Aligning machine learning models with clinician knowledge for understanding, prediction, and prevention of suicide

The National Institute of Mental Health awarded Wellesley College $707,584 under the Mental Health Research Grants program (CFDA 93.242) on August 1, 2026, to develop machine learning models that encode clinical knowledge for understanding, prediction, and prevention of suicide. The project addresses three data analysis bottlenecks in ecological momentary assessment studies of suicidal thoughts and behaviors: small sample sizes that limit deep learning applications, stochasticity in external triggering events that standard methods fail to model, and difficulty empirically testing verbal psychological theories of suicide. The research will encode qualitative clinical knowledge—intuitive knowledge difficult to mathematize—into deep learning methods to provide additional supervision, reduce reliance on limited data, incorporate clinical guidance in model discovery, and enable empirically testing psychological theories. Performance occurs in Wellesley Hills, Wellesley, Massachusetts, with a period of performance through May 31, 2031.

Abstract

Ecological momentary assessment (EMA) studies capture unprecedented data about suicidal thoughts and be- haviors (stbs) via smartphones and wearable biosensors. While promising, EMA studies are expensive to run, require a large staff, and are limited in scientific value due to three data analysis challenges. These challenges create a bottleneck for suicide research. Challenge 1: small data size. Even the largest of EMA studies yield modest sample sizes with low compliance. This bars applications of (data-hungry) deep learning (DL) methods, successful in other areas of healthcare. Challenge 2: stochasticity. Stbs are, in part, driven by stochastic (or random) external events that cannot be captured (e.g. reactions to triggering stimuli). Methods that do not model external stochasticity often learn spurious correlations, making incorrect and overconfident forecasts of future PA- tient outcomes. Challenge 3: empirically testing theories of suicide. Empirically evaluating theories with data is difficult given their verbal (non-quantitative) nature; mathematically formalized theories require an extreme level of specification detail, making IT difficult for clinical and machine learning (ML) researchers alike to determine which details capture the high-level idea of the theory. As such, current methods struggle to accurately forecast stbs, identify those at imminent risk, and test psychological theories of suicide. Our central hypothesis is that encod- ING clinical knowledge in DL models addresses key data analysis challenges. We define qualitative clinical knowledge (QCK) as clinical knowledge that is difficult to mathematize, like intuition. Encoding QCK in DL meth- ODS will provide the model additional supervision, reducing its over-reliance on limited data (challenge 1). Guided by clinical knowledge, IT will discover scientifically viable associations between EMA outcomes and unobserved external factors, avoiding spurious correlations due to stochasticity (challenge 2). Lastly, our method will pro- vide clinicians with a mechanism for encoding their knowledge into the model without specifying uninterpretable mathematical details (challenge 3). Our project has three aims. Aim 1: develop a new DL model, tailored for noisy, irregularly-sampled, partially-observed EMA data. Aim 2: identify types of QCK that clinicians can reliably provide, encode them into our DL model, and assess their predictive value. Exploratory aim 3: prototype an interactive modeling paradigm that algorithmically solicits clinicians for QCK (from aim 2) into the DL model (from aim 1). Expected outcomes: this project will yield (1) theoretically and empirically validated ML methods for use in future scientific and clinical work, (2) new insights/findings from existing EMA data, increasing the scientific yield from previous NIH-funded projects, (3) an interactive ML paradigm for embedding clinical knowledge in DL methods, useful in all mental health and healthcare contexts, and (4) highly optimized, modular open-source code, available for the broader community to apply and adapt to their needs.