Brief
Diagnostic AI and treatment-decision AI look like they belong to the same category of "medical AI," but they are structurally different problems wearing the same label. A diagnostic model — the kind that reads a retinal photograph, a chest X-ray, or an ECG strip — is solving a closed-set pattern-classification task: given an input, assign it to one of a finite, pre-defined set of labels (disease present/absent, or a specific finding category). The labels come from a fixed taxonomy that already exists in medicine (ICD codes, radiological classification scales, pathology grades), and the ground truth for training and evaluation comes from an established reference standard — expert graders, biopsy results, or a reading-center consensus. This is precisely the shape of problem at which modern deep learning excels: it is math-legible, has a target loss function, and can be validated against a held-out labeled set. That is why the first-ever FDA-authorized autonomous AI diagnostic system, IDx-DR, cleared through the De Novo pathway in April 2018, could be validated in a 900-patient pivotal trial against a Wisconsin Fundus Photograph Reading Center grading standard and post a sensitivity of 87.2% and specificity of 90.7% against that reference — a single, checkable number against a single, agreed-upon truth.
Treatment decisions do not have that structure. There is no fixed finite label set for "the right treatment," because the right treatment for a given diagnosis varies by comorbidities, prior treatment response, organ function, drug interactions, the patient's stated priorities (extend survival at the cost of side effects, or preserve quality of life), cost and insurance coverage, local formulary and staffing constraints, and how the treatment interacts with sequencing decisions the clinician has already made. Two oncologists looking at the same biopsy and staging data can reasonably prescribe different regimens and both be practicing good medicine, because the choice is a multi-objective optimization over variables that don't collapse into one number. There is no reading-center consensus for "which second-line regimen for a 71-year-old with mild renal impairment who prioritizes staying out of the hospital." This is the structural reason IBM's Watson for Oncology program — a system explicitly built to generate cancer treatment recommendations — ran into trouble: the University of Texas MD Anderson Cancer Center halted its parallel "Oncology Expert Advisor" collaboration with IBM in 2016 after the project had consumed roughly $62 million without producing a clinically usable system, and subsequent reporting in 2017–2018 documented cases where Watson for Oncology's recommendations diverged from expert consensus. The system had been trained substantially on the treatment preferences of a small number of specialist oncologists at partner institutions rather than on a representative, validated ground truth, which is exactly the failure mode you'd predict once you recognize treatment has no single label to learn.
The regulatory system has, largely by trial and error, encoded this structural distinction into law. Under the FDA's clinical decision support (CDS) framework, updated in final guidance issued January 2026, software that merely assists a clinician's judgment — providing information the clinician independently reviews rather than relying on primarily — can qualify as "non-device CDS" and fall outside direct FDA device regulation, provided it meets four statutory criteria including that it gives the clinician enough basis for a recommendation that they are not primarily relying on the AI's output alone. But software that makes an autonomous clinical decision — a diagnosis rendered without a clinician independently interpreting the image or result — falls squarely under FDA device regulation, and no autonomous AI prescribing or treatment-decision system had been cleared by the FDA as of the search underlying this analysis. IDx-DR's own regulatory history illustrates the line: the FDA required a full De Novo marketing authorization for it precisely because it was an autonomous diagnostic tool not requiring clinician interpretation, and its label is narrowly scoped to a single binary determination (referable vs. non-referable diabetic retinopathy) in adults not previously diagnosed with the condition — a scope so constrained that it sidesteps almost every source of treatment-decision complexity.
