ECG Second Opinion

Clinician-in-the-loop ECG second-opinion workflow system.

A standalone TypeScript API that orchestrates the full lifecycle of an ECG second-opinion case — from intake and quality checks through AI-assisted draft generation to mandatory clinician review, finalization, and delivery.

⚠️ Research Use Only. This repository is not FDA-cleared, not CE-marked, and not approved for clinical use. It must not be used for clinical decision-making. Every output requires review by a qualified clinician.

This repository is not an autonomous diagnostic system. It is a clinician-in-the-loop workflow layer that enforces human review before finalization or delivery.

What This Project Does

ECG Second Opinion is not an "AI that reads ECGs." It is the workflow system around the AI — the control plane that ensures every ECG case follows a strict, auditable path from submission to delivery.

Think of it as a transparent orchestrator:

A clinician or integration system submits a 12-lead ECG recording with patient alias and recording metadata
The system queues the recording for AI analysis (1D-CNN classifier)
An inference worker processes the recording — runs classification, measures confidence, computes uncertainty metrics
The system captures the AI output as a structured draft report — never as a final diagnosis
A clinical safety policy evaluates the result and raises flags for low-confidence, high-risk, or disagreement scenarios
A human clinician (cardiologist, electrophysiologist) must review the draft, add their impression, and explicitly approve or modify it
Only after clinician approval does the case move to finalization and delivery
Every step is logged, timestamped, and traceable via an append-only audit trail

The key idea: The AI generates a draft. A human makes the decision. The system enforces this boundary in code.

How It Works

The 6-State Machine

Submitted → InferencePending → AwaitingReview → Reviewed → Finalized
                 ↓
           InferenceFailed

Every arrow is a guarded transition. The aggregate rejects invalid state changes with typed domain errors.

Transition	Guard
Submitted → InferencePending	Case just created, queued for async inference
InferencePending → AwaitingReview	Inference completed, safety policy evaluated
InferencePending → InferenceFailed	Worker error captured, case not stuck
AwaitingReview → Reviewed	Clinician submitted review decision
Reviewed → Finalized	Clinician approved; no unresolved blocking safety flags

The Inference Pipeline

ECG Recording ──→ Metadata Validation
                       │
                       ▼
               Inference Worker (async)
              ┌────────┴─────────┐
              │  1D-CNN (planned) │
              │  Metadata fallback│ ← current
              └────────┬─────────┘
                       │
                       ▼
              Safety Policy Evaluation
              (8 clinical rules, AHA/ACC aligned)
                       │
                       ▼
              AwaitingReview + Safety Flags

Interoperability Exports

FHIR R4 DiagnosticReport — structured export with contained Observations, performer references, and Research-Use-Only extension
Structured ECG Report — internal JSON format with findings, recommendations, limitations, and safety flags

Current Baseline

What is implemented (backed by tests and running code):

Component	Status	Evidence
6-state case aggregate	✅ Complete	25 unit tests
3-tier auth (operator API key, reviewer JWT HS256, internal bearer)	✅ Complete	API integration tests incl. malformed, issuer, audience, role, signature, claims, timing, and token rejection cases
Zod input validation (all endpoints)	✅ Complete	12 validation tests
Clinical safety policy (8 rules)	✅ Complete	13 safety policy tests
Metadata-fallback inference service	✅ Stub	`@sota-stub` tagged
FHIR R4 DiagnosticReport export	✅ Complete	API integration tests
Structured ECG report builder	✅ Complete	API integration tests
Prometheus metrics (7 instruments)	✅ Complete	Wired in routes
Health probes (`/healthz`, `/readyz`)	✅ Complete	API integration tests
Correlation ID propagation	✅ Complete	Header tests
Fail-fast config validation	✅ Complete	Config test suite incl. reviewer audience guard
Append-only audit trail	✅ Complete	Audit trail test
Async inference with failure recovery	✅ Complete	InferenceFailed state tests
Safety flag resolution by reviewer	✅ Complete	API + unit tests incl. duplicate unresolved-flag protection
Operations summary dashboard endpoint	✅ Complete	API integration test
Docker multi-stage build (non-root)	✅ Complete	Dockerfile

