Original scientific paper

https://doi.org/10.24869/psyd.2023.62

INTERPRETABLE ESTIMATION OF SUICIDE RISK AND SEVERITY FROM COMPLETE BLOOD COUNT PARAMETERS WITH EXPLAINABLE ARTIFICIAL INTELLIGENCE METHODS

Neslihan Cansel ; Inonu University Faculty of Medicine, Department of Psychiatry, Malatya, Turkey
Fatma Hilal Yagin ; Inonu University Faculty of Medicine, Department of Biostatistics and Medical Informatics, Malatya, Turkey
Mustafa Akan ; Malatya Training and Research Hospital, Department of Psychiatry, Malatya, Turkey
Bedriye Ilkay Aygul ; Inonu University Faculty of Medicine, Department of Psychiatry, Malatya, Turkey

Abstract

Background: The peripheral inflammatory markers are important in the pathophysiology of suicidal behavior. However, methods
for practical uses haven’t been developed enough yet. This study developed predictive models based on explainable artificial intelligence
(xAI) that use the relationship between complete blood count (CBC) values and suicide risk and severity of suicide attempt.
Subjects and methods: 544 patients who attempted an incomplete suicide between 2010-2020 and 458 healthy individuals were
selected. The data were obtained from the electronic registration systems. To develop prediction models using CBC values, the data
were grouped in two different ways as suicidal/healthy and attempted/non-attempted violent suicide. The data sets were balanced for
the reliability of the results of the machine learning (ML) models. Then, the data was divided into two; 80% of as the training set and
20% as the test set. For suicide prediction, models were created with Random Forest, Logistic Regression, Support vector machines
and XGBoost algorithms. SHAP, was used to explain the optimal model.
Results: Of the four ML methods applied to CBC data, the best-performing model for predicting both suicide risk and suicide
severity was the XGBoost model. This model predicted suicidal behavior with an accuracy of 0.83 (0.78-0.88) and the severity of
suicide attempt with an accuracy of 0.943 (0.91-0.976). Lower levels of NEU, WBC, MO, NLR, MLR and, age higher levels of HCT,
PLR, PLT, HGB, RBC, EO, MPV and, BA contributed positively to the predictive created model for suicide risk, while lower PLT,
BA, PLR and RBC levels and higher MO, EO, HCT, LY, MLR, NEU, NLR, WBC, HGB and, age levels have a positive contribution to
the predictive created model for violent suicide attempt.
Conclusion: Our study suggests that the xAI model developed using CBC values may be useful in detecting the risk and severity
of suicide in the clinic.

Keywords

suicidality; violent suicide; nonviolent suicide; CBC; machine learning; Explainable Artificial Intelligence

Hrčak ID:

307222

URI

https://hrcak.srce.hr/307222

Publication date:

17.4.2023.

Visits: 491 *