A Computational Approach to Understanding the Risk Architecture of Suicidal Thoughts and Behaviors in Older Adults

Suicide is a critical global public health issue, with a notably high incidence among older people in a context of a rapidly aging population. While recent neuroimaging studies have identified certain magnetic resonance imaging (MRI) markers correlated with late-life suicide, a comprehensive examination comparing diverse MRI modalities and assessing the contributions of various demographic and clinical factors remain lacking. A major challenge in understanding the neural correlates of late-life suicide lies in its complex overlap with late-life depression. In addition to mood symptoms, cognitive decline and other age specific factors add further complexity to this association. Our project aims to elucidate the risk architecture of late-life suicide by disentangling the shared and distinct features of late-life suicide and late-life depression. Using advanced machine learning techniques, we will conduct a data driven analysis with a deep learning approach, specifically a conditional variational autoencoder (cVAE), on a large-scale open dataset (UK Biobank, n=3237) to identify features that best predict suicidal behavior independent of depression, using neuroimaging (structural MRI, resting-state MRI, and fMRI from a face-shape emotion task) and demographic/cognitive data (Aim 1). This analysis will not only characterize structural and functional MRI data and demographic factors in older adults with a history of suicidal attempt, but also offer insights into how depression and cognitive decline may interact with neuroimaging data to influence suicide risk in late life. To evaluate the generalizability of the cVAE model developed in Aim 1, we will apply it to an internal dataset (n=38) of older adults with late-life depression to assess transfer learning potential (Aim 2a). Additionally, we will explore a conventional fMRI analysis approach by applying a General Linear Model (GLM) to identify brain activation of emotion processing associated with suicide and/or depression, followed by Principal Component Analysis (PCA) to capture shared and distinct neural patterns (Aim 2b). We will use the same face-shape emotion fMRI task as in Aim 1. This analysis, independent of Aim 1, will allows us to explore how conventional and machine learning approaches contribute to understanding the neural signatures of these conditions. My training goals through this project include gaining a deeper understanding of suicide prevention, aging, and machine learning; acquire hands-on experience in large-scale data management and computational modelling; fostering collaborations with leading researchers in the field; and ultimately developing the skills needed to become an independent researcher specializing in late-life suicide. By identifying the shared and distinct features of late-life suicide and late-life depression, this study could advance our understanding of the specific neural and psychosocial factors driving suicide risk in older adults. These insights may inform targeted interventions that address unique aspects of each condition, ultimately contributing to more effective suicide prevention strategies in late life.