FDA Consultant

FDA AI Letter Response — Action Plan

The AI letter clock starts the day FDA sent it (not when you received it). Confirm the exact response deadline today.

Step 1: Categorize each deficiency (first 48 hours)

Group into: Data gaps (need new analysis), Clarification (existing docs), Regulatory interpretation (written argument).

Step 2: SE argument — predicate challenge FDA questioning SE typically means: (a) intended use doesn't match closely enough, or (b) different tech raises new S&E questions.

Your response must follow the SE framework:

"[Your device] has the same intended use as [predicate K-number] because [specific language match]. Although [your device] uses AI/ML vs. [predicate's method], this difference does not raise different questions of safety and effectiveness because [cite performance data, algorithm validation, risk analysis]."

Step 3: Clinical performance data gap Before generating new data (impossible in 90 days), exhaust existing sources:

• Reanalyze existing data with different subgroup stratification (age, sex, device type)
• Holdout test set data not used in training?
• Reader study data (physician vs. algorithm agreement)?

Step 4: Response format Each response section must:

1. Quote the exact FDA question verbatim
2. State your response clearly
3. Reference the exhibit/attachment number
4. End with: "Based on the above, [Company] respectfully requests FDA find [device] substantially equivalent to [predicate]."

Consider requesting a teleconference with the FDA reviewer before submitting — allowed once per AI letter and often clarifies acceptable evidence thresholds.

Want me to draft the SE argument section or the AI response cover letter?

Responding to an FDA Additional Information Request — 23 Questions on a 510(k)

Four months in is normal — FDA's 510(k) review clock for standard submissions is 90 days from acceptance, but AI requests stop the clock. The moment you received this, your 90-day response window started. Confirm the exact deadline from the letter header today; missing it means automatic withdrawal.

Step 1: Triage All 23 Questions Before Writing a Word

Categorize every question before drafting any response:

Do this categorization in a spreadsheet with the question number, category, owner, and draft-due date. With 23 questions, you likely have 15–16 clarifications, 4–5 data requests, and 1–3 that need careful interpretation. The 3 AI/ML validation questions almost certainly fall into "additional data needed" — but read them closely for implied design issues.

The 3 AI/ML Validation Questions — What FDA Typically Expects

FDA's AI/ML validation questions on 510(k)s follow a consistent pattern shaped by the 2021 AI/ML Action Plan and the 2023 draft guidance on predetermined change control plans. The data FDA expects:

Training and test data:

• Total dataset size, train/validation/test split ratios (and whether the split was done at the patient level, not the image/sample level — patient-level split is required to prevent data leakage)
• Dataset demographics: age, sex, race/ethnicity, relevant clinical characteristics — FDA will check for underrepresented populations
• Site diversity: number of clinical sites, geographic distribution, device/acquisition variability if applicable
• Whether test data was truly held out (locked before any model training began) or whether it was used in any hyperparameter selection

Performance metrics:

• Point estimates with 95% confidence intervals (not just point estimates)
• Performance stratified by relevant subgroups — if your device is cleared for a broad population, FDA wants to see that performance holds across demographic subgroups
• Comparison to the predicate's performance on the same or comparable dataset, or a head-to-head comparison if feasible

Algorithm description:

• Sufficient detail for FDA to understand the architecture without source code — input features, model type, output type, how the output maps to the clinical decision
• Any post-processing or thresholding applied to model outputs

If you didn't include the confidence intervals, subgroup analysis, or patient-level split documentation, those are likely what FDA is asking for. Pull your study data and generate those now.

Pre-Response Strategy: Request an Informal Meeting First

Before submitting written responses to the AI/ML questions, request an informal teleconference with the FDA reviewer. FDA explicitly encourages this for complex AI questions — the reviewer can tell you in 20 minutes what level of evidence they will accept, saving you weeks of documentation work.

How to request: Reply to the AI letter contact (or use your eSTAR submission contact) and ask for a "brief teleconference to discuss the AI/ML validation questions prior to submitting our formal response." Keep the request to one paragraph. FDA typically schedules these within 2–3 weeks.

Use the meeting to confirm: (1) whether existing data is sufficient or new studies are required, (2) acceptable formats for the additional data, and (3) whether a Predetermined Change Control Plan (PCCP) would satisfy any ongoing monitoring questions.

Response Format

The AI letter response must follow this structure for each question:

1. Restate the FDA question verbatim — do not paraphrase
2. Give a direct answer in the first sentence — do not bury the answer
3. Provide supporting data/documentation with exhibit references (e.g., "See Exhibit A, Table 3")
4. Cross-reference the original submission section where relevant context was provided

End each response section with a brief statement that the question has been fully addressed. Keep the cover letter short — one page confirming receipt, response deadline compliance, and a list of new exhibits attached.

If a Question Implies a Design Change

If any of the 3 AI/ML questions implies that your current validation approach has a fundamental gap — for example, FDA suggests your test set was contaminated by training data, or the performance does not support the intended use as claimed — this is not answerable with additional documentation. A question that implies your device does not demonstrate substantial equivalence as submitted is a signal that a design or labeling change may be needed, which would require a new SE determination.

If you see that pattern, escalate immediately and consider whether a voluntary withdrawal and resubmission (with a revised intended use or updated validation study) is faster than trying to defend the current submission.