smoothing parameter. Laplace
smoothing, often known as add-k smoothing or additive smoothing, is a method for keeping
probability estimates from having zero probabilities. It is frequently used to enhance the per-
formance of algorithms that rely on probabilistic estimations in natural language processing
and machine learning applications, such as text classification and language modeling.
In probability estimation, zero probabilities can occur when a feature has not been ob-
served in the training data. For example, in a text classification task, a word may not
appear in the training data for a particular class label, resulting in a zero probability esti-
mate for that word given the class label. This can cause problems when classifying new data
points because the zero probability can result in a zero probability for the entire data point,
regardless of the other features.
Laplace smoothing solves this problem by increasing the count of each feature by a
small constant called the alpha parameter, also known as the ”smoothing parameter” or
”smoothing factor”. As a result, there is a slight rise in the probability estimates for all
attributes, even those with zero probabilities.
We must specify a value for the smoothing parameter, which controls how much smooth-
ing is applied to the probability estimates, in order to perform Laplace smoothing. The

42
smoothing option is frequently set to 1, which increases the count of each feature by one.
This can be mathematically expressed with Formula 5 where:

• P (f eature|class) is the probability of the feature occurring given the class label.

• count(feature, class) is the number of times the feature has been observed in the train-
ing data for the class label.

• count(class) is the total number of observations for the class label.

• alpha is the smoothing parameter.

• num features is the total number of unique features in the training data.

count(f eature, class) + α

P (f eature|class) = (5)
count(class) + α · num f eatures
Laplace smoothing can help probabilistic models function better by lessening the effect
of zero probability and preventing overfitting to the training set of data. To prevent adding
too much bias to the probability estimates, it is crucial to select an acceptable value for the
smoothing parameter.

8.2 Multinomial Logistic Regression

The Multinomial logistic regression algorithm is one of the models used in the experimen-
tation for the automatic detection of potential cognitive distortions. Multinomial logistic
regression is a classification method that is used to predict a categorical dependent variable,
with multiple categories, from one or more independent variables.
It is assumed that the dependent variable in multinomial logistic regression has multiple
categories, each of which may be predicted based on the values of the independent variables.
The category with the highest probability is selected as the predicted outcome according to
the model’s estimates of each category’s likelihood.
The model is based on the assumption that the log-odds8 of the dependent variable are
a linear combination of the independent variables. This is expressed in Formula 6 where y is
the dependent variable, k is a category of y, x1 , x2 , ..., xn are the independent variables, and
b0 , b1 , b2 , ..., bn are the coefficients that are estimated by the model.

p(y = k)
!
log = b0 + b1 · x1 + b2 · x2 + ... + bn · xn (6)
1 − p(y = k)

Maximum likelihood estimation is the most used technique for calculating the beta pa-
rameter, or coefficient, in this model (MLE). This method tests several beta values repeat-
edly in search of the best match for the log odds. Logistic regression aims to maximize this
function after each of these iterations in order to determine the optimal parameter estimate.
8
The probability of success divided by the probability of failure

43
Once the optimal coefficient (or coefficients, if there are numerous independent variables) has
been identified, the conditional probabilities for each observation can be computed, logged,
and summed to obtain a predicted probability.
There are several advantages to using multinomial logistic regression, including its ability
to handle multiple categories and its ability to model the relationships between the inde-
pendent variables and the dependent variable. It’s crucial to keep in mind that the model
makes the assumption that the independent variables are unrelated to one another, which
may not always hold true in actual life.

8.2.1 penalty Hyperparameter

According to the scikit-learn API library [13] the penalty parameter is used to ”Specify
the norm of the penalty”.
Regularization is a method for avoiding overfitting in machine learning models like multi-
nomial logistic regression. Regularization is used to impose a penalty on the model’s com-
plexity, which helps to lower the variance and enhance the model’s generalization capabilities.
Regularization is accomplished in multinomial logistic regression by including a penalty term
in the objective function that is being optimized.

8.2.2 C Hyperparameter
According to the scikit-learn API library [13] the C hyperparameter is the ”Inverse of reg-
ularization strength; must be a positive float. Like in support vector machines,
smaller values specify stronger regularization”.

8.2.3 solver Hyperparameter

The solver hyperparameter in multinomial logistic regression is a parameter that determines
the algorithm used to optimize the model.

