Visualizing Rule-based Classifiers for Clinical Risk Prognosis

Abstract

Deteriorating conditions in hospital patients are a major factor in clinical patient mortality. Currently, timely detection is based on clinical experience, expertise, and attention. However, healthcare trends towards larger patient cohorts, more data, and the desire for better and more personalized care are pushing the existing, simple scoring systems to their limits. Data-driven approaches can extract decision rules from available medical coding data, which offer good interpretability and thus are key for successful adoption in practice. Before deployment, models need to be scrutinized by domain experts to identify errors and check them against existing medical knowledge. We propose a visual analytics system to support health-care professionals in inspecting and enhancing rule-based classifier through identification of similarities and contradictions, as well as modification of rules. This work was developed iteratively in close collaboration with medical professionals. We discuss how our tool supports the inspection and assessment of rule-based classifiers in the clinical coding domain and propose possible extensions.