基于YOLO-1D架构的生物医学信号检测方法研究

Research on Biomedical Signal Detection Methods Based on the YOLO-1D Architecture

摘要: 深度学习技术在生物医学信号处理中显示出了巨大的潜力，基于卷积神经网络的方法应用于心电图(ECG)和光容积脉搏波图(PPG)有很好的效果。但是，大部分方法对于复杂情况下的信号质量变化、噪声干扰以及基线漂移并不能很好地处理，造成峰值检测的准确性以及漏检率不高。本文针对这种情况提出一种新颖的深度学习框架YOLO-1D，将2D的目标检测算法迁移到1D的生物医学信号处理上，在该方法的基础上完成对心电图的复杂场景下精准的多峰检测和心律失常分类，该方法的核心思想是采用YOLO架构的一步检测方式及多尺度特征融合，通过DFL分布聚焦损失定位精准的位置；并且为了满足心电图的特殊性进行了针对心电图的基线漂移、时间扭曲、信号倒置以及MixUp的数据增强策略，检测头采用可扩展分类支路的方法来同时学习峰值检测和心律失常分类，利用加权1D非极大值抑制(1DNMS)以及容忍度匹配机制准确地量化多峰检测效果。通过大量实验发现，YOLO-1D能有效利用基线漂移数据集以及心电图额外的数据源扩充其训练集，从而使网络性能更好，且检测的准确度非常高，在噪声和干扰的大环境下性能也能保持稳定。对于该模型来说：在MIT-BIH心律失常数据库上的ECGR峰检出测试上，取得了精确率达到了96.55%、召回率为98.00%、F1值达到了97.27%的测试结果；在AAMI五分类标准下ECG的心律失常分类任务，准确率为95.2%，宏平均F1值为92.6%；而在PPG峰值检测方面，精确率为95.8%，召回率为96.3%，F1值为96.0%，相比于传统基于scipy. signal. find_peaks的方法(精确率为90.2%，召回率为92.5%，F1分数为91.3分)，在精确率、召回率以及F1分数上分别提高了5.6、3.8及4.7个百分点。之后进行了模型升级更新，模型使用更新到YOLOv11版本，在后续实验中检测精度得到了一定的提高，其中包括ECG R峰检测F1达到了97.8%，PPG峰值检测F1也达到了96.5%。另外，还尝试做了脉搏间期(PPI)回归的任务，获得其在8秒窗口检测MAE达到0.42的较好结果。消融实验表明，在采用SPP模块、SEBlock注意力机制和数据增强等完整配置下，YOLO-1D的F1值比基础架构提升了2.17个百分点。

Abstract: Deep learning techniques have demonstrated great potential in the field of biomedical signal processing, and methods based on convolutional neural networks (CNNs) have achieved remarkable results in the analysis of electrocardiogram (ECG) and photoplethysmogram (PPG) signals. However, existing methods often perform unsatisfactorily in complex scenarios such as signal quality fluctuations, noise interference, and baseline drift, which tends to reduce the accuracy of peak detection and increase the false detection rate. In this study, we propose YOLO-1D, a novel deep learning framework that successfully adapts 2D object detection algorithms to 1D biomedical signal processing tasks, enabling high-precision peak detection and arrhythmia classification. Its core idea is to leverage the one-stage detection capability and multi-scale feature fusion mechanism of the YOLO architecture, combined with Distribution Focal Loss (DFL) to achieve accurate position regression; meanwhile, tailored data augmentation strategies (including baseline drift, time warping, signal inversion, and MixUp) are designed according to the characteristics of biomedical signals. The detection head can be extended with a classification branch to realize the joint learning of peak detection and arrhythmia classification. In addition, we introduce 1D Non-Maximum Suppression (1D NMS) and a tolerance matching mechanism to more accurately evaluate the performance of multi-peak detection. Experimental results show that the YOLO-1D method can generate high-quality peak detection results and maintain stable performance even in noisy and interfering environments. Specifically: In the ECG R-peak detection task on the MIT-BIH Arrhythmia Database, the precision, recall, and F1-score reach 96.55%, 98.00%, and 97.27%, respectively; in the ECG arrhythmia classification task under the AAMI five-classification standard, the accuracy and macro-average F1-score are 95.2% and 92.6%, respectively; in the PPG peak detection task, the precision, recall, and F1-score are 95.8%, 96.3%, and 96.0%, respectively. Compared with the traditional method based on scipy.signal.find_peaks (precision: 90.2%, recall: 92.5%, F1-score: 91.3%), the proposed method achieves improvements of 5.6, 3.8, and 4.7 percentage points in precision, recall, and F1-score, respectively. In subsequent research, the model is updated to the YOLOv11 version, further improving the detection accuracy: the F1-score for ECG R-peak detection reaches 97.8%, and that for PPG peak detection reaches 96.5%. In addition, we explore the Pulse Interval (PPI) regression task and achieve a good result with a Mean Absolute Error (MAE) of 0.42 in 8-second window detection. Ablation experiments show that when YOLO-1D adopts the complete configuration including the SPP module, SEBlock attention mechanism, and data augmentation, the F1-score is 2.17 percentage points higher than that of the basic architecture.