MASTER THESIS

TITLE: Design and development of a students’ performance predicting LMS utilizing

machine learning based on mental stress level measured through a Bluetooth enabled
smart watch

MASTER DEGREE: Master's degree in Applied Telecommunications and Engineering

Management (MASTEAM)

AUTHOR: Syed Haider Ali Kazmi

ADVISOR: Cristina Barrado, Doctor Faiyaz, Angelica Reyes

DATE: June, 5th 2023

0 Master Thesis

Abstract

Stress and academic anxiety problems can negatively impact numerous

aspects of students’ lives, resulting in degrading their academic achievement,
quality of life, and social behaviour. Various research suggests that depression
is associated with lower academic performance of students. The aim of this
research is twofold. Firstly, in order to establish a correlation between students’
mental stress level and their academic performance, a dataset has been
compiled through gathering the data by conducting a survey in a university
located in Punjab, Pakistan. The questionnaires were based on measuring the
stress level of students using Perceived Stress Scale (PSS) , Cognitive
performance assessment scale, in addition to some other demographic
questions. Afterwards, this dataset has been analysed utilizing various
machine learning algorithms. The second objective was to develop an
innovative, affordable and smart performance predicting Learning
Management System that takes into account students’ mental stress while
predicting the students’ performance using machine learning models. The
technique that was used for the mental stress measurements of the students
was based on a phenomenon known as the Heart Rate Variability (HRV). A
smart watch was utilized to measure the Heart Rate Variability of the students
that was used to assess the stress level of students in academics. A Machine
Learning (ML) model was trained using various parameters that were derived
from the Heart Rate Variability. The original dataset that was used to train the
model is known as Swell dataset. The SWELL dataset consists of HRV indices
computed from the multimodal SWELL knowledge work dataset for research
on stress and user modelling. The ML model effectively made prediction about
the stress levels of the students with an accuracy of 98.1%.
CONTENTS

CHAPTER 1. INTRODUCTION.......................................................................... 1

1.1. Research motivation ....................................................................................... 1

1.2. Overall aim ........................................................................................................ 2

1.3. Research agenda ............................................................................................. 2

CHAPTER 2. THEORETICAL FRAMEWORK ................................................... 3

2.1. Machine learning .............................................................................................. 3

2.1.1. Supervised & Unsupervised machine learning ........................................................ 3
2.1.2. Decision Tree......................................................................................................... 4
2.1.3. Boosting algorithms ............................................................................................... 5
2.1.4. Light Gradient Boosting Machine............................................................................ 6

2.2. Background and related work ........................................................................ 7

2.2.1. Student performance evaluation in educational data mining ................................... 7
2.2.2. Predicting the academic performance of middle- and high-school students using
machine learning algorithms ............................................................................................ 7
2.2.3. Tracking and predicting student performance in degree programs .......................... 7

2.3. Heart Rate Variability ....................................................................................... 8

CHAPTER 3. METHODOLOGY ....................................................................... 10

3.1. Project methodology...................................................................................... 10

3.1.1. SWELL Dataset ................................................................................................... 11
3.1.1. Using Apple smart watch to measure HRV ........................................................... 12
3.1.2. Measuring stress threshold level of each individual .............................................. 12

3.2. ML model performance measures .............................................................. 13

3.2.2. Confusion Matrix .................................................................................................. 13
3.2.3. Accuracy.............................................................................................................. 14
3.2.4. Precision .............................................................................................................. 14
3.2.5. Recall .................................................................................................................. 15
3.2.6. F1 score .............................................................................................................. 15

3.3. Ethical and societal considerations............................................................ 15

CHAPTER 4. DATASET FEATURES AND ANALYSIS .................................. 16

4.1. PSS-Score ....................................................................................................... 16

4.2. Cognitive performance .................................................................................. 16

4.3. Gender .............................................................................................................. 17

4.4. Age..................................................................................................................... 18

4.5. Self-study hours ............................................................................................. 19

2 Master Thesis

4.6. Number of absentees ..................................................................................... 20

4.7. Average Grade ................................................................................................ 20

CHAPTER 5. MODEL IMPLEMENTATION AND CONFIGURATION ............. 22

CHAPTER 6. RESULTS .................................................................................. 23

6.1. First Phase: Results Survey Dataset .......................................................... 23

6.1.2. LGBM classifier feature ranking ............................................................... 24

6.1.3. LGBM classifier model measures ............................................................ 25

6.1.4. LGBM classifier confusion matrix............................................................ 26

6.1.5. Snap Boosting Classifier Feature Ranking ............................................ 28

6.1.6. Snap Boosting Classifier Model Measures ............................................ 29

6.1.7. Snap Boosting classifier confusion matrix ............................................ 30

6.2. Second phase: mental stress measurements based on HRV ............... 30

6.2.1 Testing the Decision Tree classifier model with HRV parameters ..... 32

CHAPTER 7. DISCUSSION ............................................................................. 37

7.1. Analysis and discussion of results............................................................. 37

CHAPTER 8. CONCLUSIONS ......................................................................... 38

8.1. Future work ..................................................................................................... 38

......................................................................................................................... 39

REFERENCES ................................................................................................. 40
Master Thesis 1

CHAPTER 1. INTRODUCTION
“Stress” is commonly defined as the external pressures on a person's bodily and
mental well-being, whether they be physical or psychological. Stress is a process
of interpreting and adjusting to external events, not merely a stimulus or a
reaction. Humans are frequently affected by perceived and potential stress, which
leaves them open to psychological issues and negative effects on their physical
health. Increased level of stress may cause an individual to become disorganized
and unsure about their objectives and aspirations. This could make it difficult for
them to excel in their lives and, make it difficult to manage their time well in
accordance with the situation. Several research suggests that mental stress can
degrade the performance of individuals (Kiselica et al., 1994). In order to develop
a deep understanding of such a critical issue, an in depth research has been
conducted in this thesis, additionally, with the assistance of machine learning
techniques, and utilizing artificial intelligence, a mental stress predictive model
has been generated to understand how stress can hamper academic
performance. The model thus created will provide a better insight regarding
academic stress among students and would help in a better coordination between
students, parents and teachers.

Chapter 1 gives an introduction to the thesis topic including the overall aim and
research targets. A comprehensive scientific study was applied to the concept
and the theme was broken down into an objective approach in order to evaluate
the findings in a systematic way. In order to dive deeper into the content being
presented in the thesis, Chapter 2 tries to understand the underlying theory and
the brief description associated with the technology; the fundamental algorithms
and the governing concepts related to machine learning (ML). Chapter 3 is related
to the methodology that has been followed, and tries to dissect the various
options that were taken under consideration for conducting the experiment and
study. Chapter 4 gives detail insights of the dataset and provides an overview of
the data analysis approach. In chapter 5, the advantages of using IBM Watson
studio platform have been discussed. Additionally, we discuss the steps involved
in implementing and configuring an AI machine learning experiment using IBM
Watson studio. The experiment results are documented and presented in
Chapter 6 as dedicated tables, plot diagrams and bar graphs. The thesis is
concluded in Chapters 7 and 8, with an investigative overview of the research,
drawing on conclusions and justification of thesis topic as well as
recommendations to enhance further research.

1.1. Research motivation

Although short term stress might be considered as an academic incentive, long

term chronic stress can impair performance, hamper growth and can cause
health issues. Regarding the link between stress and students' academic
performance, contentious conclusions have been published in the literature.