8.3 Support Vector Machine

A Support Vector Machine (SVM) is a type of supervised learning algorithm that can be used
for classification or regression tasks. The algorithm finds the best boundary (a hyperplane)
that separates the data into different classes. The boundary is chosen in a way that maximizes
the margin, which is the distance between the boundary and the closest data points from
each class (these points are called support vectors). The goal of SVM is to identify a
boundary that effectively divides the classes while also having the biggest margin to reduce
generalization error as shown in Figure 15.
Hyperplanes act as judgment lines for categorizing the data points. Different classes
can be given to the data points that fall on each side of the hyperplane. Additionally,
the number of features affects the hyperplane’s dimension. For instance, if there are only
two input features, the hyperplane is only a line, and if there are three input features, the
hyperplane becomes a two-dimensional plane.

44
 The hyperplane is defined by a weight vector (w) and a bias term (b). The equation
of the hyperplane is given by Formula 7 where x is a feature vector and w and b are the
parameters of the hyperplane.

w·x+b=0 (7)
The distance of a point x from the hyperplane is given by Formula 8 where ||w|| is the
norm of the weight vector.
w·x+b
distance = (8)
||w||
An SVM’s objective is to determine the hyperplane with the greatest margin—that is,
the distance between the hyperplane and the nearest data points from either class—and the
hyperplane that maximum separates the classes.
The SVM algorithm uses a method known as the ”kernel trick” to find the hyperplane.
The input data is mapped into a higher-dimensional space using the kernel method, making
it simpler to locate the hyperplane. The type of data and problem complexity determine
the kernel function that is employed. The linear, polynomial, and radial basis functions are
often employed kernel functions.
After locating the hyperplane, the SVM can be used to categorize additional data points
by determining how far they are from the hyperplane. The point is categorized as belonging
to one class if the distance is positive, and to the other class if the distance is negative.
SVMs have several advantages over other classification algorithms. They are robust to
noise and can handle high-dimensional data. They have also been extensively investigated
and employed in a wide range of applications. They also have a strong mathematical foun-
dation.

8.3.1 decision function shape Hyperparameter

Support Vector Machines (SVMs) employ the decision function to categorize data points
according to their separation from the hyperplane. Based on the trained SVM model, pre-
dictions about the class labels of new data points are made using the decision function.
The decision function is defined with Formula 9 where w is the weight vector, x is the
feature vector of the data point, and b is the bias term. The sign of the decision function
value determines the class label assigned to the data point. If the decision function value is
positive, the data point is assigned to one class, and if it is negative, it is assigned to the
other class.

f (x) = w · x + b (9)

8.4 K-Nearest Neighbour

K-Nearest Neighbors (KNN), one of the most widely used machine learning algorithms, is
a simple and easy to implement machine learning algorithm for classification and regression
issues.

45
 In the KNN algorithm, a new data point is classified based on the majority class of its
”nearest neighbors”. The number of neighbors, ”K”, is a hyperparameter that is specified by
the user.
To classify a new data point, the KNN algorithm follows these steps:

• Calculate the distance between the new data point and all the training data points.

• Select the K training data points that are closest to the new data point.

• Determine the majority class of the K nearest neighbors.

• Assign the new data point to the majority class.

8.4.1 n neighbors Hyperparameter

According to the scikit-learn API library [13] the n neighbors hyperparameter specifies the
”Number of neighbors to use by default for kneighbors queries”.

8.4.2 weights Hyperparameter

In the K-Nearest Neighbors (KNN) algorithm, the weights of the nearest neighbors can be
used to give more or less influence to certain data points when making a prediction.

8.4.3 p Hyperparameter
According to the scikit-learn API library [13] the p hyperparameter is the ”Power pa-
rameter for the Minkowski metric. When p = 1, this is equivalent to using
manhattan distance (l1), and euclidean distance (l2) for p = 2. For arbitrary p,
minkowski distance (lp) is used”.

8.5 Random Forest

Random forest is a machine learning algorithm that combines the output of multiple deci-
sion trees to reach a single result. The method combines both bias and variance reduction
techniques by constructing a large number of decision trees and then aggregating their pre-
dictions.
Decision trees are a common supervised learning method used in regression and classifi-
cation problems [19]. They are known as decision trees because they are constructed using
a tree-like structure, with leaf nodes serving as the output or prediction and inside nodes
indicating decisions based on the value of input features.
In decision trees, impurity refers to how mixed or ”impure” the data is with regard to
the target labels in a given node or subset of the data. When developing a decision tree,
the objective is to develop a model that can precisely predict the target label for a given
input. Typically, this is done by dividing the data into subsets with as pure a target label
distribution as is possible.
Decision trees frequently employ the Gini impurity, entropy, and misclassification rate
among other impurity measures.

