Towards Trustful Machine Learning for Antimicrobial Therapy Using an Explainable Artificial Intelligence Dashboard

Abstract

The application of machine learning (ML) in healthcare has surged, yet its adoption in high-stakes clinical domains, like the Intensive Care Unit (ICU), remains low. This gap is largely driven by a lack of clinician trust in AI decision support. Explainable AI (XAI) techniques aim to address this by explaining how an AI reaches its decisions, thereby improving transparency. However, rigorous evaluation of XAI methods in clinical settings is lacking. Therefore, we evaluated the perceived explainability of a dashboard incorporating three XAI methods for an ML model that predicts piperacillin plasma concentrations. The dashboard was evaluated by seven ICU clinicians using five distinct patient cases. We assessed the interpretation and perceived explainability of each XAI component through a targeted survey. The overall dashboard received a median score of seven out of ten for completeness of explainability, with Ceteris Paribus profiles identified as the most preferred XAI method. Our findings provide a practical framework for evaluating XAI in critical care, offering crucial insights into clinician preferences that can guide the future development and implementation of trustworthy AI in the ICU.