Using a Large Language Model for Postdeployment Monitoring of FDA-Approved Artificial Intelligence: Pulmonary...

Artificial intelligence (AI) is increasingly integrated into clinical workflows. The performance of AI in production can diverge from initial evaluations. Postdeployment monitoring (PDM) remains a challenging ingredient of ongoing quality assurance once AI is deployed in clinical production.

To develop and evaluate a PDM framework that uses large language models (LLMs) for free-text classification of radiology reports, and human oversight. We demonstrate its application to monitor a commercially vended pulmonary embolism (PE) detection AI (CVPED).

We retrospectively analyzed 11,999 CT pulmonary angiography studies performed between April 30, 2023, and June 17, 2024. Ground truth was determined by combining LLM-based radiology report classification and the CVPED outputs, with human review of discrepancies. We simulated a daily monitoring framework to track discrepancies between CVPED and the LLM. Drift was defined when discrepancy rate exceeded a fixed 95% confidence interval for 7 consecutive days. The confidence interval and the optimal retrospective assessment period were determined from a stable dataset with consistent performance. We simulated drift by systematically altering CVPED or LLM sensitivity and specificity, and we modeled an approach to detect data shifts. We incorporated a human-in-the-loop selective alerting framework for continuous prospective evaluation and to investigate potential for incremental detection.

Of 11,999 CT pulmonary angiography studies, 1,285 (10.7%) had PE. Overall, 373 (3.1%) had discrepant classifications between CVPED and LLM. Among 111 CVPED-positive and LLM-negative cases, 29 would have triggered an alert due to the radiologist not interacting with CVPED. Of those, 24 were CVPED false-positives, 1 was an LLM false-negative, and the framework ultimately identified 4 true-alerts for incremental PE cases. The optimal retrospective assessment period for drift detection was determined to be 2 months. A 2% to 3% decline in model specificity caused a 2- to 3-fold increase in discrepancies, and a 10% drop in sensitivity was required to produce a similar effect. For example, a 2.5% drop in LLM specificity led to a 1.7-fold increase in CVPED-negative LLM-positive discrepancies, which would have taken 22 days to detect using the proposed framework.

A PDM framework combining LLM-based free-text classification with a human-in-the-loop alerting system can continuously track an image-based AI's performance, alert for performance drift, and provide incremental clinical value.