International Journal of Computer Science Trends and Technology (IJCST) – Volume 10 Issue 5, Sep-Oct 2022

RESEARCH ARTICLE OPEN ACCESS

Machine Learning Based Suicide Ideation Prediction for

Military Personnel
Mrs R Jhansi Rani MCA [1], Manchinti Pavan Kumar Reddy [2]
[1]
Asst. Professor, Department Computer Applications
[2]
Student, Department of Computer Applications
[1], [2]
Chadalawada Ramanamma Engineering College (Autonomous)

ABSTRACT
Military personnel have greater psychological stress and are at higher suicide attempt risk compared with the general
population. High mental stress may cause suicide ideations which are crucially driving suicide attempts. However,
traditional statistical methods could only find a moderate degree of correlation between psychological stress and
suicide ideation in non-psychiatric individuals. This project utilizes machine learning techniques including logistic
regression, decision tree, random forest, gradient boosting regression tree, support vector machine and multilayer
perceptron to predict the presence of suicide ideation by six important psychological stress domains of the military
males and females. The accuracies of all the six machine learning methods are over 98%. Among them, the
multilayer perceptron and support vector machine provide the best predictions of suicide ideation approximately to
100%. As compared with the BSRS-5 score ≥ 7, a conventional criterion, for the presence of suicide ideation ≥ 1,
the proposed algorithms can improve the performances of accuracy, sensitivity, specificity, precision, the AUC of
ROC curve and the AUC of PR curve up to 5.7%, 35.9%, 4.6%, 65.2%, 4.3% and 53.2%, respectively; and for the
presence of more severely intense suicide ideation ≥ 2, the improvements are 6.1%, 26.2%, 5.8%, 83.5%, 2.8% and
64.7%, respectively.
Keywords: - Military personnel, Machine Learning, Accuracy, Improvements, Support Vector Machine.

I. INTRODUCTION identified 30% of this 70% who accounted for 81.2%

This project utilizes machine learning of subsequent administratively-recorded SAs. A
techniques including logistic regression, decision relatively small number of self-report survey
tree, random forest, gradient boosting regression tree, variables were then used to create a prediction model
support vector machine and multilayer perceptron to that identified 10% of the first-stage high-risk sample
predict the presence of suicide ideation. Most (i.e., 3% of all soldiers) at highest SA risk
nonfatal suicide attempts and suicide deaths occur (accounting for 45% of SAs in the total sample). We
among patients who deny suicidal ideation (SI) close by discussing potential applications of this
during suicide risk screenings. Little is known about approach for identifying future SI deniers at highest
risk factors for suicidal behaviors among such SA risk. In the existing work, the system is not
patients. We investigated this in a representative Characterizing High mental stress which may cause
sample of U.S. Army soldiers who denied lifetime SI suicide ideations which are crucially driving suicide
in a survey and were then followed through attempts. This system does not aim to find behavioral
administrative records for up to 45 months to learn of characteristics of suicide attempters.
administratively-recorded suicide attempts (SA). A
novel two-stage risk assessment approach was used II. RELATEDWORKS
that combined first stage prediction from MILITARY personnel are vulnerable to
administrative records to find the subsample of SI psychological stress because of tense physical
deniers with highest subsequent SA risk and then training, multiple deployments and responsibilities.
used survey reports to estimate a second-stage model The prevalence of major depression was reported,
identifying the subset of individuals in the high-risk ranging from 2.0% to 37.4% in the US military [1],
subsample at highest SA risk. 70% of survey and that of combat-related posttraumatic stress
respondents denied lifetime SI. Administrative data disorder (PTSD) was reported 2.0%-17.0% among

ISSN: 2347-8578 www.ijcstjournal.org Page 297

International Journal of Computer Science Trends and Technology (IJCST) – Volume 10 Issue 5, Sep-Oct 2022

