Life Sciences / Regulatory Review 🧬

📌 Navigate

📋 Exec Summary

Q3 2023 exposed the structural gap between AI capability and regulatory pathway. FDA tracked toward ~300 clearances/year with cumulative total approaching 700, but zero generative AI devices were cleared. Llama 2's commercial license created the first viable on-premise clinical LLM architecture (no PHI leaves the building), and health systems began evaluating deployments outside device regulation. PCCP guiding principles approached publication, and Nuance DAX Copilot expanded ambient documentation into mainstream clinical infrastructure.

📊 What Moved

📈 Trend Arcs

Arc 1: On-Premise LLM Architecture as HIPAA Strategy

Velocity: Accelerating

Meta's release of Llama 2 on July 18 with a commercial-use license created a new architectural option for healthcare AI builders that had not existed with any comparable model. Before Llama 2, a healthcare organization wanting to use a frontier-capable LLM had two options: send data to an OpenAI or Anthropic API (requiring a Business Associate Agreement, data transmission off-premise, and accepting the model provider's data retention policies), or build its own model from scratch (requiring resources available to only the largest health systems and pharma companies).

Llama 2 opened a third option: deploy a competitive open-weight model on-premise, within the health system's own infrastructure, fine-tune it on de-identified clinical data, and operate it entirely inside the firewall. No PHI leaves the building. No BAA required with a model provider. No per-token API cost at inference scale. The HIPAA architecture problem — which had been the primary friction preventing LLM adoption in clinical workflows — had a new solution.

Throughout Q3, the discussion in clinical AI circles shifted from "can we use LLMs in healthcare" to "how do we operationalize Llama 2 on-premise." Inference hardware requirements (a 70B parameter model requires multiple high-end GPUs), fine-tuning toolchains, and model governance frameworks became the engineering questions. The regulatory question — whether an on-premise Llama 2 deployment constituted a regulated medical device — remained unresolved and was largely being bracketed as a future problem.

Health system IT and clinical informatics teams began evaluating on-premise Llama 2 as infrastructure for internal tools — clinical note summarization, prior authorization drafting, discharge summary generation — that they assessed as not triggering device regulation. Whether that assessment is correct will be tested in Q4 and Q1 2024 as the FDA clarifies the boundary between software as a medical device and clinical workflow tools.

Where it stands at quarter close: On-premise LLM deployment is an active engineering project at multiple large health systems and academic medical centers. No production deployments publicly, but the architectural question has been answered: Llama 2 at 70B is a credible on-premise clinical AI foundation model.

Arc 2: The 510(k) Predicate Engine and Its Structural Limits

Velocity: Steady

The 510(k) pathway — which requires demonstrating substantial equivalence to a legally marketed predicate device — has been the workhorse for AI/ML medical device clearance since the first wave of devices in the late 2010s. In Q3 2023, that engine continued running efficiently within its established lanes: imaging AI for radiology, ECG analysis, ophthalmologic screening. These specialties have deep predicate libraries and well-understood performance benchmarks, making the substantial equivalence argument straightforward.

The structural limit became clearer over the quarter: the 510(k) pathway does not extend naturally to novel AI modalities. A generative AI device that produces free-text clinical recommendations has no cleared predicate. A multimodal model that combines imaging, lab values, and notes to generate a differential diagnosis has no cleared predicate. The predicate estate was built for narrow discriminative models; the capability frontier in Q3 2023 was generative and multimodal. The FDA's De Novo pathway — for novel devices without predicates — exists to address this, but it is slower, more expensive, and requires more clinical evidence than 510(k).

The quarter's clearance composition reinforced the pattern: the cleared devices were overwhelmingly narrow, imaging-based, 510(k) submissions. The frontier AI capability was none of those things. As long as the regulatory pathway infrastructure does not evolve alongside the capability frontier, the cleared device population will remain systematically behind what the technology can do.

Where it stands at quarter close: The 510(k) engine is producing more clearances than ever, but the devices being cleared represent Q3 2021 technology, not Q3 2023. The regulatory lag is approximately two years and widening.

Arc 3: AI Drug Discovery Moves from Claim to Evidence

Velocity: Accelerating

The Q3 2023 drug discovery story is not primarily a regulatory one — it is a clinical one, and it matters for the life sciences x AI arc because it is where the first hard evidence of AI-originated clinical benefit is being generated.

The anchor event was Insilico Medicine's INS018_055 entering Phase 2 clinical trials in June 2023 for idiopathic pulmonary fibrosis — the first small molecule discovered and designed entirely by AI to reach Phase 2. While not a Q3 event itself, its Phase 2 status was current throughout the quarter and received significant attention as the clinical AI community processed what "AI-designed drug in human trials" actually meant.

