Truecourse Health checks the real-world performance of your deployed imaging AI against the bounds in your authorized Predetermined Change Control Plan, verifies each planned model update against your Modification Protocol before rollout, and drafts the record your QA team signs.
Built for RA/QA and ML leads at radiology, ophthalmology and pathology AI companies with one to five FDA clearances.
FDA is pulling postmarket performance monitoring into the premarket conversation, and asking industry how real-world AI performance should be measured at all. Most manufacturers are answering with spreadsheets.
AI-enabled devices authorized by FDA as of early 2026, roughly double the 2022 count.
FDA AI-enabled device listof cleared AI device summaries noted any post-clearance adverse events. Only 28.2% mentioned a safety assessment.
Audit of cleared device summariesFDA opened a public request for comment on how to measure and evaluate the real-world performance of AI-enabled devices.
Crowell & Moring summaryTruecourse Health sits downstream of the tools you already use to write the plan and run your quality system. It does not replace either.
| Layer | What it covers | What it leaves open |
|---|---|---|
| PCCP authoring tools | Templates and content for writing a compliant change control plan. | Document generation. No monitoring of the deployed model. |
| AI observability platforms | Production model monitoring for MLOps teams and health systems. | Not tied to your authorized thresholds or to submission-ready output. |
| Academic drift research | Validated distribution-shift techniques for imaging models. | Published, not productized. |
| Truecourse Health | Real-world device data checked against your own PCCP bounds, planned updates verified against your Modification Protocol, and an audit-ready record handed to your eQMS. | |
FDA's final guidance on PCCPs for AI-enabled device software functions splits every plan into three parts. Each one creates work once the device ships, and between updates the cleared model still has to stay inside its bounds.
We start from what FDA already authorized, not from generic anomaly detection.
Metrics tracked, the modification protocol, and subgroup bounds go in as structured configuration.
Start narrow with DICOM metadata and FHIR observation feeds from deployment sites.
Data and performance drift, using ground-truth-light methods from the published literature, stratified by site, scanner, device version and demographics. Undersized subgroups are flagged too.
An out-of-bounds signal routes to the action trigger your own change protocol defines. A human decides; the system records it.
When you retrain or extend the model under your PCCP, the candidate runs in the background on real-world data and is checked against your Modification Protocol's acceptance criteria, overall and by subgroup.
A postmarket surveillance report and, for each update, the Impact Assessment evidence, with e-signature and traceability, ready to file into the eQMS you already run.
# device: chest X-ray triage, K-number on file metric: sensitivity lower_bound: 0.87 # from authorized PCCP window: 30d rolling subgroups: - site - scanner_vendor - age_band - sex min_subgroup_n: 200 on_breach: route_to: RA/QA lead action: evaluate_under_ACP §4.2 update_verification: mode: shadow # per Modification Protocol accept_if: sensitivity: ">= 0.89" max_subgroup_drop: 0.02 report: cadence: monthly signoff: required
Class II AI/ML software as a medical device, roughly 20 to 300 people, with one to five devices cleared or close to clearance.
FDA's final Computer Software Assurance guidance lets manufacturers lean on vendor evidence. We plan to hand you that evidence instead of making you build it.
Quality-system software, not a device.
Truecourse Health informs your decisions about your device. It never makes a clinical decision about a patient, and AI-drafted report text always goes through a human sign-off.
Truecourse Health is early. Before writing production code we are talking with RA/QA and ML leads about how they monitor deployed models today and what it costs them.
Leave your name and email and the founder will reply personally to set up a 30-minute conversation. Useful even if you never become a customer: you leave with a view of how peers are approaching PCCP monitoring.