Healthcare Dashboard — Self-Hosted Garmin + Apple Health Analytics with AI Coaching

A self-hosted, single-user health dashboard that turns your Garmin Connect and Apple Health data into a daily "how am I doing today?" score, evidence-based wellbeing alerts, and LLM-generated coaching — served privately over your own Tailscale network.

It reads the data your wearables already collect (sleep, HRV, SpO2, respiration, Body Battery, Training Readiness, resting heart rate, weight, body fat, steps, stress) and answers three questions every morning:

1. What state am I in today? — a 0–100 condition score across sleep, autonomic recovery (HRV), energy, training load, body composition, and physiology.
2. What should I watch out for? — rule-based medical alerts (sleep debt, HRV decline, low blood-oxygen, irregular circadian rhythm, medication-overuse-headache risk, barometric migraine triggers) grounded in published research.
3. What should I do? — concise, personalized daily actions from Claude, weighted by your goals and today's recovery.

⚠️ Disclaimer. This is a personal quantified-self tool, not a medical device. Scores, alerts, and AI advice are informational only and are not diagnosis or treatment. Wrist-based optical sensors (SpO2, respiration) are noisy; alerts are deliberately conservative and framed as "reasons to pay attention / see a doctor," never as a clinical verdict. Consult a qualified professional for medical decisions.

---

Why this exists

Wearables collect an enormous amount of data and then bury the useful parts. Garmin's app shows you a sleep score but not "your blood-oxygen has dropped three nights running." Apple Health stores everything but interprets nothing. This project is the missing interpretation layer: it joins both sources, scores them against personalized targets and published clinical norms, and produces a short, actionable readout — privately, on hardware you control, with no third-party health cloud.

Features

• Daily condition score (0–100) across sleep, HRV, energy (Body Battery), training load (ACWR), weight, and body fat — each with an achievement curve toward an ideal band.
• Physiology trends extracted from Garmin sleep data: blood oxygen (SpO2, avg + lowest), respiration rate, nocturnal resting HR, sleep midpoint (circadian regularity), Training Readiness, fitness age — with 28-day sparklines and week-over-week deltas.
• Evidence-based wellbeing alerts — chronic sleep deficit (Belenky 2003), HRV chronic decline (Plews 2013), recovery failure, blood-oxygen desaturation (apnea screening), elevated respiration vs. baseline, low Training Readiness streaks, irregular circadian rhythm (circular statistics), medication-overuse-headache risk (ICHD-3), and barometric pressure migraine triggers.
• LLM daily coaching (Anthropic Claude) — 1–3 concrete actions per day, personalized to your goals, equipment, injury history, and Karvonen heart-rate zones, with optional Google Calendar scheduling.
• Instrument-cluster status lamps — a car-dashboard-style icon row giving an at-a-glance read of every metric, with tap-to-expand detail.
• Life-domain scoring — extend beyond health to track meditation, learning, work, and more, with adjustable importance weights.
• Private by default — runs in Docker on your machine, exposed only inside your Tailscale tailnet via a tailscale serve sidecar (HTTPS, no public ingress).

Architecture

[iPhone Health Auto Export] ──POST/JSON──▶ [FastAPI /ingest/health-auto-export]
[Garmin Connect] ◀──python-garminconnect── [APScheduler, hourly]
                                                   │
                                                   ▼
                                   [SQLite (WAL) + scoring + Claude LLM]
                                                   │
[Browser via tailnet HTTPS] ── tailscale serve sidecar (443→80) ── nginx → FastAPI

Stack

Layer	Tech
Backend	Python 3.12, FastAPI, SQLAlchemy, APScheduler, Anthropic SDK, python-garminconnect
Frontend	React, Vite, TanStack Query, Tailwind CSS, Recharts, lucide-react
Data	SQLite (WAL)
Deploy	Docker Compose (backend, frontend, tailscale sidecar)
Secrets	1Password CLI (op run) or any .env

The science

Thresholds split into two kinds, and the code keeps them separate:

• Clinical/physiological constants (fixed for everyone): SpO2 normal ≥95% / hypoxemia <90%; resting respiration 12–18 brpm; ACWR sweet spot 0.8–1.3; HRV evaluated as an individual z-score against a 28-day baseline rather than absolute values.
• Personal targets (configurable per user): target weight, body fat, age, sex, height, resting heart rate, sleep goal, caffeine half-life, protein target.

