Attest, don't render.
Stay non-device.
The FDA CDS criteria draw a bright line between software that drives a clinical decision and software that only proves one. Land on the attestation side and you stay non-device.
The FDA non-device CDS criteria define when clinical decision support software falls outside the medical-device definition: four cumulative conditions from section 520(o)(1)(E) of the FD&C Act, refined by the FDA's Clinical Decision Support Software final guidance. The decisive one is the fourth — the software must let a clinician independently review the basis for its recommendation rather than rely primarily on it. Software that attests a decision was genuine, instead of rendering it, never engages that reliance question at all.
Device or not
is a business decision.
If your clinical-AI software is a device, you inherit premarket review, quality-system regulation, and postmarket obligations — a slower, costlier path. If it is non-device CDS, you ship on a different footing. The FDA settles that question by intended use, not by how clever the code is. Founders and regulatory leads who understand exactly where the CDS line sits can design deliberately toward the side they want — and stay there on purpose, not by accident.
Four prongs.
All must hold.
Non-device CDS must satisfy all four criteria at once. First, it does not acquire, process, or analyze a medical image or a signal from a diagnostic device. Second, it displays or analyzes medical information. Third, it offers recommendations to a health care professional. Fourth, it enables that professional to independently review the basis of the recommendation. Miss any one and the software is regulated as a device.
The one that
decides everything.
The fourth criterion is where most clinical AI fails. The software must let the clinician independently review the basis for the recommendation, so the professional does not rely primarily on it. An opaque score, a time-critical alarm, or an autonomous action leaves nothing to review — and becomes a device. RankShieldMD sidesteps the question entirely: it never renders a decision, so there is no recommendation to over-rely on.
Prove the decision.
Never make it.
Rendering produces, scores, or recommends a decision — it influences what the clinician does. Attesting proves something about a decision another system already made: that a given output truly came from a given approved model, on intact inputs, at a stated time. Because classification follows intended use, attestation-only tooling never touches the clinical conclusion and never engages the fourth-criterion reliance test.
Add integrity
without device risk.
Register a model baseline, attest a decision another system made, and hand your buyer proof they can verify — all outside the clinical pathway. Attestation, not rendering. Intended-use clean, PHI-free, non-device.
Is my clinical decision support software an FDA device?
Your clinical decision support software is a non-device only if it meets all four FDA CDS criteria at once — and the deciding one is whether a clinician can independently review the basis for its recommendation rather than rely primarily on it. The FDA non-device CDS criteria define the boundary. They come from section 520(o)(1)(E) of the Federal Food, Drug, and Cosmetic Act, added by the 21st Century Cures Act, and are interpreted in the FDA's Clinical Decision Support Software final guidance. The four conditions are cumulative: the software must not acquire, process, or analyze a medical image or a signal from an in-vitro or signal-acquisition device; it must display, analyze, or print medical information; it must provide recommendations to a health care professional; and it must enable that professional to independently review the basis for those recommendations. Miss any single prong and the software is regulated as a device. The most important thing to internalize is that this classification turns on intended use, not architecture — the same code can be device or non-device depending on the claims made for it and how it is meant to be used. That is the lever a founder or regulatory lead can actually pull.
How do you keep clinical-AI software non-device under the FDA CDS criteria?
You design and describe the software so that it satisfies all four criteria at once, and you treat intended use as a first-class engineering constraint rather than an afterthought for the labeling. Start with the mechanical prongs: keep the software off medical images and diagnostic signals, and confine it to displaying or analyzing already-captured medical information. Those are relatively easy to hold if you scope the product deliberately. The hard, decisive work sits in the third and fourth prongs together — the software may offer recommendations to a health care professional, but only if the professional can independently review the basis for them and does not rely primarily on the output. That means the recommendation and its supporting inputs and logic have to be transparent, timely enough to be checked, and framed as a suggestion rather than a directive. Equally important, the claims you make for the software must match. The FDA reads intended use partly from labeling and promotion, so describing your product as diagnosing, scoring, or driving a clinical decision can pull it inside the device definition regardless of what the code does. Keeping software non-device is therefore a discipline of both design and description: build for independent review, and claim only what keeps you on the non-device side. Where you want an integrity signal that carries no device risk at all, an attestation layer that proves provenance without offering any recommendation sits cleanly outside the four criteria. This is general information, not legal advice, and specific classification should always be confirmed with regulatory counsel.
How do you meet the fourth CDS criterion (independent review)?
