Interpretable machine learning prediction of all-cause mortality

Prior studies on all-cause mortality traditionally use linear models; however, growing field of explainable artificial intelligence (XAI) can improve prediction accuracy over traditional linear models using complex machine learning (ML) models while still revealing novel insights. We propose the IMPACT (Interpretable Machine learning Prediction of All-Cause morTality) framework that implements and explains complex, non-linear ML models by combining a tree ensemble mortality prediction model and a principled XAI technique. We apply IMPACT to the NHANES (1999-2014) dataset, which enables us to understand different subpopulations according to shorter or longer term mortality and younger and older individuals. Our IMPACT models have higher predictive accuracy than popular pre-existing mortality risk scores and biological ages. Using individualized feature importance scores, we discover novel risk predictors (e.g., arm circumference) and interactions between risk predictors (e.g., serum chloride with age and/or gender). Furthermore, IMPACT provides a novel perspective of reference intervals and may suggest that the widely accepted reference intervals for serum albumin, mean cell volume and platelet count may in fact be sub-optimal for health. Finally, in order to ensure that our models are useful to as broad of a community as possible, we develop and publish a variety of explainable risk scores usable by individuals with and without medical expertise. The predictive accuracy of IMPACT combined with the capability of discovering mortality risk predictors and complex relationships demonstrates the value and utility of XAI in epidemiologic study design.