ISSN2321-2152

[Link] .com
Vol 12, Issue 3, 2024

SUICIDAL TENDENCY DETECTION

Maturi Rajeshwari1, Revathy P 2, Dr Dileep P 3
1,2
Assistant Professor, Department of Computer Science and Engineering
3
Professor, Department of Computer Science and Engineering
1,2
Narsimha Reddy Engineering College,Kompally,Hyderabad,India.
3
Malla Reddy College of Engineering and Technology,Kompally,Hyderabad,India.

ABSTRACT:
Suicidal Tendency or the intension to kill oneself or end one’s life is a catastrophic
situation which is mostly unknown by any person in the victim’s life. Suicide has been
an intractable public health problem despite advances in the diagnosis and treatment of
major mental disorders. In many studies it is clearly evident that, victims tend to kill
themselves either to end their pain or pressure or to have a sense of relief that they are
not going to live in this world anymore. This project aims to propose a method that
helps the family, friends or the close ones of the victim to immediately detect if the
person has already started feeling the sense of depression. The main aim is to find a
strong co-relation between components in the subsystem and compare the accuracies to
build an alarming system. “Better late than never” the victim can be saved by the
proposed method and immediate treatment can be started. Unlike the existing systems,
this project aims to detect the suicidal tendencies in multiple aspects instead of focusing
on a single perspective.

[Link] capture different elements like facial

This project aims to focus on the people gestures, speech recognition and many
who have an intension to kill themselves. more. A trivial concept of incorporating
In this consequence, multi-faceted different aspects like: Facial Gestures,
method that can detect this tendency and Voice Recognition and Messaging
intimate the family, friends or the close Patterns follow the bandwagon along
ones beforehand can prove to be a boon with the technical byproducts of the
for the invention. This project tends to project. Facial gestures include unhappy
consider an electronic device particularly expressions like sad, dull, tired; voice
a mobile (as used by most of them) is the patterns include low voices that sound
key element. This device is used to dull are easy to recognize that someone is

251
 ISSN2321-2152
[Link] .com
Vol 12, Issue 3, 2024

unhappy; Texting patterns include that, it was not practically possible to

unusual texting patterns that indicate lack indicate a clear demarcation in only one
of interest in doing activities. aspect of implementation. In course of
➢ The objective of this system is that it time, numerous approaches have been
is capable of detecting suicidal proposed with the advancement in the
tendency in a specific person. Unlike technology which is to perform text
other existing system this system can mining and sentiment analysis on social
ensure the ability to focus on media platforms like reddit and twitter.
different technical aspects rather
than single one. [Link] SYSTEM
➢ The project aims to find co-relation This project aims to propose a
between the three components system that is capable of detecting
present below. suicidal tendency using multiple
➢ Facial Gesture Detection – Human approaches. Three technologies namely
Computer Interaction Human Computer Interaction, Natural
➢ Speech Recognition – Natural Language Processing and voice pattern
Language Processing analysis are taken into consideration to
➢ Messaging Patterns – Text perform a detailed execution. Later a
Tokenization through NL correlation matrix is proposed to be
implemented that is able to find strong or
[Link] SYSTEM weak correlations between the above
Since suicide is not a decision three components.
made up in a single day, many researches
had been done in the past on various [Link]:
aspects. While some of them included
how suicides occur and others concluded ➢ Data Collection Module: The data
with some methodologies on how to collection module is tasked with
detect suicidal intension. Distinct gathering relevant data related to
approaches using Human Computer suicidal tendencies. This data can be
Interaction, Natural Language sourced from various channels such
Processing and using Convolutional as surveys, questionnaires, social
Neural Networks were highly popular. media platforms, or healthcare
But these systems possessed a drawback records. The module ensures that the

252
 ISSN2321-2152
[Link] .com
Vol 12, Issue 3, 2024

collected data is comprehensive and unseen data, facilitating robust

representative of the target model development.
population, adhering to ethical
guidelines and privacy regulations. ➢ Machine Learning Module: The
machine learning module applies
➢ Data Preprocessing Module: The advanced algorithms to the training
data preprocessing module is data to construct predictive models
responsible for preparing the for detecting suicidal tendencies.
collected data for analysis and This module incorporates techniques
modeling. It encompasses several such as Convolutional Neural
tasks, including data cleaning, Networks (CNNs) and Random
feature engineering, normalization Forests. For instance, in CNN
or scaling of features, and encoding implementation, deep learning
categorical variables into numerical frameworks like TensorFlow or
representations. Additionally, it PyTorch are utilized to design and
addresses challenges such as train models capable of processing
handling missing values, outliers, diverse data types such as textual or
and imbalanced class distributions to image-based inputs. Conversely,
ensure the quality and integrity of the Random Forest implementation
dataset. involves leveraging machine
learning libraries like scikit-learn to
➢ Data Splitting Module: The data create and optimize classifiers based
splitting module divides the on decision trees, utilizing
preprocessed data into separate preprocessed features for training.
subsets for training and testing
purposes. Typically, a certain ➢ Prediction Module: The prediction
proportion of the dataset is allocated module leverages the trained
for training machine learning models, machine learning models to make
while the remainder is reserved for predictions on new or unseen data
evaluating model performance. By instances. Once the CNN and
splitting the data, the module enables Random Forest models are trained
the assessment of model using the respective machine
generalization and effectiveness on learning module, they are deployed

