Life Sciences / Regulatory Review 🧬

📌 Navigate

📋 Exec Summary

Q1 2023 delivered the adaptive AI regulatory framework: FDA published PCCP draft guidance on March 30, creating the first draftable, commentable specification for how AI/ML devices can update post-clearance. Simultaneously, GPT-4's USMLE-passing performance accelerated clinical LLM deployment without any regulatory framework, widening the gap between cleared and actually-used AI. The EU AI Act was still in negotiation before trilogue began, and the public AI/ML device list remained heavily radiology-concentrated.

📊 What Moved

📈 Trend Arcs

Arc 1: FDA Adaptive AI Framework Moves from Concept to Draft Regulation

Velocity: Accelerating

The PCCP guidance published March 30 was not spontaneous. The FDA's 2019 discussion paper on AI/ML-based SaMD had introduced the concept of a PCCP — a document pre-specifying planned model changes — as a theoretical framework for managing adaptive AI without requiring a new 510(k) for every update. The 2021 Action Plan committed the agency to developing formal guidance on PCCP. Q1 2023 saw that commitment materialize as a 24-page draft guidance available for public comment.

The draft specifies that a PCCP must include: (1) a description of planned modifications, (2) the methodology for making those modifications (training data, re-training procedures), (3) the performance objectives that define acceptable post-modification behavior, and (4) a methodology for assessing whether modifications stay within the pre-specified bounds. The guidance applies to both 510(k)s and De Novo requests, and to devices with learning algorithms that modify their own behavior based on new data.

What the guidance does not resolve: the question of when a modification is within-PCCP versus a "major change" requiring a new submission. The boundary between "modification consistent with PCCP" and "new device requiring new submission" is a judgment call that the draft guidance leaves substantially to manufacturers, with FDA review reserved for edge cases. This ambiguity is not a flaw — it is a pragmatic acknowledgment that the agency cannot predict all the ways adaptive AI will evolve — but it creates legal uncertainty that regulatory counsel will be navigating for years.

The comment period closed June 2023. Industry response was substantive: manufacturers asked for clearer bright lines on when a PCCP change triggers a new submission, requested guidance on how PCCP interacts with international regulatory requirements (CE marking, PMDA), and pushed back on the performance objective specifications as potentially too rigid for models that improve on long tail cases.

Where it stands at quarter close: Draft guidance published, in comment period, final guidance expected in 2024. PCCP is the operative framework for adaptive AI but with material open questions about scope boundaries.

Arc 2: Clinical LLM Deployment Outpaces Regulatory Clarity

Velocity: Accelerating

The deployment velocity of general-purpose LLMs in clinical settings outpaced the FDA's regulatory response throughout Q1 — and the gap widened after GPT-4's launch in March. By the end of the quarter, documented use cases in published literature and health system announcements included: clinical note summarization (Nuance DAX Express was announced in Q1; Epic and Oracle Health integrations came later), patient-facing chatbots for symptom triage, discharge summary generation, and differential diagnosis support in emergency department settings.

None of these deployments came with FDA clearance. The manufacturers of the underlying models — OpenAI, Anthropic — explicitly disclaim clinical decision-making as an intended use. Health systems deploying these tools typically categorized them as "productivity tools" or "clinical decision support" in the lowest-risk category under the FDA's clinical decision support guidance, which exempts software that "displays, analyzes, or prints medical information about a patient" and is intended for use by clinicians who can independently review its basis.

The FDA's clinical decision support guidance (finalized 2022) does carve out a category of non-device clinical decision support — specifically software that: (a) is not intended to replace clinical judgment, (b) is based on recommendations in publications or clinical guidelines, (c) is transparent enough that the clinician can independently review the basis for the recommendation, and (d) is not for life-threatening conditions. General-purpose LLMs are difficult to fit cleanly into this carve-out: they are not based on specific publications, their outputs are not always traceable to auditable sources, and there is no regulatory test for whether a given deployment is for a "life-threatening condition."

The practical result in Q1: deployment was happening, regulatory treatment was unclear, health systems were making their own risk determinations, and no enforcement action had been taken. This is a dynamic the FDA has managed before — clinical software that outpaces regulatory clarity tends to get addressed either via enforcement discretion (formal non-enforcement while guidance develops) or via guidance that retroactively characterizes what has already been deployed. It has not, as of March 31, said which approach it will take.

