Related Experiment Videos
Explainable CNN-BiLSTM Framework for Multi-Class Sleep Apnea Severity Detection Using Single-Lead ECG Signals: A
Fida'a Al-Quran1, Malik Jawarneh2, Omar Isam Al-Mrayat3
1Engineering and Artificial Intelligence Department, Al-Salt Technical College, Al-Balqa Applied University, Al-Salt 19117, Jordan.
Diagnostics (Basel, Switzerland)
|August 13, 2026
Summary
A new explainable deep learning model accurately classifies obstructive sleep apnea (OSA) severity using electrocardiogram (ECG) signals. This approach offers a computationally efficient and interpretable alternative to traditional diagnostic methods like polysomnography (PSG).
Area of Science:
- Artificial Intelligence in Medicine
- Cardiology
- Sleep Medicine
Background:
- Obstructive sleep apnea (OSA) affects over 936 million adults globally and remains underdiagnosed due to limitations of polysomnography (PSG).
- Current diagnostic methods for OSA, such as PSG, are often costly, time-consuming, and inaccessible in many healthcare settings.
- There is a need for accessible and efficient methods for OSA diagnosis and severity classification.
Purpose of the Study:
- To develop and validate a novel explainable deep learning (DL) framework for automated multi-class obstructive sleep apnea (OSA) severity classification.
- To utilize single-lead electrocardiogram (ECG) signals for OSA severity assessment.
- To enhance diagnostic interpretability and clinical trust through explainable artificial intelligence (XAI) techniques.
Main Methods:
- A hybrid CNN-BiLSTM deep learning architecture was integrated with explainable AI (XAI) techniques.
- The framework was trained and validated on a combined dataset of 220 recordings from PhysioNet Apnea-ECG and an institutional ECG dataset.
- SHAP (SHapley Additive exPlanations) was used to identify key physiological predictors of OSA severity.
Main Results:
- The proposed DL framework achieved 94.7% overall accuracy, 92.3% sensitivity, and 96.1% specificity in classifying OSA severity.
- The model demonstrated superior performance compared to conventional machine learning algorithms (SVM, Random Forest, XGBoost).
- Heart rate variability features, specifically RMSSD and pNN50, were identified as strong indicators of OSA severity, with processing time of 0.23s per 60s ECG epoch.
Conclusions:
- Explainable deep learning applied to ECG signals provides an accurate, interpretable, and computationally efficient method for assessing OSA severity.
- The developed framework shows potential for supporting OSA screening, clinical triage, and early intervention, especially in resource-limited settings.
- This approach can improve OSA diagnosis accessibility and efficiency, addressing limitations of traditional methods like PSG.