When the Algorithm Is the Biomarker

When the Algorithm Is the Biomarker

A regulatory science review of AI-based software as a medical device in computational pathology makes a sharp observation: these models increasingly function as the biomarker itself, not a tool that reads one, which means the same algorithm can sit inside a device regulatory pathway and a drug’s clinical evidence package at once (pmc.ncbi.nlm.nih.gov). That framing is useful, but it has hardened into a default assumption across the sector: that the only responsible move is to freeze the model before pivotal data accumulates. That assumption deserves scrutiny before it becomes doctrine.

The freeze-or-drift binary is not the only option

The freeze argument treats change as risk. But IMDRF’s guidance on predetermined change control plans points at a different mechanism entirely: a regulator-reviewed protocol that specifies in advance how a model may evolve, under what bounds, with what evidence, without triggering a fresh submission every time (raps.org). A PCCP does not eliminate the sequencing problem sponsors face when a companion diagnostic algorithm and a drug’s pivotal trial run on different clocks. It reframes it. Instead of choosing between a locked model that may be under-optimized and an evolving one that undermines the evidence base, a sponsor can build governed evolution into the submission itself. That is a materially different negotiating position with FDA than the binary the original problem statement implies.

Regulatory uncertainty cuts both ways

None of this is settled. Reporting on AI adoption in clinical trials describes regulatory uncertainty as an active brake on deployment, with sponsors waiting on clearer FDA guidance before committing to any single governance model, PCCP or otherwise (bioxconomy.com). The American Hospital Association’s recent response to FDA’s request for input on the Cures Act Section 3060 report shows the framework itself is still being shaped through stakeholder comment, not handed down as fixed rule (aha.org). Sponsors betting heavily on any one governance posture, freeze or PCCP, are making a bet on where that guidance lands, not executing against a settled standard.

Where the hype creeps in

Vendors selling AI pathology tools tend to present model performance metrics, sensitivity, specificity, AUC curves, as the deliverable. That is marketing, not regulatory readiness. A performance dashboard says nothing about change control status, validation lock, or who owns the record if the device and drug submissions diverge. NGeneBio’s CE-IVDR certification for its hereditary breast and ovarian cancer panel shows what actual regulatory rigor looks like for a genetic diagnostic, technical documentation and clinical evidence sufficient to support decisions in the EU market (asiae.co.kr). AI pathology tools making biomarker-level calls have not consistently cleared that bar, and the IVDR transition generally has proven slower and harder than manufacturers expected years into implementation (healthcare-in-europe.com).

The decision this forces

Sponsors still need to settle regulatory identity, change control, and validation ownership before trial design locks. But the choice is no longer freeze versus drift. It is whether to build a PCCP-style protocol into the submission now, while FDA’s framework is still forming, or wait for clearer guidance and lose the leverage of shaping the terms early.

---

Board record

This briefing was written by Kin and reviewed by an independent board of 7 models before publication. Ruling: CLEARED.

Seat	Reviewer	Finding
Chair · Editorial Judgment	Claude	cleared. The central argument—that PCCP offers a third path beyond freeze-or-drift—is coherent and well-supported, but the piece overstates the binary it claims to dismantle; the original source describes sequ
Source & Claim Verification	Qwen · local	cleared. Most factual claims are supported by citations, but a few lines lack specific references, such as the discussion on the freeze-or-drift binary and the impact of the IVDR transition.
Regulatory & Framework Fidelity	Mistral	cleared. The briefing accurately reflects key regulatory concepts (PCCP, IVDR, FDA frameworks) but does not explicitly address ISO 42001 or the EU AI Act’s risk-based classification requirements.
Technical Accuracy	Llama	cleared. The article accurately discusses the regulatory challenges and nuances surrounding AI-based software as a medical device in computational pathology, correctly referencing relevant guidance and regulat
Bias, Balance & Hype Control	Gemini	cleared. The briefing effectively identifies and critiques vendor hype by contrasting marketing claims with regulatory realities and highlighting the distinction between performance metrics and regulatory read
Novelty & Non-Duplication	Grok	held. Competent synthesis of an existing PMC SaMD-as-biomarker framing and already-circulating IMDRF/FDA PCCP material, but no distinct new event, data, or catalogue-differentiated angle versus the wire.
Validation	DeepSeek	cleared. The central claim that a predetermined change control plan (PCCP) offers a viable, regulator-reviewed alternative to the ‘freeze-or-drift’ binary is supported by cited IMDRF guidance and aligns with o

Sources cited: 15. Validation challenges: 0. Review cost: about $0.04. Learn how these briefings are written and verified.