46
 Gini impurity is a measure of the probability of misclassifying a randomly chosen element
in a set, and is defined with the Formula 10 where p(i) is the proportion of elements in the
set that belong to class i and n are the number of elements in the set.
n
Gini impurity = 1 − p(i)2 (10)
X

i=1

Entropy is a measure of the amount of uncertainty in a set, and is defined with the
Formula 11 where p(i) is the proportion of elements in the set that belong to class i and n
are the number of elements in the set.
n
Entropy = − (11)
X
p(i) · log(p(i))
i=1

Misclassification rate is simply the number of misclassified elements in a set divided by

the total number of elements in the set.

8.6 Model Performance

In this section the model performance is presented and evaluated.

8.6.1 Hyperparameters Tuning

In order to avoid or at least reduce the impact of bias, Grid Search and Cross-Validation
is used for the hyperparameters tuning. Grid search is a method for hyperparameter op-
timization that involves training and evaluating a model using a combination of different
hyperparameter values, and selecting the combination that provides the best performance.
Cross-validation is a method for evaluating the performance of a model by dividing the data
into training and validation sets, training the model on the training set, and evaluating the
model on the validation set. The data is split into k-folds in the k-fold Cross Validation
form, and the model is trained on k-1 of the folds before being tested on the last fold. The
test set is a new fold each time, and this operation is done k times. To estimate the model
performance, the results are then averaged.
It’s important to note that in every model, the k-neighbors hyperparameters from SMOTE
(see Section 7.1.2 for a detailed explanation) is additionally tuned. This parameter controls
the number of nearest neighbors used to generate synthetic samples. Increasing k neighbors
will make the synthetic samples more similar to the original samples, while decreasing
k neighbors will make the synthetic samples less similar to the original samples. The dif-
ferent values of the parameter to be tuned are the following: [5,6,7,8,9,10].
In Figure 6 we can see the results obtained from the Hyperparameters Tuning of the
different models. The Multinomial Naive Bayes has a low alpha as a best hyperparameter
value which means that the model is less smooth, in other words, the model is less likely to
assign a non-zero probability for an unseen feature in the training data. For the Multinomial
Logistic Regression has a low C as best hyperparameter value. This low value indicates that
there is less regularization which means that the model is more likely to fit the training

47
9
. On the other hand, for the Support Vector Machine a very high C value was the best
value, meaning there is more regularization. Finally, for the KNN (K-Nearest Neighbor),
the euclidean distance was the best hyperparameter (p hyperparameter = 2).

Figure 6: Results of the Hyperparameters Tuning [Own Creation].

8.6.2 Evaluation
For evaluating the performance of the different models, the analysis will be done between
four metrics: Accuracy, Precision, Recall, F1-Score. Note that the results to be analyzed
are the weighted average of the metric mentioned before to take into account the number of
instances of each class in the testing. The weighted average is specially preferred when there
is an imbalance in the testing data.
Accuracy is a more simple metric that doesn’t take into account the cost of having a
missclassification. Accuracy is expressed with Formula 12.
Number of correct predictions
Accuracy = (12)
Total number of predictions
The precision measures how many positive predictions are actually positive (True pos-
itive). This metric is particularly appropriate when the cost of False Positive is high. For
instance, in the case of spam detection, a false positive indicates that an email that’s it’s
actually important (no spam) has been classified as spam and thus the user is losing valuable
information. Precision is expressed with Formula 13.
Number of True Positive
P recision = (13)
Total predicted positive
The recall measures how many of the actual positives are predicted as positive. This
metric is the most appropriate when there is a high cost of False Negative. For example, if
a person with cancer is labelled as not having cancer, the consequences could be disastrous.
Recall is expressed with Formula 14.
Number of True Positive
Recall = (14)
Total actual positive
The F1-Score gives us a balance between Recall and Precision and is expressed with
Formula 15.
9
Fitting the training data refers to the process of adjusting a model’s parameters so that it accurately
predicts the output values for the input values in the training set.

