GitHub - scerelli/SYNAPS

SYNAPS

Self-hosted personal health monitor to see the body as complex graph

CC BY-NC-SA 4.0

Not a medical device. Synaps is a personal data aggregation tool. Nothing it shows: correlations, trends, AI analysis, or allostatic load scores constitutes medical advice or diagnosis. Always consult a qualified healthcare professional for any health issue. I assume no liability for decisions made based on information displayed by this software.

Background

A few years ago I built a small personal project, basically an health log. It's nothing fancy, just a place to store the things I find interesting to monitor about myself: blood tests, sleep, how I feel during the day. I like having a place where I can look at my own system as a whole, instead of scattered notes and PDF reports I'll never open again

Two years ago I started studying statistics as part of a master's degree I'm taking, and at some point I thought: How can I apply all this things I'm learning? This project was one of the first things that came to my mind, so I started reading papers about sleep and inflammation, about how biomarkers correlate with lifestyle, about statistical models for personal data. The more I read the more I realized there's actually a lot of interesting science behind the things I was already tracking

So I decided to take advantage of the AI tools, refactor part of the backend and add some simple and nice UI to it. My idea was to build a mathematical model, a weighted network, that connects biomarkers, conditions, lifestyle factors and environment, and from that graph try to infer in a somewhat deterministic way what's going on and what could be improved. It's not perfect and there's a lot I still want to improve, but I think it works well enough to be useful and I also have a feeling that many people may find it kinda useful

One thing I want to be VERY CLEAR about: this is not medical advice, it is not a diagnostic tool, and it should never replace a doctor. It's a personal monitoring tool and an exercise in applied statistics. All the papers and references that informed the statistical model are linked at the bottom of this page.

What it is

Synaps aggregates medical reports, daily lifestyle data, and environmental conditions into a single knowledge graph. An AI (Claude) reads the whole picture and finds correlations across domains: not just within them.

Single-user, open source, runs entirely on your own hardware.

Screenshots

Dashboard: AI reminders, quick stats, recent reports	Knowledge Graph: biomarkers, conditions, lifestyle, medications
Reports: tagged, searchable medical history	Report detail AI analysis + color-coded biomarker table
Conditions: active and resolved, with timeline	Sleep log: nightly hours and quality rating
Trends: biomarker time series with EMA and z-score

What it tracks

Statistics computed

All metrics are updated incrementally as you log data, using time-aware exponential smoothing (EWMA with continuous-time decay).

Per metric: long-τ EMA, short-τ EMA, variance, trend direction, z-score

Cross-domain correlations: Pearson r across domains (sleep, wellbeing, activity, diet, body) with lags 0, 1 and 2 days. Minimum 45 co-observations required. Autocorrelation is corrected via effective N: nEff = n × (1 − r₁) / (1 + r₁) where r₁ is the lag-1 autocorrelation of the series. Multiple comparisons corrected with Benjamini-Hochberg FDR (α = 0.05).

Trend detection: short-τ EMA (τ/3) vs. long-τ EMA. If the difference exceeds 20% of the rolling standard deviation the metric is marked increasing or decreasing

Allostatic load: normalized count of biomarker risk flags: CRP > 3 mg/L, HDL < 40 mg/dL, HbA1c > 6.5 %, waist-to-height ratio > 0.5. Score = risk flags present / biomarkers available, range 0–1. Concept from Gustafsson et al. 2023 (17-cohort IPD meta-analysis)

Health graph: Claude maps all data into typed nodes (biomarker, condition, lifestyle, environment, medication, report) and weighted edges. Linear regression on biomarker time series validates edge directions before they are stored

How the model works

Exponential weighted moving average (EWMA)

Each metric is tracked as a continuous-time exponential moving average. When a new observation arrives after Δt days:

α      = 1 − exp(−Δt / τ)
EMA    = α × value + (1 − α) × EMA_prev
emVar  = α × (value − EMA)² + (1 − α) × emVar_prev

emVar (exponential moving variance) enables real-time z-score computation and trend detection. A short-τ EMA (τ/3) compared to the long-τ EMA determines trend direction: if shortEMA > longEMA + 0.2σ the metric is marked increasing.

Within-person cross-domain correlations (N-of-1)

Daily observations across domains (sleep, wellbeing, activity, diet, body) are joined by date and analysed using Pearson r with lags of 0, 1 and 2 days. Minimum 45 co-observations required per pair. Because measurements are autocorrelated and we run many tests simultaneously, two corrections are applied:

• Autocorrelation: effective N is computed as nEff = n × (1 − r₁) / (1 + r₁) where r₁ is the lag-1 autocorrelation of the series. The Pearson t-test and confidence interval are then evaluated at nEff instead of n.
• Multiple comparisons: Benjamini-Hochberg false discovery rate correction (α = 0.05) across all tested pairs and lags

Results are labelled by confidence tier: low (n < 60), moderate (60–120), high (> 120). All findings are presented as correlational signals, not causal relationships — individual within-person patterns frequently diverge from population-level meta-analyses (Myin-Germeys et al.).

Allostatic load

A normalised count of biomarker risk flags based on the 17-cohort IPD consensus (Gustafsson et al. 2023): CRP > 3 mg/L, HDL < 40 mg/dL, HbA1c > 6.5 %, waist-to-height ratio > 0.5. Score = risk flags present / biomarkers available, range 0–1. Only flags for which data is available are included in the denominator.

Knowledge graph

Claude maps all structured data into typed nodes and weighted edges. Edge weights are validated against the statistical correlations computed above: an AI-inferred edge is down-weighted if the Pearson r for that pair is low or insignificant.

Install

Requires Docker with Compose v2

curl -fsSL https://raw.githubusercontent.com/scerelli/SYNAPS/refs/heads/main/install.sh | bash

The script:

1. Checks Docker and git are available
2. Clones the repo to /opt/synaps (override with SYNAPS_DIR=/your/path)
3. Creates a pre-configured .env (auto-generated encryption key)
4. Asks for the port (default 80) and whether to disable auth (for users behind a reverse proxy)
5. Builds and starts all services with docker compose up -d --build

Re-running the script on an existing installation does a git pull and rebuilds — no data is lost.

All services use restart: unless-stopped: on a Linux server with Docker set to start on boot, Synaps comes back up automatically after a reboot with no extra configuration needed.

Access from outside your LAN

I suggest to install netbird on your server/machine and on your phone. Once peered, open Synaps at the server's Netbird IP directly: no port forwarding, no public exposure.

Update

cd ~/synaps && docker compose pull && docker compose up -d --build

References

Possible Future Improvements