What is target architecture (planned, seams exist):

Component	Status
1D-CNN inference worker (Python, PTB-XL trained)	Planned — Wave 2
SQLite persistence layer	Planned — seam via `IEcgCaseRepository`
PostgreSQL adapter	Planned
Lead-specific abnormality detection	Planned
SCP-ECG structured report mapping	Planned
Feedback loop (clinician corrections → training data)	Planned

Architecture

┌──────────────────────────────────────────────────┐
│                   Clinician                       │
│          (Review via API / future UI)             │
└──────────────────┬───────────────────────────────┘
                   │ review / finalize (JWT auth)
┌──────────────────▼───────────────────────────────┐
│            TypeScript API (Express)               │
│                                                   │
│  ┌──────────┐ ┌───────────┐ ┌─────────────────┐  │
│  │ Routing  │ │  State    │ │  Validation     │  │
│  │ & Auth   │ │  Machine  │ │  (Zod)          │  │
│  └──────────┘ └───────────┘ └─────────────────┘  │
│  ┌──────────┐ ┌───────────┐ ┌─────────────────┐  │
│  │ Safety   │ │ Inference │ │ Health &        │  │
│  │ Policy   │ │ Service   │ │ Metrics         │  │
│  └──────────┘ └───────────┘ └─────────────────┘  │
│  ┌──────────┐ ┌───────────┐ ┌─────────────────┐  │
│  │ FHIR R4  │ │ Report    │ │ Audit           │  │
│  │ Export   │ │ Builder   │ │ Trail           │  │
│  └──────────┘ └───────────┘ └─────────────────┘  │
└──────────────────┬───────────────────────────────┘
                   │ dispatch / callback
┌──────────────────▼───────────────────────────────┐
│        Inference Worker (planned: Python)          │
│                                                   │
│  ┌─────────────────┐ ┌────────────────────────┐  │
│  │ 1D-CNN Classifier│ │ Uncertainty            │  │
│  │ (PTB-XL trained) │ │ (MC-Dropout/Ensembles) │  │
│  └─────────────────┘ └────────────────────────┘  │
└───────────────────────────────────────────────────┘

Key design decisions:

Separation of control and compute: The TypeScript API handles workflow logic only. Heavy neural network inference lives in a separate worker process.
Async inference with callback: Case creation returns 202 Accepted immediately. The inference worker calls back via an internal authenticated endpoint when done (or a failure is captured as InferenceFailed).
Claim discipline: The project distinguishes between what is implemented (backed by running code and tests), what is target architecture (planned), and what is research-informed (supported by literature but not built).

Technology Stack

Layer	Technology
Runtime	Node.js ≥ 22, TypeScript 5.8, ES2022
Framework	Express 4 with Helmet, express-rate-limit
Validation	Zod 3.24 (strict schemas for all inputs)
Auth	API key (operators), HS256 JWT with issuer/audience validation (reviewers), Bearer token (internal)
Metrics	prom-client 15 (Prometheus-compatible)
Testing	Jest 29, ts-jest, supertest
Container	Docker multi-stage build, non-root `ecg` user
Identifiers	Built-in `node:crypto.randomUUID()`

API Surface

Public Endpoints (Operator Auth: `x-api-key`)

Method	Path	Description
POST	`/api/v1/cases`	Submit ECG recording for second opinion (returns `202`)
GET	`/api/v1/cases`	List all cases
GET	`/api/v1/cases/:id`	Get case detail with full assessment
GET	`/api/v1/cases/:id/report`	Structured ECG report
GET	`/api/v1/cases/:id/exports/fhir-diagnostic-report`	FHIR R4 DiagnosticReport
GET	`/api/v1/operations/summary`	Operations dashboard summary