48
 Precision · Recall
F1-Score = 2 · (15)
Precision + Recall
As we can see in Figure 7, all the models had a better performance in the precision
metric. In other words, they have a lower False Positive rate. Models with high precision
can be considered ”conservative” since it will only predict a positive sample when it’s very
confident that is actually positive. The Support Vector Machine is particularly conservative
since it has the highest precision with a 0.75 score.
Furthermore, all of the models scored lower in the F1-Score. This could be an indication
that the models struggle in correctly identifying positive examples, in other words, there is
a low recall. This can also be caused when the class distribution is imbalanced (specially in
the training set).
Overall, the KNN (K-Nearest Neighbour) had the worst performance. This bad result
could be an indication that the dataset generated is complex 10 since KNN is better suited
for small and more simple datasets.
On the other hand, both Random Forest and Multinomial Naive Bayes curiosly had
the same performance except for the GridSearchCV where Multinomial Naive Bayes had a
slightly better score. Both models are overall the best models. The final model chosen to
integrate to the implementation of the chatbot is Random Forest. The criteria to make the
tiebreaker between the two models was done by taking into account the following:

• Complexity: Random forests are more complex than multinomial Naive Bayes, as
they involve building and training multiple decision trees. On more complicated
datasets, though, they can frequently attain higher accuracy.

• Feature Importance: The relevance of each feature in the model can be determined
by random forests 11 , which is helpful for feature selection or for comprehending the
model’s decision-making process. This kind of information is not provided by Multino-
mial Naive Bayes. This is crucial for understanding how the model ”made the decision”
or ”justified” of choosing an specific cognitive distortion.

Figure 7: Model Experimentation’s Results [Own Creation]

10
A complex dataset is one that has a large number of examples, a large number of features, or a high
degree of complexity or non-linearity in the relationships between the input variables and the output variable.
11
The importance of each feature in a random forest model can be determined by examining the amount
by which the model’s accuracy declines when the values of that feature are randomly permuted. This is
known as permutation importance, and it provides a way to quantify how important each feature is to the
predictions made by the model.

49
9 Implementation
In this section, the implementation of the project is going to be explained. The implementa-
tion has been divided between the model experimentation and the chatbot implementation
(the interface and the dialogue flow). The implementation has been done entirely in python
and Jupyter Notebooks. I have decided to use the Python programming language because it
has an extensive libraries specially for machine learning, it’s easy to learn and use and due
to the simplicity of the syntax, it’s also quicker to code.

9.1 Model Experimentation

For every model a pipeline 12 is used to preproces the data as explained in Section 7.1, using
CountVectorizer and TfidfTransformer to transform text into numbers. This is necessary
since the models can’t understand text, only numbers, as explained before. Smote is also
used to balance the data and finally the model is also added to the pipeline.
Furthermore, the GridSeacrhCV function is called to tune the hyperparameters previ-
ously chosen and evaluating the results with Cross Validation with 10 folds in order to
avoid/reduce the bias.

9.2 Chatbot Implementation

The implementation of the chatbot follows the dialogue flow shown in Figure 12 where the
user explains his situation, his thoughts and rates his mood. Then Kai uses the machine
learning model to detect the potential cognitive distortions in the user thought’s.
However, to make the dialogue flow more natural, a set of rules have been defined to
detect the intentions of the user following the rule-based chatbot design model. This is done
by provinding a database of responses and giving the chatbot a set of rules to decide on how
to choose to response from the database.
The main two libraries used for the implementation are: Natural Language Toolkit
(NLTK) and Regular Expression (RegEx). Natural Language Toolkit is a Python library
to work with human language data and Regular Expression is a sequence of characters that
specifies a search pattern in text (Python supports RegEx with the re library).
The interface has been implemented with TKinter a python framework. This framework
has been chosen because its really easy to use and has compatibility with the backend since
it’s in python too.

9.2.1 Rule-Based Chatbot

Following the rule-based methodology, the chatbot will search for specific keywords in the
user input during the conversation. The keywords are crucial to understand what does the
user intent to do. Once the intent is identified, the chatbot simply matches the intent with
the predefined response.
12
A pipeline is a tool for building and evaluating machine learning models. It’s a sequence of transforms
and a final estimator. The transofrms are applied in sequence to the input data, and the final estimator is
used to make a prediction.

