Our own paper's discussion section cites two pieces of European medical-device law and moves straight to the conclusion. This page does not move on: which regulation actually applies, where Rule 11 lands a model like ours, and what the Epic Sepsis Model already taught the field about the gap between deployed and validated. The Substack post makes the case; this page is the working file behind it.
Episode 05 of Road to CEPAS 2026 follows up on a discussion-section footnote in van den Berg et al. 2023 that cites both major pieces of European device law and then moves on. The Substack post argues that the footnote is not housekeeping — it reclassifies the false-alarm rate documented in Episode 04 from a metric into a hazard.
This page is the working file: the MDR/IVDR distinction walked through on its own terms, the Rule 11 classification ladder set out step by step, the Epic Sepsis Model as a deployed-but-unvalidated cautionary example, and the state of the AI Act overlay and the pending MDR/IVDR simplification proposal as of mid-2026.
Europe has two device regulations, and they classify by completely different logic. The word "diagnostic" pulls the reflex toward the in-vitro side — sepsis, blood cultures, diagnosis everywhere. But the deciding question isn't the subject matter; it's whether the device examines a specimen taken from the body. Our model never touches one. It runs on continuously monitored physiology and data already in the record, and produces a score that informs a clinical decision. That is software as a medical device, under the MDR — not the in-vitro regulation, despite the diagnostic-sounding subject.
This isn't pedantry. Which regulation applies decides everything downstream: what evidence is owed, whether an independent body audits the technical file, how long conformity assessment takes, what it costs.
| Governs | Applies to a model like ours? | |
|---|---|---|
| IVDR (2017/746) | Devices whose medical purpose depends on examining a specimen taken from the body | No — no specimen is examined |
| MDR (2017/745) | Standalone software and other devices, including software that informs a clinical decision from data already collected | Yes — continuous physiology, no specimen |
Both regulations are cited in Berg 2023's discussion, for completeness rather than as a live choice. Read against the definitions, only one of them governs a no-specimen physiological predictor. This page treats that as settled and moves to the question that actually has teeth: what the MDR does with software like this.
The MDR has a single rule — Rule 11, Annex VIII — that classifies software, and it is notorious for pushing almost everything upward. The logic runs by worst credible harm, not by how often the software is right. A simpler, more interpretable model with the same intended use lands in exactly the same class as a complex ensemble — a point that should sound familiar to anyone who read Episode 03.
| Trigger | Class | |
|---|---|---|
| Baseline | Provides information used in diagnostic or therapeutic decisions | IIa |
| First limb | Those decisions could cause death or irreversible deterioration | III |
| First limb | Those decisions could cause serious deterioration, short of the above | IIb |
| Second limb | Monitors vital physiological parameters, where the nature of the variation could put the patient in immediate danger | IIb |
Walking a model like Berg 2023 through this ladder: it monitors continuous vital-sign physiology; the variation it watches for is the early signature of a baby becoming septic, a state where deterioration can be immediate and severe; and it informs whether to start antibiotics in a preterm infant, where error in either direction has consequences. On a conservative reading that lands at Class IIb — very plausible given the intended use and performance claim as the paper currently states them, though a Class III argument is not impossible depending on how those are eventually worded. This is not a formal classification and none has been sought; it is the most defensible reading of the published paper against the current text of Rule 11.
Class I — the tier a manufacturer can largely self-declare — is essentially unreachable for anything that monitors vital signs to flag a fast-moving, potentially severe condition. From Class IIa upward, an independent notified body audits the quality system and reviews the technical file before anything is allowed near a patient. The thing that pushes a model over that line is not its sophistication. It's the severity of the decision it touches.
The Epic Sepsis Model is the story that should be taught alongside every sepsis-prediction paper, and it isn't ours. It was built into one of the most widely used electronic health record systems in the world and deployed across hundreds of US hospitals — about as far from "research artifact" as a model gets. An external team at Michigan Medicine then validated it independently, on hospitalisations it had never seen in training.
| Value | |
|---|---|
| Hospitalisations evaluated | 38,455 (Dec 2018 – Oct 2019, Michigan Medicine) |
| Sensitivity at deployed threshold | 33% — roughly two-thirds of sepsis cases missed |
| Positive predictive value | 12% |
| AUC | 0.63 (95% CI 0.62–0.64) |
Deployed everywhere; externally validated almost nowhere, until someone finally checked. That gap — between deployed and validated — is the gap Rule 11's performance-claim requirement exists to close before deployment, not after.
A retrospective AUC on a training-adjacent dataset, however good, is a claim about that dataset. The ESM's own developer-reported figures looked adequate before Michigan checked. Regulation is, in effect, insisting that someone external check before hundreds of hospitals — or one NICU — find out the hard way.
The AI Act (Regulation (EU) 2024/1689) layers a second regime on top of device law. An AI system that is itself a medical device — or a safety component of one — requiring notified-body assessment under the MDR generally falls into the high-risk AI category. On the Rule 11 reading above, a model like ours would. High-risk status brings obligations around data governance, representativeness of training data, transparency, human oversight, logging, and post-market monitoring, layered onto the MDR technical file rather than run as a wholly separate track.
The picture is not settled, and a research diary should say so rather than treat the law as fixed. On 16 December 2025 the European Commission published a proposal — COM(2025) 1023 final — to amend the MDR and the IVDR, including a revision to Rule 11's software-classification logic that could move more medical device software into Class I. If adopted as proposed, the AI Act's own annex would also move MDR/IVDR devices from the section that automatically triggers high-risk status into one that doesn't. This is a proposal, not adopted law. It has been submitted to the European Parliament and the Council; the Commission's own timeline points to co-legislator adoption around Q2 2027 at the earliest, and the wording is free to change during negotiation.
Everything above is the picture as it reads in mid-2026. The direction of travel is toward simplification and toward folding the AI Act into the existing device-conformity architecture rather than running two parallel assessments — but the exact class boundaries could still move before a model like this, or any model in this literature, reaches a notified body. The details will have moved again by October. The underlying logic — severity of the decision drives the obligation — won't.
One study in this literature did the thing Rule 11 implicitly asks for: a prospective randomised trial with a hard clinical outcome, not a retrospective AUC. The heart-rate-characteristics work traced back in Episode 01 ran a two-group, parallel, individually randomised trial of 3,003 very-low-birth-weight infants across nine NICUs, comparing a group where the HRC index was displayed to clinicians against a masked control. All-cause mortality fell from 10.2% to 8.1% in the displayed group (hazard ratio 0.78, 95% CI 0.61–0.99, P = .04), published in 2011.
Fifteen years later it remains the canonical prospective randomised trial in this corner of the literature. Much of the field since, ours included, has stopped at the rung marked "good ROC curve." The regulation is, in a sense, the formal version of the question that trial already answered and the rest of us have been deferring: does using this change what happens to the patient?
A retrospective AUC, however good, is a claim about a dataset. The 2011 trial is a claim about a baby. Episode 06 follows that thread properly: what a real clinical validation of a model like ours would have to look like, how the 2011 trial was actually designed, and why — as far as a Branch A search of this literature can currently tell — that trial still stands largely alone.
This page discusses and critiques the sources above; it does not reproduce them. Regulatory text and reported figures are paraphrased and pointed back to the primary source. Classification language ("very plausible," "not impossible") reflects the author's own reasoned reading of Rule 11 against the published paper, not a formal determination by a notified body.