---

Pediatric-Sleep-Apnea-SpO₂-Analysis

This repository provides tools for preprocessing SpO₂ signals, organizing data for cross-validation, and evaluating deep learning models for assessing pediatric sleep apnea severity. The SpO₂_Preprocessing_CHAT notebook automates data downloads and structuring, while ResNet_CNNBiLSTMAttention.ipynb facilitates model training and evaluation. Below are setup instructions, usage details, and file structure guidelines.

---

Table of Contents

1. Introduction
2. File and Directory Structure
3. Setup and Installation
4. Usage
   • Data Preprocessing
   • Cross-Validation Preparation
   • Model Training and Evaluation
5. Additional Notes

---

Introduction

This repository supports the research publication "Deep learning-enabled analysis of overnight peripheral oxygen saturation signals for pediatric obstructive sleep apnea severity assessment". It enables deep learning model evaluation for assessing pediatric sleep apnea severity using SpO₂ signals. The SpO₂_Preprocessing_CHAT notebook handles data extraction, event detection, and segmentation, while ResNet_CNNBiLSTMAttention evaluates ResNet and CNN-BiGRU-Attention models.

---

File and Directory Structure

Required files and directories before running the notebooks:

1. Demographic and Polysomnographic Information Files:

   • chat-baseline-dataset-0.12.0.csv
   • chat-followup-dataset-0.12.0.csv

2. XML Folder (in ./CHAT_dataset/XML):

   • Stores XML annotation files for apnea and oxygen desaturation events:

   XML/
   ├── baseline/    # Baseline group XML files
   └── followup/    # Follow-up group XML files
   

3. CSV Folder (created in Google Drive or ./CHAT_dataset/CSV):

   • Saves apnea and desaturation event details in CSV format:

   CSV/
   ├── ap_ends/     # Apnea event end times
   ├── od_nadirs/   # Oxygen desaturation nadir times
   └── od_starts/   # Oxygen desaturation start times
   

4. Processed Data Folder:

   • Stores numpy arrays with preprocessed data:

   processed_data/
   ├── AHIs.npy      # AHI values per signal
   ├── IDs.npy       # Subject IDs
   └── data.npy      # 20-minute SpO₂ segments with indexing
   

   • Each row in data.npy contains a 20-minute segment of a subject’s SpO₂ signal along with the subject index.
   • Note: Subject IDs may appear in both baseline and follow-up groups if the subject participated in both studies, but each occurrence will have a unique subject index to track individual segments accurately.

5. Folds Folder:

   • Contains preprocessed data used in this publication, organized for 3-fold cross-validation. You can download a zip file of this folder here.
   • Each fold contains training, validation, and test data arrays for evaluation:

   Folds/
   ├── fold_1/
   │   ├── OCM_testAHI1.npy 
   │   ├── OCM_testArray1.npy 
   │   ├── OCM_testIDs1.npy 
   │   ├── OCM_trainAHI1.npy 
   │   ├── OCM_trainArray1.npy 
   │   ├── OCM_trainIDs1.npy 
   │   ├── OCM_valAHI1.npy 
   │   ├── OCM_valArray1.npy 
   │   └── OCM_valIDs1.npy
   ├── fold_2/ (similar structure)
   └── fold_3/ (similar structure)
   

6. Models Folder:

   • Contains pretrained weights for each fold:

      Models/
      ├── TensorFlow/
      │   ├── attention_model_fold_1_weights.h5
      │   ├── attention_model_fold_2_weights.h5
      │   ├── attention_model_fold_3_weights.h5
      │   ├── resnet_model_fold_1_weights.h5
      │   ├── resnet_model_fold_2_weights.h5
      │   └── resnet_model_fold_3_weights.h5
      └── PyTorch/
          └── attention_model_pytorch_fold_2.pth     # PyTorch weights for attention extraction
   

---

Setup and Installation

1. Open a new notebook in Google Colab.
2. Clone the repository:

   git clone https://github.com/erfan-mtzv/Pediatric-Sleep-Apnea-SpO2-Analysis.git

3. Move .ipynb files to your Drive:

   !mkdir -p /content/drive/MyDrive/project/
   !cp -r /content/Pediatric-Sleep-Apnea-SpO2-Analysis/* /content/drive/MyDrive/project/

4. Open each notebook you need with Google Colaboratory.
5. Ensure directory paths in notebooks are updated to correctly reference all files.