Where it stands at quarter close: Deployment is widespread and accelerating, regulatory framework is ambiguous, FDA has not acted. The trajectory points toward a guidance document or enforcement action in the 12–24 month window.

Arc 3: EU AI Act Creates Divergence Risk with FDA Framework

Velocity: Accelerating

The EU AI Act was still in negotiation through Q1, before trilogue began, and was on track to create a binding AI regulatory framework that covers medical AI differently than the FDA's device-centric approach. The key divergence: the EU AI Act treats high-risk AI systems (including medical devices and in vitro diagnostics used for automated decision-making) as regulated objects regardless of whether they are sold as medical devices — the risk classification is based on intended use in a regulated context, not product registration status. The FDA's framework regulates medical devices — products marketed and sold for medical purposes with a specific intended use declared by the manufacturer.

The practical consequence of this divergence: a general-purpose LLM deployed by a hospital for clinical use might be high-risk AI under the EU AI Act (because the hospital is the deployer and the use is clinical), while the same tool might not be regulated by the FDA at all (because the manufacturer hasn't declared a medical intended use). This asymmetry creates compliance complexity for any organization operating in both markets — the entity responsible for conformity, the documentation required, and the conformity assessment process differ.

Simultaneously, the Act's general-purpose AI provisions — added late in the negotiation to address frontier models like GPT-4 — were creating their own complexity. The draft provisions required providers of general-purpose AI to maintain technical documentation, register in an EU database, and comply with copyright obligations, but the interface between these obligations and the medical device-specific high-risk AI requirements was not clearly specified. A model that is general-purpose AI (frontier model obligations) AND used in medical devices (high-risk AI obligations) — which set of requirements governs?

By March 31, the EU AI Act had not resolved these tensions. The expected finalization timeline was late 2023, with a 2-year transition period before full application.

Where it stands at quarter close: Trilogue ongoing, key provisions unresolved, final text expected late 2023. The framework will create real compliance divergence with FDA for any company operating in both jurisdictions.

🗺️ Landscape Shift

🧭 Regulatory Direction of Travel

Clearance velocity and mix: The FDA maintained a roughly linear clearance pace through Q1, with estimates of 50–60 AI/ML device clearances annually at current pace. No material acceleration or deceleration in the quarter. Radiology (particularly chest, breast, and cardiac imaging) continued to dominate. The pathway mix skewed heavily toward 510(k) — De Novo decisions on AI devices are slow because the novel classification process requires more FDA resources, and most manufacturers with viable predicates are not voluntarily choosing De Novo.

PCCP adoption trends: The draft guidance is the starting point. At quarter close, zero devices had been cleared with an approved PCCP. The pipeline of companies preparing PCCP submissions is not publicly visible, but the guidance comment traffic and industry conference discussions suggest a significant number of manufacturers are planning PCCP inclusions in 2023–2024 submissions. Adoption will be visible in 2024 clearance data.

Evidence standard evolution: The FDA's AI/ML device submissions in Q1 were requiring more prospective clinical validation data than submissions from 2019–2020. The implicit evidentiary standard has risen as the cleared device base has grown — the agency has more comparators, more understanding of failure modes, and higher baseline expectations for what "clinical validation" means for an AI device. Specifically: retrospective studies alone are no longer sufficient for most 510(k) submissions; prospective or multi-site validation is increasingly expected. Real-world performance monitoring commitments are more common as post-market conditions.

International alignment or divergence: The FDA-EU divergence on general-purpose AI classification is the headline divergence risk (see Arc 3). On the device-specific side, the IMDRF SaMD framework continues to provide alignment scaffolding between FDA, EMA, Health Canada, PMDA, and TGA — but the IMDRF framework was designed for narrow, task-specific software and does not address adaptive AI or general-purpose models. IMDRF working groups on AI were active in Q1 but had not published updated guidance.

Guidance pipeline: PCCP draft is the Q1 output. The 2023 guidance calendar (FDA) includes expected publications on: AI/ML bias and fairness in device submissions (draft expected 2023), predetermined change control plan final guidance (expected 2024), and continued policy statements on clinical decision support. The PCCP draft is the most actionable near-term output.

💰 Funding & Deal Pattern

Q1 2023 AI x life sciences funding was bifurcated: drug discovery AI attracted large rounds; clinical AI attracted smaller, more cautious capital.

