Grant: CIHR Fellowship

Suicide is preceded by suicidal thinking, which people are often hesitant to disclose. This can lead to missed opportunities for closer monitoring and intervention. To address this issue, the proposed project examines predictors of suicidal thinking that do not rely on self-report. Various predictors have been identified, such as age, sex, mental health status, genetic susceptibility, and sleep disturbances. As information on these predictors is increasingly captured in the clinic (e.g., with questionnaires, psychiatric notes, DNA sampling, actigraphy), clinical datasets are becoming larger and more complicated to analyze. One way to deal with diverse datasets is by converting them into smaller, more manageable representations, called embeddings. We will evaluate whether embeddings can represent data for 500 depressed patients over 3000 visits to The Centre for Addiction and Mental Health (CAMH). To do this, we will examine how well embeddings contribute to predictions of suicidal thinking, measured with a depression questionnaire. We will compile four datasets for each patient visit containing varying numbers of predictors of suicidal thinking. The smallest dataset will contain demographic predictors, whereas the largest dataset will contain these predictors, as well as predictors from clinical assessments, psychiatric notes, genetic susceptibility, and daily activity and sleep. From these datasets, we will generate four sets of embeddings and examine how well each set predicts suicidal thinking. If the embeddings are good representations of the datasets, prediction based on embeddings of larger datasets (with more predictors) should outperform embeddings of smaller datasets (with less predictors). Our findings have potential to inform the development of a suicide risk estimate, to complement clinical judgment when self-reported suicidal thinking is unavailable or unreliable. Improved detection may lead to opportunities for early intervention and prevention.