IEEE - 56998

Emotion Detection and Suicidal Intention Prediction

of Differently Depressed Individuals Using Machine
2023 14th International Conference on Computing Communication and Networking Technologies (ICCCNT) | 979-8-3503-3509-5/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICCCNT56998.2023.10308047

Learning Techniques
Shreya Soni Shruti Chaubey Suchita Parira
Information Science and Engineering Information Science and Engineering Information Science and Engineering
CMR Institute of Technology CMR Institute of Technology CMR Institute of Technology
Bengaluru, Karnataka Bengaluru, Karnataka Bengaluru, Karnataka
[email protected] [email protected] [email protected]

Senthil Velan S
Information Science and Engineering
CMR Institute of Technology
Bengaluru, Karnataka
[email protected]

Abstract—Facial expressions play an important role in II. EASE OF USE

conveying emotions, especially in human-machine interaction.
Automatic facial expression recognition (FER) systems have A. Framework and Usability
numerous potential applications, such as detecting mental In this section, we discuss the ease of use of our proposed
disorders, understanding human behavior, and generating synthetic framework for emotion detection and suicide prediction using
human expressions. However, achieving high recognition rates machine learning. Our framework is designed to be user-friendly,
remains a challenging task. In the literature, two popular with intuitive interfaces for data input, model training, and result
approaches for automatic FER are based on geometry and visualization.
appearance. The FER process typically consists of four stages, Firstly, the data input module of our framework is designed to
namely pre-processing, face detection, feature extraction, and accept the desired input and facial expression image of the person.
expression classification. In our project, we utilized various deep The module also includes pre-processing functions to handle
learning techniques, specifically convolutional neural networks, to missing data and standardize the data to ensure accurate model
detect seven essential human emotions: anger, disgust, fear, training.
happiness, sadness, surprise, and neutrality. Furthermore, our aim
was to predict suicidal tendencies based on the detected emotions Secondly, our framework provides easy-to-understand
since depression is the primary cause of suicide. Detecting visualization tools to display the model performance metrics, such
emotions in depressed individuals could facilitate their monitoring as accuracy, precision, recall, and F1 score. These metrics help
users evaluate the model's performance and make informed
and help prevent suicide risk by forecasting the rate of suicidal
decisions about the suitability of the model for their specific use
intentions based on their emotional state.
case.
Keywords—emotion detection, suicidal intention prediction, Finally, our framework is designed to be easily integrated into
machine learning, depression. existing mental health assessment tools and workflows. The
modular structure of our framework allows users to choose specific
I. INTRODUCTION modules for their use case, and the standard input and output
Emotion detection and suicide prediction are critical areas of formats ensure compatibility with other mental health tools.
research in the field of mental health. The use of machine learning Overall, the ease of use of our framework makes it accessible
techniques has gained significant attention in recent years due to its to a wide range of users, including mental health professionals,
ability to analyze large datasets and detect patterns that may not be researchers, and policymakers, who can leverage the power of
visible to humans. In this technical paper, we present a novel machine learning to improve early detection and intervention
approach to emotion detection and suicide prediction using strategies for individuals at risk of suicide.
machine learning algorithms. We propose a comprehensive
framework that incorporates various features such as facial
expressions, speech patterns, and physiological signals to
accurately detect emotions and predict suicidal tendencies. Our B. Maintaining the Integrity of the Specifications
framework leverages state-of-the-art machine learning techniques Maintaining the integrity of the specifications is critical in
such as deep learning and ensemble methods to achieve high ensuring the accuracy and reproducibility of the results obtained
accuracy in emotion detection and suicide prediction. The paper from our proposed framework for the detection and prediction of
also includes a detailed description of the dataset used, feature suicidal intention among depressed patients using machine
extraction methods, and model training and evaluation techniques. learning. In this section, we discuss the measures we took to
Our approach has the potential to revolutionize the field of mental maintain the integrity of the specifications of our framework.
health by providing early detection and intervention strategies for
Firstly, we used well-defined and standardized procedures for
depressed patients at risk of suicide.
data collection, preprocessing, and feature extraction to ensure that
the input data to our machine learning models were consistent and
accurate. We also carefully documented all steps involved in data