US military war veterans [2]. A meta-analysis detection by machine learning are implemented in
showed consistent results that the worldwide pooled [18]. Recognition of heart murmurs could be
prevalence of PTSD in rescue workers was up to achieved by deep neural networks [19]. In addition,
10.0% [3]. The symptoms of mental disorders Ambale-Venkatesh et al. identified the top-20 risk
developed frequently in those of continued combat factors of incident cardiovascular events by the
exposure and those of repeated deployments [4]. The random survival forest which performance was better
association between military absenteeism and mental than the conventional risk calculators [20]. Therefore,
health problems has been discussed in [5]. The rate of using machine learning and deep learning techniques
suicide attempt among active duty US Army has become an efficient and reliable tool for clinical
personnel has been increasingly higher than that in practice by physicians globally. High mental stress
the civilians [6]. According to an analysis for 27,501 may cause suicide ideations which are crucially
military participants in [7], 14.3% of survey driving suicide attempts. However, traditional
respondents reported suicide ideation and 3.0% statistical methods find merely a moderate correlation
committed suicide. In other words, 21% of those with between psychological stress and suicide ideation.
suicide ideation had suicide attempt. As is known, Machine learning could provide better performance
previous studies have revealed a relationship between of the prediction of suicide ideations. In this paper,
suicide ideation and psychological stress [8],[9]. To we utilize a large sample of the military members for
early predict the presence of suicide ideation and several machine learning techniques by taking the
further prevent the behavior of suicide are essential psychological stress dimensions into consideration to
and important in the military. With the technology predict the presence of suicide ideation. The
improvement and the availability of various kinds of schematic diagram of the proposed method in this
real world big data, artificial intelligence (AI) grows paper is illustrated in Fig. 1. A binary probabilistic
fast accordingly. The academics have made great classifier of machine learning algorithm can
efforts on the computerized algorithms to deal with determine whether the military persons, through their
big data. Machine learning, a combination of AI and questionnaires, have suicide ideations. Machine
computations, could provide accurate diagnosis of learning provides an effective manner for early
diseases and predict the outcomes [10]-[17]. For warning and prevention of suicide by automatic
instance, the circuits for seizure classification and suicide ideation detection.

Fig.1 Schematic diagram of proposed method

III. PROPOSED SYSTEM ARCHITECTURE

Gradient boosting decision tree (GBDT) is also an ensemble machine learning method and constructs
multiple additive decision tree models. The DTs fitting the gradient on pseudo residuals of previous cumulative
models are repeatedly trained to minimize mean squared error. This sequential stepwise manner combines the

ISSN: 2347-8578 www.ijcstjournal.org Page 298

International Journal of Computer Science Trends and Technology (IJCST) – Volume 10 Issue 5, Sep-Oct 2022

performance of weak learners (i.e., DT here) in an iterative fashion into a single strong learner to increase the
accuracy of prediction. Our algorithm uses the maximum tree depth as the hyper parameter to be optimized to avoid
over-fitting. Support vector machine (SVM) with linear kernel (Linear SVM) is used for our proposed method. A
data point is viewed as a 6-dimensional vector and we separate such points with a hyperplane. This linear SVM
constructs the maximum-margin hyperplane so that the distance from it to the nearest training data point of any class
(class 0 or class 1) is maximized. Random forest (RF) an ensemble machine learning technique, constructs multiple
decision trees and collects them together for classification. The training algorithm adopted in our method for random
forest is the bootstrap aggregating (bagging) technique. RF builds multiple CART models with different samples
and different initial variables. In each decision tree, a random subset of the features is taken into consideration for
splitting a node. The individual trees are not correlated with each other and thus the trees in random forest of our
method are not pruned. The final prediction result is according to the majority-votes model from the multiple DTs.
RF combines the merits of feature selection and bagging. The decision tree number is the hyper parameter to be
optimized. In the proposed system, the academics have made great efforts on the computerized algorithms to deal with
big data of suicide attempts. The system is more powerful in finding both Psychological Stress, Suicide Ideation. The six
input factors of psychological stress for machine learning include BSRS-5 score, anxiety, depression, hostility,
interpersonal sensitivity and insomnia. This paper uses six machine learning techniques including logistic regression
(LR), decision tree (DT), random forest (RF), gradient boosting decision tree (GBDT), support vector machine
(SVM) and multilayer perceptron (MLP) for the prediction of the presence of suicide ideation of the military
members. The system diagram of proposed method is illustrated in Fig. 2.

