Go to IEEE EMBS BHI 2025 Conference homepageDeveloping An Attention-Based Deep Learning Framework for Obstructive Sleep Apnea Detection Using Single-Channel Oximetry Signal
Confirmation: I have read and agree with the IEEE BHI 2025 conference submission's policy on behalf of myself and my co-authors.
Keywords: Obstructive sleep apnea, graph attention, multi-head attention, clinical decision system, oximetry signal
TL;DR: We present an interpretable deep learning model that accurately estimates AHI and classifies OSA severity using only SpO2 signals, offering a scalable alternative to full PSG.
Abstract: Obstructive Sleep Apnea (OSA) is a common yet underdiagnosed condition typically assessed using polysomnography, a resource-intensive procedure. Oximetry ($SpO_2$) offers a non-invasive, low-cost alternative for large-scale OSA screening. This study proposes an interpretable deep learning framework for estimating the Apnea-Hypopnea Index (AHI) and classifying OSA severity using only single-channel $SpO_2$ signals. The model integrates convolution, bidirectional long short-term memory, graph attention networks, and multi-head attention to capture both local and global temporal patterns. Model predictions are interpreted using class-specific attention heatmaps and residual analysis. Evaluated on two large datasets (SHHS and CFS), the model achieved strong performance, with $R^2$ values ranging from 0.868 to 0.941 and outperformed baseline models across $R^2$, F1 score, sensitivity, and precision. Confusion matrices showed high classification accuracy for No Apnea and Severe cases, while scatter plot and Bland–Altman analyses confirmed low bias and stable predictions. These results demonstrate that $SpO_2$-based models can provide accurate and scalable AHI estimation, with attention-based visualizations enhancing interpretability and supporting clinical screening without the need for full PSG.
Track: 1. Biomedical Sensor Informatics
Registration Id: 36NYQ5QVKR9
Submission Number: 355
Loading