Forecasting Prescription Efficacy

Hao-Ren Yao, Oskar Mencer, Han-Sun Chiang, Der-Chen Chang, Ophir Frieder

Abstract: The shortage of skilled healthcare professionals, i.e., physicians and nursing staff, and a growing patient population has all but mandated AI-powered treatment decision support systems. Yet, the complexity of healthcare data limits the development and utility of existing efforts, namely for such systems to effectively and efficiently leverage Electronic Health Records (EHR) by extracting actionable insights and providing interpretable clinical guidance. We present an accurate, interpretable, Minimal Viable Product (MVP) that assists clinicians in predicting personalized outcomes from among a set of ever-evolving, possible treatments of common ailments. That is, the efficacy of a particular treatment, from among multiple options, for a specific patient, given that patient’s medical history, is predicted. Thus, the treating provider is informed of, a priori of prescribing a treatment, which treatment will most-likely be effective for the given patient. By harnessing the power of self-supervised graph representation learning and kernelized binary classification, our system can not only process highly-complex medical data, namely data with high-dimensionality, heterogeneity, and temporality, but also provides interpretation via user-intuitive interfaces. The underlying methods and systems of the presented MVP are protected via multiple US patents, are licensed internationally, and after preliminary positive results, the MVP is currently under rigorous clinical evaluation by clinicians in our collaborating hospital.