Compliance Guide | Healthcare

    Section 1557 Requirements for AI Patient Care Decision Support Tools

    Under 45 CFR 92.210, as of November 2026, a covered health program or activity may not discriminate on the basis of race, color, national origin, sex, age, or disability through patient care decision support tools. It must make reasonable efforts to identify tools whose input variables measure those characteristics, and to mitigate the risk of discrimination for each tool identified. Identification and mitigation judgment sit outside any runtime tool. Runtime policy enforcement and retained evidence can show how a tool was constrained and what it did in production.

    Questions to ask about enforcement and evidence

    • Can you list every patient care decision support tool in use, automated or not, with the input variables each uses, including any measuring race, color, national origin, sex, age, or disability?
    • For each tool flagged under paragraph (b), which mitigation measures are documented, who approved them, and on what basis?
    • How do you know each measure is applied every time the tool runs in production, not only when it was reviewed?
    • Which duties depend on bias or proxy evaluation that runtime controls do not perform, and who owns them?
    • Can you show, per tool and per interaction, what the tool was permitted to do and what it did, in records retained for audit?
    • How are tool updates, new input variables and new deployments detected and returned to identification and mitigation review?

    What the rule requires

    Section 92.210 has three operative parts. Paragraph (a) prohibits a covered entity from discriminating on the basis of race, color, national origin, sex, age, or disability in its health programs or activities through the use of patient care decision support tools. Paragraph (b) creates an ongoing duty to make reasonable efforts to identify uses of those tools that employ input variables or factors measuring any of those characteristics. Paragraph (c) requires the entity, for each tool identified under (b), to make reasonable efforts to mitigate the risk of discrimination resulting from its use.

    The compliance table in Part 92 lists paragraphs (b) and (c) as applying within 300 days of July 5, 2024. Paragraph (a) is the general prohibition. The codified text uses the phrase "reasonable efforts," and that is the wording to work from when documenting your program. In the final rule, OCR replaced the term "clinical algorithms" with "patient care decision support tools" and further specified the provision's scope. HHS has described the rule as confirming that nondiscrimination obligations continue to apply to AI, clinical algorithms, predictive analytics and other tools.

    Scope: the definition is broader than AI

    The regulation defines a patient care decision support tool as any automated or non-automated tool, mechanism, method, technology, or combination thereof used by a covered entity to support clinical decision-making in its health programs or activities. An inventory limited to machine learning models can miss in-scope tools, such as scoring rules, embedded vendor features or manual methods that feed clinical decisions.

    The identification duty is triggered by input variables or factors, so an inventory needs to record what each tool takes in, not only its name. HHS OCR disability guidance also notes that tools informing resource allocation or cost-containment decisions can raise Section 1557 and Section 504 concerns, so those belong in scope reviews. A Congressional Research Service report cautions that the rule may not reach every AI use in patient care or every entity that develops health care AI. Document your own coverage analysis instead of assuming the rule applies or does not.

    Because the identification duty is ongoing, the entity should set and record its own review cadence and triggers, such as a new tool, a new input variable or a tool update. Entities with fifteen or more employees must also designate a Section 1557 Coordinator, who is a natural owner for this process alongside clinical informatics, compliance and IT.

    Which obligations depend on evaluation and which on runtime behavior

    The rule does not prescribe a technical mechanism. Mapping each duty to where it is performed keeps responsibilities clear and prevents a runtime control from being credited with work it cannot do.

    • Tool inventory and input-variable analysis (b): Organizational process. Needs inventory ownership and a re-review cadence. Runtime records can help show which tools are actually invoked, but they do not replace the analysis.
    • Judging whether outputs discriminate (c): Pre-deployment and ongoing evaluation, such as bias or proxy testing. This sits outside a runtime policy layer and needs a named owner.
    • Choosing mitigation measures (c): Documented judgment by compliance and clinical leadership, with the basis for the choice recorded.
    • Applying a mitigation measure every time a tool runs (c): Depends on runtime behavior. Enforcement in the execution path can apply the measure consistently.
    • Showing what a tool was allowed to do and did: Depends on retained evidence tied to each interaction and tool call.
    • Tool changes and new deployments: Organizational trigger process that feeds changes back into (b) and (c) reviews.

    Where runtime enforcement and evidence fit

    Mitigation is a judgment about reducing discrimination risk. Making that judgment requires evaluation work that no runtime control performs. Once the judgment is made, however, the measure still has to hold in production. Examples include limiting which data reaches a tool, restricting what a tool may do, or requiring a clinician to review an output before it is used. If a measure is only a written procedure, an auditor can reasonably ask how you know it was followed on every run.

    That is the narrow role for runtime enforcement. Policy applied in the execution path can allow, redact, block or escalate an interaction for human approval, and each decision can be retained as evidence. The rule does not require or endorse a specific runtime mechanism. This is a design option for making documented mitigation traceable to actual behavior.

    Trussed enforces policy in the execution path across models, agents, tools and MCP servers at the level of individual interactions and tool calls, and generates audit-ready evidence from that enforcement. It runs in production with healthcare organizations and maps evidence to HIPAA, including PHI handling and audit controls. Trussed does not perform bias testing, proxy analysis or model evaluation, and it does not decide whether a tool's outputs are discriminatory. Those remain your responsibility under paragraph (c).

    A practical record keeps, per tool, the declared input variables, the intended use, the mitigation measure chosen and its rationale, and the runtime policy applied. Evidence of what happened then lines up with the documented decision.

    Tradeoffs and design decisions

    Escalation for human approval makes a review requirement enforceable and leaves a record of each approval. It also adds clinician workload, so it fits best where the mitigation rationale calls for it, not as a default on every tool. Redaction can limit PHI exposure to a tool, but removing a field does not settle whether other inputs act as proxies for a protected characteristic. That question belongs to evaluation.

    OCR's 2022 proposed rule preamble gave examples of steps entities may choose: written policies and procedures on how tool information is used in decision-making, monitoring of potential impacts, and staff training. It also said algorithms should augment, not replace, clinical judgment. These were not carried into the final rule as mandates, but they are a useful reference when explaining why your measures were reasonable.

    45 CFR 92.210 obligations at a glance

    Paragraph (a)

    No discrimination through patient care decision support tools.

    Paragraph (b)

    Ongoing reasonable efforts to identify tools with protected-characteristic input variables.

    Paragraph (c)

    Reasonable efforts to mitigate discrimination risk for each tool identified.

    Timing

    Paragraphs (b) and (c) apply within 300 days of July 5, 2024.

    Map your Section 1557 mitigation measures to runtime evidence

    Talk to our team about how policy enforcement in the execution path and audit-ready evidence can support documented mitigation for AI tools in your health programs.

    Request a Demo