Predicting Post-discharge Suicidal Behavior in Bipolar Disorders Using Passive Sensing and Pattern Recognition

BACKGROUND: Bipolar disorder (BD) is a mood disorder with high recurrence and disability rates, and an estimated suicide risk 20 times higher than the general population.  This proposal aims to develop new data modeling and inference techniques that will enable more tailored clinical signal detection: examining changes within each individual, over time, to predict post-discharge suicidal behavior.  To do so, we propose integrating multimodal, high-dimensional electronic (e-) monitoring data, nonlinear techniques, and artificial intelligence classification systems. 

AIMS AND HYPOTHESES: AIMS: Aim 1: To obtain and integrate multimodal data to perform time-series analysis and to calculate entropy levels in 50 BD adults recently discharged from hospital following a suicide attempt.  Aim 2: To use machine learning techniques (Markov Brains) to distinguish participants at higher risk for a post-discharge suicide attempt based on their time-series and entropy levels (from Aim 1).   HYPOTHESES: H1: We will be able to collect enough data in 80% of our participants and to integrate multimodal data to perform time-series analysis and to calculate entropy levels. H2: Markov Brains will identify participants at higher risk for a post-discharge suicide attempt based on high (vs. low) auto-correlated time-series.

SAMPLE: 50 adults with BD recently discharged from hospital following a suicide attempt.

MEASURES: Using passive sensing (wearable called Oura ring), we will obtain objective (sleep, activity) and physiological data (heart rate variability, HRV).  Subjective data (mood, anxiety, sleep, and energy levels) will be collected using (i) self-rating scales; (ii) traditional clinician-administered scales; as well as the Beck Suicidal Ideation Scale (BSS) to document suicidal ideation or suicidal behaviors.

METHODS: We will use e-monitoring to densely sample several objective, physiological, and subjective variables in BD patients to predict post-discharge suicidal behavior. Specifically, we will integrate multimodal e-monitoring data to generate time-series and calculate entropy levels for each of these series in 50 BD patients who have been recently discharged from hospital following a suicide attempt. Objective and physiological data will be collected using a wearable; subjective data will be collected using electronic rating scales,

POTENTIAL IMPACT AND NEXT STEPS: This application challenges more traditional prediction models by conceptualizing inter- and intra-individual variability as a dynamic property of biological systems.  Digital-health technology and new analytical methods allow us to gather and analyze multimodal data and develop novel predictive models for suicide that could be sensitive and specific enough to make actionable individual clinical predictions.  

This application aligns with 2 out of the 3 AFSP priority areas: (i) Diversity (Principal Applicant is a researcher from an underrepresented background proposing research focused on understanding and preventing suicide); (ii) Evaluation of technological tools for suicide prevention (using passive sensing). This study will address the priority question of the National Action Alliance for Suicide Prevention: How can we better/optimally predict risk? If we confirm our hypotheses, the feasibility and pilot data obtained with this project will be used to prepare and submit a larger, more definitive application.  Ultimately, if we develop the capacity to predict suicidal behavior, we should be able to prevent it.