HealthTrend

Understand the trend behind your weight — not just the number on the scale.

HealthTrend is a live, open-source application for tracking body weight and estimating its underlying trajectory from noisy, irregular weigh-ins. It combines a probabilistic state-space model with uncertainty-aware forecasts, a persistent personal history, and optional Apple Health import and ongoing sync.

Open HealthTrend · Try the interactive demo · Read the method

Interactive demo with synthetic measurements. No real person's health data is shown.

What you can do

• Track your weight over time. Sign in with a one-time email code, log readings in kg or lb, edit your history, or import a CSV.
• See the underlying trend. A Kalman-filter-based estimator separates short-term fluctuations from an estimated weight trajectory and current weekly rate of change.
• Understand uncertainty. View intervals around the estimated weight, rate and 7-, 30- and 90-day forecasts, rather than treating any one prediction as certain.
• Connect Apple Health. Use iPhone Shortcuts to import up to three years of Weight samples once, then sync the most recent 14 days on a schedule. The connection uses a dedicated, revocable token; no native iOS app is required.
• Keep control of your data. Export your measurements as CSV or your account data as JSON, and delete your account and its stored data.
• Explore without signing in. Try labelled synthetic scenarios or analyse a temporary dataset through the public analysis interface.

Getting started

1. Explore without an account. Try the interactive demo or analyse your own measurements temporarily.
2. Save your history. Open the app and request a one-time email sign-in code. If new-account access is restricted, the public demo and analysis remain available without signing in.
3. Log a weigh-in, import a CSV, or follow Apple Health setup. It uses two iPhone Shortcuts: a one-time import of up to three years of Weight history and a separate 14-day recurring sync. The setup page explains where to paste the same revocable connection token into each Shortcut and how to schedule the latter.
4. Open Trend to see the estimated trajectory, rate and forecasts. Your optional goal is a display reference only; it never influences the estimate.

Product walkthrough

The screenshots below show the public analysis and import experience using synthetic data, not anyone's private health records.

Analyse your own measurements

Analyse manually entered measurements or a CSV without saving them to an account. Try the public analysis.

Understand the evidence behind a trend

The analysis separates the main conclusion from its explanation, supporting readings and detailed statistics.

Inspect uncertainty and forecasts

The estimate, current rate and forecasts include model-derived uncertainty intervals.

Preview CSV imports

CSV readings are validated and previewed before analysis. The repository includes a synthetic sample CSV.

How it works

Each weigh-in is a noisy observation of body weight, not a direct measurement of its underlying trajectory. HealthTrend uses a continuous-time Kalman filter to estimate latent weight and its current rate of change, accounting for irregular gaps between readings. The model propagates uncertainty into weight and rate estimates and into 7-, 30- and 90-day forecasts. Its parameters are documented rather than fitted to each user’s short history.

Read the full Method page · See the evaluation, including where the model underperforms

Evaluation and limitations

HealthTrend has a substantial reproducible synthetic-data evaluation, including independent numerical checks, calibration experiments, irregular sampling, parameter-identifiability studies and comparisons against simpler forecasting methods.

The results are deliberately mixed. Under the model's own assumptions, the reported 95% intervals achieved approximately nominal coverage in the tested synthetic settings (94.9% for both latent weight and velocity). However, a tuned Holt method outperformed the shipped estimator on the 30-day forecast metric in 6 of 8 tested synthetic regimes. Uncertainty coverage also deteriorated under certain model misspecifications.

The process-noise intensity (σₐ) was not identifiable from one month of observations, even when the number of weigh-ins increased tenfold. This is why HealthTrend retains documented fixed priors instead of fitting that parameter separately to each user's short history; it is not proof that the chosen priors are optimal.

The application has not been clinically validated, and the evaluation does not establish predictive accuracy on real people. It is not a medical device and does not diagnose, treat or recommend changes to health behaviour.

Read the full evaluation report, including negative findings and experimental limitations.

Architecture

Layer	Implementation
Web application	Next.js 16, React 19, TypeScript, CSS Modules, visx
API	FastAPI, Pydantic, Python, NumPy
Estimation	Deterministic state-space model and Kalman filter
Persistence	PostgreSQL, SQLAlchemy 2, psycopg 3, Alembic
Authentication	Passwordless email codes, secure HttpOnly session cookie
Apple Health	iPhone Shortcuts, source-specific revocable bearer credentials
Deployment	Vercel (web), Render (API), Neon (database), SMTP (email)

Engineering highlights

• Enforced architecture. AST-based tests enforce the api → services → core dependency direction. The deterministic numerical core has an explicit import allowlist and no framework, filesystem, clock or randomness dependencies.
• Checked API contract. The backend generates backend/openapi.json, from which the frontend TypeScript contract is generated. CI checks both generated artifacts for drift.
• Independent numerical verification. The Kalman filter was checked against a joint-Gaussian computation across an 810-case battery, with a 60-digit arithmetic check for the worst ill-conditioned case. The evaluation records disagreements rather than counting them as passes.
• Data-leakage tests. Sentinel weight values are propagated through failure paths to test that errors and logs do not reveal submitted measurements. Frontend checks enforce an explicit, narrowly scoped browser-storage policy rather than claiming that no browser storage is ever used.
• Automated validation. The latest development runs passed 1,431 backend tests and 530 frontend tests, alongside linting, type-checking and a production frontend build.

The signed-in product stores account data and weight history in PostgreSQL. The separate public analysis flow processes a submitted dataset without adding it to an account. Apple Health integration reads Weight only, sends measurements to the account's API, and does not write back to Apple Health. Connection credentials are hashed server-side and can be revoked in Settings.

See architecture and privacy for details.

Local development

Backend (Python and uv):

cd backend
uv sync --locked
# Configure the local database, auth, mailer and CORS settings; see backend/.env.example.
uv run alembic upgrade head
uv run uvicorn app.main:app --no-access-log

Frontend (Node version in frontend/.nvmrc):

cd frontend
npm ci
# Configure the API origins; see frontend/.env.example.
npm run dev

The backend must have its documented environment variables configured before migration or startup. Local development can use SQLite and a console mailer; production uses PostgreSQL and SMTP. Do not commit .env files or personal health exports.

To run checks:

# From backend/
uv run pytest -q
uv run ruff check .
uv run mypy

# From frontend/
npm run typecheck
npm run lint
npm run test
npm run build

Documentation

Browse the documentation index for the model, evaluation, architecture, privacy controls, deployment and design decisions.

Licence

MIT © 2026 Alfred Hong.