FDA PCCP AI Device Change Control Testing Guide

Statutory requirements

1. 01

   Description of modifications

   PCCP guidance, section on Description of Modifications

   The PCCP lists the specific, planned modifications, such as retraining on new data to improve performance, adapting to a new input device, or extending to a defined new subpopulation, with enough detail that FDA can evaluate them. Modifications must stay within the device's intended use and generally its indications; open-ended statements like improving the model are not reviewable.

2. 02

   Modification protocol

   PCCP guidance, section on Modification Protocol

   For each modification, the protocol describes data management practices (sources, reference standard, sequestration of test data), retraining practices, performance evaluation (metrics, statistical methods, acceptance criteria, test data), and update procedures, including how users are informed and how cybersecurity and real-world monitoring are handled. Verification and validation must be done to the same standard as the original submission.

3. 03

   Impact assessment

   PCCP guidance, section on Impact Assessment

   The manufacturer assesses the benefits and risks of each planned modification, how the protocol's controls mitigate those risks, and how modifications interact with each other. Cumulative effects across several authorized changes need explicit analysis.

4. 04

   Implementation within the quality system

   FD&C Act section 515C; 21 CFR Part 820

   Changes made under a PCCP are still design changes under the QMSR and must be documented. If a modification fails its acceptance criteria, or a change falls outside the authorized PCCP, the manufacturer cannot implement it under the PCCP and must evaluate whether a new submission is needed.

5. 05

   Guiding principles

   PCCP Guiding Principles, October 2023

   FDA, Health Canada and the MHRA describe PCCPs as focused and bounded, risk-based, evidence-based, transparent, and grounded in a total product lifecycle perspective. Each principle translates to a test practice: bounded acceptance criteria, risk-weighted test coverage, pre-specified statistical methods, user-facing change notes and continuous monitoring.

Evaluation methodologies

Testing artifacts to retain

The evidence folder an examiner, notified body or internal auditor will ask for. GONATURELLE compliance reports export most of it from signed runs.

GONATURELLE assertion rules

Each assertion maps to a requirement above. Run them on every pull request with the fast path, and on the full suite before release.

from gonaturelle import EvalSuite, assertions as A

AUTHORIZED = 'triage-model@1.0.0'   # version cleared with the PCCP
PREVIOUS = 'triage-model@1.2.0'

suite = EvalSuite.load('healthcare/triage-escalation', partition='sealed')
candidate = suite.evaluate(agent=triage_agent_v130, max_concurrency=10)

# Modification protocol, section 4.2: non-inferiority with absolute floor
for ref in [PREVIOUS, AUTHORIZED]:
    candidate.compare(baseline=ref).assert_non_inferior('sensitivity', margin=0.02, ci=0.95)
candidate.assert_metric('sensitivity', min=0.95, ci=0.95, label='emergent')

# Subgroup non-regression against each subgroup's own baseline
for group in ['age_band', 'sex', 'site']:
    candidate.slice_by(group).compare(baseline=AUTHORIZED).assert_non_inferior('sensitivity', margin=0.03)

candidate.assert_zero_contraindications()
candidate.assert_all([A.never_calls('schedule_routine_visit', when="label == 'emergent'")])

suite.record_test_set_use(reason='PCCP modification M-2: retraining on 2026 data')
candidate.export_manifest('dhf/pccp/M-2_v130.json', sign=True, protocol='PCCP-MP-02')

Audit checklist