Consistent research has revealed that students with exceedingly higher levels of
stress had poor grade point averages (Deng et al., 2022). Depending on the
2 Master Thesis

levels and factors, stress may or may not impede academic performance (Deng
et al., 2022). Students may experience exceeding stress and feelings of
academic burnout if they are unable to timely manage and finish their work load
in the designated time. Additionally, students who are under a lot of stress have
a tendency to put off tasks like finishing projects on time and meeting deadlines
(Lin et al., 2020). Naturally, this will have an impact on their ability to study and
the calibre of their work. This study is more inclined towards a general category
of classification which classifies how students’ academic performance is affected
through stress.

The motivation behind this research work is to assist students and their
teachers/parents to develop a better understanding of determining the impact of
mental stress in an academic environment. Teachers/parents would be able to
provide extra support to their students that would give these students a fair
chance to achieve success.

1.2. Overall aim

Through the use of analysis of higher education data gathered through a survey,
this project aims to assist students in achieving better academic performance
while categorizing their mental stress. In terms of students’ mental stress and
academic performance, it would be self-explanatory if a ML-modelled thematic
framework could be developed to assist university academic staff as well as
students in determining if a particular student is tapping his/her maximum
academic potential. The primary objective is to establish a correlation between
the students´ perceived stress and their academic performance. The secondary
aim of this quantitative study is to predict students mental stress while taking into
account their Heart Rate Variability (HRV).

The method section provides a more deep explanation of the ML parameters

being used. This study intends to assist university students and faculty members
in gauging the developments of the students’ stress and forecast future academic
outcomes and scores.

1.3. Research agenda

The main agenda and the focus of the study is broken down further into the
following objectives:
1. Determining the legitimacy of using ML algorithms for predicting student
performance in academics.
2. Determining if there is a possible correlation between students’ academic
performance and their mental stress.
3. Taking advantage of IBM Watson ML platform for implementing the
particular ML model that delivers the best accuracy, precision, recall and
F1 score when fed with data collected from university students.
4. Implementing a ML model to predict the mental stress using Heart Rate
Variability.
5. How accurately can we measure HRV from a wrist wearable device like a
smart watch?
Master Thesis 3

CHAPTER 2. THEORETICAL FRAMEWORK

Machine learning consists of a considerable amount of literature and research
studies on numerous algorithms. The main objective of this domain is to tackle
complex issues, hence, can be a daunting subject to approach.

2.1. Machine learning

Being a sub domain of artificial intelligence, machine learning targets problems

in the same approach as humans do i.e. formulating a computational ML
algorithms to assess the outcome of a challenge or problem. The basic driving
principle of a ML model is to mimic the neural networks enmeshed in the human
brain and making calculated conclusions through the information fed to the neural
network model. Figure 1, elaborates the various domains of the artificial
intelligence (William et al., 2021).

Fig. 1 Sub-domains of Artificial Intelligence

2.1.1. Supervised & Unsupervised machine learning

Supervised learning is a branch of machine learning and artificial intelligence that

utilizes labeled datasets to train algorithms for accurate data classification or
outcome prediction. In the training process, the algorithm adjusts its weights
based on input data until it achieves an appropriate model fit, typically through
cross-validation (IBM, n.d.). This methodology provides organizations with a
4 Master Thesis

potent tool to tackle various real-world issues, including spam filtering, image
recognition, speech analysis, and medical diagnosis, provided there is an ample
supply of labeled data (IBM, n.d.).

Supervised machine learning can be categorized into two primary types:

regression and classification. Classification problems involve mapping input data
to discrete class labels, such as distinguishing between "cat" or "dog," "malignant"
or "benign," or "spam" or "non-spam." Several popular classification algorithms
are available, including Naive Bayes, K-Nearest Neighbors, Decision Trees, and
Support Vector Machines. Regression problems, on the other hand, deal with
mapping inputs to continuous numerical outputs, like predicting the price of a
house. Regression analysis aims to establish relationships between independent
and dependent variables. Common regression algorithms include Linear
Regression, Ridge Regression, Lasso Regression, and Neural Network
Regression (Pykes, 2021).

Unsupervised learning, in contrast, is employed to uncover functional patterns,

associations, and relationships within data. It can also be used for dimensionality
reduction and data grouping. Clustering, a data mining technique, involves
identifying similarities and dissimilarities among uncategorized data points and
grouping them accordingly. The granularity and size of these groups are typically
determined using the K-number, with K-means being a popular clustering
algorithm.

2.1.2. Decision Tree

Decision trees are non-parametric supervised learning algorithms that can be

used for both classification and regression tasks. They have a hierarchical, tree-
like structure consisting of a root node, branches, internal nodes, and leaf nodes
(IBM, 2023).

Fig. 2 A typical Decision Tree (IBM, 2023)

Decision trees are a type of machine learning algorithm that uses a tree-like
structure to make decisions as depicted in figure 2. The algorithm starts with a
Master Thesis 5

root node, and branches off into internal nodes or decision nodes, which evaluate
available features to form homogenous subsets represented by leaf nodes. For
example, when deciding whether to go surfing, one might use a decision tree with
rules like "Is the temperature warm enough?" and "Is the wind too strong?". The
following decision rules can be followed to make the possible choice (IBM, 2023):

Fig. 3 Decision Tree initialization (IBM, 2023)

The flowchart, shown in figure 3, structure also creates an easy to digest

representation of decision-making, allowing different groups across an
organization to better understand why a decision was made (IBM, 2023).

To construct a decision tree, the algorithm follows a divide-and-conquer approach

to find the optimal split points in the data for classifying records into specific class
labels. This iterative process continues until either all records or the majority of
records are correctly classified. It is preferable to have smaller trees as they can
result in pure leaf nodes, while larger trees are more prone to overfitting and data
fragmentation. To address these concerns and reduce complexity, pruning
techniques can be employed. The accuracy of the decision tree model can be
assessed through cross-validation or by creating an ensemble using the random
forest algorithm, which combines multiple decision trees to improve predictive
performance (IBM, 2023).

2.1.3. Boosting algorithms

Boosting is a technique used to enhance the accuracy of weak learner algorithms

by transforming them into stronger learning algorithms. In boosting, each tree is
trained on a modified version of the original dataset. For instance, the AdaBoost
Algorithm assigns equal weights to each tree and trains them. After evaluating
the outcomes of the initial tree, the algorithm increases the weights of
observations that are challenging to classify and decreases the weights of those
6 Master Thesis

that are easier to classify. This iterative process allows subsequent trees to
classify observations that were previously misclassified. The final ensemble
model's predictions are determined by the weighted sum of the predictions made
by the preceding tree models (IBM, 2023).

2.1.4. Light Gradient Boosting Machine

Figure 4, shows the scheme that is adopted by typical decision tree algorithms
where the growth of the trees is level-wise.

Fig. 4 A symmetric level-wise tree growth strategy in which each node in a

particular level has child nodes, thus, developing an additional layer of depth
(Microsoft Corporation, 2023)

On the other hand, for Light Gradient Boosting Machine (LGBM) ,developed by
Microsoft, the growth of the trees is leaf-wise, as shown in figure 5. For growth,
LGBM would choose leaf with maximum delta loss.