You make the basis of every recommendation visible, checkable, and non-coercive, so a clinician can reach the same conclusion on their own instead of deferring to the software. The fourth criterion is the reliance test: the professional must be able to independently review the basis for a recommendation and must not rely primarily on it to make a clinical decision. That single sentence is what separates most non-device CDS from regulated devices. To meet it, surface the inputs and the reasoning behind each recommendation in plain, inspectable form — the data considered, the sources, and the logic — not just a conclusion or a bare number. Avoid the patterns that quietly defeat independent review: opaque scores a clinician cannot interrogate, time-critical alarms that must be acted on before anyone could review anything, autonomous actions that execute without a human in the loop, and directive language that reads as an order rather than a suggestion. Any of those means the professional would rely primarily on the software, which fails the criterion and makes the software a device. RankShieldMD approaches the fourth criterion from a different angle entirely: because it attests decisions rather than rendering them, it produces no recommendation for a clinician to over-rely on, so the reliance question never arises for the attestation layer. Whether you are building CDS that must satisfy the fourth criterion directly, or adding integrity tooling that stays outside it, the underlying principle is the same — the clinician's independent judgment has to remain in control. As always, confirm the specifics with FDA or regulatory counsel for your product.
How do you add AI integrity tooling without triggering device regulation?
You keep the tooling's intended use limited to verifying what happened — which model ran, on which inputs, at what time — and keep it out of the clinical pathway that produces the recommendation. Integrity and attestation tooling only risks becoming a device when its intended use crosses into producing, scoring, or driving a clinical decision. Verification-only tooling does none of that. It records and proves statements about a decision another system already made; it offers no recommendation, so there is nothing for a clinician to rely on, and the fourth-criterion reliance test simply does not apply. Because classification follows intended use, such tooling can sit alongside a regulated device, or alongside non-device CDS, without becoming a device itself and without entering the device's clinical pathway. RankShieldMD is engineered to hold exactly this position. It registers a cryptographic baseline of an approved model, then at decision time seals digests of the model, the inputs, and the output to an externally-anchored, post-quantum-signed transparency log — PHI-free, because raw identifiers are rejected at the guard and never enter the ledger. The output is an evidence package a buyer, an auditor, or a reviewer can verify independently. Adopting it gives clinical-AI and SaMD teams an independent integrity signal for their models without inheriting any new regulatory burden, and without RankShieldMD ever entering the clinical decision. It supports your compliance posture; it does not make you compliant, and it is not itself a clearance. Confirm the classification of any specific configuration with regulatory counsel.
What is the difference between attesting and rendering a clinical decision?
Rendering means producing the clinical decision — scoring a patient, recommending a course of action, or driving what the clinician does next. Attesting means proving something about a decision another system already made, without ever touching the clinical conclusion. The distinction is the entire basis for staying non-device, because FDA classification turns on intended use. A system that renders has an intended use of informing or driving a clinical decision, which puts the fourth-criterion reliance question squarely in play: can the clinician independently review it, or will they rely primarily on it? A system that attests has an intended use of verifying provenance and integrity — proving that a specific output truly came from a specific approved model, on intact inputs, at a stated time. It makes no recommendation and expresses no clinical opinion, so there is nothing for a clinician to rely on and the reliance test never arises. Think of it as the difference between a doctor writing the prescription and a notary certifying that the prescription is genuine and unaltered — the notary never practices medicine. This is why RankShieldMD is non-device by construction. It attests; it never renders, scores, or recommends. Clinical judgment stays with the clinician and the clinical system, while RankShieldMD holds only the cryptographic proof about the decision. That boundary is deliberate, and we hold to it honestly: it keeps the technology safe from encouraging over-reliance, and it keeps its intended use clearly outside the device definition. None of this is a substitute for a regulatory determination of your own product.
What we are careful never to claim.
This is not legal advice
The CDS criteria here are general educational information, not legal or regulatory advice. Device classification is fact-specific and depends on intended use. Consult FDA or qualified regulatory counsel for the classification of a specific product.
It supports compliance
RankShieldMD produces verifiable integrity evidence that supports FDA submissions and postmarket monitoring. It does not make an organization compliant, and it is not itself an FDA clearance.
It attests, never renders
It proves a decision was genuine; it never produces, scores, or recommends one. It is PHI-free by construction — digests only, with raw identifiers rejected at the guard before anything is sealed.
Ask RankShieldMD about the FDA CDS non-device line.
What are the FDA non-device CDS criteria?
They come from section 520(o)(1)(E) of the FD&C Act, refined by the FDA’s Clinical Decision Support Software final guidance. Software is non-device CDS only if it meets all four: it does not acquire, process, or analyze a medical image or signal; it displays or analyzes medical information; it offers recommendations to a health care professional; and it enables that professional to independently review the basis for the recommendation rather than relying primarily on it.