Fig.2 Flow chart of proposed method

To solve the phenomenon of different dynamic ranges for the six input variables, we apply the normalization of
Min-Max scaling, to normalize input data into the interval 0~1. Min-Max normalization executes a linear
transformation on the original data. Each of the actual data d of feature x is mapped to a normalized value adjusted
in the range of 0 to 1. 10-fold cross validation is utilized in this paper. The data numbers illustrated by one fold are
detailed. For the prediction of the presence of any suicide ideation (suicide ideation ≥ 1), the numbers for training
and test sets are 3191 (class 0: 3080, class 1: 111) and 355 (class 0: 338, class 1: 17), respectively. For the prediction

ISSN: 2347-8578 www.ijcstjournal.org Page 299

International Journal of Computer Science Trends and Technology (IJCST) – Volume 10 Issue 5, Sep-Oct 2022

of more severely intense suicide ideation (suicide ideation ≥ 2), the numbers for training and test sets are 3191 (class
0: 3157, class 1: 34) and 355 (class 0: 347, class 1: 8), respectively. This imbalance in the dataset between class 0
and class 1 is obvious. This problem is addressed by applying the synthetic minority over-sampling technique
(SMOTE). The training data for class 1 are pre-processed by SMOTE to 3080 and 3157 for the two predictions,
respectively. Support vector machine (SVM) with linear kernel (Linear SVM) is used for our proposed method. A
data point is viewed as a 6-dimensional vector and we separate such points with a hyperplane. This linear SVM
constructs the maximum-margin hyperplane so that the distance from it to the nearest training data point of any class
(class 0 or class 1) is maximized. If the training set is not linearly separable, soft-margin SVM allows the fat
decision margin and some outliers are inside or on the wrong side of the margin. Our method adopts soft-margin
SVM, which minimizes training error traded off against margin. Regularization strategy with a constraint by
regularization term aims to fit training set data and avoid over-fitting. ℓ2-norm is utilized in SVM for our method.
The regularization hyperparameter is optimized in our algorithm to control overfitting.
Random forest (RF), an ensemble machine learning technique, constructs multiple decision trees and
collects them together for classification. The training algorithm adopted in our method for random forest is the
bootstrap aggregating (bagging) technique. RF builds multiple CART models with different samples and different
initial variables. In each decision tree, a random subset of the features is taken into consideration for splitting a node.
The individual trees are not correlated with each other and thus the trees in random forest of our method are not
pruned. The final prediction result is according to the majority-votes model from the multiple DTs. RF combines the
merits of feature selection and bagging. The decision tree number is the hyperparameter to be optimized. Gradient
boosting decision tree (GBDT) is also an ensemble machine learning method and constructs multiple additive
decision tree models. The DTs fitting the gradient on pseudo residuals of previous cumulative models are repeatedly
trained to minimize mean squared error. This sequential stepwise manner combines the performance of weak
learners (i.e., DT here) in an iterative fashion into a single strong learner to increase the accuracy of prediction. Our
algorithm uses the maximum tree depth as the hyperparameter to be optimized to avoid over-fitting. Multilayer
perceptron (MLP) consists of an input layer, hidden layers and an output layer for our algorithm. In fully connected
MLP, each node in one layer connects with a certain weight to every node in the following layer. In the forward
propagation, the signal flow moves from the input layer through the hidden layers to the output layer. Learning is
carried out through backward propagation. The loss function consists of cross entropy and ℓ2-norm regularization to
prevent over-fitting. The optimizer Adam is adopted in our method. Besides regularization hyperparameter, the
numbers of hidden layers, neurons and iterations are also used as the hyperparameters to be optimized in our MLP
method.