Fig. 5 Leaf wise tree growth in which only the node having the highest delta loss is split.
Successive splitting occurs only on one side of the tree, thus, resulting in an
asymmetrical tree (Microsoft Corporation, 2023).

In terms of memory consumption and computational speed, LGBM can

substantially outperforms XGB (Ke et al., 2017).
Master Thesis 7

2.2. Background and related work

Numerous machine learning (ML) projects have been developed for forecasting
student performance. Some of the research work, related to the theme of this
thesis, are elaborated in the following sections

2.2.1. Student performance evaluation in educational data mining

Two machine learning models were created and a comparative analysis was
performed by researchers (Ahmed et al., 2021). The models being ANN model
and the Random Forest. TensorFlow was integrated at the backend for both the
models. They tried to develop the models in order to predict the academic
success using students previous academic evaluation and geographical data.
They achieved the required results by investigating that ANN can out preform the
Random Forest model by using a sizeable amount of data. The increased interest
among researchers in applying data mining techniques to evaluate student data
served as a prerequisite for this study. Future implications could include using
recurrent neural networks (RNN) for identifying students who are about to drop-
out from college.

2.2.2. Predicting the academic performance of middle- and high-

school students using machine learning algorithms

In their study (Rajendran et al., 2022), the authors utilized machine learning
techniques to develop models for predicting the academic performance of high
school students. The models took into account various socio-demographic
factors such as age, gender, obesity, average household income, family size, and
marital status of parents, as well as school-related variables like type of gender
education and academic level, and student-related variables such as stress and
lifestyle. The output variable considered in the models was the students' GPA.
The results showed that the gradient boosting method outperformed other
techniques, followed by random forest, in terms of generating better predictions.
The analysis of the models led to the conclusion that maintaining a health-
conscious lifestyle has a positive correlation with academic performance, while
the presence of stress has a negative impact. Nevertheless, the impact of gender
was not identified as a significant predictor of a student's academic performance.

2.2.3. Tracking and predicting student performance in degree

programs

This paper (Xu et al., 2017) addresses several new challenges in the field. The
authors propose an innovative approach for predicting the future performance of
students in degree programs based on their current and past performance. To
construct base predictors, they develop a course clustering method using a latent
factor model. They also introduce an ensemble-based progressive prediction
architecture to incorporate the evolving performance of students into the
prediction process. These data-driven methods can complement other
pedagogical approaches and provide valuable information for academic advisors.
8 Master Thesis

The information can be used to recommend subsequent courses to students and

implement pedagogical interventions if necessary. Furthermore, the findings of
this study have implications for curriculum design in degree programs and the
formulation of education policies. However, it should be noted that the study does
not account for students' performance and stress specific to a particular college
program, which according to the authors, could be significant factors influencing
the outcomes.

2.3. Heart Rate Variability

The second part of the thesis focuses on a technique used to measure the mental
stress of the students, which is based on a phenomenon called Heart Rate
Variability (HRV). HRV analysis is a tool increasingly utilized for non-invasive
analysis of the Autonomous Nervous System (ANS) in the human body. Its
analysis and contextual application have gained importance due to its sensitivity
to both physiological and psychological environmental factors. Altered HRV
measurements are extensively utilized for monitoring the arrhythmic
dysregulation of the Autonomous Nervous System. Additionally, HRV
measurements are employed to monitor and assess sleep patterns, stress levels,
drowsiness, and the effects of prolonged strenuous exercise training (Colom et
al., 2010).

Fig. 6 RR interval representing the distance between two successive heartbeats

HRV refers to the irregular fluctuations in the duration between successive

heartbeat intervals, which have been found to be influenced by stress in scientific
literature. Specifically, the RR interval, representing the distance between two
consecutive heartbeats (R waves of their QRS complexes), is directly associated
with HRV, as shown in Figure 6 (Colom et al., 2010).
Master Thesis 9

There are various approaches to reducing chronic stress and improving HRV.
One potential method involves modulating the autonomic nervous system. During
periods of stress, sympathetic nervous system responses become more
dominant, triggering the "fight or flight" response. This heightened sympathetic
activity leads to an increase in heart rate and a decrease in HRV. Conversely, the
parasympathetic nervous system, responsible for the "rest and digest" state,
becomes less active and restricted during stressful times (Welltory, 2023).

Low HRV: “ If the intervals in-between your heartbeats are relatively

equal, then you are in a fight or flight state and your HRV is quite low.”

High HRV: “If the interval length variates and you are in a more relaxed
state then your HRV is high. This is mostly associated with good recovery.”
10 Master Thesis

CHAPTER 3. METHODOLOGY
In this chapter project methodology that has been followed in the thesis has been
elaborated.

3.1. Project methodology

The study was divided in three main phases.

The first phase consisted of a literature study. The study looked at previous
related work in the form of research articles, surveys, journals, and e-books. This
process was done to familiarize the reader on the current state-of-the-art ML
techniques and to show a research gap, to justify the current research that was
being conducted. Google Scholar was used to find these resources.

In the second phase, in order to establish a correlation between perceived

mental stress and students´ academic performance a dataset was formed based
on a survey that was conducted by asking 298 students to fill a survey form
(attached in the Annex). All of the students were enrolled at Government College
University, Lahore, and were studying in the first semester of their undergraduate
degree. The stress levels of the students were measured using a Perceived
Stress Scale (PSS) which is a classic stress assessment instrument. “The tool,
while originally developed in 1983, remains a popular choice for helping us
understand how different situations affect our feelings and our perceived stress.
The questions in this scale ask about your feelings and thoughts during the last
month. Individual scores on the PSS can range from 0 to 40 with higher scores
indicating higher perceived stress.” (NH Dept. of Administrative Services, n.d.)

As per the Perceived Stress Scale (PSS) criteria, following are the thresholds
defined (NH Dept. of Administrative Services, n.d.):

• Scores ranging from 0-13 would be considered low stress.

• Scores ranging from 14-26 would be considered moderate stress.
• Scores ranging from 27-40 would be considered high perceived stress.

The survey form consisted of five direct questions i.e., asking the students to
specify their age, gender, self-study hours, number of times they skipped their
school day and grades in the last three graded activities. The stress and cognitive
performance levels of the students were gauged through the two psychological
assessment scales (Perceived Stress Scale and Cognitive assessment scale, as
attached in the appendix).

The third phase was an experimental phase. In this phase, deep analysis was
performed with the help of a Bluetooth enabled Apple smart watch. A total of four
students were considered for this experimental stage. Informed consent was
obtained from each of the four students. The age group of the four students varied
between 20 – 23 years.
Master Thesis 11

The experimental phase was conducted during the mid-semester exams of the
students. In the experimental phase, the perceived mental stress and Heart Rate
Variability (HRV) was measured with the help of the Perceived Stress Scale
(PSS) and Apple smart watch, respectively. The students were asked to wear the
Apple watch during their exams.

3.1.1. SWELL Dataset

The model was trained using various parameters that were derived from the Heart
Rate Variability. These HRV parameters were extracted from an original dataset,
known as Swell dataset, that was used to train the model. The SWELL dataset
consists of HRV indices computed from the multimodal SWELL knowledge work
dataset for research on stress and user modelling (Hazer-Rau et al., 2020).

