Radiology Reimagined: Advancing Clinical Practice Through AI Innovation (2025)-"AI Continuous Monitoring - What Is It and How Can It Be Done?"

"AI Continuous Monitoring - What Is It and How Can It Be Done?" – Dr. Amine Korchi

Pdf Summary

This document by Amine Korchi, M.D., discusses the critical importance and practical approaches to continuous monitoring of artificial intelligence (AI) systems in radiology post-deployment. AI performance can degrade over time due to factors such as changes in input data, equipment updates, patient demographic shifts, and model drift, posing patient safety risks and regulatory compliance challenges (e.g., FDA, EU AI Act).<br /><br />Three main AI monitoring types are described: <br />1. Direct performance monitoring compares AI outputs to ground truth (expert labels) but requires labeled data and is often delayed.<br />2. Indirect statistical monitoring assesses changes in input data distributions, feature importance, and AI output patterns without requiring ground truth, serving as early warning signals.<br />3. Operational monitoring tracks system reliability metrics like turnaround time, uptime, and failure rates, focusing on workflow integration rather than diagnostic accuracy.<br /><br />Currently, radiology AI monitoring is mostly absent or limited to manual, reactive, and non-continuous methods, such as retrospective reviews or user case flagging. Continuous monitoring—real-time, ongoing surveillance—is needed to detect performance drops early, enabling timely corrective actions, especially for adaptive or autonomous AI. However, continuous monitoring faces challenges including lack of ground truth, resource demands, integration complexity, governance, alarm fatigue, and equity concerns.<br /><br />The American College of Radiology (ACR) recently launched Assess-AI, a voluntary national quality registry enabling continuous AI performance monitoring by aggregating image metadata, patient demographics, AI results, and radiology report text from multiple sites. This provides benchmarking and case-level deep-dive capabilities to enhance quality assurance.<br /><br />Furthermore, developer-built systems employing statistical drift detection (e.g., via high-dimensional embeddings and AI output monitoring) show promise for scalable continuous surveillance. Effective monitoring requires clear thresholds, roles, escalation pathways, regular review, and clinician engagement to ensure safety and sustainability.<br /><br />In summary, continuous AI monitoring is essential to maintain safety, effectiveness, and trust in radiology AI tools. While direct ground-truth assessment is ideal but often infeasible in practice, indirect methods offer practical early warning signals. Emerging registry programs and vendor solutions are advancing continuous monitoring from concept to reality, but successful implementation demands multi-stakeholder collaboration, infrastructure, and governance. Radiologists play a key role in feedback and governance to achieve safe, reliable AI integration in clinical practice.

Keywords

Artificial Intelligence in Radiology

Continuous AI Monitoring

AI Performance Degradation

Direct Performance Monitoring

Indirect Statistical Monitoring

Operational Monitoring

AI Regulatory Compliance

Assess-AI Registry

AI Model Drift Detection

Radiologist Role in AI Governance