IV. RESULTS AND DISCUSSION

The performance comparisons of each machine learning technique with the BSRS-5 score ≥7 [9] assessed
by accuracy, sensitivity, specificity, precision, F1 score, r value, are under curve(AUC) of ROC curve and the AUC
of PR curve are shown as the average of the 10-fold cross validation. The p values for all techniques are <0.0001.
The accuracies for all of the six machine learning methods are over 98%. The results show that the MLP and SVM
provide the best predictions. The second is the LR, and then the three tree-based methods. The RF and GBDT yield
overall better results than that of the DT. The prediction results for suicide ideation ≥ 1 are better than those of
suicide ideation ≥ 2. The ROC curves and the PR curves and the corresponding AUCs are shown in Fig. 3, for one
of the 10-fold cross validation. The areas under ROC and PR curves reach to 1 for MLP and SVM. The tendency is
similar with the results. The learning curves for the six proposed methods for suicide ideation ≥ 1 are shown in Fig.
4. The learning curves depict an improvement in F1 score when there are changes in the number of iteration (logistic
regression, support vector machine and multilayer perceptron) or tree depth (decision tree, random forest, gradient
boosting regression tree). All curves for the six proposed methods exhibit convergence well. The distributions of
variable importance for six input parameters except MLP are shown in Fig. 5. The average values of feature
importance of the 10-fold cross validation ar also shown. All of the five machine learning techniques take the BSRS-
5 score as the most important variable and the proportion is even larger for the case of suicide ideation ≥ 2. For both

ISSN: 2347-8578 www.ijcstjournal.org Page 300

International Journal of Computer Science Trends and Technology (IJCST) – Volume 10 Issue 5, Sep-Oct 2022

of DT and GBDT, the anxiety dimension has less importance compared to other three methods. For the screening
instrument to predict suicide ideation proposed in [8], the sensitivities for psychiatric group, community group and
general medical group are 83.76%, 21.57%, and 10.57%, respectively, and the specificities for the three groups are
72.17%, 99.49%, and 99.88%, respectively. As compared with the BSRS-5 score ≥7 in [9], a conventional criterion,
for the presence of suicide ideation ≥ 1, the proposed algorithms can improve the performances of accuracy,
sensitivity, specificity, precision, the AUC of ROC curve and the AUC of PR curve up to 5.7%, 35.9%, 4.6%,
65.2%, 4.3% and 53.2%, respectively; and for the presence of more severely intense suicide ideation ≥ 2, the
improvements are 6.1%, 26.2%, 5.8%, 83.5%, 2.8% and 64.7%, respectively. Instead of only considering the BSRS-
5 score in screen of suicide ideation like [8] and [9], our algorithm additionally takes the five psychopathological
domains which are related to the BSRS-5 score, i.e., anxiety, depression, hostility, interpersonal sensitivity and
insomnia as the input variables, and our schemes incorporating machine learning techniques provide better results
than those of [8] and [9]. In addition, we add several critical physiological data including age, sex, body height,
body weight, waist circumference, heart rate, systolic blood pressure, diastolic blood pressure and physical activity
on the initial inputs of the BSRS-5 score and related five psychopathological domains in our proposed models. We
find that the performances are only improved by incorporating these nine physiological data into the model of
logistic regression. All of the performances of logistic regression regarding accuracy, sensitivity, specificity and
precision reach 100% for suicide ideation ≥ 1 and 99.9% for suicide ideation ≥ 2. However, for the other five
machine learning methods, the performances are not getting better with additional inputs of these physiological data.
As the incidence of suicide attempts is relatively low, a meta-analysis reveals that the utility of suicide ideation for
predicting later suicide is limited by low positive predictive value and modest sensitivity [41]. Machine learning
techniques for the BSRS-score and related five psychopathological domains can be aimed for later suicide attempts
in future work.