The original dataset consists of data that was captured using the following means
(Hazer-Rau et al., 2020);

• Computer interactions, via a computer logging tool

• Facial expressions, via a webcam
• Body postures, via a Kinect 3D camera
• Physiology (ECG and skin conductance), via body sensors

As the third phase of this research work revolves around the idea of predicting
mental stress from HRV (gathered through a smartwatch) of the students, hence,
the model was developed using only the features related to the HRV (in Time
Domain) that are as following (Shaffer & Ginsberg, 2017);

• RR: Interval between two heartbeats

• MEAN_RR: mean of RR-interval
• MEDIAN_RR: median of RR-interval
• SDRR: standard deviation of the RR-intervals
• RMSSD: It’s a measure for how much variation there exists in the heart
rate. In a healthy heart, there is a natural variation, which is due to a
balance between the sympathetic nervous system (SNS) and
parasympathetic parts (PSNS) of the Autonomous Nervous System (Singh
et al., 2018). If your body experiences stress, then the sympathetic system
will activate, to prepare for fight or flight behaviour, and your heartrate will
increase.
• SDSD: the standard deviation of the differences between successive NN
intervals
• HR: Heart Rate
• pNN25: The number of pairs of successive RR-intervals that differ by
more than 25 ms (normal RR-intervals are often called NN-intervals)
• Output condition: Stress/no stress
12 Master Thesis

Table 1. HRV parameters calculations (Kim et al., 2022)

Table 1, shows the expressions used to calculate the HRV parameters in time
domain. Once the HRV data of the four students was extracted from the Apple
smart watch, these expressions were used to calculate the HRV parameters and
fed to the trained (with SWELL dataset) ML model to evaluate the performance
of the model.

3.1.1. Using Apple smart watch to measure HRV

Different wearable watches and gadgets have been studied and evaluated for HR
estimation measurement. For a total of one minute of effective granularity, the
Apple Watch has the best performance estimation (Hernando et al., 2018). This
app stores the raw RR values, with a precision of centi-seconds, in the user’s
Personal Health Record, accessible to be exported in XML format using Apple’s
Health App (Hernando et al., 2018).

3.1.2. Measuring stress threshold level of each individual

According to Dr Andrea Dinardo, different individuals have different threshold

levels as shown in figure 7. She further elaborates that “Thresholds are more
individualistic while tipping points are more universal. Thresholds vary from
person to person (e.g Type A vs Type B), situation to situation (e.g Work or
Personal Reasons), and are based on individual strengths and challenges
(Dinardo, 2016).
Master Thesis 13

Fig. 7 How thresholds affect an individual stress response (Dinardo, 2016)

To establish a baseline stress measurement indicator, four different students

underwent a pre-experimental phase in which research methods were employed.
This phase consisted of two consecutive 20-minute stages: a relaxation stage
and a stress stage, separated by a 1-minute break. In the relaxation stage, the
students were instructed to listen to pleasant orchestral music. During the stress
stage, they were required to complete an online version of the Stroop test, an
attentional test that involves identifying the ink colour of words while disregarding
their literal meaning (Hernando et al., 2018).

3.2. ML model performance measures

In this thesis, the typical parameters that have been used in evaluating the
performance of a classification ML model based are as following;

3.2.2. Confusion Matrix

The confusion matrix is a summary of the classification model's performance in

predicting outcomes across different classes. Vertical axis represents the
predicted labels by the model and the horizontal axis represents the actual labels
in the dataset.
14 Master Thesis

Fig. 8. Confusion matrix consisting of True Positives (TP), False Positives (FP), False
Negatives (FN), Ture Negatives (TN) (Suresh, 2021)

3.2.3. Accuracy

Accuracy of an algorithm is represented as the ratio of correctly classified

(TP+TN) instances to the total number of instances (TP+TN+FP+FN) (Singh et
al., 2021).

3.2.4. Precision

Precision is represented as the ratio of correctly classified instances with positive

class (TP) to the total instances predicted to have the positive class (TP+FP)
(Singh et al., 2021).
Master Thesis 15

3.2.5. Recall

Recall metric is defined as the ratio of correctly classified instances with positive
(TP) to total number of instances who have actually positive class(Singh et al.,
2021).

3.2.6. F1 score

F1 score is also known as the F Measure. The F1 score states the equilibrium
between the precision and the recall (Singh et al., 2021).

3.3. Ethical and societal considerations

Informed consent was obtained from all the participants while conducting this
research work. None of the participants were minor or underage. The survey as
well as the details of the participating four students in this research work is kept
anonymous for privacy reasons. The purpose of the research was restated to all
of the participants. Moreover, while conducting the survey, special permission
was taken from the departmental director of the university.

This research will assist students and their teachers/parents to develop a better
understanding of determining the impact of mental stress in an academic
environment. Teachers/parents would be able to provide extra support to their
students that would give these students, irrespective of their gender, a fair chance
to achieve success.
16 Master Thesis

CHAPTER 4. DATASET FEATURES AND ANALYSIS

The dataset was compiled by gathering the responses of 298 students. It has
seven distinct features which are as discussed in this chapter.

4.1. PSS-Score

Figure 9 depicts the perceived stress of the students. This feature is based on
score obtained from a Perceived Stress Scale (PSS) questionnaire.

As mentioned in the chapter 3, following the criteria given by Perceived Stress

Scale, scores ranging from 0-13, 14 – 26 and 27 – 46 were considered as low
stress level, medium stress level and high stress level, respectively. In the
dataset, the three levels of stress i.e., low stress, moderate stress and high
stress, were codified in three numerical classes i.e., 0, 1 and 2, respectively.

Fig. 9 PSS score calculated for all the 298 students. Among these, 70 students
had no stress, 171 students had medium stress level and 57 students had high
stress level

4.2. Cognitive performance

Cognitive ability refers to the capacity of the human brain to process, store and
retrieve information. It also refers to innate functions of the brain which include
attention, memory and reasoning ability. According to Sternberg and Sternberg
(2009), it is the essential psychological component for people to successfully
complete an activity.
Master Thesis 17

In order to measure the cognitive performance level of the students, a cognitive

performance assessment scale was utilized. This scale is based on 25 questions.
“The most straightforward way to score the scale is simply to add up the ratings
of the 25 individual questions, yielding a score from 0-100. Scores on the scale
predict episodes of absent-mindedness in both the laboratory and everyday life,
including slow performance on focused attention tasks, traffic and work accidents,
and forgetting to save one's data on the computer.” (UC Berkeley, 2011).

Fig. 10 Cognitive performance indicator

Figure 10 shows the cognitive performance score of the students. This score (0-
100) has been codified ranging from a scale of 0 -10, with 0 being the lowest
(worst) score and 10 being the highest (best). On vertical axis we can observe
the cognitive performance score of the students whereas on the horizontal axis
the number of students.

4.3. Gender

Both male and female students were considered in the survey. Male participants
were assigned a “0” coded value, whereas, the female participants were
represented by “1” as shown in figure 11.
18 Master Thesis

Fig. 11 Gender classification of the students

4.4. Age

This feature represents the age group of the students taking into account only the
number of years. The age of the students was not codified.