50
 The code has the following structure:

1. Importing Libraries: The first step is importing all the required libraries. The
re library is the package that handles regular expressions as mentioned before. The
wordnet from the NLTK library is also going to be used. Wordnet is a lexical English
database that defines semantics relationships between words (useful to find the meaning
of the words, synonyms, antonyms, etc ...). The main purpose of wordnet for the project
is to build up a dictionary of keywords to the keywords. This will allow me to avoid
manually expand having to introduce every possible alternative word a user could use
to match the specific keyword. The itertools is a module that allows us to handle the
iterators in an efficient way.

2. Building the Keyword List: This part consists of building the list of keywords that
the chatbot needs to look for.

3. Building a dictionary of Intents: Once we have the list of keywords, the next step
is building a dictionary of intents to match intents with keywords.

4. Defining a dictionary of responses: Consists on giving a list of predefined responses

for each intent.

5. Matching Intents and Generating Responses: In this part the user input is
taken and evaluated to see if there is any keyword. This is done with the RegEx search
function. If there is no keyword match, the dialogue flow is continued (see Figure 12).

9.3 Dialogue example

In this section the chatbot is tested. There are two tests, one testing that the dialogue
flow works correctly and the other one tests if the chatbot correctly detects the keywords
previously designed and responds accordingly.

9.3.1 Dialogue Flow Test

In order to test the chatbot, an script has been written. The script is based on an example
given in A Therapist’s Guide to Brief Cognitive Behavioral [1], a book. The example is
about a person of advanced age who feels he has nothing to offer to their family because the
person was out of breath while playing with their grandchild.

Identifying Potential Cognitive Distortions

As we can see in Figure 8, the conversation is started by Kai by greeting and asking
how the user feels. After receiving the user’s answer, the chatbot detects that the feeling
is negative with the aid of the SentimentAnalyzer of the NLTK library. Then the Kai asks
the user to explain more and what was the person thinking at the time. This last person
is really important since is the automatic thought that is going to be used for the model to
detect the potential cognitive distortion.

51
 Figure 8: Screenshot showing the first part of the Dialogue flow [Own Creation]

Challenging Cognitive Distortions

As we can see in Figure 9, the second part of the dialogue starts which consists in chal-
lenging the Potential Cognitive Distortion by asking questions to make the user rationalize
and think of an alternative way of seeing things. This is a fundamental part of the CBT
therapy. Towards the end of the conversation, Kai tells the user the potential cognitive
distortion detected by the model and gives some tips. If the mood is not improved, which
means that the user doesn’t feel any better after the session, a relaxation exercise is provided.

Figure 9: Screenshot showing the second part of the Dialogue flow [Own Creation]

52
9.3.2 Rule-Based Test
As seen in Figure 10, the chatbot correctly detects the keywords and responds accordingly.

Figure 10: Screenshot showing how the chatbot correctly detects the the keywords and
responds accordingly [Own Creation]

53
10 Conclusions
The main objective of the project was designing and developing a chatbot to support therapy
for people who have mild to moderate sympotms of mental health issues. To achieve that a
through research has been done for the design of the dialogue flow following the Cognitive
Behavioral Therapy (CBT) methodology.
Finding an appropriate dataset for training the machine learning models for the auto-
matic detection of potential cognitive distortions in text presented a significant challenge
for the project. Unfortunately, there isn’t a publicly accessible dataset, thus it was decided
to generate one by gathering examples of cognitive distortions from reliable sources. The
resulting dataset was small and could have been biased. SMOTE and cross validation were
used to resolve this.
Furthermore, a review the different hyperparameters that could significantly impact the
performance of the various models for model experimentation has been done for the hy-
perparameter tuning. We can conclude that the objective of detecting cognitive distortions
was accomplished based on the performance that the model experimentation (see Figure 7)
appears to have achieved.
Finally the chatbot development was done in python and the GUI with TKinter, a
python’s framework. Additionaly, to make the dialogue flow more dynamic, a set of rules
have been defined, to make the chatbot respond accordingly when it detects that the user is
doing certain actions, for instance greeting.