14th ICCCNT IEEE Conference

Authorized licensed use limited to: SRM Institute of Science and Technology. Downloaded on August 06,2024 at 15:49:20 UTC from IEEE Xplore. Restrictions apply.
July 6-8, 2023
IIT - Delhi, Delhi
 preparation, including the sources of the data, any preprocessingIEEE - 56998
predict suicidal tendencies. Our proposed framework for emotion
steps taken, and the exact parameters used for feature extraction. detection and suicide prediction using machine learning builds on
the existing research in this area and aims to provide a
Secondly, we used rigorous validation techniques to evaluate comprehensive and accurate approach for early detection and
the performance of our machine learning models. We used cross- intervention strategies for individuals at risk of suicide.
validation techniques to avoid overfitting and ensure that the
models generalize well to new data. We also performed sensitivity
analysis to assess the robustness of our models to variations in
input data, model parameters, and feature selection. IV. PROPOSED WORK
Thirdly, we maintained a clear and detailed description of our
model architecture, including the type of machine learning A. Emotion Detection Model
algorithms used, the number of layers and nodes in the network,
and the activation functions used in the model. We also provided The data was collected from kaggle in csv format. There are four
detailed descriptions of the hyper-parameters used in the model and basic steps to be performed for implementing emotion detection.
the procedure for selecting them.
i. Preprocessing
Finally, we made our framework and all source code publicly
ii. Face registration
available to ensure transparency and facilitate replication of our
results by other researchers. We provided detailed documentation iii. Facial feature extraction
of the code, including instructions for running the framework,
information on dependencies, and descriptions of the output files. iv. Emotion classification

Overall, we took measures to maintain the integrity of the Description about all these processes are given below:
specifications of our framework by using standardized procedures i. Preprocessing: Preprocessing refers to a set of image
for data collection, rigorous validation techniques, detailed operations at the most fundamental level of abstraction, where both
documentation of the model architecture and hyperparameters, and input and output are intensity images. Common preprocessing steps
making our framework and source code publicly available. involve various operations, including:
a. Reduce the noise
III. RELATED WORK b. Convert the Image To Binary/Grayscale.
In this section, we present a literature survey of the existing c. Pixel Brightness Transformation.
research on emotion detection and suicide prediction using
machine learning techniques. d. Geometric Transformation

