Personalized Stroke Risk Stratification in Atrial Fibrillation: Integrating Probabilistic Graphical Models and High-dimensional EHR Data for Interpretable Prognosis - PROJECT SUMMARY/ABSTRACT Implementation of contemporary strategies to reduce stroke related to atrial fibrillation (AF) is limited by (1) ru- dimentary stroke risk stratification tools and (2) disparities in care and outcomes. There remains a critical need for personalized, socially aware, equitable stroke risk prediction among patients with AF, in order to optimally implement contemporary stroke-prevention therapies. The overarching goal is to build on recent R56 support to develop a portable, equitable, personalized risk-stratification tool to improve stroke prevention among pa- tients with AF. The objectives are to: (i) discover new stroke risk factors for patients with AF, incorporating so- cial determinants of health (SDoH) with millions of health record covariates, using an innovative comorbidity discovery framework (Poisson Binomial Comorbidity [PBC]); (ii) combine these new risk factors with estab- lished ones using machine-learning (ML), in order to determine their network structure and provide explainabil- ity; and (iii) develop, deploy, and test a personalized stroke risk stratification tool for AF patients across differ- ent health systems in a disparity-aware fashion. The central hypothesis is that stroke risk stratification can be improved through methods that: leverage all available data, including SDoH; capture and quantify synergies among known and newly discovered risk factors in socioeconomic context; and can be ported to other health systems, adapting to different populations. The rationale for this project is that current AF-related stroke risk management lacks the precision and awareness required to optimally implement treatments because it does not adequately account for (1) population diversity, (2) SDoH and disparities, (3) synergistic interactions among risk factors, and (4) novel, emerging risk factors. The central hypothesis will be tested by pursuing three spe- cific aims: 1) Discover new clinical and socioeconomic relationships that influence stroke risk in patients with AF; 2) Develop a socially conscious, ML-based machinery for calculating personalized stroke risk among pa- tients with AF; and 3) Benchmark an ML-based stroke risk stratification across a diverse cohort of health sys- tems within PCORnet and discover biases and drivers of downstream care disparities. In the first aim, the PBC approach will be used to leverage large datasets that include SDoH, in order identify new risk markers. The second aim will focus on building novel, Probabilistic Graphical Models (PGMs) to understand the impact of SDoH on AF-related stroke risk. In the third aim, the models will be tested across a diverse set of healthcare systems to understand portability, diversity, and bias. The research proposed in this application is innovative because it (1) leverages uniquely-available data on SDoH, (2) employs a much more powerful and portable analytic approach to understand risk; and (3) is designed with an eye towards understanding and reducing dis- parities and bias in risk prediction models. The proposed research is significant because it will improve care across the spectrum of patients with AF, while at the same time addressing disparities and bias. Ultimately, the results will yield a more personalized and equitable approach to stroke prevention in AF.