In July, Nvidia invested $50 million in Recursion Pharmaceuticals, pairing Recursion's phenomics-based biology foundation models with Nvidia's BioNeMo platform — a generative AI cloud specifically designed for drug discovery applications. This represented the first significant infrastructure investment in AI drug discovery as a compute-intensive discipline analogous to frontier LLM training: requiring large compute commitments, specialized hardware, and a platform strategy.

The arc across Q3 was a shift in the standard of evidence being demanded of AI drug discovery companies. The question moved from "does your AI identify targets faster?" (a claim that can be made without clinical data) to "does the molecule your AI designed work in humans?" (a question that takes five to ten years to answer and is now being answered, at least at Phase 2). Investors, pharma partners, and the scientific community began distinguishing between AI companies with molecules in clinical trials and AI companies with pipeline claims.

Where it stands at quarter close: One AI-designed compound is in Phase 2 (Insilico/INS018_055). Several others are in Phase 1. The clinical evidence generation arc is active but will not produce efficacy readouts until 2024-2025 at the earliest. BioNeMo represents the infrastructure layer being built for the next generation of AI-designed therapeutics.

🗺️ Landscape Shift

The competitive and regulatory map across Q3 2023 was shaped by the gap between what AI could do and what the FDA had a pathway to clear.

The most significant structural change in the quarter was the creation of a viable on-premise architecture for clinical LLM deployment (Llama 2) while the regulatory boundary between tool and device remained undefined. Health systems and clinical AI companies spent Q3 testing the question: which LLM applications are regulated medical devices and which are clinical workflow software? The FDA's clinical decision support guidance from 2022 provides some framework but does not directly address LLMs. That ambiguity is both an opportunity (deploy as a tool while the pathway is undefined) and a risk (the FDA can issue a warning letter or request a 510(k) retroactively if the product is later deemed to meet the device definition).

🧭 Regulatory Direction of Travel

Clearance velocity and mix: The FDA's total AI/ML-enabled device list was approaching 700 cumulative clearances by September 30. New clearances in 2023 were running at a pace that would produce 250-300 additional clearances for the full year — a rate of approximately one per working day. The mix was stable: 510(k) accounted for approximately 97% of clearances; radiology accounted for approximately 80% of the specialty mix. De Novo clearances were single digits. PMA clearances for AI/ML software were effectively zero. This distribution is a function of where predicates exist, not where clinical need is greatest.

PCCP adoption trajectory: The Predetermined Change Control Plan framework was in final development at quarter close, with the FDA/Health Canada/MHRA joint guiding principles expected in October. PCCP represents the FDA's bet on how to handle adaptive AI: pre-specify the change envelope at the time of clearance rather than requiring new submissions for every model update within that envelope. No device had been cleared using a PCCP as of September 30, but several submissions were reportedly in the pipeline incorporating PCCP elements. The framework's practical adoption will depend on how the FDA reviews the first PCCP-inclusive submissions — specifically, what level of specificity in the change protocol is required, and what post-market performance monitoring triggers a mandatory update to the PCCP.

Evidence standard evolution: What the FDA was asking for in AI/ML 510(k) submissions in Q3 2023 was evolving in the direction of more real-world performance data rather than test set benchmarks. The 2021 AI/ML action plan and subsequent guidance documents had increasingly emphasized the importance of testing on demographically diverse populations, prospective performance monitoring post-clearance, and algorithmic transparency sufficient for physician users to understand the model's operating range. These are not formal requirements in the way that a 510(k) performance benchmark is a requirement, but submissions that addressed them were moving faster than submissions that did not.

International alignment: The FDA-Health Canada-MHRA collaboration on PCCP guiding principles was the quarter's most significant international regulatory development. It established a trilateral framework for thinking about adaptive AI — a meaningful advance on the prior state of three separate regulatory bodies developing parallel but uncoordinated frameworks. EMA was not part of this specific collaboration, though EU AI Act negotiations were proceeding in parallel in Brussels.

Guidance pipeline: The PCCP guiding principles were the primary near-term publication expected. The FDA had signaled that further guidance on good machine learning practice (GMLP) and software as a medical device (SaMD) performance assessment would follow in 2024. No formal guidance directly addressing LLMs or generative AI in regulated medical applications was in the published pipeline as of September 30.