Fig. 12 Age of students in the survey

As shown in the figure 12, frequently the students falls in the age group of 18 to
21 years. There were also a small number of students who didn’t specify their
age, depicted as missing in the figure. The missing field were left blank in the
dataset on which the model was trained.
Master Thesis 19

4.5. Self-study hours

This feature depicts the total number of hours spent by any student on his/her
self-study at home or on campus.
As shown in figure 13, approximately 120 students spent between 2 to 4 hours a
day, studying by themselves doing homework and other academic related tasks.
About 45 students preferred not to answer this question, depicted as missing in
the figure. The missing field were left blank in the dataset on which the model
was trained.

Fig. 13 Average number of hours spent by students, each day, on self-study

20 Master Thesis

4.6. Number of absentees

This feature represents the number of days any student was absent from the
class (i.e., absent for the whole working day) during the past three months in the
on-going semester.

Fig. 14 Absence of students during the last three months

As depicted in figure 14, about 110 students skipped 0 to 3 lecture days, being
the most frequent. About 38 students preferred not to answer this question,
shown as missing.

4.7. Average Grade

The performance of the students has been analysed through this feature.
Students were asked to mention their scores in terms of percentage in their last
three graded class activities. Then the average of these three graded class
activities was taken and classified in three coded values i.e., 0, 1 and 2,
corresponding to percentage score falling in three categories i.e., 0-50 %, 51-
70% and 71- 100 %, respectively, representing 0 the lowest and 2 the best
academic performance.
Master Thesis 21

Fig. 15 Average grade of students participating in the survey

As shown in figure 15, 170 students had an normal performance in the last three
graded activities with the score falling between 51-70%, thus, classified as 1.
Around 66 students had poor performance with a score falling between 0-50 %,
classified as 0. Approximately, 62 students had a good score with average grade
between 71- 100 %.
22 Master Thesis

CHAPTER 5. MODEL IMPLEMENTATION AND

CONFIGURATION
In this study, IBM Watson Studio has been utilized to build and deploy machine
learning models. Watson Studio empowers data scientists, developers and
analysts to build, run and manage AI models, and optimize decisions anywhere
on IBM Cloud Pak for Data.

The following steps are to be followed to develop a machine learning model on

the IBM Watson Machine Learning Platform.

• In the first step, Watson Studio is deployed in the IBM cloud.

• Two resources i.e., IBM Cloud Object Storage and IBM Machine Learning
service instance are created while being linked to the Watson studio
space.
• Afterwards, the user creates a new project space in the Watson studio.
• Upload the data set in .csv format so the data can be fed to the ML model.
• Create the ML experiment that will automatically analyse the given tabular
data and will generate the candidate model pipelines customized for the
predictive modelling.
• Selection of output variable that we need the model to predict. As this
study consisted of building two distinct models, thus, two out variables
were considered, i.e., average grade and stress level. The output variable
i.e., average grade has a multiclass with 0, 1 and 2, corresponding to
percentage score falling in three categories i.e., 0-50 %, 51-70% and 71-
100 %, respectively.
• 90 percent of the dataset was used to train the model whereas the 10%
data was used to test the model accuracy.
• Choice of selecting algorithms either by the user or by the machine. IBM
Watson machine gives the option to choose from eleven different
algorithms which are appropriate for any kind of data and may include
multiple variables under investigation.
• Run the experiment that will make the progress map available for review.
The progress map includes several important steps including reading the
data set, splitting the holdup data (10% of the total), reading the remaining
90% data, pre-processing, model selection, and then it moves to selected
algorithms for further processing.
• Once the experiment completes, it will provide multiple tables showing the
feature ranking, performance measures, confusion matrix, etc,.
Master Thesis 23

CHAPTER 6. RESULTS
The dataset, compiled from both the survey based on the questionnaire and the
Apple smart watch was analysed, separately, using IBM Watson ML experiment.
The study was divided in two phases as described in the following sections

6.1. First Phase: Results Survey Dataset

Fig. 16 Progress map showing different stages in IBM ML experiment.

respectively.

Figure 16 shows the detailed progress map of the ML model building using IBM
Watson studio. The experiment consists of various steps as shown in the figure.
In the first step, the dataset is read by the machine. Then the dataset is split into
training and testing data. Afterwards, the data is cleaned in the Pre-processing
stage. In the last step, model that is best suited for the dataset, in terms of
accuracy, is selected. The users have the option to set the numbers of models
that they want to observe in the output. These models are fine-tuned by applying
feature engineering and hyperparameter optimization.

The two models that gave the best performance measure were chosen, namely,
LGBM classifier and Snap Boosting Machine Classifier, with the Cross Validation
accuracy of 71.4% and 71.0%. Feature engineering and Hyperparameter
optimization was applied to further improve the accuracy of the two models, thus
generating different pipelines as shown in figure 16.
24 Master Thesis

6.1.2. LGBM classifier feature ranking

Table 2 shows the ranking based on importance for various features of the
dataset. As recommended by the IBM machine, any feature with the importance
above 65% should be considered while the feature having importance below this
threshold may be neglected for the optimum results.

Table 2. Feature ranking using LGBM classifier

As depicted in the Table 2, the four most important features using the LGBM
classifier comes out to be, the total number of days any student skipped the class
lectures, the perceived stress scale score, the cognitive performance and the age
of the students. This shows us that stress level of the students, with a feature
importance of 99%, has an important role in predicting the average grades of the
students.
Master Thesis 25

6.1.3. LGBM classifier model measures

Table 3, shows various model measures of the LGBM classifier. The accuracy
of the LGBM classifier is 71.4%, impacted by the size of dataset and missing
values in the dataset, could be improved further by increasing the dataset size.

Table 3. LGBM classifier model measures The Cross Validation accuracy score
of the LGBM classifier comes out to be 71.4%

As depicted in table 3, we have a multi-class classification, consequently, IBM

adopted the averaging methods for F1 score calculation, resulting in a set of
different average scores i.e., weighted and macro F1 measures (Leung, 2022).

By taking the arithmetic mean of all the per-class F1 scores, macro F1 score is
calculated. This method treats all classes equally regardless of their support
values (Leung, 2022). Since, we have an imbalanced dataset, hence, the F1
macro score of the model should be given importance, which in our case comes
out to be 0.713, signifying above average performance of the model.
26 Master Thesis

6.1.4. LGBM classifier confusion matrix

The accuracy of the model is derived from the Confusion Matrix. Tables 4, 5 and
5 shows the confusion matrix of the stress levels of one class compared with the
other two classes.

Table 4. Confusion matrix predicting “0” stress level (no stress) against the other
two levels i.e., “1” (medium stress) and “2” (high stress)

From table 4, it can be observed that the overall accuracy of predicting “0” stress
level is 72.4%, with True Positives (i.e., individuals who had no stress and were
correctly identified by the ML model) are equal to 5 and False Negatives (i.e.,
individuals who had stress and were incorrectly identified as not having stress by
the ML model) are also equal to 5. Similarly, the True Negatives (i.e., individuals
who had stress and were correctly identified by the ML model) are equal to 16
and False Positives (i.e., individuals who had no stress and were incorrectly
identified as having stress by the ML model) are also equal to 3.
Master Thesis 27

Table 5. Confusion matrix predicting “1” stress level (medium) against the
other two levels i.e., “0” (no stress) and “2” (high stress)

From table 5, it can be observed that the overall accuracy of predicting “1” stress
level is 65.5%, compared with the “0” (no stress) and “2” (high stress) output
classes.

