Human-guided Rule Learning for ICU Readmission Risk Analysis

Published: 29 Jun 2024, Last Modified: 04 Jul 2024KDD-AIDSH 2024 PosterEveryoneRevisionsBibTeXCC BY 4.0
Keywords: Interactive machine learning, Probabilistic rules, Human-in-the-loop, Machine learning for healthcare
Abstract: Interactive machine learning systems that can incorporate human feedback for automatic model updating have great potential use in critical areas such as health care, as they can combine the strength of data-driven modeling and prior knowledge from domain experts. Designing such a system is a challenging task because it must enable mutual understanding between humans and computers, relying on interpretable and comprehensible models. Specifically, we consider the problem of incorporating human feedback for model updating in rule set learning for the task of predicting readmission risks for ICU patients. Building upon the recently proposed Truly Unordered Rule Sets (TURS) model, we propose a certain format for feedback for rules, together with an automatic model updating scheme. We conduct a pilot study and demonstrate that the rules obtained by updating the TURS model learned from ICU patients' data can empirically incorporate human feedback without sacrificing predictive performance. Notably, the updated model can exclude conditions of rules that ICU physicians consider clinically irrelevant, and thus enhance the trust of physicians.
Submission Number: 13
Loading