Atman Health Inc. (Rahul Deo)

Atman Health, Inc. comment on FDA-2026-N-7874. Full comment attached.

Atman Health develops ATMAN-HF, prescription software that manages adults with heart failure between clinic visits. It titrates guideline-directed medical therapy, triages exacerbations and adjusts diuretics without per-decision clinician approval. On the two-axis framework it is an action-taking function with severe consequences of error. No language model in ATMAN-HF makes or changes a clinical decision. Every decision is computed by a deterministic function from the structured record and a versioned knowledge base built from FDA-approved labeling and heart failure guidelines. Language models hold the conversation and turn speech, chart notes and prescription instructions into structured data, checked before use. For hybrid devices like this, evidence belongs where errors arise.

Q1. We support the two-axis framework and propose one more dimension: whether the generative component selects the action or supplies an input to a deterministic function that does. Risk stays where the action places it. The failure becomes a wrong input, such as a misheard dose, and controls can target it. We support traceability as a dimension: every ATMAN-HF decision records its data and the labeling or guideline rule behind it.

Q5. Drift can be prevented by construction. The action set is enumerated, the engine composes clinical sentences that are spoken verbatim, and the model writes only non-clinical lines under a permission set that forbids naming any medication, dose or clinical fact. Element S.2 (scope maintenance) tests that this holds.

Q6. Estimate under-escalation and over-escalation separately, each with its own confidence bound, and judge each against the severity of its harms under ISO 14971. A missed exacerbation can mean hospitalization or death. An unneeded escalation means emergency visits and lost trust. A pooled accuracy figure hides the direction that matters.

Q7, Q9. Evaluate the final device, with evidence organized by component. The deterministic decision layer is verified at every decision boundary against labeling and guidelines. Each generative step is measured against its own source: captured answers against call audio, chart extraction against the source passage, prescription normalization against the original instruction. Acceptance criteria follow from the hazardous situation each error creates. End-to-end clinical confirmation completes the evidence.

Q11, Q14. Output adjudication cannot see input errors: reviewers of the record the device used cannot tell that a dose was misheard, so confirmation should check each generative step against its primary source. Device-alone performance can be measured inside a supervised study. The device’s order is recorded before the clinician acts and adjudicated independently, and the device’s unmodified path is replayed where the clinician changed an order. The reference standard is conformance to guidelines and labeling, judged by independent clinicians.

Q15. The comparator is care without the device. In CHAMP-HF, fewer than 1% of patients reached target doses of all guideline classes over 12 months. Remote navigator titration with an encoded algorithm raised guideline therapy within 3 months while usual care did not change (Desai, MacRae et al., JAMA Cardiology 2020).

Q13. Ground synthetic personas in real patient records, run them on a different model family from the device, and compare their speech with real transcripts. Capture accuracy should rest on real patient speech.

Q18. Greater postmarket reliance fits the generative steps if each meets premarket accuracy criteria and is re-benchmarked in the field. The deterministic layer can be verified completely before marketing.

Q20. Validate a supervisory agent by seeding known errors of each class into its stream and reporting detection rate and time by class. It should share no models or vendors with the device it supervises.

Q22. Knowledge base updates driven by labeling or guidelines within the indication, and model replacement within a fixed input and output contract, suit a PCCP with re-benchmarking. A change that widens what a model can output, or adds a drug class or selection logic, calls for a new submission.

Q24. Pin model versions, run sentinel benchmarks in production to catch silent vendor changes, require contractual notice of change, and re-benchmark any replacement before release.

Rahul Deo, MD, PhD, Chief Medical Officer and Chief Product Officer, Atman Health, Inc.

Comment of Atman Health, Inc. on Docket FDA-2026-N-7874
To: Dockets Management Staff, Food and Drug Administration
Docket: FDA-2026-N-7874, Considerations for the Regulation of Generative AI-Enabled Medical Devices:
Discussion Paper and Request for Feedback

From: Atman Health, Inc.
Date: October 19, 2026

Atman Health and ATMAN-HF
Atman Health develops ATMAN-HF, prescription software that manages adults with heart failure
between clinic visits. It titrates guideline-directed medical therapy, detects and triages heart failure
exacerbations, and initiates and adjusts diuretic therapy. It orders the laboratory studies these functions
require and acts without per-decision clinician approval. On the discussion paper's two-axis framework,
ATMAN-HF is an action-taking function with severe consequences of error: it prescribes medications
that lower blood pressure, slow the heart and raise potassium.

No language model in ATMAN-HF makes or changes a clinical decision. Every decision is computed by a
deterministic function from the patient's structured record and a versioned knowledge base built from
FDA-approved labeling and the heart failure guidelines. Language models hold the telephone
conversation and turn unstructured sources (speech, chart notes, prescription instructions) into
structured data, which is checked before the engine uses it. The deterministic engine has been in clinical
use as decision support since 2022. ATMAN-HF is being developed under the ARPA-H ADVOCATE
program.

Our comments concern high-consequence, action-taking devices in which generative models supply the
inputs and deterministic logic makes the decision. The discussion paper addresses generative models
that produce the clinical output. For hybrid devices, the evidence belongs at the points where their
errors arise.

Risk assessment
Question 1. Decision authority of the generative component. We support the two-axis framework and
propose one further dimension: whether the generative component selects the action or supplies an
input to a deterministic function that selects it. The device's risk stays where its actions and their
consequences place it. The failure mode changes: a generative error becomes a wrong input, such as a
misheard dose or a missed symptom, and controls and evidence can be built against that failure.