Reviewer Endpoints (JWT Auth: `Authorization: Bearer <jwt>`)

Method	Path	Description
POST	`/api/v1/cases/:id/review`	Submit clinician review
POST	`/api/v1/cases/:id/safety-flags/:flagCode/resolve`	Resolve a safety flag
POST	`/api/v1/cases/:id/finalize`	Finalize case

Internal Endpoints (Bearer Auth)

Method	Path	Description
POST	`/api/internal/inference-callback`	Inference worker callback

Operational Endpoints (No Auth)

Method	Path	Description
GET	`/`	API identity and route map
GET	`/healthz`	Liveness probe
GET	`/readyz`	Readiness probe
GET	`/metrics`	Prometheus metrics

Security

Three-tier authentication: Operator API key for case management, HS256 JWT for clinician review (with role-based access), internal bearer token for inference worker callbacks.
Timing-safe comparison: All secret comparison uses crypto.timingSafeEqual.
JWT claim validation: Reviewer tokens validate alg, iss, aud, exp, nbf, sub, and role before access is granted.
Input validation: Every endpoint validates input via Zod schemas with strict mode — no extra fields accepted, and classifier probabilities must sum to 1.0 ± 0.001.
Rate limiting: Applied to all API endpoints.
Helmet: Security headers enabled by default (CSP, HSTS, etc.).
CORS: Not enabled — add per deployment requirements.
Non-root container: Docker runs as unprivileged ecg user.
No PHI in logs: Patient aliases only, never real identifiers.
Clock skew tolerance: JWT validation includes configurable clock skew window for distributed deployments.

Getting Started

# Install
npm install

# Copy environment config
cp .env.example .env

# Development mode (hot reload)
npm run dev

# Build & run
npm run build
npm start

# Docker
docker compose up --build

Example: Full Workflow

# 1. Submit case
curl -X POST http://localhost:3100/api/v1/cases \
  -H 'Content-Type: application/json' \
  -H 'x-api-key: ecg-operator-dev-token-change-me-0001' \
  -d '{
    "recording": {
      "recordingId": "ptbxl-00001",
      "patientAlias": "Patient-001",
      "recordingDate": "2024-01-15T10:00:00.000Z",
      "samplingFrequencyHz": 500,
      "leadCount": 12,
      "durationSeconds": 10,
      "samplesPerLead": 5000,
      "sourceDataset": "PTB-XL"
    },
    "clinicalQuestion": {
      "questionText": "Rule out myocardial infarction",
      "urgency": "routine"
    }
  }'

# Response: { "caseId": "...", "status": "InferencePending", "message": "..." }

# 2. Wait for async inference, then review (requires reviewer JWT)
curl -X POST http://localhost:3100/api/v1/cases/{caseId}/review \
  -H 'Content-Type: application/json' \
  -H 'Authorization: Bearer <reviewer-jwt>' \
  -d '{
    "decision": "accepted",
    "clinicalNotes": "NSR confirmed. No acute ST changes."
  }'

# 3. Finalize
curl -X POST http://localhost:3100/api/v1/cases/{caseId}/finalize \
  -H 'Content-Type: application/json' \
  -H 'Authorization: Bearer <reviewer-jwt>' \
  -d '{
    "outcome": "delivered",
    "finalSummary": "Normal sinus rhythm confirmed by cardiologist."
  }'

Testing

Tests:       95 passed, 95 total
Test Suites: 5 passed, 5 total

  tests/config.test.ts          — Config validation and production safety guards (9 tests)
  tests/cases.test.ts           — Aggregate state machine and safety-flag invariants (25 tests)
  tests/safety-policy.test.ts   — Clinical safety rules (13 tests)
  tests/validation.test.ts      — Zod schema validation and probability invariants (12 tests)
  tests/api.test.ts             — API integration and auth boundary coverage (36 tests)

npm test                # Run all tests
npm run test:watch      # Watch mode
npm run lint            # ESLint over src/ and tests/
npm run typecheck       # TypeScript type-check without emit