253
 ISSN2321-2152
[Link] .com
Vol 12, Issue 3, 2024

to predict suicidal tendencies based

on features extracted from new data
points. This module serves as the
final stage in the pipeline, enabling
the application of predictive models
in real-world scenarios to identify
individuals at risk of suicidal
behavior and facilitate timely
intervention and support. Results of classification results using

The method proposed in this project first features based on tweets. When we use

identifies the suicidal keywords, and then only features based on tweets, below

we use these keywords to extract tweets table shows the accuracy, precision and

from Twitter using Twitter Streaming recall, and Completion Speed of our

Application Programming Interface model in recognizing suicidal posts on

(API).After that we pre-process the data tweeter. A 10-fold cross validation was

or extract features from the text used in the experiments. Table shows that

[Link] SVM and Decision Tree the Decision Tree classifier achieves the

with three types of weight optimizers are best accuracy of 97.89%, precision of

imposed on the dataset and at last we 97.89%, recall 97.38% and Naïve Bayes

determine the model efficiency on the classifier complete a speed of 5.35

basis of accuracy score, precision, recall second which is the fastest speed among

and F1 [Link] are retrieved from the remaining algorithms.

twitter data and intensity of suicidal

tendency is calculated based on the
weightage given to the word. Among the
two types of machine learning approach
we followed supervised learning for this
proposed method. A flowchart that
represents the proposed model is as
follows

Fig.1. Output results.

254
 ISSN2321-2152
[Link] .com
Vol 12, Issue 3, 2024

Peoples are taking lots of stress due to terms of accuracy, precision, recall and
today’s competitive environment in FSCORE.
almost all fields like working area and Note: u r asking to record voice and facial
educational area diverting peoples into expression to detect depression or
stress which sometime causes them to suicidal tendency but we don’t any senor
take suicide steps. To overcome from this or devices to record so things so we are
problem author of this paper using using depression and suicidal dataset
PATIENTS dataset to train machine from KAGGLE which contains columns
learning algorithms and then this trained to detect depression. From facial
model can take patient current status expression suicide depression can be
values as input and then predict whether detected as patient can be depress for
that patient is having any suicidal some other reason not for suicide so we
thoughts. If suicidal thoughts detected need to used accurate suicide or
then doctors will take necessary steps to depression related dataset to detect such
help patients in recovering from those tendency so we choose below dataset
thoughts. Below is the dataset screen used in this
To implement this project author has project
used KAGGLE suicidal attempt and
stress dataset and then train this dataset
with deep learning algorithm called CNN.
This dataset contains NON-NUMERIC
characters which will not understand by
MACHINE LEARNING algorithms so In above dataset first row contains dataset
author applying Neural Machine column names and remaining rows
Translation algorithm called Natural contains dataset values and in above
Language Processing (NLP) to translate dataset we can see some values are
all non-numeric characters to numeric numeric and some are non-numeric and
characters which will understand by this non-numeric characters will be
machine learning algorithms. translate to numeric format by using NLP
In this paper we have trained existing technique. NLP will assign numeric ID to
Random Forest and propose CNN each unique non-numeric characters and
algorithm to train dataset and then this ID’s will be used to train ML
evaluate both algorithms performance in algorithms. In below screen we can see

255
 ISSN2321-2152
[Link] .com
Vol 12, Issue 3, 2024

dataset contains ‘suicidal_attempt’ and

‘stressed’ column.

In above screen dataset loaded and we

can see some records from dataset and
In above dataset screen you can see
dataset contains some non-numeric
dataset contains ‘attempt_suicide’ and
characters and to translate them first click
‘stressed’ column. We will used above
on ‘Preprocess Dataset’ button to remove
dataset to train ML algorithms.
missing values and then replace with 0
To run project double click on ‘[Link]’
file to get below screen

In above screen we can see all missing

In above screen click on ‘Upload Suicide
data is replaced with 0 and we can see
Attempt & Stressed Dataset’ button to
dataset contains total 469 records. In
upload dataset and to get below screen
graph we can see total patients with and
without suicidal thought. In above graph
X-axis represents YES and NO values
and y-axis represents total counts of YES
and NO patients. YES means patients has
suicidal thoughts and NO means patients

In above screen selecting and uploading has no suicidal thoughts. Now close

‘Suicidal’ dataset and then click on above graph and then click on ‘Machine

‘Open’ button to load dataset and to get Translation & Features Extraction’

below screen button to translate all dataset NON-

256
 ISSN2321-2152
[Link] .com
Vol 12, Issue 3, 2024

NUMERIC features to NUMERIC

features.

