Real-time Monitoring and Correction of Clinical Decision Support Systems using Artificial Intelligence

This project develops artificial intelligence (AI)/machine learning (ML) methods for real time monitoring and updating of clinical decision support (CDS) systems to improve model reliability across patient populations. As AI/ML systems mature and their use in CDS expands, their potential to influence health outcomes grows. A key concern is the reliance on retrospective patient data, which often under-characterizes certain diseases and contains unwarranted differences in treatment and outcomes that naïve models may reproduce. This has motivated methods that embed performance constraints into models to mitigate undesirable behavior. Current approaches face persistent challenges including performance degradation with changes in clinical practice, maintaining consistent performance across patient populations without reducing accuracy, and limited ability to adapt to existing workflows. This project develops AI-based postprocessing methods to address these gaps, focusing on resource- and time-constrained settings where AI/ML tools help prioritize patient care. Our primary application is an ML system that predicts inpatient admissions in the emergency department (ED) to improve patient flow. We also evaluate generalization across endpoints, sites, and time. The central hypothesis is that continuous monitoring and updating of model performance can ensure reliable CDS operation. In the first two aims, we train models on retrospective data and assess their impact on ED attending decisions using a silent prospective study. In the third aim, we extend postprocessing methods for multicalibration to real-time operation, enabling adaptation to health data streams. The fourth aim implements a real-time AI auditing and calibration system for predicting patient disposition in the emergency department at Boston Children’s Hospital. This work supports the National Library of Medicine’s vision of sustainable computational infrastructure and seeks to reduce systematic performance degradation in CDS systems that may influence care decisions.