ApSense: Data-driven Algorithm in PPG-based Sleep Apnea Sensing

This repository contains the official implementation of the paper: "ApSense: Data-driven Algorithm in PPG-based Sleep Apnea Sensing"

Overview

ApSense is a novel deep learning approach for detecting Obstructive Sleep Apnea (OSA) events using fingertip Photoplethysmography (PPG) signals. The model demonstrates superior performance in OSA event recognition, particularly on high-variance datasets, while maintaining a relatively small model size.

Data Preparation

1. Data EDF Extraction

MESA Dataset

1. Request access to the MESA dataset
2. Download the polysomnography (PSG) EDF files
3. Use the provided 1_EDF_File_Extract_MESA.ipynb script to extract PPG signals:

HeartBEAT Dataset

1. Request access to the HeartBEAT dataset
2. Download the PSG EDF files
3. Use the provided 1_EDF_File_Extract_HeartBEAT.ipynb script to extract PPG signals:

2. Feature Extraction & Annotation

PPG Feature Extraction

Run the feature extraction script to process the raw PPG signals: 2_Data_prep_MESA.ipynb and 2_Data_prep_HeartBEAT.ipynb This script extracts the following features:

• Pulse Wave Amplitude (PWA)
• PP Interval (PPI)
• Derivative of PWA (dPWA)
• Derivative of PPI (dPPI)
• Systolic Phase Duration (SPD)
• Diastolic Phase Duration (DPD)
• Pulse Area (PA)

Annotation Process

The annotation process:

1. Segments PPG signals into 60-second windows with 50% overlap
2. Marks OSA events based on specialist annotations
3. Aligns annotations with PPG windows

Model Training and Evaluation

Training

Train the ApSense model using:

python -u main.py \
	--dataset mesa \
	--model $YOUR_MODEL \
	--dataset_dir $PATH_TO_YOUR_PROCESSED_DATASET \
	--log_dir $LOG_DIR \
	--weight_dir $WEIGHT_DIR \
	--subsampling \
	--gpu GPU_NUMBER \
	> "stdout/mesa_${YOUR_MODEL}_aug.log" &

Configuration options in models/dsepnet.py:

• Number of RVarDSepBlocks (M)
• LSTM Variational Dropout settings
• Branch CNN configuration

Evaluation

Evaluate the trained model:

python evaluate.py --dataset mesa \
                   --model $YOUR_MODEL \
                   --dataset_dir $PATH_TO_YOUR_PROCESSED_DATASET 

The evaluation script reports:

• Accuracy
• Macro F1 Score
• Sensitivity
• Specificity
• AUROC

Citation

If you find this work useful, please cite our paper:

@article{choksatchawathi2024apsense,
  title={ApSense: Data-Driven Algorithm in PPG-Based Sleep Apnea Sensing},
  author={Choksatchawathi, Tanut and Sawadwuthikul, Guntitat and Thuwajit, Punnawish and Kaewlee, Thitikorn and Mateepithaktham, Thee and Saisaard, Siraphop and Sudhawiyangkul, Thapanun and Chaitusaney, Busarakum and Saengmolee, Wanumaidah and Wilaiprasitporn, Theerawit},
  journal={IEEE Internet of Things Journal},
  year={2024},
  publisher={IEEE}
}

Acknowledgments

We thank the MESA and HeartBEAT studies for providing the datasets used in this research.