ISSN: 2347-8578 www.ijcstjournal.org Page 301

International Journal of Computer Science Trends and Technology (IJCST) – Volume 10 Issue 5, Sep-Oct 2022

Fig.3 Receiver Operating Characteristic Curves and Precision Recall Curves

Fig.4 A set of learning curves for each of the six machine learning methods

Fig.5 Feature Importance for the machine learning methods in the proposed algorithm

ISSN: 2347-8578 www.ijcstjournal.org Page 302

International Journal of Computer Science Trends and Technology (IJCST) – Volume 10 Issue 5, Sep-Oct 2022

V. FUTURE SCOPE AND Dandeker, R. J. Rona, M. Hotopf, S. Wessely

CONCLUSION “What are the consequences of deployment to Iraq
and Afghanistan on the mental health of the UK
Our study uses machine learning techniques
armed forces? A cohort study,” The Lancet, vol.
for several psychological stress dimensions training on
375, no. 9728, pp.1783-1797, May 2010.
the prediction of suicide ideation. This paper utilizes
[5] N.S. Tzeng, C. K. Chen, T. S. Wang, H. A. Chang,
six machine learning techniques to predict the
Y. C. Kao, H. W. Yeh, W. S. Chiang, and S. Y.
presence of any or more severely intense suicide
Huang, “Forensic psychiatric evaluation for
ideation of military personnel. Normalization of input
military absenteeism in Taiwan,” Journal of the
data and imbalanced classification strategy facilitate
American Academy of Psychiatry and the Law,
the prediction of machine learning methods. The
vol.44, pp.352-358, September 2016.
experimental results show that the techniques of MLP
[6] K. E. Bachynski, M. Canham-Chervak, S. A.
and SVM provide the best performance for the two
Black, E. O. Dada, A. M. Millikan, and B. H.
predictions. Most of the proposed machine learning
Jones, “Mental health risk factors for suicides in
techniques take the BSRS-5 score to be the most
US Army, 2007–2008,” Injury Prevention, vol.
pivotal variable and the five psychopathological
18, pp.405-412, December 2012.
dimensions are also adopted as the features to improve
[7] K. L. Zuromski, S. L. Bernecker, P. M. Gutierrez,
the screening. As compared with the prior study using
T. E. Joiner, A. J. King, H. Liu, J. A. Naifeh, M.
the BSRS-5 score only, the machine learning
K. Nock, N. A. Sampson, A. M. Zaslavsky, M. B.
techniques can improve the performances of predicting
Stein, R. J. Ursano, and R. C. Kessler,
suicide ideation. This work can substantially help to
“Assessment of a risk index for suicide attempts
screen out the military personnel at high suicide risk
among US Army soldiers with suicidal ideation,”
for suicide prevention.
JAMA Network Open, vol.2, no.3, e190766,
March 2019.
REFERENCES
[8] F.W. Lung and M.B. Lee, “The five-tem brief
[1] A. M. Gadermann, C. C. Engel, J. A. Naifeh,M. K. symptom rating scale as a suicidal ideation
Nock, M. Petukhova, P. N. Santiago, W. screening instrument for psychiatric inpatients and
Benjamin, and A. M. Zaslavsky, and R. C. community residents,” BMC Psychiatry, vol.8,
Kessler, “Prevalence of DSM-IV major no.53, July 2008.
depression among U.S. military personnel: meta- [9] C. C. Ma, and Y. M. Tai, “Cut-off values of five-
analysis and simulation,” Military Medicine, vol. item brief symptom rating scale in evaluating
177, no. 8, pp.47-59, August 2012. suicidality among military recruits,” Taiwanese
[2] L. K. Richardson, B. C. Frueh, and R. Acierno, Journal of Psychiatry, vol. 28, no. 2, pp.109-120,
“Prevalence estimates of combat-related PTSD: June 2014.
critical review,” Australian and New Zealand [10] L. Han, S. Luo, J. Yu, L. Pan, and S. Chen, “Rule
Journal of Psychiatry, vol. 44, no. 1, pp.4-19, extraction from support vector machines using
January 2010. ensemble learning approach: An application for
[3] W. Berger, E. S. F. Coutinho, I. Figueira, C. diagnosis of diabetes,” IEEE Journal of
Marques-Portella, M. P. Luz, T. C. Neylan, C. R. Biomedical and Health Informatics, vol. 19, no. 2,
Marmar, and M. V. Mendlowiz, “Rescuers at risk: pp. 728–734, March 2015.
a systematic review and meta-regression analysis [11] M. Shahin, B. Ahmed, S. T. B. Hamida, F. L.
of the worldwide current prevalence and Mulaffer, M. Glos, and T. Penzel, “Deep learning
correlates of PTSD in rescue workers,” Social and insomnia: assisting clinicians with their
Psychiatry and Psychiatric Epidemiology, vol. 47, diagnosis,” IEEE Journal of Biomedical and
no. 6, pp.1001-1011, June 2012. Health Informatics, vol. 21, no. 6, pp. 1546 -
[4] N. T. Fear, M. Jones, D. Murphy, L. Hull, A. 1553, November 2017.
CIversen, B. Coker, L. Machell, J. Sundin, C. [12] J. Shi, X. Zheng, Y. Li, Q. Zhang, and S. Ying,
Woodhead, N. Jones, N. Greenberg, S. Landau, C. “Multimodal neuroimaging feature learning with
multimodal stacked deep polynomial networks for

