Identification of Suicidal Ideation in the Canadian Community Health Survey—Mental Health Component Using Deep Learning

Abstract

Introduction: Suicidal ideation (SI) is prevalent in the general population, and is a risk factor for suicide. Predicting which patients are likely to have SI remains challenging. Deep Learning (DL) may be a useful tool in this context, as it can be used to find patterns in complex, heterogeneous, and incomplete datasets. An automated screening system for SI could help prompt clinicians to be more attentive to patients at risk for suicide.Methods: Using the Canadian Community Health Survey—Mental Health Component, we trained a DL model based on 23,859 survey responses to classify patients with and without SI. Models were created to classify both lifetime SI and SI over the last 12 months. From 582 possible parameters we produced 96- and 21-feature versions of the models. Models were trained using an undersampling procedure that balanced the training set between SI and non-SI; validation was done on held-out data.Results: For lifetime SI, the 96 feature model had an Area under the receiver operating curve (AUC) of 0.79 and the 21 feature model had an AUC of 0.77. For SI in the last 12 months the 96 feature model had an AUC of 0.71 and the 21 feature model had an AUC of 0.68. In addition, sensitivity analyses demonstrated feature relationships in line with existing literature.Discussion: Although further study is required to ensure clinical relevance and sample generalizability, this study is an initial proof of concept for the use of DL to improve identification of SI. Sensitivity analyses can help improve the interpretability of DL models. This kind of model would help start conversations with patients which could lead to improved care and a reduction in suicidal behavior.

Editorial: Methods and interventions to predict and tackle suicide risk

“…The first is that, due to the variability in predictors selected by models across datasets, there remains significant value in research focusing on identifying both general and population-specific risk factors. This literature allows us to interrogate models generated using machine learning ( 10 , 11 ). The second is that machine learning models perform well on some key metrics ( 12 ) but will likely require health systems to do the hard work of data harmonization– likely driven by the literature on risk factors– to produce large datasets, which will allow for further model optimization.…”