Ten-year prediction of suicide death using Cox regression and machine learning in a nationwide retrospective cohort study in South Korea

Background: Death by suicide is a preventable public health concern worldwide. The aim of this study is to investigate the probability of suicide death using baseline characteristics and simple medical facility visit history data using Cox regression, support vector machines (SVMs), and deep neural networks (DNNs). Method: This study included 819,951 subjects in the National Health Insurance Service (NHIS)–Cohort Sample Database from 2004 to 2013. The dataset was divided randomly into two independent training and validation groups. To improve the performance of predicting suicide death, we applied SVM and DNN to the same training set as the Cox regression model. Results: Among the study population, 2546 people died by intentional self-harm during the follow-up time. Sex, age, type of insurance, household income, disability, and medical records of eight ICD-10 codes (including mental and behavioural disorders) were selected by a Cox regression model with backward stepwise elimination. The area of under the curve (AUC) of Cox regression (0.688), SVM (0.687), and DNN (0.683) were approximately the same. The group with top.5% of predicted probability had hazard ratio of 26.21 compared to that with the lowest 10% of predicted probability. Limitations: This study is limited by the lack of information on suicidal ideation and attempts, other potential covariates such as information of medication and subcategory ICD-10 codes. Moreover, predictors from the prior 12–24 months of the date of death could be expected to show better performances than predictors from up to 10 years ago. Conclusions: We suggest a 10-year probability prediction model for suicide death using general characteristics and simple insurance data, which are annually conducted by the Korean government. Suicide death prevention might be enhanced by our prediction model.