Where do the CDS criteria actually come from?
They are statutory. The 21st Century Cures Act amended the FD&C Act in 2016 to exclude certain CDS software from the device definition at section 520(o)(1)(E). The FDA then issued its Clinical Decision Support Software guidance to explain how it interprets each of the four prongs. The statute sets the boundary; the guidance shows how the FDA reads it.
Does meeting three of the four criteria make software non-device?
No. All four criteria must be satisfied at once. Miss any single one — including the fourth independent-review prong — and the software falls back inside the device definition and is regulated accordingly. The criteria are cumulative, not a menu.
What is the fourth CDS criterion?
The fourth criterion requires the software to enable the health care professional to independently review the basis for its recommendation, so the professional does not rely primarily on the software to make a clinical decision. In practice the recommendation and its underlying inputs and logic must be transparent enough that a clinician can reach the same conclusion on their own.
Why is the fourth criterion the one that decides device status?
Because it captures the real regulatory concern: whether a clinician can second-guess the software or is effectively led by it. The first three prongs are relatively mechanical. The fourth is about reliance. Software that a clinician cannot independently check — a time-critical alarm, an opaque score, or an autonomous action — fails it and becomes a device.
What makes software fail the independent-review test?
Opacity and time pressure. If the recommendation arrives without its basis, is framed as a directive, drives a time-critical decision, or is scored by a model the clinician cannot inspect, the professional cannot independently review it and would rely primarily on the output. That fails the fourth criterion and pulls the software into device territory.
What is the difference between attesting and rendering a clinical decision?
Rendering means producing, scoring, or recommending the decision itself — the software influences what the clinician does. Attesting means proving something about a decision another system already made: that a specific output truly came from a specific approved model, on intact inputs, at a stated time. Attestation never touches the clinical conclusion.
Why does attesting rather than rendering keep software non-device?
Because FDA classification turns on intended use. Attestation’s intended use is to verify provenance and integrity, not to inform or drive a clinical decision. It offers no recommendation for a clinician to rely on, so the reliance question at the heart of the fourth criterion never arises. The intended use sits outside the device definition.
Does RankShieldMD render, score, or recommend a decision?
No. It attests. It seals a digest of the model, the inputs, and the output to a verifiable log so anyone can later prove the decision was genuine. It never produces the decision, scores a patient, or recommends a course of action — that judgment stays with the clinician and the clinical system.
Is integrity or attestation tooling a medical device?
Not when its intended use is to verify provenance rather than inform a clinical decision. Tooling that proves a decision came from an approved, un-tampered model — without producing, scoring, or driving that decision — sits outside the device definition. RankShieldMD is built to stay squarely on that side of the line.
Can I add AI integrity tooling without triggering device regulation?
Yes, if the tooling attests rather than renders. Keep its role to proving what happened — which model ran, on what inputs, when — and out of the clinical pathway that produces the recommendation. Because classification follows intended use, verification-only tooling can sit alongside a device without becoming one.
Does software architecture determine FDA device status?
No. Classification turns on intended use, not architecture. The same code can be device or non-device depending on the claims made for it and how it is meant to be used. This is why the same attestation engine stays non-device: its intended use is to verify integrity, never to render a clinical decision.
How do marketing claims affect CDS device classification?
Heavily. The FDA reads intended use partly from labeling and promotional claims. Describe software as making, scoring, or driving a clinical decision and you can pull it into the device definition regardless of what the code does. RankShieldMD is careful to claim only attestation of provenance and integrity — never diagnosis, scoring, or recommendation.
Is this page legal or regulatory advice?
No. This is general educational information about the FDA CDS criteria, not legal or regulatory advice. Device classification is fact-specific and depends on your product’s intended use. Consult FDA or qualified regulatory counsel for the classification of any specific product before you rely on any position.
Does RankShieldMD make our software FDA compliant?
No software makes an organization compliant. RankShieldMD supports compliance by producing verifiable integrity evidence your submissions and postmarket monitoring can rely on. Classification, clearance, and compliance remain your organization’s responsibility, established with your regulatory team and counsel.
Does the attestation layer ever see patient data?
No. It is PHI-free by construction. It seals digests of the model, inputs, and output; raw patient identifiers are rejected at the guard and never enter the log. Adopting it shrinks your PHI footprint rather than expanding it, and it stays outside the clinical decision entirely.
Add an integrity signal that stays on the non-device side.
Attest a decision, hand your buyer proof they can verify, and keep clinical judgment where it belongs.