A self-supervised learning framework with hierarchical residual cross fusion network for sleep apnea detection

Sleep apnea is a common but frequently underdiagnosed respiratory disorder that poses serious health risks, including cardiovascular diseases and cognitive impairments. Despite recent advances in deep learning-based detection approaches, most methods heavily rely on large-scale, high-quality labeled data, which are costly and labor-intensive to acquire due to the need for expert annotation. To overcome this limitation, we propose a self-supervised learning framework with hierarchical residual cross fusion network (SSL-HRCNet) for sleep apnea detection using single-lead ECG signals. The framework operates through a two-stage pipeline: self-supervised pre-training followed by supervised fine-tuning. During pre-training, we apply a contrastive learning strategy that treats two physiologically meaningful transformed views, R-R intervals (RRIs) and R-peak amplitudes (R-peaks), extracted from the same ECG segment as positive pairs to learn cardiopulmonary representations from unlabeled signals. The feature encoder is further enhanced by hierarchical learnable residual blocks, which employ stacked depthwise-pointwise convolutional layers with progressively expanding receptive fields and learnable residual connections, facilitating multi-scale temporal modeling. During fine-tuning, an attention-based cross fusion module adaptively integrates the complementary representations of RRIs and R-peaks, improving the model's ability to discriminate apnea events. Extensive evaluations on the Apnea-ECG dataset demonstrate that SSL-HRCNet achieves competitive accuracy using only 10% labeled data, and outperforms existing approaches under full supervision with an accuracy of 91.91% and sensitivity of 89.97%. Moreover, the representations learned during pre-training on Apnea-ECG transfer well to the UCDDB dataset, demonstrating robustness to domain shifts and adaptability to clinical applications.