Project Structure

ecg-second-opinion/
├── src/
│   ├── index.ts                # Entry point, graceful shutdown
│   ├── app.ts                  # Express application (11 routes)
│   ├── cases.ts                # Root aggregate — 6-state machine
│   ├── case-contracts.ts       # All TypeScript interfaces and types
│   ├── case-presentation.ts    # Report builders, DTO mappers
│   ├── case-exports.ts         # FHIR R4 DiagnosticReport export
│   ├── case-repository.ts      # In-memory repository + audit trail
│   ├── validation.ts           # Zod schemas for all API inputs
│   ├── safety-policy.ts        # 8-rule clinical safety evaluation
│   ├── inference-service.ts    # Metadata-fallback stub (@sota-stub)
│   ├── config.ts               # Environment-based configuration
│   ├── metrics.ts              # Prometheus metrics (7 instruments)
│   ├── health.ts               # /healthz and /readyz probes
│   ├── correlation.ts          # X-Correlation-Id middleware
│   ├── express-request.d.ts    # Express Request augmentation declarations
│   ├── auth-common.ts          # Shared auth utilities
│   ├── operator-auth.ts        # API key middleware
│   ├── reviewer-auth.ts        # JWT HS256 middleware
│   └── internal-auth.ts        # Bearer token middleware
├── tests/
│   ├── config.test.ts          # Config guards and production-safe defaults
│   ├── cases.test.ts           # 25 state machine and flag-invariant tests
│   ├── safety-policy.test.ts   # 13 safety policy tests
│   ├── validation.test.ts      # 12 schema and probability tests
│   └── api.test.ts             # 35 integration tests
├── Dockerfile                  # Multi-stage, non-root
├── docker-compose.yml
├── .env.example                # All environment variables documented
├── tsconfig.json
├── eslint.config.mjs           # ESLint flat config
├── jest.config.js
├── package.json
├── LICENSE                     # MIT
├── CONTRIBUTING.md
├── SECURITY.md
├── CODE_OF_CONDUCT.md
└── README.md

Diagnostic Categories

Following the PTB-XL labeling scheme and SCP-ECG standard:

Code	Category	Description
NORM	Normal sinus rhythm	No significant abnormalities
MI	Myocardial infarction	ST-elevation/depression, Q-wave patterns
STTC	ST/T change	Non-specific ST segment or T-wave changes
CD	Conduction disturbance	Bundle branch blocks, AV blocks
HYP	Hypertrophy	Left/right ventricular hypertrophy

Safety Policy

The DefaultEcgClinicalSafetyPolicy evaluates every inference result against 8 clinical rules aligned with AHA/ACC/HRS guidelines:

#	Rule	Trigger	Severity	Blocks
1	LOW_CONFIDENCE	Confidence band `low`	Warning	No
2	INSUFFICIENT_DATA	Confidence band `insufficient_data`	Critical	Yes
3	SIGNIFICANT_DISAGREEMENT	AI disagrees with original interpretation	Warning	No
4	STAT_MI_DETECTED	MI detected on stat-priority recording	Critical	Yes
5	SHORT_RECORDING	Duration < 10s	Warning	No
6	LOW_SAMPLING_RATE	Frequency < 100 Hz	Warning	No
7	HIGH_EPISTEMIC_UNCERTAINTY	Epistemic entropy > 0.7 (MC-Dropout)	Critical	Yes
8	NO_XAI_EXPLANATION	No interpretability artifacts available	Info	No

Blocking flags prevent delivery even after clinician review. They must be explicitly resolved by a reviewer before the case can be finalized as delivered.