A. Emotion Detection ii. Face Registration: The computer technology of face

registration is utilized in diverse applications to recognize human
Emotion detection has been studied extensively using various faces in digital images. In the face registration process, the first
techniques, including machine learning algorithms. One of the step involves identifying faces in the image using a set of landmark
commonly used techniques for emotion detection is facial points known as "face localization" or "face detection". The next
expression analysis. A study by Liu et al. (2019) [1] proposed a step entails geometrically normalizing the detected faces to align
deep learning-based framework for facial expression recognition. with a template image in a process called "face registration".
They achieved an accuracy of 94.68% on the Facial Expression
Recognition (FER) 2013 dataset. Another study by Zhang et al. iii. Facial Feature Extraction: Facial feature extraction is a
(2020) [2] proposed a multimodal emotion recognition system that crucial step in face recognition that involves identifying specific
combines facial expression, speech, and physiological signals. The regions, points, landmarks, or contours in a given 2D or 3D range
proposed system achieved an accuracy of 78.8% on the AffectNet image. This process generates a numerical feature vector from the
dataset. registered image. Various features can be extracted during this
step, including but not limited to:
B. Suicide Prediction
a. Lips
Suicide prediction is a challenging task due to the complex
nature of the problem. Machine learning techniques have been used b. Eyes
to develop predictive models for suicide risk assessment. One of c. Eyebrows
the commonly used techniques for suicide prediction is natural
language processing (NLP). A study by Coppersmith et al. (2018) d. Nose tip
[3] proposed a machine learning-based approach that uses NLP to
predict suicide risk from social media data. The proposed approach iv. Emotion Classification: During the third stage of the
achieved an accuracy of 80.6% on the Suicide Watch subreddit process, the algorithm endeavors to categorize the faces into one of
dataset. Another study by Cao et al. (2021) [4] proposed a suicide the seven fundamental emotions. Paul Ekman, an American
risk prediction model that combines demographic and clinical data psychologist and professor emeritus at the University of California,
with machine learning algorithms. The proposed model achieved San Francisco, is a notable figure in the research of emotions and
an accuracy of 92.9% on a large-scale Chinese dataset. their correlation with facial expressions. He is credited with
developing an "atlas of emotions" that comprises over ten thousand
C. Emotion Detection and Suicide Prediction: facial expressions.
Emotion detection and suicide prediction are closely related Our purpose is to improve the accuracy of facial expression
areas of research in the field of mental health. Several studies have classification by using a new CNN architecture. As a first step,
proposed frameworks that combine multiple modalities to detect after preparing the database we fixed the batch size input of CNN
emotions and predict suicidal tendencies. A study by Dinesh et al. architecture to 165 × 165 then we trained the architecture with fine
(2019) proposed a deep learning-based framework that combines tuning by Visual Geometry Group (VGG) model to generate the
facial expressions, speech, and physiological signals for emotion first model. In the second step to improve the classification we
detection and suicide prediction. The proposed framework repeat the training of our CNN architecture but the fine tuning here
achieved an accuracy of 86.8% on the AffectNet dataset and an is achieved with the obtained first model, and finally we get our
accuracy of 93.4% on the Suicide Prevention dataset. final model. After detection of the human emotions with and
without mask we go on predicting their intentions based on the
Overall, the literature survey shows that machine learning emotion detected.
techniques have been used successfully to detect emotions and

14th ICCCNT IEEE Conference

Authorized licensed use limited to: SRM Institute of Science and Technology. Downloaded on August 06,2024 at 15:49:20 UTC from IEEE Xplore. Restrictions apply.
