Exploratory Research Project - ADAPT

ADAPT (EXPLORATORY PROJECT): SUMMARY/ABSTRACT Significance: Machine learning-based risk algorithms have transformational potential to improve suicide risk identification. However, the lack of large-scale validations, transfer guidance, and automated learning-based adaptation impedes adoption in clinical practice. This project aims to address this translation gap by systematically assessing and improving a suicide risk algorithm’s generalizability and adaptability from an original development setting to a new healthcare system. Investigators: The transdisciplinary team has comprehensive expertise in applying advanced machine learning techniques on electronic health record (EHR) data for predictive modeling and prevention analytics (Liu, Aseltine, Simon), studying clinical diagnosis, prognosis and treatment of serious mood disorders and suicide (Rothschild), identifying and assessing suicide risk (Simon), and promoting health services delivery redesign through technology and implementing informatics tools in clinical settings (Gerber). Innovation: This pioneering study will comprehensively evaluate and improve the generalizability and adaptability of an evidence-based suicide risk algorithm in different contexts. The team will build a unified pipeline of Automated, Data-driven, AdaPtable, and Transferable learning for suicide risk prediction (ADAPT). The versatile ADAPT tool will be accessible to non-expert users and compatible with EHR common data model standards, providing a scalable, interpretable and sustainable solution to risk algorithm translation across different clinical contexts. Moreover, we will design an advanced deep learning approach for suicide risk prediction and evaluate its effectiveness on generalizability and adaptability. Approach: The proposed study aims to assess the generalizability of the Mental Health Research Network (MHRN) risk algorithm and explore transfer and ensemble learning to adapt a previously learned model from original data sources into a tailored one optimized for a new health system (Aim 1); develop a unified pipeline, ADAPT, to integrate data preprocessing, model assessment and adaptation, model interpretation, and automated learning; explore how ADAPT’s results can be used to help match individuals to a range of intervention approaches where specialized or intensive treatment is reserved for those with the highest risk (Aim 2); design an innovative deep learning approach and test its effectiveness using ADAPT (Aim 3a); engage stakeholders to better understand potential barriers and facilitators to implementation, iteratively improve ADAPT’s usability, acceptability, and feasibility through their feedback using validated scales (Aim 3b). Environment: The UMass Chan Medical School (UMass) has proven its ability to support this ambitious study by its success with numerous NIMH-funded systems-based suicide prevention studies. Impact: The study holds great potential for promoting the implementation of an evidence-based EHR suicide risk algorithm in clinical practice. Paired with effective interventions, it will enable improved suicide prevention.