Table 6. Confusion matrix predicting “1” stress level (medium) against the other
two levels i.e., “0” (no stress) and “2” (high stress)

Table 6 depicts that the overall accuracy of predicting “2” (high stress level) is
79.3%, compared with the “0” (no stress) and “1” (medium stress) output classes.
28 Master Thesis

6.1.5. Snap Boosting Classifier Feature Ranking

The ranking based on importance of the features, while using Snap Boosting
classifier, is shown in Table 7.

Table 7. Feature ranking using Snap Boosting classifier

The most important features using the Snap Boosting classifier comes out to be,
the cognitive performance of the students. The importance of the perceived stress
in this case is only 16%, which implies that this models is not suitable for
predicting the average grades of the students based on their stress level.
Master Thesis 29

6.1.6. Snap Boosting Classifier Model Measures

Table 8 shows the performance parameters of the Snap Boosting model.

Table 8. Snap Boosting classifier model measures. The Cross Validation

accuracy score comes out to be 71.0%

From table 8, it can be deduced that Snap Boosting model has more or less
similar performance measures as compared to LGBM classifier model. But, this
model is not recommended to predict the grades of the students based on their
stress levels as it shows high correlation for cognitive performance, only.
30 Master Thesis

6.1.7. Snap Boosting classifier confusion matrix

As depicted in table 9, the Hold-out accuracy of the Snap Boosting classifier is

69.0%.

Table 9. Confusion matrix of Snap Boosting classifier

6.2. Second phase: mental stress measurements based on HRV

Using the modified version of the SWELL dataset, a model was built and
deployed on the IBM Watson studio. The top performing model in this case came
out to be Decision Tree Classifier.

Table 10. Confusion matrix of Decision Tree classifier

Table 10, depicts the confusion matrix of the Decision Tree classifier. It can be
seen that the accuracy of the classifier comes out to be 98.5% which proves that
the model is best suited for predicting the stress level of the individuals using
HRV parameters.
Master Thesis 31

Table 11. Model evaluation measures of the Decision Tree classifier

Table 11 shows the model evaluation measures in terms of performance. Both

holdout and cross validation scores are shown. The holdout and cross validation
accuracies are 98.5% and 98.1%, respectively. The model has a F1 scores of
0.987 and 0.982 corresponding to the holdout score and cross validation score,
respectively, which signifies that the model has an excellent performance
measures in terms of precision and recall, thus, resulting in a well-balanced
model.
32 Master Thesis

Table 12. Feature ranking of Decision Tree classifier model. The features of the
dataset are ranked according to their corresponding correlation in predicting the
stress

As shown in table 12, the most important feature having the highest correlation
in predicting the stress is median of RR intervals between two heartbeats.

6.2.1 Testing the Decision Tree classifier model with HRV parameters

After developing this model, the next step was to gather the data related to HRV
of four students, in order to confirm the viability of using this model for HRV
measured through the Apple smart watch.

In parallel to measuring the stress levels of the students with the Apple smart
watch, the stress score of the students was also measured with the Perceived
Stress Scale.
Master Thesis 33

Table 13. the PSS score, cognitive assessment (CA) score, age and grades
achieved i.e., performance in the exams of the four students

Grade CA
Age PSS score
level score
Sample 1 Sample 2 Sample 3
Student 1 22 1 29 14 ( moderate stress) 09 (no stress) 13 ( moderate stress)
Student 2 23 1 47 16 (moderate stress) 23 ( high stress) 20 (moderate stress)
Student 3 20 1 39 19 (moderate stress) 16 ( moderate stress) 20 (moderate stress)
Student 4 21 2 58 26 (high stress) 29 (high stress) 30 (high stress)

For each of the four students, three samples (during the examination time period)
after every 3 days were taken with the help of PSS. From table 13, It can be
observed that student 4 (an outlier) was constantly under high stress during the
examination days, but irrespective of the high stress level, achieved an above
average grades in the three exams. This may be due to the high cognitive
performance score.

The HRV measurements were taken in two phases. In the first phase, the
measurements were taken in order to determine the baseline threshold level of
HRV parameters for the four students, whereas, the second phase was
implemented during the mid-term examination days.

Table 14. HRV parameters measured during relaxing stage of the four students

Measurements during relax stage

HRV Parameters Student 1 Student 2 Student 3 Student 4
SDRR 140.97 90.37 62.76 517.53
MEAN_RR 885.15 881.75 809.62 923.28
MEDIAN_RR 853.76 893.4 811.18 617.79
RMSSD 15.55 15.72 19.21 9.96
SDRR_RMSSD 9.063 5.74 3.26 51.93
HR 69.49 68.8 74.56 81.34
pNN25 11.13 11.8 20.2 1.2

The HRV, during the relax stage, of the four students was recorded with the help
of the Apple smart watch and the corresponding parameters were calculated as
shown in table 14.
34 Master Thesis

Table 15. HRV parameters measured during stress stage

Measurements during stress stage
HRV Parameters Student 1 Student 2 Student 3 Student 4
SDRR 81.31 84.49 57.88 199.96
MEAN_RR 939.42 898.186 848.62 793.61
MEDIAN_RR 948.35 907.0 851.57 720.74
RMSSD 12.96 16.3 14.14 14.786
SDRR_RMSSD 6.27 5.182 4.093 13.523
HR 64.36 67.45 71.04 79.49
pNN25 5.6 13.06 8.13 8.26

Table 15 illustrates the HRV measurements calculated during the stress stage,
of the four students.

Considering the results obtained, in tables 14 and 15, it can be observed that
the median value of the RR interval increases during the stress stage of the
students whereas the heart rate (HR) decreases.

After calculating the HRV parameters during both the relax and stress stages,
these parameters were fed into the previously deployed model on IBM cloud and
the model made predictions about the stress level of the students. As depicted in
table 16, all the predictions were made correctly.

Table 16. Predictions made by the Decision Tress Classifier model, deployed on
the IBM Cloud
Master Thesis 35

Similarly, the HRV recording was taken into account before and during the mid-
term examination of the four students as shown in tables 17 and 18, respectively.

Table 17. HRV measurements during the examination days of the students (measurements
taken before the start of the exam)

Measurements (before exam)

HRV Parameters Student 1 Student 2 Student 3 Student 4
SDRR 63.96 70.86 65.096 119.70
MEAN_RR 802.85 704.35 958.968 738.15
MEDIAN_RR 804.29 695.00 958.817 725.53
RMSSD 14.2348 20.419 11.787 13.936
SDRR_RMSSD 4.4938 3.470 5.522 8.589
HR 75.22 86.01 62.86 83.06
pNN25 6.8 23.26 4.0 6.2

Table 18. HRV measurements during the examination days of the students
(measurements taken while the students were taking their exam)
Measurements during exam
HRV Parameters Student 1 Student 2 Student 3 Student 4
SDRR 385.751 124.941 141.86 32.84
MEAN_RR 1017.31 696.782 1028.88 792.47
MEDIAN_RR 970.72 710.4 1020.40 796.62
RMSSD 15.868 15.723 15.2304 7.556
SDRR_RMSSD 24.30 7.946 9.31422 4.347
HR 68.096 89.159 59.452 75.84
pNN25 10.466 8.06 10.4 0.4
36 Master Thesis

After calculating the HRV parameters during both before and during exam stages,
these parameters were fed into the previously deployed model on IBM cloud and
the model made predictions about the presence of mental stress in the students.