In above screen with existing random

forest algorithm we got 89% accuracy
In above screen we can see complete and now click on ‘Predict Suicidal
dataset is translated to numeric data and Attempt from Test Data’ button to upload
in below two lines we can see dataset test data and then CNN will predict
using 614 records to train CNN whether test patient records has any
algorithms and using 154 records for suicidal and NO suicidal thoughts.
testing CNN performance. Now train and
test data is ready and now click on ‘Train
Propose CNN Algorithm’ button to train
CNN with above dataset and to get below
output

In above screen we are selecting and

uploading ‘[Link]’ file and then
click on ‘Open’ button to load test data
and to get below prediction result

In above screen in square brackets we can

In above screen we can see with CNN we see patient test records values and after

got 92% accuracy and now click on arrow symbol ‘====>’ we can see

‘Train Existing Random Forest prediction result from CNN as whether

Algorithm’ button to train existing patient has suicidal thoughts or NOT.

Random Forest algorithm on same data You can scroll down above screen to get

and calculate accuracy all predicted records like below screen

257
 ISSN2321-2152
[Link] .com
Vol 12, Issue 3, 2024

shortcomings, like eventually having to

use human intervention to approve the
predictions by the ML model, it’s limited
to detecting suicidal tendencies and
depression and contextual analysis as the
data collected doesn’t come with prior
background explanation. The amount of
Now click on ‘Comparison Graph’
text keeps increasing with the
button to get below graph
popularization of social networking
services. And suicide detection and
prevention remain a crucial task in our
modern society. It is therefore essential to
develop new methods to detect online
texts containing suicidal ideation in the
In above graph x-axis represents hope that suicide can be prevented. In the
algorithm names and y-axis represents project an automatic recognition of
accuracy, precision, recall and FSCORE suicidal posts is presented using machine
in different colour bars. In above graph learning techniques As a future scope we
we can see CNN is performing well can implement some more algorithms to
compare to existing Random Forest improve the efficiency and quality of
algoritham. work and we can implement the same
project on videos using image processing
[Link] techniques.
Machine Learning applied to detect [Link]
suicide intention and depression amongst [1]. Hayes, L.M., 2013. Suicide
individual is effective as traditional prevention in correctional facilities:
approaches are hindered by factors like Reflections and next steps. International
face-to-face conversation and shyness to journal of law and psychiatry 36, 188–
express themselves. In the future I would 194
like to delve more into context analysis [2]. S. Lee et al., "Detection of a Suicide
via the retweet history and given links to by Hanging Based on a 3-D Image
external sites. Even though ML seems to Analysis," in IEEE Sensors Journal, vol.
be a better method, yet it entails some

258
 ISSN2321-2152
[Link] .com
Vol 12, Issue 3, 2024

14, no. 9, pp. 2934-2935, Sept. 2014. doi: Recognition", 2017 IEEE 29th
10.1109/JSEN.2014.2332070. International Conference on Tools with
[3]. Calderon-Vilca, H. D., Wun-Rafael, Artificial Intelligence (ICTAI).
W. I., & MirandaLoarte, R. (2017), doi:10.1109/ictai.2017.00109.
"Simulation of suicide tendency by using [8]. Kim, Y., Kim, M., Goo, J., & Kim,
machine learning", 2017 36th H. (2018), "Learning Self-Informed
International Conference of the Chilean Feature Contribution for Deep Learning-
Computer Science Society (SCCC). Based Acoustic Modeling." IEEE/ACM
doi:10.1109/sccc.2017.8405128. Transactions on Audio, Speech, and
[4]. Hu, Z., Hu, Y., Wu, B., & Liu, J. Language Processing, Vol.26(11), 2204–
(2017), "Hand Pose Estimation with 2214. doi:10.1109/taslp.2018.2858923
CNN-RNN", 2017 European Conference [9]. Crombez, N., Caron, G., Funatomi,
on Electrical Engineering and Computer T., &Mukaigawa, Y. (2018),
Science (EECS). "Reliable Planar Object Pose
doi:10.1109/eecs.2017.91. Estimation in Light Fields From Best
[5]. Kamel, A., Sheng, B., Yang, P., Li, Subaperture Camera Pairs." IEEE
P., Shen, R., & Feng, D. D. (2018), "Deep Robotics and Automation Letters,
Convolutional Neural Networks for Vol.3(4),pp.3561–3568.
Human Action Recognition Using Depth doi:10.1109/lra.2018.2853267.
Maps and Postures", IEEE Transactions [10]. Jianhong Wang,1 Tian Lan,,
on Systems, Man, and Cybernetics: "Spatio-temporal Aware Non-negative
Systems, pp.1–14. Component Representation for Action
doi:10.1109/tsmc.2018.2850149. Recognition ", IEEE Transactions On
[6]. Shotton, J., Sharp, T., Kipman, A., Parallel And Distributed Systems, Vol.
Fitzgibbon, A., Finocchio, M., Blake, A., 26, no. 9, pp. 2520-2533, 2015.
Moore, R. (2013)," Real-time human
pose recognition in parts from single
depth images", Communications of the
ACM, Vol.56(1), 116.
doi:10.1145/2398356.2398381.
[7]. Hu, L., & Xu, J. (2017)," Body Joints
Selection Convolutional Neural
Networks for Skeletal Action

259