Implementation Guide | Healthcare

    Logging Section 1557 discrimination risk mitigation for clinical AI as audit evidence

    Section 1557 clinical AI audit evidence is a record, kept per interaction, of what policy applied to a clinical AI tool and what the tool was permitted to do: allow, redact, block, or escalate for human approval. Capture those decisions with the policy version, the reason and any human outcome. Store them as protected evidence under HIPAA 164.312(b), with retention and residency set deliberately. Policies and committee minutes show intent. This record shows mitigation operated.

    Assemble an audit-ready evidence set

    1. 1

      1. Fix the tool inventory

      List tools meeting the § 92.210 definition and mark which run through an enforced execution path.

    2. 2

      2. Map evidence to obligations

      For each record type, document the 164.312(b) requirement and the mitigation activity it supports.

    3. 3

      3. Set retention and residency

      Decide retention period and storage location with counsel, and record the decision in policy.

    4. 4

      4. Assign ownership

      Name owners for log review, evidence store access and responses to audit or OCR requests.

    5. 5

      5. Test retrieval

      Pick an interaction at random and reconstruct the policy, version, decision and human outcome.

    6. 6

      6. Package separately

      Present runtime evidence alongside, not in place of, testing, bias assessment and governance records.

    What the rule covers and why documentation alone falls short

    The Section 1557 final rule, published in May 2024, prohibits discrimination on the basis of race, color, national origin, sex, age, or disability in certain health programs and activities. Within it, 45 CFR 92.210 addresses patient care decision support tools. HHS OCR replaced the earlier term "clinical algorithms" with that phrase and defined it 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. The definition reaches tools that affect the care patients receive. The compliance date for § 92.210(b) and (c) is within 300 days of the rule's effective date. As of November 2026, no amendment, delay or repeal of § 92.210 is reported.

    Most covered entities can produce a policy, a committee charter or a pre-deployment review. Those artifacts show that someone considered discrimination risk. They do not show what a tool was allowed to do in a specific interaction, or that a mitigation step actually fired. That gap is the subject of this guide. The design below is editorial implementation guidance for generating operational evidence. It is not a statement of what OCR will accept.

    Start with scope. Inventory every tool that meets the § 92.210 definition, including non-automated ones. Only the subset that is AI-driven and handles patient data in an execution path is a candidate for runtime evidence. The rest needs evidence from other controls.

    Categories of evidence and where each comes from

    Treat Section 1557 mitigation evidence as four distinct categories, because they come from different controls and answer different questions.

    The first is governance evidence: policies, ownership, committee decisions. The second is pre-deployment evidence: testing and model-level bias assessment. The third is runtime enforcement evidence: records of policy decisions applied to individual interactions and tool calls. The fourth is human oversight evidence: who reviewed an escalated action and what they decided.

    Runtime enforcement logs cover the third and part of the fourth. They do not substitute for the first two. A record that a tool call was blocked says nothing about whether the underlying model is biased, and it should not be presented as if it does. Keeping the categories separate makes the evidence set easier to defend, because each artifact is claimed only for what it can show.

    The runtime category is the one most organizations lack. Application logs and system event logs typically record that something happened. They rarely record which policy governed the action, which version of it, and why the outcome was allow rather than block.

    Specify the events and fields to record

    Record a decision at the point where the policy is applied, so each record reflects what the AI was permitted to do on that interaction. Capturing it downstream, from application output, risks recording what happened without the rule that governed it. The field list below is editorial design guidance. Adapt it to your tool inventory and counsel's review.

    Capture four decision types: allow, redact (for example, protected attributes or PHI removed from an input or output), block, and escalate for human approval. Each record should carry the same core fields so a reviewer can query across them. Store a reference to the interaction rather than raw PHI wherever the reference is sufficient to reconstruct the event from the source system.

    • Time and identity: Timestamp, tool name and version, and the system or user context that initiated the interaction.
    • Policy reference: Policy identifier and version in force at that moment, so the decision traces to a specific rule.
    • Decision and reason code: Allow, redact, block or escalate, with a structured reason rather than free text.
    • Human review: Reviewer identity, time of review and outcome for escalated items.
    • Interaction reference: A pointer to the interaction, not a copy of the PHI it contained.
    • Logging status: Changes to logging configuration, including who changed it and when.

    Apply HIPAA audit controls and PHI handling to the evidence

    HIPAA 164.312(b) requires hardware, software, and/or procedural mechanisms that record and examine activity in information systems that contain or use electronic protected health information. HHS lists audit controls as a required specification, not an addressable one. The standard is technology-neutral, so the question is not which product but whether the mechanism records and permits examination of activity. An AI system that handles ePHI falls within its scope.

    Two consequences follow. First, the evidence store may itself contain or reference ePHI, so treat it as an information system under 164.312(b) with its own access controls and its own audit trail. Record who reads, exports or alters evidence, and restrict who can disable logging. Second, minimize what each record holds. The interaction-reference approach above reduces exposure while keeping the record useful.

    HHS's January 2025 Security Rule proposal describes audit trails, event logs and access reports as records of activity and notes that regular review helps detect unusual patterns. As of November 2026 that proposal should be treated as proposed, not binding. Its purpose is also security-oriented, so it does not define Section 1557 events. You need to add those event definitions yourself.

    NIST SP 800-66r2 prompts several decisions: where audit information will reside, for example on a separate server, what audit trail data and monitoring procedures are needed, how often to review logs based on risk assessment and risk management, and whether an audit trail exists for all accesses to the information. Keep evidence apart from the system it describes, and set review frequency from your own risk analysis.

    Evidence quality, scope limits and where enforcement fits

    Judge your evidence against four criteria. Completeness: every in-scope interaction produces a record, including allowed ones, since absence of a block is only meaningful if allows are recorded too. Immutability: records cannot be altered or silently removed, and changes to logging are themselves recorded. Traceability: each decision links to the policy version in force. Retrievability: a reviewer can find the record for a given interaction without reconstructing it from several systems.

    Runtime policy enforcement in the execution path is one way to meet these criteria, because the decision and the record come from the same event. Trussed enforces policy in the execution path across models, agents, tools and MCP servers, applying allow, redact, block, rate limit or escalate for human approval to individual interactions and tool calls. That enforcement generates audit-ready evidence mapped to HIPAA. It can be deployed self-managed in the customer cloud, with configurable retention and data residency. Trussed runs in production with healthcare organizations.

    The boundaries matter as much. Runtime evidence does not cover model lineage, dataset provenance or end-to-end tracing of agentic workflows. It does not replace pre-deployment testing, model-level bias assessment or governance documentation. Those controls remain necessary, and each needs its own evidence. A defensible Section 1557 file shows all of them, with each claimed only for what it demonstrates.

    From mitigation policy to audit-ready evidence

    Policy

    A versioned rule for how a clinical AI tool may handle inputs, outputs and tool calls.

    Runtime decision

    Allow, redact, block or escalate, recorded for each interaction.

    Human escalation

    Reviewer identity and outcome attached to the same record.

    Protected evidence store

    Access-controlled, retained and located according to documented parameters.

    Review your clinical AI evidence design

    Walk through which interactions, decisions and human escalations your current controls evidence, and where runtime enforcement in the execution path can close the gap.

    Talk to an Expert