MobileNetV2: A lightweight classification model for home-based sleep apnea screening

TL;DR

This work tackles the challenge of accessible, low-cost OSA screening by deploying a lightweight MobileNetV2 that fuses ECG-derived sleep-stage cues with respiratory-event analysis to estimate the apnea-hypopnea index ($AHI$). The method leverages ECG spectrograms for sleep staging and respiration-derived features to detect breathing abnormalities, with $AHI$ computed via a rule-based approach from respiratory signals. Validations on Apnea-ECG, UCDDB, and MIT-BIH datasets show high performance: overall OSA detection accuracy of $0.978$, respiratory-event ROC-AUC of $0.98$, and sleep-stage ROC-AUC $>0.85$ on UCDDB, along with a strong correlation between real and predicted $AHI$ ($r=0.8831$). These results indicate a promising path toward portable, wearable sleep apnea screening that could reduce screening costs and broaden access to home-based monitoring. The work also highlights areas for improvement, such as enhancing discrimination for Mild/Moderate OSA and exploring hybrid architectures and model optimization for deployment on resource-limited devices.

Abstract

This study proposes a novel lightweight neural network model leveraging features extracted from electrocardiogram (ECG) and respiratory signals for early OSA screening. ECG signals are used to generate feature spectrograms to predict sleep stages, while respiratory signals are employed to detect sleep-related breathing abnormalities. By integrating these predictions, the method calculates the apnea-hypopnea index (AHI) with enhanced accuracy, facilitating precise OSA diagnosis. The method was validated on three publicly available sleep apnea databases: the Apnea-ECG database, the UCDDB dataset, and the MIT-BIH Polysomnographic database. Results showed an overall OSA detection accuracy of 0.978, highlighting the model's robustness. Respiratory event classification achieved an accuracy of 0.969 and an area under the receiver operating characteristic curve (ROC-AUC) of 0.98. For sleep stage classification, in UCDDB dataset, the ROC-AUC exceeded 0.85 across all stages, with recall for Sleep reaching 0.906 and specificity for REM and Wake states at 0.956 and 0.937, respectively. This study underscores the potential of integrating lightweight neural networks with multi-signal analysis for accurate, portable, and cost-effective OSA screening, paving the way for broader adoption in home-based and wearable health monitoring systems.

Figures (13)

• Figure 1: preprocess for ECG signal
• Figure 2: Overnight SpO2 data for subject 19 in the UCDDB database
• Figure 3: Four examples of contaminated ECG segments, where: the red dot represents the maximum value of the 10-second segment and the blue asterisk represents the minimum value of the 10-second segment.
• Figure 4: Two examples of ECG segments with extracted QRS points.
• Figure 5: Four images are provided to compare the synchronization between respiratory waveforms and EDR signals. (a) and (b) illustrate comparisons for abnormal respiratory events, while (c) and (d) show comparisons for normal respiratory events.