Nowadays, deep learning-based models have been widely developed for atrial fibrillation (AF) detection in electrocardiogram (ECG) signals. However, owing to the inevitable over-fitting problem, classification accuracy of the developed models severely differed when applying on the independent test datasets. This situation is more significant for AF detection from dynamic ECGs. In this study, we explored two potential training strategies to address the over-fitting problem in AF detection. The first one is to use the Fast Fourier transform (FFT) and Hanning-window-based filter to suppress the influence from individual difference. Another is to train the model on the wearable ECG data to improve the robustness of model. Wearable ECG data from 29 patients with arrhythmia were collected for at least 24 h. To verify the effectiveness of the training strategies, a Long Short-Term Memory (LSTM) and Convolution Neural Network (CNN)-based model was proposed and tested. We tested the model on the independent wearable ECG data set, as well as the MIT-BIH Atrial Fibrillation database and PhysioNet/Computing in Cardiology Challenge 2017 database. The model achieved 96.23%, 95.44%, and 95.28% accuracy rates on the three databases, respectively. Pertaining to the comparison of the accuracy rates on each training set, the accuracy of the model trained in conjunction with the proposed training strategies only reduced by 2%, while the accuracy of the model trained without the training strategies decreased by approximately 15%. Therefore, the proposed training strategies serve as effective mechanisms for devising a robust AF detector and significantly enhanced the detection accuracy rates of the resulting deep networks.

如今，基于深度学习的模型已被广泛开发用于心电图 (ECG) 信号中的房颤 (AF) 检测。然而，由于不可避免的过拟合问题，所开发模型的分类精度在应用于独立测试数据集时存在严重差异。这种情况对于动态心电图的 AF 检测更为重要。在这项研究中，我们探索了两种潜在的训练策略来解决 AF 检测中的过度拟合问题。第一个是使用快速傅里叶变换（FFT）和基于汉宁窗的滤波器来抑制个体差异的影响。另一种是在可穿戴心电数据上训练模型，以提高模型的鲁棒性。收集了 29 名心律失常患者的可穿戴 ECG 数据至少 24 小时。为了验证训练策略的有效性，提出并测试了基于长短期记忆 (LSTM) 和卷积神经网络 (CNN) 的模型。我们在独立的可穿戴 ECG 数据集以及 MIT-BIH 心房颤动数据库和 PhysioNet/Computing in Cardiology Challenge 2017 数据库上测试了该模型。该模型在三个数据库上分别达到了 96.23%、95.44% 和 95.28% 的准确率。关于每个训练集的准确率比较，结合提出的训练策略训练的模型的准确率仅下降了 2%，而没有训练策略训练的模型的准确率下降了大约 15%。所以，