FRISTS: High-Performing Interpretable Medical Prediction

Sophia Lin, Xinyu Dong, Fusheng Wang

Abstract: Current medical prediction models struggle with three key challenges: (1) lackluster performance on real-world health records, (2) reliance on non-routine tests (e.g., ECGs or blood tests), and (3) model uninterpretability, which prevents adoption. A key challenge is that increasing transparency often decreases the model’s performance. We present FRISTS (Feature-Ranked Interpretable Sequential Time Series), a novel medical prediction approach that uses patients’ previous medical history records to predict future diagnoses. FRISTS combines time series-based recurrent neural networks, e.g. LSTMS, AI-guided feature selection, and a Shapley-inspired permutation method that enables interpretability. We applied our model to 18 million health records in the Cerner Health Facts database to predict heart failure, which yielded a six-fold increase over decision trees and a significant increase over LSTM while capturing more condition-specific features. Since FRISTS is extendable to any prediction task on electronic health records, it accelerates the adoption of machine learning methods in AI-assisted healthcare that achieve both high real-world performance and interpretability.