Characterizing trending features in time-series prediction of clinical event onset

Abstract

The recent advances in electronic health records (EHR) systems have made it possible for AI algorithms to predict adverse clinical events using real-time patient health data. However, most existing algorithms apply a static prediction to a fixed time window of patient measurements, often failing to consider the dynamic shifts in the risk profile over time. In this research, we employ L1 trend filtering analysis to pinpoint piecewise linear trends in risk profiles and examine the predictive power of trend-based features in the time-series prediction of clinical events. Identifying these trends aids in recognizing the most predictive features and sets the stage for determining whether including these features can enhance prediction performance.