Second-Look AI at Scale: Safer, Smarter Pharmacy Workflows

Predictive Real-World DeploymentPredictive Clinical Outcomes

Authors:Diana S.M. Buist; Annie Y. Ng; Bryan Haslam, et al.

Journal:Journal of the American College of Radiology

DOI:10.1016/j.jacr.2025.12.015

Scaling human-in-the-loop AI across ~580k exams increased cancer detection (0.99/1,000) and precision (0.55/100) in real-world use. This offers a blueprint for pharmacy: using "second-look" queues to catch errors without adding noise—provided you have the oversight to fix the AI when it inevitably drifts.

December 4, 2025

Quick Take

A vertically integrated Learning Accelerator Framework scaled a multi-stage AI breast screening workflow across 579,583 exams and was associated with an absolute increase in cancer detection rate (CDR) of 0.99 cancers per 1,000 exams (95% CI 0.59–1.42) and an absolute increase in positive predictive value (PPV) of 0.55 cancers per 100 recalls (95% CI 0.30–1.03).

Results from the large observational pre–post evaluation suggest that comparable benefits were achieved across prespecified breast-density and race/ethnicity subgroups, addressing common concerns regarding algorithmic bias.

For pharmacy leadership, this study validates the "Safeguard" model: rather than replacing clinicians, AI routes discordant "normal" checks to a second human reviewer, a workflow that can be adapted for high-risk medication verification audits.

Why it Matters

Healthcare systems currently face a "translation gap" where AI models perform well in research but degrade in clinical reality due to model drift, unrepresentative training data, and rigid workflows.

In the high-volume environment of inpatient pharmacy, static rule-based alerts cause fatigue (90-95% override rates), while "black box" AI solutions risk becoming obsolete the moment a formulary change or drug shortage alters the data landscape.

This research demonstrates that sustainable AI requires an organizational architecture—vertical integration between developer and clinician—that treats model maintenance, drift detection, and retraining as continuous operational necessities rather than one-time software purchases.

What They Did

The study utilized a vertically integrated network (RadNet provider + DeepHealth developer) to implement a four-component framework: an Integrated Data Registry, Continuous Technology Stack, Adaptive Clinical Services, and an Iterative Learning Loop.

Researchers deployed a multi-stage workflow where an AI "Safeguard" reviewed mammograms initially deemed normal by a radiologist; if the AI flagged the exam as suspicious, it was routed to a "signed pending release" queue for a second expert human review.

The framework underwent staged validation, moving from retrospective testing (>25,000 exams) to a single-site pilot (41,000 exams) and finally a large-scale pre-post evaluation (579,583 exams) across five practices and 96 radiologists.

Real-time feedback loops were tested during a hardware upgrade event, where changes in scanner resolution caused model drift (bounding boxes grew larger), triggering an immediate retraining and redeployment cycle within three months.

What They Found

In the large-scale implementation (N=579,583), the AI-driven workflow achieved a 21.6% relative increase in cancer detection rate compared to the standard of care, finding approximately one additional cancer for every 1,000 women screened.

The "Safeguard" pilot phase (N=41,000) identified 41 additional cancers that had been missed by the initial human reader, establishing a "hit rate" of roughly 1 in 1,000 reviews for the second-look workflow.

Precision improved alongside sensitivity; the Positive Predictive Value (PPV) increased by 10.55 cancers per 100 recalls in the large-scale analysis, indicating the system reduced false-positive noise relative to the increased detection rate.

The framework successfully managed model drift; when new imaging hardware caused a 30.8% increase in bounding box size (false positive pixel area), the integrated feedback loop allowed developers to retrain and validate a fix that reduced the variance to -7.3% within a single quarter.

Takeaways

Adopt the "Safeguard" workflow for medication safety: Instead of using AI to interrupt front-line verification, route AI-flagged "verified" orders to a background queue for a second pharmacist review to catch errors without slowing primary distribution.

Demand "Continuous Learning" clauses in vendor contracts: The study proves that environmental changes (like hardware upgrades or formulary switches) break models; vendors must demonstrate a rapid, data-driven process for retraining on local data.

Build an Integrated Data Registry that captures the "Why": To retrain models effectively, pharmacy informatics teams must capture not just the override, but structured data on downstream outcomes (e.g., rescue agent use, lab values) to establish ground truth.

Monitor for drift weekly, not annually: Establish governance committees that review AI performance metrics (alert volume, acceptance rate) continuously to detect when models lose alignment with clinical reality.

Strengths and Limitations

Strengths:

The study leverages a massive dataset (>500,000 exams) and a "hard" endpoint (biopsy-proven malignancy), providing a higher level of evidence than typical AI studies relying on small, retrospective cohorts.

The vertical integration model demonstrates a closed-loop system where clinical feedback directly triggers technical retraining, offering a practical blueprint for the "Learning Health System."

Limitations:

The observational pre-post design cannot rule out temporal confounders, such as improvements in imaging technology independent of the AI intervention.

The specific success metrics (biopsy results) are cleaner than pharmacy outcomes (adverse drug events), making direct transferability of the "ground truth" feedback loop more complex for medication management.

There is a commercial conflict of interest, as the study authors are employed by or affiliated with the organizations (RadNet/DeepHealth) marketing the solution.

Bottom Line

This study moves clinical AI from "toy datasets" to operational reality, proving that a human-in-the-loop "second look" workflow combined with rapid retraining infrastructure is the viable path for high-reliability error detection in complex healthcare environments.