Cleared at 93%. Deployed at 64%. Clinical AI drifts

The gap

An FDA-cleared AI triage tool for pulmonary embolism looked strong in its clearance paperwork and much weaker in the ward. In a preprint posted on medRxiv, researchers report that the deployed sensitivity of the Aidoc BriefCase PE triage tool was 64.1%, against 93% sensitivity reported in the FDA 510(k) summary. The deployed figure comes with a 95% confidence interval of 61.5-66.6% — well below the cleared benchmark, not a borderline miss. m:…2203

The gap is the story. Clearance measures a model under controlled conditions; deployment measures it against real scans, real sites and real drift. The study is a post-market check, and it finds the tool's real-world performance significantly lower than the number on its clearance. m:…2203

Where it misses

Sensitivity varies by where the clot is. The tool caught 88.6% of central pulmonary embolisms, 63.1% of lobar or segmental cases, and 33.3% of subsegmental cases. The tool's sensitivity falls as the embolus gets more peripheral — and subsegmental cases are where it is weakest. m:…2203

Performance also varies by site. Sensitivity differed by 13 percentage points between the two highest-volume sites — 68.7% against 55.7%. The same tool, the same task, and a 13-point gap depending on where it runs. A gap that size between sites running the same tool suggests the difference is not the model alone but the scans, the workflow, or the patients each site sees. m:…2203

The drift

Performance also moves over time. At one site, the false-positive rate fell 40% per year while sensitivity fell 15% per year. The tool was getting quieter and less sensitive at the same time — a pattern the study describes as temporal drift, and one that a one-off clearance check cannot catch. m:…2203

That is the core problem with cleared-but-unmonitored AI: nobody re-runs the clearance test every month, so a slow decline can run for a year before anyone notices. The pattern is only visible because someone kept measuring after deployment. A one-off clearance check cannot catch a decline that unfolds over a year. A tool that is cleared once and never re-checked can drift while everyone assumes the clearance number still holds. m:…2203

The watchers

Other preprints this week build the kind of monitoring the triage study implies is missing. AToP, an explainability framework for Transformer-based predictions from electronic health records, converts per-token Integrated Gradients into structured temporal pattern narratives — visit-anchored stories about why a model made a call, rather than a ranked list of tokens. In its test, the underlying Transformer model trained on MIMIC-IV v3.1 achieved an AUROC of 0.77 and PR-AUC of 0.34 for 30-day unplanned readmission prediction. m:…7967

AToP also surfaces rare, high-attribution markers that standard rankings miss — including psychosis, self-harm overdose, suicidal ideation, and severe MDD — and identifies risk-increasing patterns such as psychiatric sequences and oncologic progressions. MedCheckpoint takes a different angle: it automates postmarketing pharmacovigilance with LLM agents that enforce study-design locks and source-traceable grounding, outperforming baseline coding agents in report quality across five dimensions, with higher run-to-run consistency — and it prevents misleading failures on infeasible comparisons or unmeasurable outcomes. m:…7967 m:…3779

Neither system fixes a drifting triage tool on its own. But together they point at the missing layer in clinical AI: continuous, explainable monitoring after deployment, so that a 64.1% does not pass for a 93% unnoticed. m:…2203