💰 Funding & Deal Pattern

Life sciences x AI capital flows in Q3 2023 reflected a bifurcated market: drug discovery AI continued to attract significant investment, while clinical AI (device-adjacent) faced a more cautious environment driven by regulatory uncertainty and the absence of cleared LLM products.

Insilico Medicine, having reached Phase 2, was in a stronger fundraising position than at any point in its history. The Phase 2 status provided evidence-of-concept that clinical development was possible, not merely claimed.

The most active segment below drug discovery was ambient clinical documentation — companies building physician workflow tools (note generation, prior authorization, coding assistance) that were explicitly positioned outside device regulation. The combination of immediate commercial deployment opportunity and lower regulatory overhead made this a more attractive investment segment in Q3 than device-regulated clinical AI.

🔍 The Counter-Narrative

• The consensus: 80% radiology concentration proves FDA is failing to address clinical AI broadly. The reality: The 510(k) pathway produces clearances where predicates exist. Radiology has hundreds; psychiatry has zero. The correct frame is "other specialties have not yet built a predicate estate." The solution is earlier De Novo submissions in underrepresented specialties, not regulatory criticism of the existing mechanism.

• The consensus: Zero generative AI clearances means FDA is blocking generative AI adoption in medicine. The reality: No company had submitted a 510(k) or De Novo for a generative AI device with sufficient clinical evidence. FDA cannot clear devices that haven't been submitted. This is a product development and evidence generation problem, not a regulatory obstruction problem.

📐 Builder's Benchmark

Clearance timelines by pathway (estimated, Q3 2023):

The practical implication for builders: if you are developing a device in a specialty with an established AI predicate estate (radiology, cardiology ECG, ophthalmology), 510(k) is the viable path with an 18-24 month total timeline from design lock to clearance. If you are developing a generative AI device or a device in a specialty without AI predicates, you are in De Novo territory — plan for 24-36 months and considerably more clinical evidence.

Reimbursement: This remains the most significant post-clearance friction for clinical AI. CMS Category B IDE codes and new technology add-on payments (NTAPs) exist but are not systematically available for AI/ML diagnostic devices. Several cleared AI devices were operating in commercial markets with payer-by-payer coverage determination, which creates unpredictable revenue models. The AMA's work on CPT codes for AI-enabled services was ongoing but had not produced broad coverage as of Q3 2023.

Clinical validation study designs being accepted: For 510(k) AI/ML submissions, the FDA was increasingly emphasizing prospective testing (not just retrospective test set performance), reader studies with appropriate statistical powering, and demographic diversity in validation cohorts. Studies relying entirely on retrospective datasets without prospective validation were receiving more AI3 feedback requests. The minimum bar was rising even within the established pathway.

PCCP adoption: Zero devices cleared using a PCCP as of September 30. First PCCP-inclusive submissions reportedly in process. Practical guidance on what level of change specification is sufficient remains the outstanding unknown.

👀 What to Watch

1. PCCP guiding principles publication (October 2023): The FDA/Health Canada/MHRA document is expected in October. The critical detail to watch: whether the document specifies what a "performance specification" must include for a manufacturer to pre-authorize model retraining. This detail determines whether PCCP is a practical pathway or a framework waiting for operationalization.

2. First PCCP-inclusive 510(k) submission: The first manufacturer to submit a 510(k) incorporating a PCCP will define the reference point for what's acceptable. Watch the FDA's 510(k) database for submissions with PCCP elements in Q4 2023 and Q1 2024.

3. INS018_055 Phase 2 interim data: Insilico Medicine's AI-designed IPF compound entered Phase 2 in June. Interim safety and early efficacy data would be expected to emerge in conference presentations in late 2023 or early 2024. The first clinical signal from an AI-designed drug will be the most consequential data point for the AI drug discovery sector since the Phase 2 entry itself.

4. Nuance DAX Copilot health system adoption metrics: Microsoft/Nuance has not disclosed deployment numbers for DAX Copilot, but industry reporting has suggested significant expansion throughout 2023. Any quantitative disclosure of health system deployments or physician adoption rates would establish the baseline for what "ambient AI documentation at scale" looks like — and inform how other vendors and health systems evaluate similar tools.

5. FDA's response to clinical LLM deployment inquiries: Device developers and health systems have been submitting informal Q-Submissions (pre-submission meetings) to FDA asking how clinical LLM applications should be regulated. If the volume of these inquiries reaches a threshold that prompts FDA to issue a general response or FAQ — as it did with earlier waves of clinical AI — that signal would emerge in Q4 2023 or Q1 2024. Watch the FDA docket and any CDRH public statements at conferences (RSNA in late November is the relevant venue).

📎 Sources

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