Pulse Oximetry and Sleep Apnoea Detection using Deep Learning
Team
- E/19/324, Bimbara Rathnayake, e19324@eng.pdn.ac.lk
Table of Contents
Introduction
With the increasing availability of wearable health-monitoring devices, single-channel oximetry has emerged as a promising non-invasive tool for continuous health assessment. This project explores the use of deep learning to detect obstructive sleep apnoea (OSA) from overnight SpO₂ signals, using only pulse oximetry data, potentially enabling low-cost and remote diagnosis.
Problem Statement
Obstructive Sleep Apnoea (OSA) is a widespread but underdiagnosed condition. Polysomnography (PSG) is the current gold standard but is costly, time-consuming, and not widely accessible. This project aims to verify the performance of a deep learning-based model (OxiNet) in estimating the Apnoea-Hypopnea Index (AHI) from SpO₂ data and compare it with simpler machine learning approaches like linear regression and feature-engineered models.
Methodology
- Dataset: Publicly available sleep study databases from sleepdata.org (e.g., SHHS, MESA, MROS, CFS, UHV).
- Baseline Models:
- Linear Regression using ODI (Oxygen Desaturation Index)
- CatBoost Regressor using Oximetry Biomarkers (OBMs)
- Deep Learning Model:
- OxiNet: A hybrid CNN + CRNN architecture trained to estimate AHI.
- Trained using SHHS1 and evaluated on multiple external databases.
- Evaluation:
- Regression: ICC, Bland-Altman plots
- Classification: F1-Macro Score across 4 OSA severity levels
- Explainability: Feature Occlusion method
Results and Discussion
- OxiNet outperformed traditional ML models with an ICC of 0.96 and F1-Macro of 0.84 on SHHS1.
- Baseline models missed 21% of moderate-severe OSA cases, while OxiNet missed only 0.2%.
- Robustness was analyzed across demographics (age, sex, ethnicity) and comorbidities (e.g., COPD).
- Performance dipped on minority groups, highlighting the importance of representative training data.
Future Work
- Explore novel embedding techniques (e.g., wavelet, autoencoders, transformers).
- Implement real-time OSA detection pipeline.
- Evaluate model on newer datasets and different pulse oximeter hardware.
- Address bias and fairness issues across populations.