Research program
From Simulation to Vendor and Hospital Governance
Each study answers the question the last one raised: detect drift without labels, correct it at the site, then govern it inside a hospital.
-
Study 1 · Detect
Label-free drift detection in lung-nodule AI
154
extra radiologist reads per 1,000 cases, while AUROC stayed flat
In a simulation of 23,851 NLST participants, scanner and protocol shifts eroded the model's rule-out specificity without moving AUROC. Label-free monitoring caught simulated scanner-vendor drift before outcome labels arrived.
Manuscript under review
-
Study 2 · Correct
Not every alarm is harm, not every harm alarms
Undetected Harm Attenuation
A follow-up NLST drift study found that scanner shift can trigger alarms without hurting performance, while some harmful shifts stay silent on output-only monitoring.
Analysis complete
-
Study 3 · Govern
Drift governance in a hospital screening program
1,500+
low-dose CT screening subjects across 3 annual rounds
A prospective study at Chang Gung Memorial Hospital testing a stable-scanner arm and a new-scanner arm with a locked lung-nodule AI. Each alert runs through a three-layer evidence ladder and a CAPA workflow.
Planned