The handoff from diagnosis to treatment is where the two systems' feedback loops diverge most sharply, and where failure risk concentrates. A diagnostic model's feedback loop is short and empirical: a false negative or false positive on a retinal scan or a mammogram is discoverable relatively quickly through follow-up imaging, biopsy, or clinical deterioration, and that discrepancy can be fed back into a labeled dataset to retrain or recalibrate the model. A treatment-decision system's feedback loop is long, noisy, and confounded: whether a chosen regimen was "right" may not be knowable for months or years, is entangled with the patient's adherence, comorbid events, and subsequent treatment changes, and even a bad outcome doesn't prove the decision was wrong (a low-probability adverse event can follow a statistically sound choice). This is a case example of a broader distinction in the clinical-evidence hierarchy: diagnosis works against a surrogate-like, immediately verifiable output, while treatment efficacy is judged against outcomes that unfold over a much longer and more confounded horizon — closer to how Phase II surrogate endpoints (tumor shrinkage, biomarker change) often fail to predict the Phase III clinical outcome they're meant to proxy for. Failures compound at the handoff itself: an AI system correctly diagnosing a condition but embedded in a workflow that auto-suggests a single "most likely" treatment path risks laundering the appearance of clinical certainty from the diagnostic step into the treatment step, where no comparable certainty exists — precisely the ambiguity the FDA's 2026 CDS guidance is trying to police by scrutinizing whether a tool's output format lets a clinician exercise independent judgment or subtly forecloses it.
Components (7)
Closed-set label taxonomy (diagnosis side)
Provides the finite, pre-agreed set of categories (disease present/absent, severity grade) that a diagnostic model is trained and scored against, making the problem mathematically tractable.
Reference standard / ground truth grader
Supplies the objective label used both to train the model and to measure its accuracy — for IDx-DR this was Wisconsin Fundus Photograph Reading Center certified grading, the benchmark the pivotal trial measured against.
Autonomous diagnostic AI system
Renders a diagnostic determination without requiring a clinician to independently interpret the underlying image or signal, which is the feature that triggers full FDA device regulation rather than the CDS exemption.
Multi-variable treatment-optimization space
Represents the open-ended set of comorbidities, prior therapies, drug interactions, and sequencing dependencies that any treatment recommendation must account for, with no single correct combination.
Patient values and preference input
Introduces a subjective, non-standardizable variable (risk tolerance, quality-of-life priority) into the treatment decision that has no analog in a diagnostic classification task.
Clinical decision support (CDS) regulatory boundary
The FDA's four-criteria test (does not process images/signals directly, and gives the clinician enough basis that they are not primarily relying on the output) determines whether a treatment-support tool is exempt from device regulation or must be cleared as a medical device.
Training-data provenance and representativeness
Determines whether a treatment-recommendation model reflects a validated population-level standard of care or merely the idiosyncratic preferences of the small clinician group whose cases were used to train it, as occurred in the Watson for Oncology program.
How It Works (9 steps)
1Diagnostic model trained against fixed label set
A machine-learning model is trained on a large corpus of labeled examples (images, signals, or structured data) where each example maps to one of a finite set of pre-defined diagnostic categories, using a reference-standard grading process to assign the labels.
AI developerreference-standard reading center or pathology lab
Why this step: Without a finite label set and an agreed ground truth, there is no objective function to train against or metric to validate performance.
2Prospective validation against reference standard
The trained model is tested in a preregistered clinical trial against the same reference standard used for training, generating sensitivity/specificity/PPV/NPV metrics that can be independently audited.
SponsorFDA reviewersclinical trial sites
Why this step: Diagnostic accuracy claims require prospective, preregistered validation because retrospective or non-blinded testing can inflate performance estimates.
3FDA classifies device vs. non-device function
The FDA determines whether the software's diagnostic or treatment-related output falls under device regulation by applying Non-Device CDS criteria — chiefly whether the software processes images/signals directly and whether it gives the clinician enough basis to avoid primary reliance on the output.
FDA Center for Devices and Radiological Health (CDRH)sponsor
Why this step: This classification determines whether the tool needs premarket clearance (510(k)/De Novo/PMA) or can market as clinical software without device authorization.
4Autonomous diagnostic tools receive device authorization
Tools that render a diagnostic determination without requiring clinician interpretation of the underlying image are treated as regulated devices and go through 510(k), De Novo, or PMA review with a narrowly scoped indication for use.
FDAdevice sponsor
Why this step: Autonomous determination removes the clinician's independent-judgment safeguard, so the FDA substitutes premarket evidentiary review for that safeguard.
