ObjectiveTo develop a transfer learning-based classification model for sleep apnea syndrome using electrocardiogram (ECG) data, increasing its classification accuracy and clinical applicability.MethodsBased on the Apnea-ECG and MIT-BIH polysomnographic databases, with respiratory signals as input, we applied a Butterworth low-pass filter for denoising, and constructed an original data set. To address the problem of insufficient respiratory signal data, a model training method based on a transfer learning approach was proposed: first, ECG signals with a large sample size were used for model pre-training, and then they were fine-tuning for respiratory signals, finally fulfilling binary classification or multi-class classification tasks. A cascade model combining residual network and bidirectional long short-term memory network was proposed, which performed better in capturing the timing features of signals and improving classification performance. Additionally, the performance of this model was made comparative analysis with those of various classic convolutional neural networks.ResultsThrough comparative experiments, it was found that employing transfer learning approach could accelerate model convergence and improve the model's overall performance. Validated on the test set, the proposed cascade model demonstrated a favorable performance in both binary classification and multi-class classification tasks, achieving an accuracy of 95.43% on the binary classification task and 91.26% on the multi-class classification task.ConclusionThis study offers novel insights into the design of disease classification models under small-sample conditions, and validates the effectiveness of transfer learning in sleep apnea syndrome classification, thereby demonstrating its potential clinical utility.