NightWatch - Sleep Stage Classifiers

Datasets and models for sleep stage prediction.

Setup

This project uses Poetry for dependency management. Follow the steps below to set up the project.

1. Install Poetry:

   curl -sSL https://install.python-poetry.org | python3 -

2. Install Dependencies:

   poetry update && poetry install

Download and prepare the Sleep-Accel dataset

The Sleep Accel dataset contains motion and heart rate measurements collected from 31 subjects using an Apple Watch, and it is available from here.

Datasets are stored under /data. To download and prepare Sleep-Accel, use the data/download.sh script. By default, the script creates a sleep-accel directory in the current working directory, where the dataset samples are downloaded and prepared.

cd data/
./download.sh

For convenience, all other scripts assume that the dataset is under ./data/sleep-accel. To override the target directory, use the --data-dir argument:

./data/download.sh --data-dir some/other/path

The script executes three stages: download, prepare, and cleanup. To execute each stage independently, run the script for the specific stage:

cd data/
./download.sh cleanup

Extract Features

The dataset contains polysomnography (PSG) annotated samples, sampled every 30 seconds. To compute features for each PSG sample and prepare sample sequences for sequence classification, use the build_seq_ds tool:

poetry run build_seq_ds

By default, this tool assumes the Sleep-Accel dataset resides under ./data/sleep-accel and creates a new directory under ./data/sleep-accel-seq containing sequences for each user with a specific time length, in minutes. By default, 10-minute length sequences are extracted with a stride of 5 minutes. To change the window size or stride, e.g., to 1 hour sampled every minute, do:

poetry run build_seq_ds --window-size-min 60 --window-stride-min 1

For convenience, the data loaders assume the sequence dataset resides under ./data/sleep-accel-seq. To change the source and target directories, use the --sleep-accel-dir and --target-dir arguments:

poetry run build_seq_ds --window-size-min 60 --window-stride-min 1 --sleep-accel-dir raw/sleep/accel --target-dir ./data/sleep-accel-1h-every-1min

Execute poetry run build_seq_ds --help to check usage documentation. The supported parameters are:

• --sleep-accel-dir: The directory containing the raw sleep acceleration data. (Default: ./data/sleep-accel)
• --target-dir: The directory where the processed sleep acceleration sequence dataset will be saved. (Default: ./data/sleep-accel-seq)
• --window-size-min: The input window size in minutes. (Default: 10)
• --window-stride-min: The input stride in minutes. (Default: 5)
• --train-test-split: The ratio of training to testing data. (Default: 0.9)

Performing Experiments

To train sequence classifiers on the PSG-annotated data, use the train_seq_cls tool:

poetry run train_seq_cls --batch_size 16 --max_epochs 100

This will use the data from ./data/sleep-accel-seq to train a default 1-layer LSTM sequence encoder followed by an MLP classifier. To experiment with different model configurations, pass in the hyperparameters to the train_seq_cls tool. Supported parameters are:

• --ds_dir: Path to the dataset directory. (Default: ./data/sleep-accel-seq)
• --model_path: Path for saving model checkpoints and training logs. (Default: ./exps)
• --num_units: Number of hidden units in the LSTM. (Default: 128)
• --num_layers: Number of LSTM layers. (Default: 1)
• --dropout: Dropout rate. (Default: 0.2)
• --lr: Learning rate for the optimizer. (Default: 1e-3)
• --max_epochs: Number of epochs to train the model. (Default: 20)
• --batch_size: Batch size for training. (Default: 32)

Results

Window Size	Window Stride	Num Layers	Units	ACC
60	1	2	32	81.4
60	1	4	32	83.7