LAUSR

Acute heart failure (AHF) is a complex disease with heterogeneous manifestations and adverse outcomes. The interpretation of machine-learning risk scores is vital to support clinical decisions. Individualized Feature Importance (IFI) was designed to attribute changes in risk scores to clinical features and help contrast decision trajectory for a patient against those of patient subgroups that received distinct clinical decisions. Score Confidence Interval (SCI) was developed to quantify certainty in the prediction, which further encourages clinicians’ interpretation. Study was based on retrospective data from 25 hospitals in the US of 20,640 adult patients, with 87% discharged home (Class 0) and 13% transferred to the ICU or died in hospital (Class 1). IFI is based on Shapley Value, based on which SCI was designed to capture the variation in score if input features are missing. These methods were applied to previously developed risk score for AHF patients in the wards; however, they can be applied to any risk score. The SCI is wide at the beginning of the stay and narrows down towards the end as more clinical measurements become available, indicating the risk score is relatively certain at the end (Fig. 1a). IFI values (Fig. 1b) show how selected features drive changes in the risk score. To aid decision-making at the latest time, top missing features are prompted (Fig. 1c). Decision trajectories show the way top features drive the risk score (Fig. 1d), that this patient is at higher risk to discharge (Fig. 1e) and is more similar to ICU-transfers (Fig. 1f). Fig. 1g shows SCI improves risk score performance by abstaining uncertain cases from decision-making. IFI apportions risk score to clinical measurements. SCI reduces false alarm rates. By providing clinical context, they have the potential to enhance incorporation of risk scores in the clinical workflow to aid medical decisions by identifying patients at risk for deterioration and determining appropriate levels of care.