Table 19. Predictions made by the Decision Tress Classifier

model, deployed on the IBM Cloud

From table 13, we observed that using the PSS scale, it was determined that
student 4 was constantly under high stress during the examination days, but
observing the prediction made in table 19, for student 4, the DT classifier model
predicted “no stress” (even though the student was undertaking the exam).

The correlation between the physiological stress measurement and the

subjective perception of stress could be further explored by including more test
subjects under study.
Master Thesis 37

CHAPTER 7. DISCUSSION

7.1. Analysis and discussion of results

The first objective of this research work was to determine the feasibility of using
ML algorithms by establishing a correlation between students’ academic
performance and their mental stress level. To this extent this research seems to
be effective as the two classifier ML models that are developed tends to achieve
an accuracy of 71% approximately. The low accuracy is due small size of
dataset. The accuracy could be enhanced, significantly, taking into account a
larger dataset. Furthermore, when we consider the feature ranking of the two ML
classifiers, it can be observed that Perceived Stress Score (PSS) feature plays a
significant role in predicting the academic performance of the students,
considering the LGBM classifier. The higher the PSS score, the lower the
academic performance of the students.

The second research question was to explore the benefit of using IBM Watson
ML platform. Implementing this research by utilizing IBM Watson ML cloud based
platform proved to be a very viable and powerful solution as one can easily utilize
the automatic experimental approach of the IBM ML without the need of extensive
coding. In this regard, steps to build and implement the ML models using IBM
auto AI experiment has also been mentioned in this thesis. Top performing ML
model was chosen on the basis of accuracy in addition to various parameters
like, F1 Score, Precision, Recall, etc.

The third research objective was to the implementation of a ML model to predict

the mental stress using Heart Rate Variability. In this context, several research
papers and reference sources were considered to determine the parameters of
Heart Rate Variability (HRV). There are three main classifications of HRV
parameters, namely, time domain, frequency domain and non-linear parameters.
In this research only the time domain parameters are considered and a ML model
has been developed utilizing the time domain parameters as features for the
dataset. The ML model was trained and tested using the SWELL dataset.

The SWELL dataset was compiled by retrieving information through an

experimental setup (Koldijk et al., 2014). The experimental setup included twenty
five individuals, and they were monitored while performing usual work related
tasks that included making presentations, report writing, browsing information
and responding to e-mails (Koldijk et al., 2014). The working environment of these
25 individuals were controlled with interruptions, time pressure and various
stresses (Koldijk et al., 2014). The data was compiled based on monitoring the
facial expressions from camera recordings, computer logging, body movements
from Kinect 3D sensor, and skin conductance from body sensors (Koldijk et al.,
2014). As per the researchers, “the participants’ subjective experience on work
load, mental efforts and emotional stress was assessed with pre documented
questionnaires as a central focus for ground truth. The resulting dataset on
working behaviour and affect is a valuable contribution to several research fields,
such as work psychology, user modelling and context aware systems.” (Koldijk
et al., 2014).
38 Master Thesis

The original SWELL dataset was altered by taking into account only the HRV time
domain parameters as mentioned in the previous section. The top performing
model in our case came out to be Decision Tree Classifier with a Cross Validation
accuracy of 98.1%.

After developing this model, the next step was to gather the data related to HRV
of four students, in order to confirm the viability of using this model for HRV
measured through Apple smart watch. As far as the question of how accurately
the smart watch measures the HRV is concerned, the results of this analysis as
described in the previous chapter, confirms the suitability of using Apple smart
watch.

Furthermore, it was deduced that whenever the mean value of the RR interval
increases, the heart rate decrease and consequently the stress level increases.
Additionally, it was observed that the physiological stress measurement can
deviate from subjective perception of stress. In another study it was revealed that
there might be a genetic moderation in the association between resting state HRV
and perceived stress (Looser et al., 2023).

CHAPTER 8. CONCLUSIONS
This research work utilized a quantitative approach to analyse the impact of
stress on heart rate variability (HRV) using machine learning techniques. The
findings revealed a positive association between higher stress levels and
increased HRV, consequently, degrading the academic performance of the
students as proved from the first phase of this study. In this context, Snap
Boosting model had more or less similar performance measures as compared to
LGBM classifier model. But, this model is not recommended to predict the grades
of the students based on their stress levels as it shows high correlation for only
cognitive performance.

The stress level was determined by extracting features from HRV analysis, and
a classification technique was employed using threshold values derived from the
training dataset. Performance accuracy measures were employed to assess the
outcomes. Consequently, this study suggests that stress influences HRV, thus
establishing its potential as an objective tool for assessing stress in academic
settings.

8.1. Future work

Following are some of the recommendations given in regard to the future scope
of this study;
• The accuracy of the model, developed to predict the academic
performance of the students, could be further enhanced if the size of the
dataset is increased i.e., conducting survey on a larger scale.
• The dataset that was gathered with the help of a survey based on
responses from the students, seems to be less efficient. Steps can be
Master Thesis 39

taken to reduce the human error due to negligence of the students taking
part in the survey.
• The correlation between the physiological stress measurement and the
subjective perception of stress can be further explored by including more
test subjects under study.
• A smart Learning Management System (LMS) can be designed that can
give smart recommendations to the students on ways to improve their
academic performance & to cope up with the mental stress suggestions
like breathing exercises etc. Data mining could be further applied while
considering solutions for students with different performance/stress levels.
Figure 17 shows proposed structure of a smart LMS database.

Fig. 17 Proposed Database structure for LMS

40 Master Thesis

References
Analysis of Autonomic Nervous System. (n.d.). HRV - measuring parameter -

ans analysis. Retrieved April 3, 2023, from https://www.ans-

analysis.com/hrv/hrv-measuring-parameter.html

Colom, R., Karama, S., Jung, R. E., & Haier, R. J. (2010). Human intelligence

and brain networks. Dialogues in clinical neuroscience. Retrieved from

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3181994/

Deng, Y., Cherian, J., Khan, N. U. N., Kumari, K., Sial, M. S., Comite, U.,

Gavurova, B., & Popp, J. (2022). Family and Academic Stress and

Their Impact on Students' Depression Level and Academic

Performance. Frontiers in psychiatry, 13, 869337.

https://doi.org/10.3389/fpsyt.2022.869337

Dinardo, D. A. (2016, October). What's Your stress threshold? Thriving Under

Pressure. Retrieved from

https://drandreadinardo.com/2016/10/05/whats-your-stress-threshold/

Hazer-Rau, D., Meudt, S., Daucher, A., Spohrs, J., Hoffmann, H., Schwenker,

F., & Traue, H. C. (2020, April 17). The UULMMAC database-a

multimodal affective corpus for affective computing in human-computer

interaction. MDPI. Retrieved from https://www.mdpi.com/1424-

8220/20/8/2308

Hernando, D., Roca, S., Sancho, J., Alesanco, Á., & Bailón, R. (2018, August

10). Validation of the apple watch for heart rate variability

measurements during relax and mental stress in healthy subjects.

