Skip to content
GONATURELLE
United States (federal)

FDA SaMD Clinical AI Testing & Validation Guide

When clinical AI software meets the device definition, FDA expects validation evidence that a test set independent of training represents the intended patient population, that performance holds across subgroups, and that the human-AI team performs as intended. The evidence lives inside design controls, so every evaluation run needs version traceability.

Jump to assertionsAudit checklist
Reviewed by Nandini Joshi, Head of Regulatory Science · updated 1 Nov 2026
FDA SaMD
Instrument
FDA Software as a Medical Device (SaMD) requirements for AI-enabled device software functions
Primary citation
21 CFR Part 820 (QMSR); GMLP Guiding Principles (2021); FDA draft guidance on AI-enabled device software functions (January 2025)
Authority
U.S. Food and Drug Administration, Center for Devices and Radiological Health (CDRH)
Status
As of November 2026: the GMLP guiding principles (October 2021) and the final PCCP guidance (December 2024) stand; the January 2025 lifecycle guidance for AI-enabled device software functions was issued in draft, and teams should check FDA's guidance database for a final version. The QMSR, aligning Part 820 with ISO 13485:2016, took effect 2 February 2026. FDA revised its Clinical Decision Support Software guidance on 6 January 2026.
HealthcareLife Sciences

Statutory requirements

  1. 01

    Decide first whether the function is a device

    FD&C Act sections 201(h) and 520(o)(1)(E)

    Section 201(h) of the FD&C Act defines a device, and section 520(o) excludes certain software, including some clinical decision support that lets a health care professional independently review the basis for a recommendation. FDA's January 2026 CDS guidance revised how it reads those criteria. The classification answer drives everything else: a non-device documentation assistant and a device-level triage tool carry very different evidence burdens.

  2. 02

    Design controls and validation under the QMSR

    21 CFR Part 820; ISO 13485:2016 clause 7.3

    The Quality Management System Regulation incorporates ISO 13485:2016 by reference. Design and development verification and validation must show that the device meets user needs and intended uses under defined operating conditions, and software validation must be documented. For AI, the validation protocol, test data, acceptance criteria and results are design history records, not notebook outputs.

  3. 03

    Good Machine Learning Practice

    GMLP Guiding Principles, October 2021

    FDA, Health Canada and the MHRA published 10 guiding principles in October 2021. Several are directly about testing: clinical study participants and data sets are representative of the intended patient population (principle 3), training data sets are independent of test sets (principle 4), testing demonstrates device performance during clinically relevant conditions (principle 8), and deployed models are monitored for performance with retraining risks managed (principle 10). Principle 7 puts the focus on the performance of the human-AI team.

  4. 04

    Marketing submission content for AI-enabled functions

    FDA draft guidance, AI-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations (January 2025)

    The January 2025 draft lifecycle guidance recommends that submissions describe the model, data management (collection, cleaning, annotation, sequestration of test data), performance validation with subgroup analyses, bias analysis, and a performance monitoring plan. It also emphasises transparency to users, including information about limitations and the data used to validate the device.

  5. 05

    Change control after clearance

    FD&C Act section 515C; FDA final PCCP guidance (December 2024)

    Modifications to a cleared AI function may need a new submission unless covered by an authorized Predetermined Change Control Plan. A PCCP specifies the planned modifications, the modification protocol (including verification and validation methods and acceptance criteria) and an impact assessment. Without one, every model update reopens the question of whether a new 510(k) is needed.

Evaluation methodologies

Sequestered, independent test sets

Hold out test data that was never used for training, prompt tuning or few-shot examples, ideally from different sites or time periods. Record access controls on the test set; if engineers can iterate against it, it stops being independent. GONATURELLE packs keep a sealed partition that only scores, never returns case text to the developer loop.

Subgroup performance with confidence intervals

Report sensitivity, specificity, PPV and NPV (or task-appropriate metrics) by age, sex, race and ethnicity where available, acquisition site and clinically relevant comorbidities. Small subgroups need intervals, not point estimates, so reviewers can see where evidence is thin.