10.1 Future Work

Since the quality of the dataset affects the model’s performance, it is essential to increase the
dataset’s size and quality by gathering more representative examples. This might be done
with the use of crowdsourcing, by asking experts to provide examples, or by asking them to
label phrases on open forums like Reddit.
Additionally, as the project developed, I came to understand that there are situations in
which more than one cognitive bias could be present in a single sentence and that some cog-
nitive biases are more connected to one another than others, making it harder to distinguish
between them. To classify a sentence into more than one cognitive distortion, it would be
interesting to perform a multilabel text classification. Additionally, it would be interesting
to revise the different categories of cognitive distortions and establish a ”super group” that
aggregates all related cognitive distortions.

10.2 Reflexions
I was inspired to do this project because I thought it would be interesting to use AI in health
care. After completing this project, I have come to the conclusion that the hype surrounding
machine learning has blinded me. In reality, machine learning models are highly dependent
on data, both in terms of quantity and quality.
It is particularly challenging to apply AI successfully in industries like healthcare, where
data is limited and there are specific regulations with which we must comply. Another
challenge is that most machine learning models end up being ”black boxes” (it is difficult to

54
understand how the model makes the decision based on data). Furthermore, the fact that
we must train the models can be time-consuming, exceedingly error-prone, and frequently
results in failure due to poor data quality. In fact, according to TechRepublic [20], about
85% of AI projects fail in bussines settings and most causes include bad data quality and
problems with data labelling.
If we look at the definition of AI, is a discipline that strives or looks for designing computer
systems that mimic human intelligence without the need of human intervention (except for
the fact that you have obtain the data, clean it and then in some cases manually label
it). While working on the project, I couldn’t help but wonder whether babies and young
children needed to see many dogs in order to recognize them. How do people identify things?
Understanding how our brains truly function and how we learn things is perhaps the first
step toward developing ”real” AI. After all, using massive amounts of data to train the
models may not be the best strategy; perhaps there is another way. Perhaps neuroscience
holds the key to real AI development.

55
A 15 Major Cognitive Distortions

Figure 11: 15 major cognitive distortions by PositivePyschology, a website [4]

56
B Dialogue Flow

Figure 12: Chatbot’s Dialogue Flow [Own Creation]

57
 development is far more complex than t
compromising the functioning of MD with malicious intent).
software lifecycle process. This is depict
As for the RMS, an Information Security Management
System (ISMS) should be implemented, in which a where this typical software lifecycle proce
within the MDSW EU regulated environment
cybersecurity management plan must be defined and
the previously mentioned general and
implemented, whereby all related risks are identified and
C EU Regulatory Environment for MDSW
requirements.Develop-
adequate risk mitigation measures established. It is
ment
important to consider that cybersecurity risk mitigation

!"# !"#$%3&+2
!45/6-"745$89-:03
!"#-!)*$02,0,
!"#-!)*$02,%0
!45/6-"745$89-:%, !"#-!)*-!)))$%2'3'
45*9$0,%'1%%

$"# 2,(.1&.351)+,
;*'/
!"#$%&'(% 45*9$0,%'1%<
!)*$+,,,% !"#-!)*$0(,,,>?@AB@?C
)*-./$+,,,01% !)*-95$+%,,%121%
!"#$%2003 !"#$0((''
!"#-."$0203+ "')/+./'/3. !45/6-*DE)/$89-:<,
!45/6-"745$89!-:%0 @#AB@C2-DEFGE $.=5)1.>./+&
!45/6-;)$89-:&3

#678'1.
9):.3,3*.
!)*$<03,&
!)*$+03,&
%&'()*)+,- ?.I*6,>./+
!45/6-4*-:32
?.&)0/ 2*)/)3'*-
./0)/..1)/0 .4'*5'+)6/
!)*$<03<< 45*9$0,0,1%
!"#$'0&%10%, !"#$%&%22
;:"!-;;4!$=)(2 !45/6-"745$89-:&%
;;4!$.!/2,
<.&+)/0 26H)/0

Figure 1 | EU Regulatory Environment for MDSW Development.

