Abstract
Abstract
Study Objectives
Current multimodal approaches for automated sleep apnea detection usually assume continuous signal availability, although sensor dropout is common in clinical practice. We evaluated probability-level late fusion across signal configurations and missing-data conditions to determine whether performance depends more on signal complementarity or modality count.
Methods
We trained random forest base classifiers on EEG, ECG, SpO₂, and abdominal effort features from 148 patients. Each classifier generated three-class posterior probabilities, which were concatenated and input to a meta-classifier for final epoch-level prediction. Unreliable SpO₂ epochs were excluded only from SpO₂ base-model fitting but retained during meta-classifier training and testing using uniform placeholder probabilities [0.333, 0.333, 0.333] plus the isBad quality flag. Selected unimodal and multimodal configurations were evaluated in 37 held-out patients using macro-F1 as the primary metric. Bootstrap confidence intervals stratified by SpO₂ quality assessed robustness under dropout.
Results
Abdominal effort alone achieved macro-F1 = 0.996. SpO₂ + abdominal effort achieved the highest macro-F1 = 0.997, only marginally above abdominal effort alone. ECG + EEG outperformed SpO₂ + ECG and SpO₂ + EEG despite lacking direct respiratory or oxygenation information. SpO₂ + EEG showed degraded performance (macro-F1 = 0.739; hypopnea precision = 0.203). Adding SpO₂ to ECG + EEG reduced performance (0.872 vs 0.899). Performance remained stable despite 30.7% SpO₂ dropout.
Conclusions
Within probability-level late fusion, performance depended more on signal complementarity than modality count. Quality-aware probability integration enabled robust classification under realistic SpO₂ dropout without retraining, but fixed-epoch late fusion limited exploitation of temporally misaligned oximetry information.