sleep-detection/README.md

sleep-detection

A small ML pipeline that turns wrist accelerometer + heart rate data into per-epoch sleep stage predictions (hypnogram building blocks).

Built to support InfiniTime PR #2304, which adds sleep tracking to InfiniTime (PineTime firmware) but stops at raw data logging. This repo explores whether a small offline-trained classifier can turn that raw data into sleep stages and eventually an on-device hypnogram.

Repo layout

• src/sleep_detection/ — data loading, epoching/features, model, training loop
• scripts/ — supporting/experimentation scripts
• plots/ — output plots

Datasets

Not included in this repo — download separately:

• PhysioNet: Motion and heart rate from a wrist-worn wearable and labeled sleep from polysomnography (Walch et al., 2019) — primary dataset used so far.
• PhysioNet: A Multi-Night Instantaneous Heart Rate and Accelerometry Dataset with EEG Sleep Stage Labels — larger, newer, same sensor modalities.

Getting started

This is just a regular python project. I recommend you work in an editable virtual environment during development.

cd sleep-detection
python3 -m venv .venv
source .venv/bin/activate
python3 -m pip install -e .

The uv project manager should also work just fine if you prefer to use that.

Datasets

Not included in this repo, but a convenience download script exists. Note that this will take a while because the physionet servers are deliberately quite slow.

./download.sh

Usage

As an example training a model on the sleep-accel dataset subject 1066528:

sleep-detection \
   -a physionet.org/files/sleep-accel/1.0.0/motion/1066528_acceleration.txt \
   -H physionet.org/files/sleep-accel/1.0.0/heart_rate/1066528_heartrate.txt \
   -l physionet.org/files/sleep-accel/1.0.0/labels/1066528_labeled_sleep.txt \
   --output model.pt

Help Wanted

I am not sure if the way I am feeding / dividing the training data up is done appropriately and if the model is too simple for this kind of dataset. I would love some help from someone who have experience and proper insight into how to train models using pytorch. Feel free to message me, submit an issue, fork the project etc.