SCH: Machine learning for personalized preventative intervention in perinatal depression

PROJECT SUMMARY (See instructions}: The goal for this project is to develop machine learning (ML) to accelerate the discovery of scalable, interpretable, and personalized preventative interventions for perinatal depression. Approximately 15% of pregnant individuals experience perinatal depression, which can have devastating long-term

consequences. Suicide is a leading cause of death among new mothers in the U.S. However, individualized preventive interventions are not routinely offered at present due to lack of routine screening practices and limited resources. Machine learning offers an opportunity to improve mental health services

during the perinatal period by identifying patients who would benefit from specific preventative interventions. We will develop fundamental advances in ML techniques for the discovery of personalized interventions as well as advances in the social science of incorporating domain and lived experience into

algorithmic systems. Our specific aims bridge prediction with the adaptive experimentation needed to identify personalized interventions. In Aim 1, we will develop methods which use existing historical data to lay the groundwork for a randomized experiment of interventions, including to robustly inform which

variables to measure and how to set an initial allocation policy based on those variables. In Aim 2, we will elicit domain expertise from clinicians and lived experience expertise from perinatal individuals via semistructured interviews which will inform both the requirements for a trustworthy and implementable ML

system and a structured representation of clinical expertise that can be incorporated to initialize a ML policy together with historical data. Finally, in Aim 3, we will synthesize these products into an integrated framework for online learning to discover personalized preventative interventions. The key component of

this framework is continued interaction with patients to provide intermediate feedback and accelerate convergence to a high-quality policy for allocating preventative interventions.