Multimodal fusion of SpO₂ time series and clinical data for explainable classification of obstructive sleep apnea severity
Year: 2026
Type: article
Abstract: Obstructive Sleep Apnea Syndrome (OSAS) is a sleep disorder whose diagnosis and severity classification rely primarily on polysomnography (PSG); however, its use in large scale screening programmes is limited by high costs and restricted accessibility. In this study, we propose an interpretable machine learning framework for the automatic classification of OSAS severity by integrating overnight oxygen saturation (SpO₂) time series data with clinical and hematological variables at the feature level. In this retrospective study, 78 individuals who had undergone PSG were classified into four categories: simple snoring, mild, moderate, and severe OSAS. Continuous fingertip SpO₂ signals recorded at 16 Hz were divided into non overlapping 60 s segments. Time domain statistics and frequency domain features based on power spectral density estimated using the Welch method were extracted from each segment. Segment level features were aggregated at the individual level and combined through feature level (early) fusion with body mass index, Epworth Sleepiness Scale score, and selected haemogram parameters. The resulting feature vectors were evaluated using five fold stratified cross validation with Random Forest (RF), Extra Trees (ET), and k-Nearest Neighbour classifiers (kNN). Among the evaluated models, ET achieved the best performance, yielding an accuracy of 89.83% ± 5.45, a macro F1-score of 89.88% ± 5.60, and an ROC-AUC of 0.9881 ± 0.0131. In ten-times repeated five-fold cross-validation, an accuracy of 85.53% ± 2.40 was achieved, whereas nested cross-validation yielded an accuracy of 85.38% ± 1.93, indicating a modest reduction in performance under more rigorous validation. Ablation analyses showed that the principal improvement in performance resulted from the addition of PSD-based features together with BMI and ESS, whereas hematological parameters provided only limited incremental benefit in terms of accuracy and macro F1-score. SHAP-based explainability analysis showed that time- and frequency-domain features derived from the SpO₂ signal made the greatest relative contribution to model predictions. Clinical variables showed a moderate contribution, whereas hematological variables contributed to a lesser extent. These findings suggest that integrating SpO₂-derived features with clinical data may support the automated and interpretable classification of OSAS severity and complement clinical assessment.
Source: Scientific Reports
Language: English
Subfield: Physiology (medical)
Field: Medicine
Domain: Health Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7224105694