Notable correctness details: the sleep midpoint (a standard circadian marker) is a cyclic quantity, so regularity is computed with circular statistics — a naive linear standard deviation breaks for anyone whose midpoint crosses midnight. Medication-overuse risk counts distinct medication days (per ICHD-3), not doses. See docs/superpowers/specs/ for the design rationale and SPEC.md for the full spec.

Quick start

Prerequisites: Docker, a Garmin Connect account, an Anthropic API key, a Tailscale account, and (optionally) an iPhone with Health Auto Export for Apple Health metrics.

# 1. Configure secrets and personal profile
cp .env.example .env        # then edit: ANTHROPIC_API_KEY, GARMIN_*, TS_AUTHKEY, HAE_INGEST_TOKEN
                            # and your profile: USER_AGE, TARGET_WEIGHT_KG, WEATHER_LATITUDE, ...

# 2. Build & start (backend + frontend + tailscale sidecar)
docker compose up -d --build

# 3. First Garmin login (interactive, handles MFA)
docker compose exec backend python -m app.cli garmin-login

# 4. Open the dashboard inside your tailnet
#    https://<hostname>.<tailnet>.ts.net

The repo includes 1Password-based helper scripts (bin/up.sh, bin/verify.sh) for resolving secrets via op run. If you don't use 1Password, a plain .env works the same way.

Personalize your profile

Your biometric targets and training profile live in backend/app/config.py as an example profile and are overridable via environment variables. Set at least:

Variable	Meaning
USER_AGE, USER_SEX, USER_HEIGHT_CM, USER_RESTING_HR	drives Karvonen HR zones & coaching
TARGET_WEIGHT_KG, TARGET_BODY_FAT_PCT	body-composition scoring
WEATHER_LATITUDE, WEATHER_LONGITUDE, WEATHER_LOCATION_LABEL	barometric migraine monitoring
MEDITATION_TARGET_MIN	meditation life-domain goal

Apple Health (optional)

In Health Auto Export, add a REST API automation:

• URL: https://<hostname>.<tailnet>.ts.net/ingest/health-auto-export
• Header Authorization: Bearer <your HAE_INGEST_TOKEN>
• Data Type: Health Metrics (All), Format: JSON v2, Summarize: ON, Cadence: 6 h

Development

# Backend (Python 3.12)
cd backend
uv venv --python 3.12 .venv
uv pip install --python .venv/bin/python -e '.' --group dev
.venv/bin/python -m pytest
.venv/bin/python -m ruff check app/ tests/

# Frontend
cd frontend
npm install
npm run dev      # http://localhost:5173 (proxies /api to :8000)
npm run build

CLI

docker compose exec backend python -m app.cli <command>

Command	Purpose
garmin-login	Interactive Garmin Connect login (for MFA)
sync-garmin	Manual pull from Garmin
recompute [YYYY-MM-DD]	Recompute a day's score (default: today)
regenerate-advice [YYYY-MM-DD]	Force-regenerate LLM advice

Tailscale ACL

To use the tag:healthcare auth-key tag, add a tagOwner in your tailnet ACL:

{
  "tagOwners": {
    "tag:healthcare": ["your-email@example.com"]
  }
}

License

MIT © 2026 nagamine-git

---

_{Keywords: self-hosted health dashboard, Garmin Connect API, Apple Health integration, HRV tracking, sleep analysis, SpO2 monitoring, Body Battery, Training Readiness, quantified self, FastAPI, React, Tailscale, Anthropic Claude, LLM health coaching, circadian rhythm, wearable data analytics, personal health record, open source health app.}