Scientific Foundation

Based on: Mosin S.G. (2024) "Neural network diagnosis of the cardiovascular diseases based on data-driven method", Software & Systems, 37(1), pp. 122–130. doi: 10.15827/0236-235X.142.122-130.

Key findings from the paper:

Architecture: 1D multi-layer convolutional neural network for ECG time series classification
Method: Data-driven approach — no manual PQRST feature extraction needed
Dataset: PTB-XL — 21,837 12-lead ECG recordings, 10s at 100 Hz (1000 samples/lead)
Best result: 3-layer CNN with smaller pooling window achieves 85.66% MI detection accuracy
Activation: ELU (supports negative values inherent in ECG signals)
Optimizer: Mini-batch Adam with batch normalization
Classification: Softmax output layer for category probability distribution

The current codebase implements the metadata-fallback mode. Full 1D-CNN inference is planned for Wave 2.

Regulatory Positioning

This project lives in the clinical decision support workflow space, but it should not be positioned as relying on the FDA non-device CDS exclusion for a future clinically deployed ECG-signal-analysis product.

Under 21 U.S.C. § 360j(o)(1)(E), the non-device CDS exclusion does not cover software functions intended to acquire, process, or analyze a medical image or a pattern or signal from a signal acquisition system. FDA's January 2026 final guidance on Clinical Decision Support Software clarifies that this boundary matters when software analyzes signal data while supporting diagnosis or treatment decisions.

That is relevant here because the target production architecture includes 1D-CNN analysis of ECG signal data. A clinically deployed version should be planned as a device-regulated SaMD pathway, not marketed as a non-device CDS exemption claim. The enforced Reviewed state before Finalized still matters, but as a human-oversight control, not as a substitute for device regulatory obligations.

Current status: Research-use-only prototype. Not submitted for FDA review. Not CE-marked. Not approved for clinical use.

Strategic regulatory workstreams for future clinical deployment:

U.S. pathway analysis: 510(k) vs De Novo depending on intended claims, predicates, and risk framing
AI/ML lifecycle planning, including PCCP-style change governance for future model updates
GMLP-aligned lifecycle evidence, including drift monitoring and post-market performance review
EU AI Act preparation for medical-deployment documentation, human oversight, and data-governance obligations

Roadmap

Sprint 1 — Infrastructure hardening: replace setImmediate() dispatch with a durable queue, bounded retry policy, dead-letter or replay path, and durable PostgreSQL-backed persistence
Sprint 2 — Regulatory and lifecycle compliance: document SaMD-oriented positioning, add GMLP work products, define PCCP-style update governance, and prepare EU AI Act evidence surfaces for medical deployment
Sprint 3 — Clinical interoperability: add SNOMED CT / LOINC coding, strengthen FHIR profile declarations, and plan HL7 aECG / DICOM Waveform support for enterprise environments
Sprint 4 — Real inference stack: implement the Python 1D-CNN worker, MC-Dropout uncertainty, Grad-CAM-style explainability, and drift-aware monitoring around PTB-XL-trained models

Community

Issues: GitHub Issues
Contributing: See CONTRIBUTING.md
Security: See SECURITY.md
Code of Conduct: See CODE_OF_CONDUCT.md
Support: See SUPPORT.md
Citation: See CITATION.cff

License

MIT — see LICENSE.

ECG Second Opinion — Система Второго Мнения по ЭКГ

Система второго мнения по ЭКГ с обязательным контролем врача-кардиолога.

Автономный TypeScript API, который управляет полным жизненным циклом ЭКГ-консультации — от приёма записи и проверки качества через AI-анализ до обязательного врачебного заключения, финализации и выдачи.

⚠️ Только для исследовательских целей. Репозиторий не имеет FDA-clearance, не имеет CE-marking и не предназначен для клинического применения. Система не может использоваться для принятия клинических решений. Каждый результат требует проверки квалифицированным врачом.

