JMIR Publications

ABSTRACT

Background:

Suicide is a major public health concern whilst the prediction of suicidal behaviors, remains challenging.

Objective:

To construct explainable and applicable algorithms using machine learning approach for suicidal behaviors prediction.

Methods:

Based on the prospective cohort of UK Biobank, we included 223 eligible cases with suicide attempts or deaths (suicidal behaviors), according to inpatient hospital or death register data, within 1 year from baseline and randomly selected controls (1:20) without such a condition, n=4,460. We similarly identified 833 cases with suicidal behaviors 1-6 years from baseline and 16,600 corresponding controls. Based on 143 input features, we applied bagged balanced lightGBM with stratified 10-fold cross-validation to construct prediction models. The external-validity of the established models was assessed among 50,310 individuals who participated in the UK Biobank repeated assessments.

Results:

Suicidality cases were on average 56 years, with equal sex distribution. The application of these models in the external validation dataset demonstrated good model performance (the area under the receiver operating characteristic curve = 0.919 and 0.892). The simplified models with top 20 important features were considered optimal for further application, based on which we found individuals in the top quintile of predicted risk accounting for 91.7% and 80.7% of all suicidality cases within 1 year and during 1-to-6 years, respectively. We further obtained comparable predication accuracy when applying these models to sub-populations with different genetic susceptibilities to suicidality.

Conclusions:

We established a machine learning-based algorithm for predicting both short- and long-term risk of suicidality with high accuracy, across populations of varying genetic risk for suicide, highlighting a cost-effective method of identifying individuals with high-risk of suicidality.