Artificial Intelligence-Based Prediction Models for Suicide Risk Assessment and Prevention: A Retrospective Observational Study

Abstract

Background: Suicide remains one of the leading causes of preventable mortality worldwide and represents a major public health challenge. Despite advances in psychiatric assessment and mental health services, accurately identifying individuals at imminent risk of suicide remains difficult because suicidal behavior arises from a complex interaction of biological, psychological, social, and environmental factors. Conventional clinical risk assessment methods primarily rely on clinician judgment and standardized screening tools, which often demonstrate limited predictive accuracy and considerable inter-observer variability. Recent developments in artificial intelligence (AI) and machine learning (ML) have enabled the integration of large-scale hospital medical records, demographic variables, psychiatric history, medication profiles, laboratory findings, and behavioral characteristics to generate individualized suicide risk predictions. AI-driven predictive models have demonstrated superior performance compared with traditional statistical approaches by identifying complex nonlinear relationships and hidden interactions among multiple risk factors. However, evidence regarding the applicability of these models in routine clinical practice remains limited, particularly in retrospective hospital-based settings.
Objective: To evaluate the performance of artificial intelligence-based prediction models for suicide risk assessment and prevention using retrospective clinical data and to identify the most influential predictors associated with high suicide risk.
Materials and Methods: A retrospective observational study was conducted over four months using medical records of 100 patients who underwent psychiatric evaluation at PMCH, Patna. Demographic characteristics, psychiatric diagnoses, previous suicide attempts, substance use disorders, family history of mental illness, medication adherence, hospitalization history, depression severity scores, anxiety scores, and psychosocial variables were extracted from hospital medical records maintained in the Medical Record Department (MRD). Missing values were managed through multiple imputation where appropriate. Four predictive models—Logistic Regression, Random Forest, Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost)—were developed and internally validated using five-fold cross-validation. Model performance was evaluated using accuracy, sensitivity, specificity, precision, F1-score, area under the receiver operating characteristic curve (AUC), and calibration statistics. Statistical significance was considered at p <0.05.
Results: Among the 100 patients included, 36 were categorized as high suicide risk based on documented psychiatric evaluation. Previous suicide attempt, severe depressive symptoms, substance use disorder, medication non-adherence, unemployment, and multiple psychiatric hospitalizations were significantly associated with high suicide risk (p<0.05). XGBoost demonstrated the highest predictive performance with an AUC of 0.94, sensitivity of 91.7%, specificity of 89.1%, accuracy of 90.0%, and F1-score of 0.90. Random Forest achieved an AUC of 0.91, while SVM and Logistic Regression demonstrated AUC values of 0.88 and 0.84, respectively. Feature importance analysis identified previous suicide attempt, depression severity score, medication adherence, substance use disorder, psychiatric hospitalization history, and age as the strongest predictors.
Conclusion: Artificial intelligence-based prediction models demonstrated excellent discrimination in identifying patients at elevated suicide risk. Tree-based ensemble algorithms, particularly XGBoost, outperformed conventional logistic regression by effectively modeling complex clinical interactions. Integration of AI-assisted decision support into psychiatric practice may facilitate early identification of vulnerable individuals, optimize allocation of mental health resources, and strengthen suicide prevention strategies. Prospective multicenter validation studies are warranted before widespread clinical implementation.