Detecção automática da apneia do sono em sinais de ECG utilizando aprendizado de máquina

Abstract

Obstructive Sleep Apnea (OSA) is a respiratory disorder associated with high cardiovascular morbidity and mortality, as well as significant costs for healthcare system. The gold standard diagnostic test, polysomnography, is expensive, time-consuming, and requires specialized infrastructure, which limits its large-scale use. In this context, the single-lead electrocardiogram (ECG) emerges as a promising alternative for automated screening, as it indirectly captures cardiorespiratory changes associated with apneic events. This dissertation proposes an automated method for OSA detection based on nocturnal ECG signals from the public Apnea-ECG (PhysioNet) database. The raw signal is processed through a preprocessing pipeline that includes noise detection and removal, cascaded filtering (median filter, Savitzky-Golay, and high-pass Butterworth) and Z-score normalization, while preserving ECG morphology. Then, 60 seconds segments are transformed into time-frequency spectrograms using STFT and enhanced with the CLAHE technique, generating 2D images used as input to modern deep learning architectures: ConvNext-Tiny, ConvNextV2-tiny, Swin transformer-Tiny. The imbalance between apnea/normal classes is handled with focal loss function, and ensemble strategies among models are also evaluated. Accuracy, sensitivity, specificity, and F1-score are used to assess performance In the experimental evaluation, the proposed ensemble with weighted voting achieved an accuracy of 95,9%, sensitivity of 97,4%, specificity of 93,5%, and F1-score of approximately 97%, outperforming most related works on the Apnea-ECG database and highlighting the potential of the approach for automatic OSA screening systems.