Что делает проект

ECG Second Opinion — это не «ИИ, который читает ЭКГ». Это система управления рабочим процессом вокруг ИИ — контрольная плоскость, которая гарантирует, что каждый случай ЭКГ следует строгому, проверяемому пути от подачи до выдачи.

Клиницист или интеграционная система подаёт 12-канальную ЭКГ-запись с псевдонимом пациента и метаданными записи
Система ставит запись в очередь на ИИ-анализ (1D-CNN классификатор)
Инференс-воркер обрабатывает запись — классификация, оценка уверенности, метрики неопределённости
Система фиксирует результат ИИ как черновой структурированный отчёт — не как финальный диагноз
Политика клинической безопасности оценивает результат и выставляет флаги для случаев с низкой уверенностью, высоким риском или расхождением
Врач-кардиолог обязан проверить черновик, добавить своё заключение и явно подтвердить или изменить его
Только после врачебного одобрения случай переходит к финализации и выдаче
Каждый шаг логируется, помечается временем и отслеживается через неизменяемый журнал аудита

Ключевая идея: ИИ генерирует черновик. Решение принимает человек. Система обеспечивает эту границу в коде.

Машина состояний

Submitted → InferencePending → AwaitingReview → Reviewed → Finalized
                  ↓
            InferenceFailed

Каждый переход защищён инвариантами. Агрегат отклоняет недопустимые изменения состояния с типизированными доменными ошибками.

Текущий baseline

Компонент	Статус	Доказательство
Агрегат с 6 состояниями	✅ Реализован	25 юнит-тестов
3-уровневая аутентификация	✅ Реализована	Интеграционные тесты
Zod-валидация (все эндпоинты)	✅ Реализована	12 тестов валидации
Политика безопасности (8 правил)	✅ Реализована	13 тестов
Метаданные-фолбэк инференс	✅ Стаб	`@sota-stub`
FHIR R4 DiagnosticReport экспорт	✅ Реализован	Интеграционные тесты
Prometheus-метрики (7 инструментов)	✅ Реализованы	Подключены в роутах
Неизменяемый журнал аудита	✅ Реализован	Тест аудита

Технологический стек

Слой	Технология
Среда выполнения	Node.js ≥ 22, TypeScript 5.8, ES2022
Фреймворк	Express 4 + Helmet + express-rate-limit
Валидация	Zod 3.24
Аутентификация	API-ключ, HS256 JWT с проверкой issuer/audience, Bearer-токен
Метрики	prom-client 15
Тестирование	Jest 29, ts-jest, supertest
Контейнер	Docker multi-stage, непривилегированный пользователь

API

Публичные маршруты (авторизация: `x-api-key`)

Метод	Путь	Описание
POST	`/api/v1/cases`	Подать ЭКГ-запись для второго мнения
GET	`/api/v1/cases`	Список случаев
GET	`/api/v1/cases/:id`	Детали случая
GET	`/api/v1/cases/:id/report`	Структурированный отчёт
GET	`/api/v1/cases/:id/exports/fhir-diagnostic-report`	FHIR R4 экспорт

Маршруты рецензента (JWT: `Authorization: Bearer <jwt>`)

Метод	Путь	Описание
POST	`/api/v1/cases/:id/review`	Врачебное заключение
POST	`/api/v1/cases/:id/finalize`	Финализация случая

Быстрый старт

npm install
cp .env.example .env
npm run dev

Тестирование

Тесты:    95 пройдено, 95 всего (5 наборов)

npm test

Научная основа

Мосин С.Г. (2024) «Нейросетевая диагностика заболеваний сердечно-сосудистой системы на основе метода, управляемого данными», Программные продукты и системы, 37(1), с. 122–130. doi: 10.15827/0236-235X.142.122-130.

Регуляторное позиционирование

Проект находится в области workflow-систем для клинической поддержки решений, но его нельзя описывать как систему, которая в продакшене будет полагаться на FDA-исключение для non-device CDS.

