TF-HyperNet: Sleep apnea detection using dual-stream time–frequency fusion and a soft-margin hyperbox decision head

Abstract

Sleep apnea (SA) is a prevalent sleep-related breathing disorder that poses a substantial threat to public health. Deep learning-based ECG detection methods hold considerable promise for clinical applications. However, real-world clinical monitoring is complicated by both class imbalance and heterogeneous pathological signal morphologies. To address these challenges, we propose TF-HyperNet, a classification framework that integrates multiscale time–frequency representations with a soft-margin hyperbox decision head (SHBDH). The framework employs a dual-stream time–frequency encoder to simultaneously extract multiscale temporal and spectral representations from ECG-derived RRI and AMP sequences, yielding complementary temporal and frequency representations. For classification, a multi-center mapping module partitions abnormal samples with high intra-class variability into multiple subsets. The SHBDH then defines subclass-specific hyperbox decision regions with explicit lower and upper bounds and aggregates their subclass containment margins using LogSumExp to obtain class-level scores. To reduce decision-region overlap under imbalanced data distributions, we further introduce dimension-wise overlap regularization based on the mean overlap length across individual feature dimensions. By penalizing cross-class overlap along individual feature dimensions, this regularization promotes more separable subclass decision regions. Across 10 repeated experiments on the PhysioNet Apnea-ECG Dataset, TF-HyperNet achieved an average classification accuracy of 92.55 ± 0.21 % , a sensitivity of 90.08 ± 0.42 % , and a specificity of 94.08 ± 0.27 % . These findings indicate that TF-HyperNet provides an effective approach to ECG-based SA detection under the evaluated benchmark setting.