AI 輔助醫療診斷（符合 EU AI Act 規範）

osop_version: "1.0"
id: "eu-ai-act-high-risk-medical-diagnosis"
name:"AI 輔助醫療診斷（符合 EU AI Act 規範）"
description: "用於 AI 輔助醫療診斷的高風險 AI 工作流程。依據 EU AI Act 第 19 條對高風險 AI 系統的要求，包含強制性人工監督、風險評估與核准關卡。
"
  High-risk AI workflow for AI-assisted medical diagnosis.
  Includes mandatory human oversight, risk assessment, and approval gates
  as required by EU AI Act Article 19 for high-risk AI systems.
version: "1.0.0"
tags:
  - compliance
  - eu-ai-act
  - high-risk
  - medical
  - human-oversight

metadata:
  regulation: "EU AI Act (Regulation 2024/1689)"
  risk_classification: "high-risk"
  article_19_compliant: true
  data_retention_months: 60
  responsible_entity: "Example Hospital AI Department"

nodes:
  - id: "patient_data_ingestion"
    type: "api"
    subtype: "rest"
    name: "匯入病患資料"
    description: >
      Receive patient medical records, lab results, and imaging data
      from the hospital information system. Validate data completeness
      and format before processing.
    security:
      risk_level: "high"
      data_classification: "sensitive-medical"
      encryption: "AES-256"
      access_control: "role-based"

  - id: "data_validation"
    type: "cli"
    subtype: "script"
    name: "驗證輸入資料"
    description: >
      Run validation checks on patient data: schema conformance,
      required fields, data range checks, and anomaly detection.
      Reject incomplete or malformed records.
    security:
      risk_level: "medium"
      data_classification: "sensitive-medical"

  - id: "ai_diagnosis"
    type: "agent"
    subtype: "llm"
    name: "AI 初步診斷"
    description: >
      AI model analyzes patient data (medical history, lab results,
      imaging) and produces a preliminary diagnosis with confidence
      scores, differential diagnoses, and supporting evidence.
    security:
      risk_level: "critical"
      data_classification: "sensitive-medical"
      model_governance: "approved-clinical-model"

  - id: "risk_assessment"
    type: "agent"
    subtype: "llm"
    name: "臨床風險評估"
    description: >
      Evaluate the AI diagnosis against clinical risk thresholds.
      Flag cases where confidence is below threshold, where the
      diagnosis involves life-threatening conditions, or where
      the AI identifies conflicting indicators.
    security:
      risk_level: "critical"
      data_classification: "sensitive-medical"

  - id: "physician_review"
    type: "human"
    subtype: "review"
    name: "醫師審查"
    description: >
      Licensed physician reviews the AI recommendation, risk
      assessment, and supporting evidence. The physician makes
      the final clinical decision. This step is mandatory and
      cannot be bypassed.
    security:
      risk_level: "critical"
      data_classification: "sensitive-medical"
      mandatory: true
      bypass_allowed: false

  - id: "approval_gate"
    type: "human"
    subtype: "input"
    name: "核准關卡"
    description: >
      Final approval checkpoint before the diagnosis is recorded
      in the patient record. Requires explicit physician sign-off.
      Rejected cases are returned for further review.
    security:
      risk_level: "critical"
      mandatory: true
      bypass_allowed: false

  - id: "record_diagnosis"
    type: "api"
    subtype: "rest"
    name: "記錄最終診斷"
    description: >
      Write the approved diagnosis to the patient's electronic
      health record. Include the AI recommendation, physician
      decision, and full audit trail reference.
    security:
      risk_level: "high"
      data_classification: "sensitive-medical"
      audit_trail: true

  - id: "error_handler"
    type: "agent"
    subtype: "llm"
    name: "錯誤處理"
    description: >
      Handle failures at any stage. Log the error, notify the
      responsible physician, and escalate if patient safety
      may be affected.
    security:
      risk_level: "high"

edges:
  - from: "patient_data_ingestion"
    to: "data_validation"
    mode: "sequential"

  - from: "data_validation"
    to: "ai_diagnosis"
    mode: "conditional"
    when: "validation.status == 'passed'"

  - from: "data_validation"
    to: "error_handler"
    mode: "conditional"
    when: "validation.status == 'failed'"
    label: "Invalid input data"

  - from: "ai_diagnosis"
    to: "risk_assessment"
    mode: "sequential"

  - from: "ai_diagnosis"
    to: "error_handler"
    mode: "fallback"
    label: "AI model failure"

  - from: "risk_assessment"
    to: "physician_review"
    mode: "sequential"

  - from: "risk_assessment"
    to: "error_handler"
    mode: "fallback"
    label: "Risk assessment failure"

  - from: "physician_review"
    to: "approval_gate"
    mode: "sequential"

  - from: "approval_gate"
    to: "record_diagnosis"
    mode: "conditional"
    when: "approval.decision == 'approved'"

  - from: "approval_gate"
    to: "physician_review"
    mode: "conditional"
    when: "approval.decision == 'request_revision'"
    label: "Returned for further review"

  - from: "approval_gate"
    to: "error_handler"
    mode: "conditional"
    when: "approval.decision == 'rejected'"
    label: "Diagnosis rejected"

  - from: "record_diagnosis"
    to: "error_handler"
    mode: "fallback"
    label: "Failed to write to EHR"