ISSN: 2347-8578 www.ijcstjournal.org Page 303

International Journal of Computer Science Trends and Technology (IJCST) – Volume 10 Issue 5, Sep-Oct 2022

diagnosis of Alzheimer’s disease,” IEEE Journal Bluemke, and J. A. C. Lima, “Cardiovascular

of Biomedical and Health Informatics, vol. 22, no. event prediction by machine learning: the multi-
1, pp. 173-183, January 2018. ethnic study of atherosclerosis,” Circulation
[13] B. Lei, Peng Yang, Y. Zhuo, F. Zhou, D. Ni, S. Research, vol. 121, no. 9, pp. 1092-1101, October
Chen, X. Xiao, and T. Wang, “Neuroimaging 2017.
retrieval via adaptive ensemble manifold learning
for brain disease diagnosis,” IEEE Journal of
Biomedical and Health Informatics, vol.23, no.4,
pp.1661-1673, July 2019.
[14] G. M. Lin, M. J. Chen, C. H. Yeh, Y. Y. Lin, H.
Y. Kuo, M. H. Lin, and M. C. Chen, S. D. Lin, Y.
Gao, A. Ran, and C. Y. Cheung, “Transforming
retinal photographs to entropy images in deep
learning to improve automated detection for
diabetic retinopathy,” Journal of Ophthalmology,
vol. 2018, Article ID 2159702, September 2018.
[15] X. Du, R. Tang, S. Yin, Y. Zhang, and S. Li,
“Direct segmentation-based full quantification for
left ventricle via deep multi-task regression
learning network,” IEEE Journal of Biomedical
and Health Informatics, vol. 23, no. 3, pp. 942-
948, May 2019.
[16] A. Akay and H. Hess, “Deep learning: current and
emerging applications in medicine and
technology,” IEEE Journal of Biomedical and
Health Informatics, vol. 23, no. 3, pp. 906-920,
May 2019.
[17] M. J. Chen, P. H. Yang, M. T. Hsieh, C. H. Yeh,
C. H. Huang, C. M. Yang, and G. M. Lin,
“Machine learning of PM2.5 and PM10
concentrations to relate with outpatient visits for
upper respiratory tract infections in Taiwan: a
nationwide analysis,” World Journal of Clinical
Cases, vol.6, no.8, pp.200-206, August 2018.
[18] L. Feng, Z. Li and Y. Wang, “VLSI design of
SVM-based seizure detection system with on-chip
learning capability,” IEEE Transactions on
Biomedical Circuits and Systems, vol. 12, no. 1,
pp. 171-181, February 2018.
[19] J. P. Dominguez-Morales, A. F. Jimenez-
Fernandez, M. J. Dominguez-Morales, and G.
Jimenez-Moreno, “Deep neural networks for the
recognition and classification of heart murmurs
using neuromorphic auditory sensors,” IEEE
Transactions on Biomedical Circuits and Systems,
vol. 12, no. 1, pp. 24–34, February 2018.
[20] B. Ambale-Venkatesh, X. Yang, C. O. Wu, K.
Liu, W. G. Hundley, R. McClelland, A. S. Gomes,
A. R. Folsom, S. Shea, E. Guallar, D. A.

ISSN: 2347-8578 www.ijcstjournal.org Page 304