1038. Cardiovascular event risk prediction and temporo-spatial management toolbox for high risk patients

Authors: Clemens M. Schirmer, MD, PhD, FAANS ; Suresh Sampangiraman (Wilkes Barre, PA)


Geisinger is a integrated health system with a large population at risk for cardiovascular and stroke (CVA) events. Clinicians in everyday practice are lacking robust tools to assess patient based CVA event risks and the healthcare system is lacking tools to identify high-risk patients for targeted interventions. Previous work in this area relies on small datasets that are highly curated and results are difficult to generalize or utilize in daily practice. 


Taking a two-pronged approach we developed a methodology to automatically recalculate the ASCVD risk score for every patient touchpoint in the electronic medical record and display the resultant temporal risk assessment both for individual patients and map high-risk patients to spatial location in the delivery network and identify high-risks patients for targeted management interventions. We also developed a machine learning model to incorporate an increased number of un-weighted feature vectors compared to the ASCVD score. 


A total of 44 million patient encounters since 2013 was used to create a snapshot of the population at risk and several iterations of random forest and support vector machine models were compared for an accuracy of 75%. CVA risk scores can be mapped, smoothed and displayed in realtime on a geospatial information display dashboard which allows the user to focus on population level features and drill down to individual patients. 


We conclude that supervised CVA risk prediction models relying on objective feature vectors are possible and achieve a moderate accuracy. Further work is required to identify additional features of the existing medical record that can be parsed with minimal burden on data cleaning to enhance the model accuracy. Multi-level temporospatial risk score visualization has been achieved and can serve as an addition to the toolkit for both individual management decisions and population-management tools.