We also support traceability to primary sources as a dimension. In ATMAN-HF every decision records the
data it used and the labeling or guideline rule behind it, so any action can be reconstructed and
reviewed.

Question 5. Conversational drift. A conversational device can be kept from migrating toward actiondirecting output by construction. In ATMAN-HF the action set is enumerated in advance, clinical
sentences are composed by the engine and spoken verbatim, and the conversational model writes only
non-clinical lines under a fixed permission set that forbids naming any medication, dose or clinical fact.
The intended use is then fixed by the action set across every conversational trajectory, and
benchmarking element S.2 (scope maintenance and boundary adherence) tests that the constraint holds
under adversarial and multi-turn input.

Question 6. Under-escalation and over-escalation. Each direction of error should be estimated
separately, with its own confidence bound, and judged against the severity of the harms it leads to
under ISO 14971. A missed heart failure exacerbation can lead to hospitalization or death. An unneeded
escalation leads to emergency visits, burden on the care team and lost patient trust. A single pooled
accuracy figure hides the direction that matters. Where the escalation criteria are deterministic, the
generative contribution is recognizing what the patient reported, and that recognition can be measured
against the call audio.

Premarket evaluation
Questions 7 and 9. Evidence matched to architecture. We support evaluating the final device as
deployed. For devices in which generative models are confined to defined steps, the evidence can be
organized by the component responsible for each benchmarking element:

• The deterministic decision layer, verified conventionally. Its inputs are structured and bounded,
so every rule can be tested at its decision boundaries against labeling and guidelines, for
example the potassium value at which a mineralocorticoid receptor antagonist is withheld.
Elements E.1 (clinical knowledge and task fidelity) and E.2 (information gathering and clinical
analysis), and the calculations in E.3 (quantitative and measurement analysis), attach here.
• Each generative step, measured against its own source: captured answers against the call audio,
chart extraction against the source passage, and prescription normalization against the original
instruction. The acceptance criteria for each step follow from the hazardous situation its error
would create.
• The conversational layer, benchmarked for elements S.1 (safety-critical recognition and
escalation), S.2, E.4 (communication quality and user comprehension), R.1 (robustness,
reliability and reproducibility) and R.2 (subgroup performance), including accents, dialects and
lower health literacy.
• End-to-end clinical confirmation of the whole device.

Questions 11 and 14. Clinical confirmation for autonomous action.

• Input errors are invisible to output adjudication. Adjudicators who review the record the device
acted on cannot see that a dose was misheard. Clinical confirmation for these devices should
include a prespecified accuracy check of each generative step against its primary source.
• AI-alone performance can be measured inside a supervised study. The device's order is recorded
before the clinician acts, and independent clinicians adjudicate the device's decision. Where the
clinician changes an order, the device's unmodified path is replayed, so sequences of decisions
are judged as well as single decisions. Patients are protected by clinician review while the
evidence describes the device acting alone.
 • For guideline-directed treatment, the reference standard is conformance to the guidelines and
FDA-approved labeling, judged by two independent clinicians with a third resolving
disagreement. Each decision is classified as appropriate, an acceptable alternative, or an error,
and each error by the hazardous situation it would create. Agreement with the treating clinician
is recorded as a signal.

Question 15. Comparator. For a device that executes guideline titration, the comparator is the care
patients receive without it. In heart failure that care leaves most patients short of target doses: in
CHAMP-HF, fewer than 1% of patients reached target doses of all guideline medication classes at any
point over 12 months. Remote titration by navigators following an encoded algorithm increased
guideline therapy within 3 months, while usual care did not change (Desai, MacRae et al., JAMA
Cardiology 2020). Effectiveness measured against each patient's baseline therapy, or against usual care,
shows what the device adds.

Question 13. Synthetic patients. Simulated patients played by language models exercise conversational
behavior at scale. Three safeguards address the risk the paper raises. Ground each persona in a real
patient record, so its answers stay consistent with real clinical data. Run the persona on a different
model family from the device. Compare persona speech with real transcripts before relying on the
results. Capture accuracy should rest on real patient speech.

Postmarket monitoring and change control
Question 18. Premarket and postmarket evidence. Greater reliance on postmarket evidence fits the
generative steps of a high-consequence device, whose real-world inputs cannot be fully sampled before
marketing, provided each step meets its premarket accuracy criteria and is re-benchmarked in the field.
The deterministic decision layer can be verified completely before marketing, and its evidence belongs
there.

Question 20. Supervisory agents. A supervisory agent's reliability can be measured by seeding known
errors of each class, wrong inputs and wrong decisions, into the stream it monitors, and reporting its
detection rate and time to detection for each class. The independence the paper asks of expert
adjudicators applies equally to a supervisory agent: it should share no models or vendors with the
device it supervises. The ARPA-H ADVOCATE program connects each device to a supervisory agent
through a standard interface, which supports this evaluation.

Question 22. Categories of change.

• Knowledge base updates driven by new labeling or guideline content within the indication. Each
change can be verified against its source before release, and these suit a PCCP.
• Replacement of a generative model within a fixed input and output contract and a closed
vocabulary. These suit a PCCP, with release only after the step passes its re-benchmark.
• A change that widens what a generative component can output, or that adds a new drug class or
new selection logic. These call for a new submission.
Question 24. Third-party model changes. A manufacturer can pin each model version and run a sentinel
subset of each step's benchmark continuously in production, so a silent vendor change appears as a shift
in results. A contract can bind the model developer to notice of deprecation and change. Any
replacement model is re-benchmarked against the premarket criteria before release.

Contact
Rahul Deo, MD, PhD Chief Medical Officer and Chief Product Officer, Atman Health, Inc.