Согласно 21 U.S.C. § 360j(o)(1)(E) и финальному руководству FDA по Clinical Decision Support Software от января 2026 года, исключение не распространяется на ПО, которое предназначено для получения, обработки или анализа медицинского изображения либо сигнала из системы съёма физиологических данных.

Это важно для данного проекта, потому что целевая продакшен-архитектура включает 1D-CNN анализ ЭКГ-сигналов. Поэтому клинически развёрнутая версия должна планироваться как device-regulated SaMD, а не как non-device CDS. Жёсткая обязательность стадии Reviewed перед Finalized остаётся важной, но уже как механизм human oversight, а не как замена регуляторному пути.

Текущий статус: Исследовательский прототип. Не подавался на FDA review. Не имеет CE-marking. Не допущен к клиническому применению.

Дорожная карта

Sprint 1 — Инфраструктурное усиление: заменить setImmediate() на durable queue, добавить bounded retry / dead-letter path и перейти к долговечной PostgreSQL-персистентности
Sprint 2 — Комплаенс и жизненный цикл модели: зафиксировать SaMD-путь, подготовить GMLP-артефакты, PCCP-подобное управление обновлениями и evidence-поверхность под EU AI Act для медицинского применения
Sprint 3 — Медицинская интероперабельность: добавить SNOMED CT / LOINC, усилить FHIR profile declarations и спланировать HL7 aECG / DICOM Waveform для enterprise-интеграций
Sprint 4 — Реальный inference stack: Python 1D-CNN worker, MC-Dropout, Grad-CAM/XAI и drift-aware monitoring для моделей на PTB-XL

Лицензия

MIT — см. LICENSE.

Name		Name	Last commit message	Last commit date
Latest commit History 4 Commits
.github		.github
docs		docs
src		src
tests		tests
.env.example		.env.example
.gitignore		.gitignore
CITATION.cff		CITATION.cff
CODE_OF_CONDUCT.md		CODE_OF_CONDUCT.md
CONTRIBUTING.md		CONTRIBUTING.md
Dockerfile		Dockerfile
LICENSE		LICENSE
README.md		README.md
README.ru.md		README.ru.md
SECURITY.md		SECURITY.md
SUPPORT.md		SUPPORT.md
docker-compose.yml		docker-compose.yml
eslint.config.mjs		eslint.config.mjs
jest.config.js		jest.config.js
package-lock.json		package-lock.json
package.json		package.json
tsconfig.json		tsconfig.json

Folders and files

Latest commit

History

Repository files navigation

ECG Second Opinion

Table of Contents

What This Project Does

How It Works

The 6-State Machine

The Inference Pipeline

Interoperability Exports

Current Baseline

Architecture

Technology Stack

API Surface

Public Endpoints (Operator Auth: x-api-key)

Reviewer Endpoints (JWT Auth: Authorization: Bearer <jwt>)

Internal Endpoints (Bearer Auth)

Operational Endpoints (No Auth)

Security

Getting Started

Example: Full Workflow

Testing

Project Structure

Diagnostic Categories

Safety Policy

Scientific Foundation

Regulatory Positioning

Roadmap

Community

License

ECG Second Opinion — Система Второго Мнения по ЭКГ

Что делает проект

Машина состояний

Текущий baseline

Технологический стек

API

Публичные маршруты (авторизация: x-api-key)

Маршруты рецензента (JWT: Authorization: Bearer <jwt>)

Быстрый старт

Тестирование

Научная основа

Регуляторное позиционирование

Дорожная карта

Лицензия

About

Resources

License

Code of conduct

Contributing

Security policy

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Public Endpoints (Operator Auth: `x-api-key`)

Reviewer Endpoints (JWT Auth: `Authorization: Bearer <jwt>`)

Публичные маршруты (авторизация: `x-api-key`)

Маршруты рецензента (JWT: `Authorization: Bearer <jwt>`)

Packages