Interpretable Predictive Models for Healthcare via Rational Multi-Layer Perceptrons

Abstract

The healthcare sector has recently experienced an unprecedented surge in digital data accumulation, especially in the form of electronic health records (EHRs). These records constitute a precious resource that information systems researchers could utilize for various clinical applications, such as morbidity prediction and risk stratification. Recently, deep learning has demonstrated state-of-the-art empirical results in terms of predictive performance on EHRs. However, the blackbox nature of deep learning models prevents both clinicians and patients from trusting the models, especially with regard to life-critical decision making. To mitigate this, attention mechanisms are normally employed to improve the transparency of deep learning models. However, these mechanisms can only highlight important inputs without sufficient clarity on how they correlate with each other and still confuse end users. To address this drawback, we pioneer a novel model called Rational Multi-Layer Perceptrons (RMLP) that is constructed from weighted finite state automata. RMLP is able to provide better interpretability by coherently linking together relevant inputs at different timesteps into distinct sequences. RMLP can be shown to be a generalization of a multi-layer perceptron (that only works on static data) to sequential, dynamic data. With its theoretical roots in rational series, RMLP’s ability to process longitudinal time-series data and extract interpretable patterns sets it apart. Using real-world EHRs, we have substantiated the effectiveness of our RMLP model through empirical comparisons on six clinical tasks, all of which demonstrate its considerable efficacy.