Sustainable Algorithms for Generalizable and Effective (SAGE) Learning Prediction Systems That Deliver on the Promise of AI for All Patients and Communities

About This Grant

PROJECT SUMMARY / ABSTRACT Temporal changes in clinical practice, patient populations, and information systems degrade performance of artificial intelligence (AI) and machine learning (ML) models. Lack of model generalizability across patient contexts—be they clinical, geographic, or sociodemographic—creates performance gaps that can lead to differences in access to care and unevenly distribute algorithmic benefits. Failing to proactively address performance drift and differences in performance across patient subgroups risks patient safety, undermines user trust, and fails to deliver on the promise of predictive analytics based either traditional ML or novel large language models (LLMs) to improve patient and population health. Learning prediction systems (LPS), an extension of the learning health system paradigm of data-driven continuous improvement, would conduct post- deployment surveillance to collect evidence of model success or deterioration and recommend changes that sustain prospective model performance, both overall and within patient subgroups. Existing model maintenance methods focus on population-level performance and have yet to explore how model updates may generalize across variable patient contexts and impact subgroup performance. Our central objective is to design LPS that promote sustainable AI/ML decision support tools (AI-DST) to ensure all patients benefit from the AI-enabled transformation of healthcare. Using data from the Department of Veterans Affairs (VA) and Vanderbilt University Medical Center (VUMC), we will apply complex simulation studies and real-world evaluations across 3 clinical domains to advance novel LPS metrics and methods that sustain generalizable and effective delivery of clinical AI/ML. In Aim 1, we characterize the complex spectrum of temporal changes in performance gaps between patient contexts and whether/how model updating practices and learning algorithms (ML and LLMs) impact performance gap drift. In Aim 2, we develop and benchmark novel methods to characterize and detect exacerbation of performance gaps between patient subgroups, as well as updating strategies that foster restoration of overall and subgroup performance. In Aim 3, we extend LPS methods in support of small patient subgroups, such as those served by rural healthcare centers, where limited sample sizes present unique challenges to effectively monitoring performance, detecting deterioration, and training updates. In Aim 4, we disentangle changes in outcomes associated with effective AI-DST from dataset shift, establishing LPS methods to monitor and update models while accounting for feedback interference, including monitoring of decision consistency across patient contexts. We will evaluate these methods prospectively in a deployed AI-DST preventing postpartum hemorrhage at VUMC. With expertise in informatics, AI, data science, ethics, and clinical care, our team is well-positioned to have a major impact on the development and implementation of LPS that sustain AI-DST for all patients, propelling the adoption of responsible AI/ML.