Figure 13: EU Regulatory Environment for MDSW Development. [5]

11 Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with regard
personal data and on the free movement of such data, and repealing Directive 95/46/EC (General Data Protection Regulation)
12 Directive (EU) 2016/1148 of the European Parliament and of the Council of 6 July 2016 concerning measures for a high common level of sec

information systems across the Union

© Asphali

58
WHITE PAPER
The EURegulatory Environment of Medical DeviceSoftwareDevelopment

D Standard and guidance documents useful to demon-

strate MDSW compliance with MDR
A comprehensive list of the currently available standards and guidance documents recommended to be followed by
MDSWdevelopers to achieve compliance with applicable GSPRs, as well as their latest versions, is shown below:

Requirement Standard or Guidance Title

EN ISO 13485:2016/AC:2018 (*) Medical devices - Quality management systems - Requirements for
regulatory purposes
ISO/IEC 25020:2019 Systems and software engineering — Systems and software
Quality Requirements and Evaluation (SQuaRE) — Quality measurement
framework
Quality
ISO/IEC 25012:2008 Software engineering -- Software product Quality Requirements and Evaluation
Management (SQuaRE) -- Data quality model
System (QMS) ISO/IEC/IEEE 15939:2017 Systems and software engineering — Measurement process
MDCG 2019-11 Qualification and classification of software - Regulation (EU) 2017/745 and
Regulation (EU) 2017/746
IMDRF/SaMD WG/N23 FINAL: 2015 Software as a Medical Device (SaMD): Application of Quality Management System
IMDRF/SaMD WG/N10 FINAL:2013 Software as a Medical Device (SaMD): Key Definitions
EN ISO 14971:2019 (*) Medical devices - Application of risk management to medical devices
IEC 80001-1:2010 Application of risk management for IT-networks incorporating medical devices — Part
(series) 1: Roles, responsibilities and activities
EC/TR 80002-1:2009 Medical device software — Part 1: Guidance on the application of ISO 14971 to medical
device software
Risk Management ISO/TS 25238:2007 Health informatics — Classification of safety risks from health software
EN ISO 15223-1:2016 (*) Medical devices - Symbols to be used with medical device labels, labelling and
System (RMS) information to be supplied - Part 1:
General requirements
IMDRF/SaMD WG/N12 FINAL:2014 "Software as a Medical Device": Possible Framework for Risk Categorization and
Corresponding Considerations
IMDRF/AE WG/N43 FINAL:2020 & IMDRF terminologies for categorized Adverse Event Reporting (AER): terms,
Annexes terminology structure and codes
EN ISO 14155:2020 (*) Clinical investigation of medical devices for human subjects - Good clinical practice

Clinical
MDCG 2020-1 Guidance on clinical evaluation (MDR) / Performance evaluation (IVDR) of medical
Evaluation device software
IMDRF/SaMD WG/N41 FINAL:2017 Software as a Medical Device (SaMD): Clinical Evaluation
ISO/IEC 27000:2018(en) Information technology — Security techniques — Information security management
(series) systems — Overview and vocabulary
ISO 27799:2016 Health informatics — Information security management in health using ISO/IEC 27002

Cybersecurity
IEC/CD 81001-5-1 (draft 2021) Health software and health IT systems safety, effectiveness and security — Part 5-1:
Security — Activities in the product lifecycle
MDCG 2019-16 rev.1 Guidance on cybersecurity for medical devices
IMDRF/CYBER WG/N60FINAL:2020 Principles and Practices for Medical Device Cybersecurity
IEC 62366-1:2015 (*) Medical devices - Application of usability engineering to medical devices
ISO 9241-210:2010 Ergonomics of human-system interaction - Human-centered design for interactive
Usability systems
ANSI/AAMI HE75:2009/(R)2018 (*) Human factors engineering- Design of medical devices
AAMI TIR50:2014 (*) Post-market surveillance of use error management
EN 62304:2006/AC:2008(*) Medical device software - Software life-cycle processes
IEC 82304-1:2016 Health software — Part 1: General requirements for product safety
Software lifecycle ISO/IEC 14764:2006 Software Engineering — Software Life Cycle Processes — Maintenance
IMDRF/MC/N35 FINAL:2015 Statement regarding Use of IEC 62304:2006 "Medical device software -- Software life
cycle processes"
(*) Although they are not software-specific, these standards are highly relevant for the development of MDSW.

Table 1 | Standard and guidance documents useful to demonstrate MDSW compliance with MDR.
Table 6: Standard and guidance documents useful to demonstrate MDSW compliance with
MDR. [5]

© Asphalion S.L. 2020 | P. 6

59
E Number of examples grouped by cognitive distor-
tions

Figure 14: Number of examples grouped by cognitive distortions [Own Creation]

F Support Vector Machine Hyperplanes

Figure 15: There are numerous alternative hyperplanes that might be used to divide the
two groups of data points (left image). Finding a plane with the greatest margin—that is,
the greatest separation between data points from both classes—is our goal (right image) in
SVM. Images obtained from Towards Data Science [6], a website.