5Treatment-recommendation systems face the open-set problem
A treatment-support system attempts to map a diagnosis plus patient context onto a recommended course of action, but the output space is not a finite pre-agreed taxonomy — it must weigh comorbidities, prior treatment history, and patient preference simultaneously, with no single correct answer to train or score against.
AI developertreating clinicianpatient
Why this step: Without a closed label set or single ground truth, the system cannot be validated the same way a diagnostic classifier is, and its recommendations are inherently context-dependent judgments rather than checkable classifications.
6Training data reflects local practice, not universal ground truth
Treatment-recommendation models are typically trained on the documented preferences and case histories of a limited set of clinicians or institutions, which the model then reproduces as if it were an objective standard of care.
AI developerpartner hospital's specialist clinicians
Why this step: Absent a universal ground truth for 'correct treatment,' the model's outputs are only as generalizable as the population of cases and clinicians it was trained on, which can silently narrow or bias its recommendations when deployed elsewhere.
7Handoff point: diagnostic certainty projected onto treatment ambiguity
A clinician or workflow may treat a validated diagnostic AI's confidence score as if it extends to the treatment recommendation that follows, even though the treatment step carries a fundamentally different and unquantified uncertainty.
treating cliniciancare workflow/EHR system
Why this step: This is the structural failure point: the rigor of the diagnostic step does not transfer to the treatment step, and conflating them risks overconfidence in a recommendation that has not been validated the same way.
8Clinician exercises independent judgment (or fails to)
For non-device CDS to remain outside FDA regulation, the clinician must not rely primarily on the AI's treatment output — they must independently evaluate the basis for the recommendation against the patient's full context.
treating clinician
Why this step: This is the regulatory and clinical safeguard substituting for the lack of a closed-set ground truth in treatment decisions; removing it (automation bias, time pressure) is where documented treatment-AI failures have occurred.
9Long, confounded outcome feedback loop
Whether a treatment recommendation was appropriate may only become apparent months or years later, entangled with adherence, intervening comorbid events, and subsequent treatment changes, making it far harder to feed outcomes back into model retraining than for a diagnostic model.
treating clinicianpatientAI developer (post-market surveillance)
Why this step: Fast, clean feedback loops are what allow diagnostic models to be iteratively validated and recalibrated; treatment models lack this, which limits how quickly errors can be detected and corrected.
What Makes It Work
Closed-set classification vs. open-set optimization
Diagnosis maps an input to one of a finite number of pre-agreed labels with a known ground truth, which is the exact problem shape deep learning is optimized to solve; treatment requires weighing an open, patient-specific set of variables against no single correct answer, which is a fundamentally different computational and epistemic problem.
Reference-standard grounding
Diagnostic AI can be trained and audited against an external, agreed-upon grading standard (reading centers, biopsy, established scales); treatment decisions have no equivalent independent arbiter of correctness, since appropriate treatment varies legitimately by patient values and context.
Regulatory reliance-on-output test
The FDA's non-device CDS criteria hinge on whether a clinician can independently evaluate the basis for a recommendation rather than relying on it primarily; this test operationalizes the diagnosis/treatment distinction into a legal boundary determining which tools face premarket device review.
Feedback-loop latency and confounding
Diagnostic errors surface relatively quickly and cleanly through follow-up testing, enabling rapid model correction; treatment-decision errors surface slowly and are confounded by adherence and subsequent care changes, which slows and weakens the corrective feedback available to improve the model.
Where It Breaks (4)
Treatment-recommendation model trained on narrow, non-representative clinician preferences
Consequence: The model reproduces the idiosyncratic practice patterns of a small training population as if they were a validated universal standard, producing recommendations that diverge from broader clinical consensus when deployed elsewhere — the pattern documented in the Watson for Oncology program, where the parallel MD Anderson collaboration was halted in 2016 after consuming roughly $62 million without a clinically usable system.