July 6-8, 2023
IIT - Delhi, Delhi
 IEEE - 56998  image : image of the face of the patient in
jpeg/jpg/png format
 age : age of the patient (numeric)
 relationship status : relationship status of the
patient (married/single/divorced)
 gender : gender of the patient
(male/female/transgender)
 depression level : depression level of the patient as
specified by the doctor (acute/moderate/ severe)
 working status : working status of the patient
(employed/unemployed)
 no of days : no. of. days the patient is admitted in
the hospital (numeric)
Fig. 5.1. Steps in suicidal Intention Prediction Model  social interaction level : the level of interaction a
person has in his personal life (numeric)
 admission_status : whether the patient is admitted
B. Suicidal Intention Prediction Model in the hospital ( yes/no)
Following are the steps taken to generate a model using  drugs/alcohol : whether the patient is drug/alcohol
Machine Learning: addicted (yes/no)
1. Dataset creation Following is an example of the input given by the user and
2. Data preprocessing the prediction of our model as output based on the input:
3. Data visualization
4. Training the data using the most efficient ML algorithm
1. Dataset Creation Input photo:
Creation of the dataset is the most challenging part of the entire
project as a lot of research was needed to be done on various case
studies of depressed patients. After all the research, the following
are the dependent features of the dataset that were proved to be
most valuable factors affecting the suicide intentions in depressed
patients: emotion, age, relationship status, gender, depression level,
working status, no of days of admission in the hospital, social
interaction level, whether admitted in hospital, drugs/alcohol
consumption. All these are the factors that are considered to predict
the rate of suicide risk. The final column in the dataset is the
intention column that is the dependent feature. The dependent
feature was evaluated based on the knowledge gained by various Fig. 5.2 user input of the face
case studies.
age : 30
2. Data preprocessing
relationship status : unmarried
The data preprocessing is an important step to filter the relevant
data from the large dataset. In this step all the fields with null gender : male
values were replaced by the desired or the most relevant values.
depression level : severe
Label Encoding was done for the categorical data.
working status : unemployed
3. Data visualization
no of days : 50
Data was visualized using the heatmap function of the seaborn
library to figure out the most relevant independent features. various social_interaction_level : 2
graphs were plotted to determine the relation between the columns.
admission_status : yes
4. Training and testing
drugs/alcohol : yes
The data was splitted into training and testing datasets in 80:20
ratio. The DecisionTreeClassifier model was imported and used to
train the training dataset. The Decision Tree model showed the output:
accuracy of 80% on testing the model.
emotion : Anxiety
Suicide risk : High
C. Emotion detection and Suicidal intention prediction based on
the emotion
The image of the face and various desired data about the depressed The image of the face and various desired data about the
patient was taken as input from the user. The image was passed to depressed patient was taken as input from the user as shown in the
the emotion detection model to detect the emotion of the person Fig. 5.1. The image was passed to the emotion detection model to
and the resultant dominant emotion with all the other inputs was detect the emotion of the person and the resultant dominant
passed to the intention prediction model to predict the rate of emotion with all the other inputs was passed to the intention
suicidal intention of the person on that specific emotion. prediction model to predict the rate of suicidal intention of the
person on that specific emotion. The dominant emotion of the
patient is anxiety as detected by the emotion detection model and
the suicidal intention rate predicted by the intention prediction
V. RESULTS AND DISCUSSIONS modelis high. The decision tree and logistic regression algorithms
The input is taken from the user in the following format: were used to train the suicidal intention prediction model and the