---

Usage

1. Data Preprocessing

SpO₂_Preprocessing_CHAT Notebook:

• Open SpO₂_Preprocessing_CHAT.ipynb in Google Colab.
• Set the webpage number (1-5) and dataset group (baseline or followup).
• Execute cells to download .edf files, detect apnea/desaturation events, and save processed data in .npy format in Google Drive:
  • AHI arrays: Apnea-Hypopnea Index values per signal.
  • ID arrays: Subject IDs.
  • Data arrays: 20-minute segmented SpO₂ signals.

2. Cross-Validation Preparation

• After preprocessing all pages for both groups, upload processed data for cross-validation.
• Merge .npy arrays into structured files for n-fold cross-validation in the Folds folder.

3. Model Training and Evaluation

The ResNet_CNNBiLSTMAttention.ipynb notebook includes:

• ResNet for feature extraction (TensorFlow).
• CNN-BiGRU-Attention for interpretability (TensorFlow and PyTorch).

Steps:

1. Load Pretrained Weights:

   • Upload pretrained weights from ./Models for each fold.

2. Evaluate Model Performance:

   • Run evaluations on cross-validation data with metrics like 4-class accuracy, RMSE, and R².

3. Attention Score Visualization:

   • Load .pth weights for CNN-BiGRU-Attention in PyTorch.
   • Extract and plot attention scores to observe model focus on SpO₂ signal features.

---

Additional Notes

• Carefully follow the table of contents and comments in the notebooks for effective usage.
• Interpretability: Attention visualizations offer insights into model focus on SpO₂ signal features.

This repository is a complete toolkit for processing SpO₂ data and evaluating deep learning models for pediatric sleep apnea research with interpretable model output.

---

PSG Variables from the CHAT Dataset for Validating the Labeling Process

The following table summarizes key polysomnographic (PSG) variables related to sleep stages and apnea events from the CHAT dataset:

Category	Variable	Description
Obstructive Apneas	oanbp3	Total number of obstructive apneas in NREM (supine), ≥ 3% desaturation, with/without arousal.
	oanop3	Total number of obstructive apneas in NREM (non-supine), ≥ 3% desaturation, with/without arousal.
	oarbp3	Total number of obstructive apneas in REM (supine), ≥ 3% desaturation, with/without arousal.
	oarop3	Total number of obstructive apneas in REM (non-supine), ≥ 3% desaturation, with/without arousal.
Hypopneas	hnrbp3	Total number of hypopneas in NREM (supine), > 50% flow reduction, ≥ 3% desaturation, with/without arousal.
	hnrop3	Total number of hypopneas in NREM (non-supine), > 50% flow reduction, ≥ 3% desaturation.
	hrembp3	Total number of hypopneas in REM (supine), > 50% flow reduction, ≥ 3% desaturation.
	hrop3	Total number of hypopneas in REM (non-supine), > 50% flow reduction, ≥ 3% desaturation.
Central Apneas	carbp3	Total number of central apneas in REM (supine), ≥ 3% desaturation, with/without arousal.
	carop3	Total number of central apneas in REM (non-supine), ≥ 3% desaturation, with/without arousal.
	canbp3	Total number of central apneas in NREM (supine), ≥ 3% desaturation, with/without arousal.
	canop3	Total number of central apneas in NREM (non-supine), ≥ 3% desaturation, with/without arousal.
Mixed Apneas	marbp3	Total number of mixed apneas in REM (supine), ≥ 3% desaturation, with/without arousal.
	marop3	Total number of mixed apneas in REM (non-supine), ≥ 3% desaturation, with/without arousal.
	manrbp3	Total number of mixed apneas in NREM (supine), ≥ 3% desaturation, with/without arousal.
	manrop3	Total number of mixed apneas in NREM (non-supine), ≥ 3% desaturation, with/without arousal.

---

Repository Description:
Repository for "Deep learning-enabled analysis of overnight peripheral oxygen saturation signals for pediatric obstructive sleep apnea severity assessment". It includes data preprocessing, segmentation, and deep learning model evaluation pipelines for SpO₂-based pediatric OSA severity assessment.