Beyond the benchmark dataset: Real-world generalizability and regulatory challenges in medical AI

Описание: Presented by the MICCAI SIG for Challenges
October 10, 2025
Speakers:
Ghada Zamzmi, Regulatory Scientist and AI Researcher, Heartflow
Jean Feng, Associate Professor, University of California, San Francisco and the UCSF-UC Berkeley Joint Program in Computational Precision Health
Abstract:
Innovation is essential to advancing AI in healthcare, but when innovation is pursued without consideration for deployment and regulatory challenges, it often leads to solutions that consume significant resources yet fail to reach clinical use.In the first portion of this webinar, Ghada Zamzmi will introduce a regulatory-driven approach to AI design and development—one that integrates safety, effectiveness, regulatory science throughout the entire AI lifecycle from pre-market data collection, model development, evaluation, clinical validation, and post-market monitoring. Ghada will highlight regulatory science–informed research practices that can accelerate the safe and successful deployment of AI technologies. The second portion of the webinar, led by Jean Feng, will focus on the post-market phase, where AI systems encounter real-world variability. Even models that perform well during development may degrade after deployment, often affecting some subgroups more than others. Jean Feng will present a systematic diagnostic framework, SHIFT, designed to detect and address performance drift, drawing on case studies such as an acute care needs prediction model and an intraoperative acute kidney injury risk model.Together, these talks will demonstrate how bridging the gap between innovation, deployment challenges, and regulatory science enables the development of medical AI products that are not only cutting-edge, but also safe, effective, and deliver lasting value to patients and healthcare systems.