PSG-Free Multi-View Facial Imaging and Attention-Based Fusion for OSA Severity Classification

This is the official implementation of the paper:
"PSG-Free Multi-View Facial Imaging and Attention-Based Fusion for OSA Severity Classification" Submitted to Expert Systems with Applications (ESWA).

---

📌 Overview

This repository provides the code for a two-step training strategy and an attention-based fusion model to classify Obstructive Sleep Apnea (OSA) severity using multi-view facial images. Our approach eliminates the need for Polysomnography (PSG) by leveraging deep learning architectures (ResNet, EfficientNet, ViT, Swin) to identify severe OSA (AHI ≥ 30).

🚀 Key Features

• Multi-View Fusion: Integration of Frontal, Lateral, and Neck views.
• Attention-Based Fusion: Adaptive weighting of features from different views for enhanced classification.
• Reproducibility: Controlled random seeds for both batch shuffling and data augmentation (spatial & spectral consistency).

---

🛠 Installation

# Clone the repository
git clone https://github.com/dongyyyyy/Image-based-OSA-Classification.git
cd Image-based-OSA-Classification

# Run single-view training 
sh shell_files/train/benchmark/train_single_view_image.sh

• If you want to change the model architecture and hyper-parameters, please change the shell file!

# Run multi-view training
sh shell_files/train/benchmark/train_attention_mutli_view_image_three.sh

• In the shell file, you have to insert "model weight path" to load pretrained model weight parameter that is trained using 'single-view' image.

📂 Project Structure

.
├── config/            # config files directory (argparse)
├── model/             # Backbone architectures (ResNet, EfficientNet, ViT, Swin, multi-view attention)
├── utils/             # Data augmentation, dataloader and utility functions
├── shell_files        # shell files to train the model
├── train              # Train the model
└── test               # Verify the model

@article{
  title={PSG-Free Multi-View Facial Imaging and Attention-Based Fusion for OSA Severity Classification},
  author={Dongyoung Kim, Yunhee Woo, Jihoon Park, Jaemin Jeong, Seunghun Oh, Youngwoong Ko, Il-Hwan Lee, Dong-Kyu Kim and Jeong-Gun Lee},
  journal={Expert Systems with Applications},
  year={2026.03.01}
}