MDPI. Retrieved from https://www.mdpi.com/1424-8220/18/8/2619/htm

Master Thesis 41

IBM Watson. (n.d.). IBM Watson Studio - Overview. IBM. Retrieved April 3,

2023, from https://www.ibm.com/cloud/watson-studio

IBM. (2023). What is a decision tree?. Retrieved April 3, 2023, from

https://www.ibm.com/topics/decision-

trees#:~:text=A%20decision%20tree%20is%20a,internal%20nodes%20

and%20leaf%20nodes

What is supervised learning? IBM. (n.d.). Retrieved April 6, 2023, from

https://www.ibm.com/topics/supervised-learning

IBM. (2022). What is boosting? Retrieved from

https://www.ibm.com/topics/boosting

Kiselica, M. S., Baker, S. B., Thomas, R. N., & Reedy, S. (1994). Effects of

stress inoculation training on anxiety, stress, and academic perfor-

mance among adolescents. Journal of Counseling Psychology, 41(3),

335–342. https://doi.org/10.1037/0022-0167.41.3.335

Koldijk, S., Sappelli, M., Verberne, S., Neerincx, M. A., & Kraaij, W. (2014). The

swell knowledge work dataset for stress and User Modeling Research.

Proceedings of the 16th International Conference on Multimodal

Interaction. https://doi.org/10.1145/2663204.2663257

Kim, H.-G., Cheon, E.-J., Bai, D.-S., Lee, Y. H., & Koo, B.-H. (2018, February

28). Stress and heart rate variability: A meta-analysis and review of the

literature. Psychiatry Investigation. Retrieved from

https://www.psychiatryinvestigation.org/journal/view.php?doi=10.30773

%2Fpi.2017.08.17

Kurtis Pykes - The Difference Between Classification and Regression in

Machine Learning, Published in Towards Data Science, Aug, 7, 2021

42 Master Thesis

Kim, Y., Yoon, H. Y., Kwon, I. K., Youn, I., & Han, S. (2022). Heart rate

variability as a potential indicator of cancer pain in a mouse model of

peritoneal metastasis. Sensors, 22(6), 2152.

https://doi.org/10.3390/s22062152

Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., & Liu, T.-Y. (2017, December

4). LightGBM: A highly efficient gradient boosting decision tree. ACM.

https://dl.acm.org/doi/10.5555/3294996.3295074#sec-ref

Lin, X.-J., Zhang, C.-Y., Yang, S., Hsu, M.-L., Cheng, H., Chen, J., & Yu, H.

(2020, June 3). Stress and its association with academic performance

among dental undergraduate students in Fujian, China: A cross-

sectional online questionnaire survey - BMC Medical Education.

BioMed Central. Retrieved April 3, 2023, from

https://bmcmededuc.biomedcentral.com/articles/10.1186/s12909-020-

02095-4

Leung, K. (2022, September 13). Micro, Macro & weighted averages of F1

score, clearly explained. Medium.

https://towardsdatascience.com/micro-macro-weighted-averages-of-f1-

score-clearly-explained-

b603420b292f#:~:text=The%20macro%2Daveraged%20F1%20score,r

egardless%20of%20their%20support%20values.&text=The%20value%

20of%200.58%20we,score%20in%20our%20classification%20report

Looser, V. N., Ludyga, S., & Gerber, M. (2023). Does heart rate variability

mediate the association between chronic stress, Cardiorespiratory

Fitness, and working memory in young adults? Scandinavian Journal of

Master Thesis 43

Medicine & Science in Sports, 33(5), 609–618.

https://doi.org/10.1111/sms.14308

Metabolic Meals. (2021, June 20). A beginner's Guide to Heart Rate Variability

(HRV) - metabolic meals. Blog. Retrieved April 3, 2023, from

https://blog.mymetabolicmeals.com/hrv-guide/

Microsoft Corporation. (n.d.). LightGBM. Features - LightGBM 3.3.5.99.

Retrieved April 3, 2023, from https://lightgbm.readthedocs.io/en/lat-

est/Features.html#references

NH Dept. of Administrative Services . (n.d.). Perceived stress scale - new

hampshire. Retrieved April 3, 2023, from

https://www.das.nh.gov/wellness/Docs/Percieved%20Stress%20Scale.

pdf

Padilha, T. P. P., & Catrambone, R. (2021). USE OF THE TENSORFLOW

FRAMEWORK TO SUPPORT EDUCATIONAL PROBLEMS: A

SYSTEMATIC MAPPING. International Conference on Education.

https://doi.org/10.17501/24246700.2021.7133

Rajendran, S., Chamundeswari, S., & Sinha, A. A. (2022). Predicting the aca-

demic performance of middle- and high-school students using machine

learning algorithms. Social Sciences & Humanities Open, 6(1), 100357.

https://doi.org/10.1016/j.ssaho.2022.100357

Singh, P., Singh, N., Singh, K. K., & Singh, A. (2021). Diagnosing of disease

using machine learning. Machine Learning and the Internet of Medical

Things in Healthcare, 89–111. https://doi.org/10.1016/b978-0-12-

821229-5.00003-3
44 Master Thesis

Shi, Y., & Qu, S. (2022). The effect of cognitive ability on academic

achievement: The mediating role of self-discipline and the moderating

role of planning. Frontiers in Psychology, 13.

https://doi.org/10.3389/fpsyg.2022.1014655

Shaffer, F., & Ginsberg, J. P. (2017, September 28). An overview of heart rate

variability metrics and norms. Frontiers in public health. Retrieved April

3, 2023, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5624990/

Singh, N., Moneghetti, K. J., Christle, J. W., Hadley, D., Plews, D., & Froelicher,

V. (2018). Heart Rate Variability: An Old Metric with New Meaning in

the Era of using mHealth Technologies for Health and Exercise Training

Guidance. Part One: Physiology and Methods. Arrhythmia &

electrophysiology review, 7(3), 193–198.

https://doi.org/10.15420/aer.2018.27.2

Suresh, A. (2021, June 22). What is a confusion matrix?. Medium.

https://medium.com/analytics-vidhya/what-is-a-confusion-matrix-d1c0f8feda5

Welltory, W. I. (2023, March 18). Stress and HRV - what's The connection:

Welltory app. Welltory. Retrieved April 3, 2023, from

https://welltory.com/stress-and-hrv/

Williams, S., Layard Horsfall, H., Funnell, J. P., Hanrahan, J. G., Khan, D. Z.,

Muirhead, W., Stoyanov, D., & Marcus, H. J. (2021). Artificial

Intelligence in Brain Tumour Surgery-An Emerging

Paradigm. Cancers, 13(19), 5010.

https://doi.org/10.3390/cancers13195010

Xu, J., Moon, K. H., & van der Schaar, M. (2017). A machine learning approach

for tracking and predicting student performance in degree programs.

Master Thesis 45

IEEE Journal of Selected Topics in Signal Processing, 11(5), 742–753.

https://doi.org/10.1109/jstsp.2017.2692560

UC Berkeley . (2011). Cognitive failures questionnaire - university of California,

Berkeley. Open Computing Facility . Retrieved from

https://www.ocf.berkeley.edu/~jfkihlstrom/ConsciousnessWeb/Meditatio

n/CFQ.htm
46 Master Thesis

Annex
Master Thesis 47
48 Master Thesis

The Cognitive Assessment Questionnaire

The following questions are about minor mistakes, which everyone makes from
time to time, but some of which happen more often than others. We want to know
how often these things have happened to your in the past 6 months. Please circle
the appropriate number
Master Thesis 49
50 Master Thesis

Survey questionnaire