🔍 The Counter-Narrative

• The consensus: PCCP speeds up AI device development by reducing re-submissions. The reality: PCCP front-loads specification work that previously happened informally or not at all. Manufacturers must pre-specify retraining methodology, data requirements, performance bounds, and degradation triggers before submission. The net effect on time-to-market is unclear; the effect on development discipline is positive. "Speed-up" is marketing for the framework; the accurate description is "upfront specification burden in exchange for post-market iteration flexibility."

• The consensus: 200 cleared AI devices proves a maturing AI device market. The reality: 80% of those clearances are radiology — the market is a single-specialty, single-modality ecosystem. For startups building outside chest X-ray interpretation or breast density analysis, the predicate ecosystem is sparse, evidentiary precedents are limited, and the regulatory path is less established than the headline number implies.

📐 Builder's Benchmark

Median clearance timelines by pathway (estimates, Q1 2023):

• 510(k) for AI/ML SaMD: 12–18 months from submission to clearance (includes typical RTA cycle and review time)
• De Novo for AI/ML SaMD: 18–30 months (novel classification, more intensive review)
• PMA for AI in high-risk indications: 3–5 years (limited AI device PMAs; most AI avoids PMA-level risk claims)
• Breakthrough Device designation (if granted): can compress by 30–50% — meaningful for companies that qualify

Reimbursement coverage rates for AI-enabled devices:

• Cleared AI radiology tools with CPT codes: handful with established reimbursement (e.g., Viz.ai, RapidAI in stroke) — single-digit percentage of cleared devices have stable, specific reimbursement
• Most cleared AI tools: billed under existing procedure codes for the underlying procedure, without AI-specific payment increments
• De novo CPT code creation timeline: typically 2–4 years from clearance to new CPT code, with additional 1–2 years to coverage determination

Clinical validation study designs being accepted (Q1 2023 FDA standard):

• Minimum for most 510(k) AI submissions: multi-reader, multi-case study (MRMC) for diagnostic imaging, OR prospective evaluation against reference standard for non-imaging AI
• Single-site retrospective studies: increasingly insufficient alone; FDA requesting supplementary prospective or multi-site data
• Labeling claims supported by validation: narrow and literal — the cleared indication is exactly what the validation covers, nothing broader
• Subgroup analysis: FDA increasingly requesting pre-specified performance data by race, sex, age, and relevant patient characteristics for AI devices post-2021

Time from submission to clearance (AI devices, 2022–Q1 2023 cohort):

• 510(k) first cycle clearance (no major deficiency): approximately 90–180 days
• 510(k) with major deficiency (typical for novel AI): add 90–180 days per additional review cycle
• Average for AI/ML SaMD with novel claims: 2–3 review cycles, total 12–18 months

PCCP adoption rates:

• Formal PCCP submissions: zero approved at quarter close (guidance was in draft)
• Companies actively preparing PCCP submissions: not publicly quantifiable, estimated at dozens based on industry conference engagement with the draft

👀 What to Watch

1. PCCP draft guidance comment period close (June 2023) — Industry comments will reveal where the specification burden is highest and which provisions manufacturers find most ambiguous. The volume and nature of comments will shape the final guidance. Watch for comments from major medical device manufacturers (Siemens Healthineers, GE Healthcare, Philips) as proxies for what large-market players will and will not accept.

2. FDA statements on LLMs in clinical settings — No formal FDA communication as of March 31. Any FDA speech, guidance, or Q&A document in Q2 addressing GPT-4 or general-purpose LLMs in clinical use will immediately become the operative framework for health system compliance teams. Watch CDRH director communications and Digital Health Center of Excellence publications.

3. EU AI Act trilogue conclusion — Expected late 2023 but political pressure may accelerate. The GPAI provisions and the medical device interface are the two open technical questions. Watch for European Parliament plenary vote timeline as the leading indicator of trilogue completion.

4. First PCCP submission announced — When the first company announces a 510(k) submission that includes a PCCP, it will reveal the practical interpretation of the draft guidance and set a de facto template. Watch 510(k) database and press releases from AI device companies with products already in the market.

5. CMS coverage decisions for AI devices — No specific Q2 dates, but the AMA CPT editorial panel meets periodically and CMS coverage decisions on AI-specific CPT codes will be the commercial enabler for cleared-but-uncompensated tools. Any CMS coverage decision for an AI-enabled diagnostic tool is a signal about the reimbursement model's viability.

📎 Sources

Key references for this quarter. Links provided where available; historical entries may reference publications by title and date.