Implication of machine learning models versus traditional models for the prediction of suicidal thoughts or ideation in west of Iran; data mining approaches on a population-based cross-sectional study

Abstract

Objective To identify the effective factors in suicidal thoughts or ideations by comparing several classification data mining methods and logistic regression (LR). Method This was a secondary data analysis conducted on data from a cross-sectional study involving 1500 individuals selected using multi-stage stratified cluster random sampling in the urban area of Ilam City during 2023. The data was collected by a standardized questionnaire. Five classification methods, including decision tree (DT), random forest (RF), support vector machine (SVM), neural networks, and LR, were used to identify the effective factors in the suicide thought or ideation. Results Data from 1370 individuals were analyzed. The SVM model outperformed others in most indicators, with 77.9 sensitivity, 95.3 negative predictive value, and the highest balanced accuracy (79.3). Its precision-recall AUC, along with LR, was about 60 higher than other models. In contrast, the DT model showed superior specificity (99.9), positive predictive value (50), positive likelihood ratio (5.60), and negative likelihood ratio (0.99). Across DT, RF, and SVM, the main predictors of suicidal ideation were suicide attempt history, tiredness of life, BMI, and age. Conclusion AI models specifically SVM and DT outperform traditional ones for detecting suicidal ideation. Key predictors include a history of suicide attempts, being tired of life, depression, and anxiety, highlighting areas for health policymakers to focus on in prevention strategies.