Clinical Expanded Mode: Collaborative Intelligence

Clinical Expanded Mode: Collaborative Intelligence

This document defines the technical architecture for Expanded Mode, where licensed physicians are integrated into the system's reasoning loop. It implements a Collaborative Intelligence (CI) model that ensures AI speed and human judgment work in harmony.

1. The "Clinical Sandwich" Workflow

We do not replace the physician; we "sandwich" them between two layers of AI-driven clinical decision support (CDS).

flowchart TD
    subgraph topLayer ["1. Pre-Visit Synthesis (AI)"]
        HS["Health State vN"] --> Intake["Automated Intake"]
        Intake --> PS["Problem Situation"]
        PS --> Popper["Popperian Assessment Engine"]
        Popper --> Draft["Draft Assessment & Plan"]
    end

    subgraph middleLayer ["2. Clinical Judgment (Human)"]
        Draft --> Review["Clinician Review Dashboard"]
        Review --> Validation["Validation / Override"]
        Validation --> Final["Authorized Care Plan"]
    end

    subgraph bottomLayer ["3. Post-Visit Execution (AI)"]
        Final --> Tracker["Adherence Tracking"]
        Tracker --> Debug["Debug Loop (Weekly)"]
        Debug --> UpdateHS["Update Health State"]
    end

    UpdateHS -.-> HS

1.1 The Top Layer (Automated Preparation)

Before the physician sees the user, the Popperian Assessment Engine (see 09-Popperian-Assessment-Engine.md) builds a comprehensive summary. The clinician receives an "Assessment Pack" containing:

  • The surviving "Hard-to-Vary" explanation.
  • Key refutation facts (why alternative theories were killed).
  • A draft clinical note (SOAP format).

1.2 The Middle Layer (The Clinical Moment)

The physician reviews the Pack. Their time is spent on High-Order Judgment, not data hunting.

1.3 The Bottom Layer (Adherence & Debug)

Once the physician authorizes a plan, the AI manages the 24/7 "Muscle" work—tracking biomarkers and alerting the clinician only if a biological bottleneck or red flag is detected.

2. Adversarial Human-In-The-Loop (AHITL)

To maintain high safety standards, the system uses an Adversarial Check when a human overrides an AI recommendation.

  • Respectful Challenge: If a clinician overrides a "Survivor Theory" with a diagnosis that the AI previously refuted based on hard data (e.g., clinician diagnoses Iron Deficiency, but the AI knows Ferritin is normal), the system surfaces a Challenge Artifact.
  • Structure: "Clinician, the system previously refuted [Diagnosis X] because of [Data Y]. Do you wish to override this refutation? Please provide clinical rationale."
  • Audit: Both the override and the rationale are logged in the versioned Health State for peer review and liability protection.

3. The Clinician Dashboard (The Glass Box)

The clinician dashboard is designed for Trust Calibration.

| Component | Design Principle | Purpose | | :--- | :--- | :--- | | Reasoning Trace | Visual Chain of Thought | Shows the multi-agent debate (Conjecturer vs. Falsifier). | | Evidence Drawer | One-Click Grounding | Links every claim to a Tier 1 source or a user biomarker. | | Confidence Bounds | Visual Uncertainty | Uses error bars/ranges to signal where data is thin. | | Epistemic Humility | The "IDK" Flag | Highlights areas where the system cannot rule out a risk. |

4. Technical Integration: The Clinical Mirror

The Clinical Mirror is a secure, HIPAA-compliant projection of the user's Health State optimized for medical workflows.

  • FHIR/HL7 Mapping: All Assessment Engine outputs are converted into standard medical codes (ICD-10, CPT, LOINC) to ensure compatibility with external EHRs.
  • Bi-Directional Sync:
    • Inbound: Physician notes from EHR flow back into the Health State.
    • Outbound: Validated AI assessments flow into the EHR as clinician-signed notes.

5. Regulatory & Safety Guardrails

5.1 FDA CDS Compliance

The system is explicitly built to comply with the 21st Century Cures Act exemptions for Clinical Decision Support:

  1. Transparency: The clinician can see all inputs and the reasoning logic.
  2. No Autonomous Decision: The system suggests; the human decides.
  3. Source-Grounded: Recommendations are grounded in peer-reviewed clinical guidelines.

5.2 Medication Guardrail

While the AI can suggest lifestyle adjustments, Pharmacotherapy remains a human-only domain.

  • AI Role: Flags potential drug-lifestyle interactions (e.g., "This exercise dose may lower glucose too much given current Insulin dose").
  • Physician Role: Renders the prescription and adjusts the dose.

6. North Star

"The physician feels like they have an elite medical team supporting them, allowing them to practice at the very top of their license."

Back to Epistemology All Documents