Reader and human-AI team studies

When the clinician remains in the loop, measure the team: clinician performance with and without the device, time to decision, and override behaviour. For generative clinical documentation, pair this with expert rubric scoring of omissions and fabricated findings.

Clinically relevant stress conditions

Test the conditions principle 8 targets: incomplete records, conflicting medication lists, atypical presentations, pediatric dosing and abbreviations. These are the edge-traps where a final-answer score can stay flat while the trajectory goes wrong.

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.

evidence/fda-samd-clinical/8 items
  • 01Device determination memo covering section 520(o) and the January 2026 CDS guidance criteria
  • 02Software validation protocol with acceptance criteria approved before execution
  • 03Test data management record: sources, sequestration, annotation protocol, reviewer credentials and agreement
  • 04Performance report with subgroup analyses and confidence intervals
  • 05Bias analysis and documented limitations for labeling and user-facing transparency
  • 06Traceability matrix linking requirements, hazards, test cases and results
  • 07Post-market performance monitoring plan and drift thresholds
  • 08Signed run manifests that tie each result to model, prompt and dataset versions

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.

fda_samd_validation.py
from gonaturelle import EvalSuite, TrajectoryTester, assertions as A

# GMLP principle 4: sealed partition, never used during prompt tuning
suite = EvalSuite.load('healthcare/triage-escalation', partition='sealed')
results = suite.evaluate(agent=triage_agent, max_concurrency=10)

# Acceptance criteria from the approved validation protocol
results.assert_metric('sensitivity', min=0.95, ci=0.95, label='emergent')
results.assert_metric('specificity', min=0.80, ci=0.95, label='emergent')
results.assert_zero_contraindications()

# GMLP principle 3: subgroup performance with intervals
for group in ['age_band', 'sex', 'site']:
    results.slice_by(group).assert_metric('sensitivity', min=0.92, ci=0.95, min_support=40)

# Principle 8: clinically relevant conditions in the trajectory
TrajectoryTester(results).assert_order('fetch_medication_list', 'recommend_disposition')
results.assert_all([
    A.never_calls('schedule_routine_visit', when="label == 'emergent'"),
    A.judge(rubric='triage-omission-v2', threshold=0.90),
])

results.export_manifest('dhf/validation_run.json', sign=True, protocol='VAL-PRT-014')

Audit checklist

FDA SaMD readiness0/10

FDA SaMD testing questions

Is an ambient clinical documentation agent a medical device?

It depends on intended use and function. Software that only transcribes and summarizes for clinician review is often outside the device definition, while software that analyzes data to drive a diagnosis or treatment recommendation can fall inside it. Get a documented determination from regulatory counsel; this guide is not legal or medical advice.

What changed with the January 2026 CDS guidance?

FDA revised its interpretation of the non-device CDS criteria, including how it treats time-critical decision support and risk scores, and at the same time withdrew its adoption of the IMDRF SaMD clinical evaluation guidance. Commentators note it did not expressly change FDA's approach to AI, so re-run your device determination rather than assume an outcome.

Can GONATURELLE results go into a 510(k) or De Novo submission?

The validation summaries are formatted to fit design history files and submission sections, but GONATURELLE does not provide FDA clearance and does not replace clinical judgment. Your regulatory team decides what evidence to submit and remains accountable for it.

Related guides

  1. 1FDA PCCPFDA PCCP AI Device Change Control Testing Guide · United States (federal)
  2. 2ONC HTI-1 DSIONC HTI-1 Decision Support Transparency Testing Guide · United States (federal)
  3. 3HIPAAHIPAA PHI Leakage Prevention Testing Guide for AI Agents · United States (federal)
  4. 4EU AI ActEU AI Act High-Risk AI Testing & Validation Guide · European Union

Run FDA SaMD assertions in CI.

Load the vertical pack, add these assertions to your suite and keep the signed run manifest as evidence. 14-day free production trial, no card.

Start free trialBook a technical demo