60
References
[1] J. A. Cully and A. L. Teten. A Therapist’s Guide to Brief Cognitive Behavioral Therapy.
Department of Veterans Affairs South Central MIRECC, Houston, 2008.
[2] J. Calvo. 2022. Aprendizaje por transferencia: NLP Blog Europeanvalley. [online]
Europeanvalley. es. Available at: https: // www. europeanvalley. es/ noticias/
aprendizaje-por-transferencia-nlp/ [Accessed, 26, September 2022.
[3] Introduction — Machine Learning — Google Developers — develop-
ers.google.com. https://developers.google.com/machine-learning/guides/
text-classification. [Accessed 26-Sep-2022].
[4] MA. Courtney E. Ackerman. CBT Techniques: 25 Cognitive Behavioral Ther-
apy Worksheets — positivepsychology.com. https://positivepsychology.com/
cbt-cognitive-behavioral-therapy-techniques-worksheets/. [Accessed 27-Sep-
2022].
[5] Asphalion - Scientific and Regulatory Affairs consultancy — asphalion.com. https:
//www.asphalion.com. [Accessed 06-Dec-2022].
[6] Rohith Gandhi. Support Vector Machine — Introduction to Machine Learn-
ing Algorithms — towardsdatascience.com. https://towardsdatascience.com/
support-vector-machine-introduction-to-machine-learning-algorithms-934a444fca47.
[Accessed 01-Jan-2023].
[7] Payscale - Salary Comparison, Salary Survey, Search Wages — payscale.com. https:
//www.payscale.com. [Accessed 06-Oct-2022].
[8] Mental health — who.int. https://www.who.int/health-topics/mental-health#
tab=tab_1. [Accessed 25-Sep-2022].
[9] The treatment gap in mental health care - PubMed — pubmed.ncbi.nlm.nih.gov. https:
//pubmed.ncbi.nlm.nih.gov/15640922/. [Accessed 11-Jan-2023].
[10] Michiyo Hirai and George A. Clum. A meta-analytic study of self-help interventions for
anxiety problems. Behavior Therapy, 37(2):99–111, June 2006.
[11] Benjamin Shickel, Scott Siegel, Martin Heesacker, Sherry Benton, and Parisa Rashidi.
Automatic detection and classification of cognitive distortions in mental health text.
Automatic Detection and Classification of Cognitive Distortions in Mental Health Text,
9 2019.
[12] Ignacio de Toledo Rodriguez, Giancarlo Salton, and Robert Ross. Formulating auto-
mated responses to cognitive distortions for cbt interactions. 2021.
[13] Daniel M Low, Laurie Rumker, John Torous, Guillermo Cecchi, Satrajit S Ghosh, and
Tanya Talkar. Natural language processing reveals vulnerable mental health support
groups and heightened health anxiety on reddit during covid-19: Observational study.
Journal of medical Internet research, 22(10):e22635, 2020.

61
[14] Degree final project — FIB - Barcelona School of Informatics —
fib.upc.edu. https://www.fib.upc.edu/en/studies/bachelors-degrees/
bachelor-degree-informatics-engineering/degree-final-project. [Accessed
29-Nov-2022].

[15] Writing a GDPR-compliant privacy notice (template included) - GDPR.eu — gdpr.eu.

https://gdpr.eu/privacy-notice/. [Accessed 05-Dec-2022].

[16] Art. 13 GDPR - Information to be provided where personal data are col-
lected from the data subject - GDPR.eu — gdpr.eu. https://gdpr.eu/
article-13-personal-data-collected/. [Accessed 05-Dec-2022].

[17] Chapter 3 (Art. 12-23) Archives - GDPR.eu — gdpr.eu. https://gdpr.eu/tag/

chapter-3/. [Accessed 05-Dec-2022].

[18] CO2 emissions per kWh in Spain - Nowtricity — nowtricity.com. https://www.

nowtricity.com/country/spain/. [Accessed 10-Jan-2023].

[19] API Reference — scikit-learn.org. https://scikit-learn.org/stable/modules/

classes.html#module-sklearn.naive_bayes. [Accessed 01-Jan-2023].

[20] Why 85 https://www.techrepublic.com/article/why-85-of-ai-projects-fail/.

[Accessed 13-Jan-2023].

62