Scalable drift monitoring in medical imaging AI

Abstract

Abstract The integration of artificial intelligence (AI) into clinical medical imaging has advanced diagnostics but also poses challenges inmaintaining long-term reliability as input data change. To address this, we present MMC+, a modular framework for scalabledrift monitoring that builds on our CheXstray framework for real-time drift detection using multi-modal data concordance (MMC). MMC+ provides a flexible architecture that enables cost-effective continuous performance monitoring in real-world healthcare settings. The framework’s capabilities include aggregating information from multiple data streams from statistical tests ordistance metrics, leveraging foundation model embeddings for generalization across sites without requiring site-specific training, and quantifying uncertainty to better communicate drift significance. Validated with data from Massachusetts General Hospitalduring the COVID-19 pandemic, MMC+ effectively detects significant data shifts that correlate with performance changes and indicates when AI systems may deviate from acceptable performance bounds. The code for MMC+ is publicly available (Link: https://github.com/FJDorfner/MedImaging-ModelDriftMonitoring/tree/MMC_plus), promoting broader adoption of AI solutions in clinical settings. Our framework significantly advances drift detection while promoting transparency and collaboration in clinical AI research.

The authors' abstract, as published at the source. npj Digital Medicine, 2026 · DOI ↗