14th ICCCNT IEEE Conference

Authorized licensed use limited to: SRM Institute of Science and Technology. Downloaded on August 06,2024 at 15:49:20 UTC from IEEE Xplore. Restrictions apply.
July 6-8, 2023
IIT - Delhi, Delhi
 IEEE - 56998
decision tree model was found to be more accurate with the testing aims at providing a tool that will automatically detect the suicide
data as decision tree algorithm fits well with the numerical data. In risk rate of a person based on various emotions. This project can be
the context of predicting the suicidal intention of depressed used to monitor the depressed patients and their suicidal intentions
patients, a decision tree algorithm may have some advantages over from even remote places.
logistic regression.
Firstly, decision trees can handle non-linear relationships
between features and the target variable, which may be important REFERENCES
when dealing with complex mental health data that may have non-
linear relationships between variables.Secondly, decision trees can
[1] Abir Fathallah, Lofti Abdi, and Ali Douik, “ Facial Expression
handle missing values in the data, which can be common in mental
Recognition via Deep Learning”.
health datasets. Logistic regression, on the other hand, requires
imputation or removal of missing values before the model can be [2] Dilbag Singh, “ Human Emotion Recognition System.
trained. Thirdly, decision trees can handle both continuous and [3] Daun Shin, Kyundgo Kim, Seung-Bo Lee, Changwoo Lee, Ye
categorical data, while logistic regression is better suited for Seul Bae, Won ik cho, Min Ji Kim, C. Hyung Keun Park, Eui
continuous data.Fourthly, decision trees can handle interactions Kyu Chie, Nam Soo Kim, and Young Min Ahn, “Detection of
between variables, which may be important when predicting the Depression and Sucide Risk Based on Text From Clinical
suicidal intention of a patient, as there may be complex interactions Interviews using Machine Learning” [National Library of
between different risk factors. Medicine].
Lastly, decision trees can provide interpretable results [4] Ayaan Haque, “NLP for Sucide and Depression Identification
that can help clinicians identify important risk factors for suicidal with NoisyLabels.
intention. Logistic regression may not provide as much [5] https://worldpopulationreview.com/country-rankings/
interpretability, as it is a linear model and does not provide a clear
depression-rates-by-country.
visual representation of the decision-making process. Of course,
the choice of algorithm ultimately depends on the characteristics of [6] https://m.economictimes.com/magazines/panache/one-in-
the data and the specific problem at hand. It's important to carefully seven-indians-between-15-24-years-of-age-feel-depressed-
consider the strengths and weaknesses of each algorithm before lack-interest-in-doing-things-shows-unicef-report/
making a decision. articleshow/86807417.cms.
[7] https://mindvoyage.in/depression-in-india-latest-statistics/
#:~:text=According%20to%20a%20report%2C%20in,the
%20cases%20stand%20at%2056%2C675%2C969.
VI. CONCLUSIONS AND FUTURE DIRECTIONS
[8] https://www.who.int/india/health-topics/depression.
The project "Emotion Detection and Suicidal Intention [9] Nie, W., Bao, Y., Zhao, Y., & Liu, A. (2023). Long Dialogue
Prediction of Depressed Patients using ML" aims to develop a Emotion Detection Based on Commonsense Knowledge
machine learning model that can detect emotions and predict Graph Guidance. IEEE Transactions on Multimedia.
suicidal tendencies in depressed patients based on their facial
expressions. The project utilizes a Convolutional Neural Network [10] Alsubai, S. (2023). Emotion Detection Using Deep
(CNN) for emotion detection and a Decision tree algorithm for Normalized Attention-Based Neural Network and Modified-
suicidal intention prediction. The model was trained and tested Random Forest. Sensors, 23(1), 225.
using the FER-2013 and 90,000 datasets, which include facial [11] Karilingappa, K., Jayadevappa, D., & Ganganna, S. (2023).
images with different emotional expressions. Human emotion detection and classification using modified
The project has important implications for the field of Viola-Jones and convolution neural network. IAES
mental health, as it provides a potential tool for early detection and International Journal of Artificial Intelligence, 12(1), 79.
intervention in depressed patients who may be at risk for suicidal [12] Dessai, A. U., & Virani, H. G. (2023). Emotion Detection and
behavior. This project represents an important step towards using Classification Using Machine Learning Techniques. In
machine learning techniques for mental health assessment and Multidisciplinary Applications of Deep Learning-Based
intervention. Further research and development are necessary to Artificial Emotional Intelligence (pp. 11-31). IGI Global.
improve the accuracy and reliability of the models, as well as to
[13] Hung, L. P., & Alias, S. (2023). Beyond Sentiment Analysis:
address ethical and privacy concerns related to the use of personal
A Review of Recent Trends in Text Based Sentiment Analysis
data in mental health assessments.This project can be enhanced
further by adding more independent features in the dataset that and Emotion Detection. Journal of Advanced Computational
have a great impact in the suicide intentions of the people to Intelligence and Intelligent Informatics, 27(1), 84-95.
increase the accuracy. The real time data can be used from the [14] Qaqish, E., Aranki, A., & Etaiwi, W. (2023). Sentiment
hospital and fed to this model to predict the results. This project analysis and emotion detection of post-COVID educational
has a lot of scope as it can be used to monitor the depressed Tweets: Jordan case. Social Network Analysis and Mining,
patients from remote to prevent suicides through CCTV cameras 13(1), 1-11.
that can record the facial expressions and the model can predict the [15] Gupta, S., Singh, A., & Ranjan, J. (2023). Multimodal,
suicidal intentions based on the emotions detected. multiview and multitasking depression detection framework
The prediction of suicidal risk of depressed patients endorsed with auxiliary sentiment polarity and emotion
based on their emotion helps to prevent the suicides that are very detection. International Journal of System Assurance
common in depressed patients when they are alone. This project Engineering and Management, 1-16.

14th ICCCNT IEEE Conference

Authorized licensed use limited to: SRM Institute of Science and Technology. Downloaded on August 06,2024 at 15:49:20 UTC from IEEE Xplore. Restrictions apply.
July 6-8, 2023
IIT - Delhi, Delhi