Safeguard: Prospective validation against diverse, external patient populations and independent clinical review before deployment; the FDA's Non-Device CDS Criterion 4 requiring sufficient explanatory basis for clinician review.
Automation bias at the diagnosis-to-treatment handoff
Consequence: Clinicians extend trust earned by a validated diagnostic tool's accuracy to a treatment recommendation generated by the same or an adjacent system, even though the treatment step has not been validated with the same rigor, risking under-scrutinized treatment choices.
Safeguard: Regulatory requirement that non-device CDS tools not be primarily relied upon, and interface design that clearly separates diagnostic confidence from treatment-recommendation uncertainty.
Misclassifying an autonomous treatment-decision tool as non-device CDS
Consequence: A tool that in practice forecloses independent clinical judgment (e.g., presenting a single confident output without adequate basis) could operate without the premarket safety review that autonomous decision-making would otherwise require, exposing patients to unvalidated treatment logic.
Safeguard: The FDA's January 2026 revised CDS guidance sharpening the four-criteria device/non-device boundary, though this boundary was described by legal commentary as still requiring case-by-case mapping of each software function against the criteria.
Slow, confounded feedback loops delaying error detection in treatment models
Consequence: Because treatment outcomes unfold over months or years and are confounded by adherence and intervening events, a systematically flawed treatment-recommendation model can remain in use far longer before poor performance is detected, compared to a diagnostic model whose errors surface quickly.
Safeguard: Structured post-market surveillance and outcome registries, though these remain less standardized for treatment-decision AI than for diagnostic AI performance tracking.
Facts & Figures (6)
The claims behind this analysis, each with its verification status — including what is contested, unverified, or could not be established.
By the end of 2025, the FDA's cumulative tracker listed 1,451 authorized AI/ML medical devices, with 295 new authorizations in 2025 alone and 62% of the 2025 cohort classified as diagnostic in nature.
The overwhelming device-authorization volume in diagnostic/imaging AI versus near-zero autonomous treatment-decision clearances is direct evidence of the structural asymmetry this explainer describes.
✓ GROUNDED
The FDA's first-ever autonomous AI diagnostic authorization, IDx-DR, cleared via De Novo in April 2018, was validated in a 900-patient pivotal trial against Wisconsin Fundus Photograph Reading Center grading, achieving sensitivity of 87.2% and specificity of 90.7%.
This establishes the concrete template for how diagnostic AI is validated against a single, objective reference standard — the mechanism underlying the diagnosis side of this explainer's core distinction.
✓ GROUNDED
The University of Texas MD Anderson Cancer Center halted its 'Oncology Expert Advisor' collaboration with IBM Watson in 2016, after the project had consumed approximately $62 million without producing a clinically usable system.
This is the canonical documented failure case illustrating why open-ended treatment recommendation is a fundamentally harder AI problem than closed-set diagnosis, grounding the treatment-side failure modes.
✓ GROUNDED
The FDA issued a revised final Clinical Decision Support Software guidance dated January 29, 2026 (superseding an interim version issued January 6, 2026), which sharpens the four statutory criteria distinguishing non-device CDS from regulated device CDS under Section 520(o)(1)(E) of the FD&C Act.
This is the current, operative regulatory mechanism that formally encodes the diagnosis/treatment distinction into device-classification law, directly anchoring the CDS regulatory-boundary component and step.
✓ GROUNDED
The FDA has explicitly stated that IDx-DR, as an autonomous AI tool guiding retinopathy care without a clinician's interpretation, was not exempt under the 21st Century Cures Act's software exclusion and required full marketing authorization.
This shows the specific regulatory trigger — autonomy without clinician interpretation — that forces device-level review, which is the operative test separating diagnostic autonomy from treatment-support software.
✓ GROUNDED
As of the sources reviewed, no autonomous AI prescribing or treatment-decision service has been cleared by the FDA.
This confirms that the structural difficulty of treatment decisions has translated into an empirical regulatory gap, directly supporting the claim that treatment-decision AI lags diagnostic AI in real-world deployment.
✓ GROUNDED