融合频域注意力与时空特征的睡眠呼吸事件分类识别模型OA

Accurate classification of sleep respiratory events is crucial for diagnosing Obstructive Sleep Apnea(OSA)and Central Sleep Apnea(CSA).However,existing methods often suffer from high-frequency noise interference and insufficient coverage of pathologically sensitive frequency bands during time-frequency feature extraction and fusion.To address this issue,we propose CBFA,a classification algorithm that integrates frequency-domain attention with spatio-temporal features.The model adopts a hi-erarchical deep neural network architecture,where an improved CNN14 extracts discriminative frequency-domain features and BiLSTM captures long-range temporal dynamics.A frequency-domain attention module(FreqAttention)is further introduced to enhance representations within the 0~6000 Hz pathological band.Experiments on the PSG-Audio multimodal dataset demon-strate that CBFA algorithm effectively suppresses high-frequency noise and improves the representation of low-frequency fea-tures.Ablation studies confirm the contribution of BiLSTM to temporal modeling,increasing CSA recall from 0.71 to 0.83(+12 percentage points),and the benefit of FreqAttention for OSA detection,raising recall from 0.88 to 0.96(+8 points).Compared with existing approaches such as VGG_LSTM and CNRR,CBFA achieves superior performance,with an accuracy of 92%,a macro-average F1 score of 0.82,and a weighted F1 score of 0.93.The results indicate that CBFA algorithm can effectively reduce the impact of high-frequency noise,enhance the representation of pathologically sensitive frequency bands,and achieve superior performance in classifying OSA and CSA.This provides an efficient and reliable solution for automated audio-based detection of sleep respirat-ory events.