一种基于深度特征融合的可解释性12导联心电图自动诊断模型研究

doi:10.12122/j.issn.1673-4254.2026.01.23

摘要/Abstract

摘要：

目的提升12导联心电图（ECG）自动诊断的准确性和可信度。方法提出了一种基于深度特征融合的12导联ECG自动诊断模型（MRHL-ECGNet）。该模型包含多尺度特征提取前端、ResNet-34、全局特征混合模块及时间序列分析模块，首次将Hyena Hierarchy卷积算子应用于12导联心电图自动诊断任务中，以高效捕捉ECG中的长程依赖关系，并显著降低模型计算复杂度。同时采用基于积分梯度（IG）的可解释性分析技术，实现MRHL-ECGNet决策依据可视化。使用CPSC2018数据集对MRHL-ECGNet进行训练和测试，并采用多项定量评价指标与评估实验对MRHL-ECGNet进行全面评估。结果在测试集上对9种类别ECG的分类任务中，MRHL-ECGNet的准确率、AUC值、F1分数、精确率和召回率分别达到0.972、0.983、0.864、0.873和0.857，均优于其他对比模型，且在GPU上对单样本输出诊断结果所需的时间为0.007s，在CPU上也仅需0.156s，内存占用为67.196MB。结论本研究提出的MRHL-ECGNet不仅具有卓越的分类性能，还具备轻量化及可解释性的特点，在临床ECG辅助诊断中具有较高的应用价值。

关键词: 心电图自动诊断, 深度学习, Hyena Hierarchy卷积算子, 模型可解释性

Abstract:

Objective To enhance the accuracy and reliability of 12-lead electrocardiogram (ECG) automatic diagnosis. Methods Herein we propose a 12-lead ECG automatic diagnosis model based on deep feature fusion (MRHL-ECGNet), which consists of a multi-scale feature extraction front-end, ResNet-34, a global feature mixing module, and a time-series analysis module. The Hyena Hierarchy Convolution Operator was applied to the 12-lead ECG automatic diagnosis task for more efficient capture of long-range dependencies while reducing computational complexity. Integrated Gradients (IG)-based interpretability analysis technology was used to achieve visualization of the decision-making basis of MRHL-ECGNet. The CPSC2018 dataset was used to train and test MRHL-ECGNet, and its performance was assessed using multiple quantitative evaluation indicators and evaluation experiments. Results In the 9-class ECG classification task on the test set, MRHL-ECGNet achieved an accuracy of 0.972, an AUC of 0.983, an F1 score of 0.864, a precision of 0.873, and a recall of 0.857, all surpassing other comparative models. This model only took 0.007 s to output a diagnosis for a single sample on a GPU and 0.156 s on a CPU, with a memory footprint of 67.196 MB. Conclusion The proposed MRHL-ECGNet model demonstrates excellent classification performance in 12-lead ECG automatic diagnosis with a lightweight design and good interpretability, and thus has great potential for clinical application in ECG-aided diagnosis.

Key words: electrocardiogram automatic diagnosis, deep learning, Hyena Hierarchy Convolution Operator, interpretability of model

卢学麒, 陈华元, 吴秋岑, 温耀棋, 刘国光, 陈超敏. 一种基于深度特征融合的可解释性12导联心电图自动诊断模型研究[J]. 南方医科大学学报, 2026, 46(1): 208-218.

Xueqi LU, Huayuan CHEN, Qiucen WU, Yaoqi WEN, Guoguang LIU, Chaomin CHEN. Evaluation of an interpretable 12-lead ECG automatic diagnosis model based on deep feature fusion[J]. Journal of Southern Medical University, 2026, 46(1): 208-218.

图/表 11

参考文献 31

[1]	Berkaya SK, Uysal AK, Gunal ES, et al. A survey on ECG analysis[J]. Biomed Sig Proc Control, 2018, 43: 216-35. doi：10.1016/j.bspc.2018.03.003
[2]	Chou's Electrocardiography in Clinical Practice[M]. Elsevier Science Health Science div,2008 . doi：10.1016/b978-141603774-3.10027-9
[3]	Hong SD, Zhou YX, Shang JY, et al. Opportunities and challenges of deep learning methods for electrocardiogram data: a systematic review[J]. Comput Biol Med, 2020, 122: 103801. doi：10.1016/j.compbiomed.2020.103801
[4]	Priori SG, Blomström-Lundqvist C, Mazzanti A, et al. 2015 ESC Guidelines for the management of patients with ventricular arrhythmias and the prevention of sudden cardiac death: The Task Force for the Management of Patients with Ventricular Arrhythmias and the Prevention of Sudden Cardiac Death of the European Society of Cardiology (ESC). Endorsed by: Association for European Paediatric and Congenital Cardiology (AEPC)[J]. Eur Heart J, 2015, 36(41): 2793-867. doi：10.1093/eurheartj/ehv316
[5]	Hannun AY, Rajpurkar P, Haghpanahi M, et al. Cardiologist-level arrhythmia detection and classification in ambulatory electro-cardiograms using a deep neural network[J]. Nat Med, 2019, 25(1): 65-9. doi：10.1038/s41591-018-0268-3
[6]	Yang XZ, Zhang XY, Yang MY, et al. 12-Lead ECG arrhythmia classification using cascaded convolutional neural network and expert feature[J]. J Electrocardiol, 2021, 67: 56-62. doi：10.1016/j.jelectrocard.2021.04.016
[7]	Vaswani A, Shazeer NM, Parmar N, et al. Attention is all you need[C]//Neural Information Processing Systems., 2017
[8]	Dosovitskiy A, Beyer L, Kolesnikov A, et al. An image is worth 16x16 words: Transformers for image recognition at scale[J]. arxiv preprint arxiv: 2010. 11929, 2020.
[9]	Zhou HY, Zhang SH, Peng JQ, et al. Informer: beyond efficient transformer for long sequence time-series forecasting[J]. Proc AAAI Conf Artif Intell, 2021, 35(12): 11106-15. doi：10.1609/aaai.v35i12.17325
[10]	Lee J, Yoon W, Kim S, et al. BioBERT: a pre-trained biomedical language representation model for biomedical text mining[J]. Bioinformatics, 2020, 36(4): 1234-40. doi：10.1093/bioinformatics/btz682
[11]	Zhang SY, Lian C, Xu BR, et al. A token selection-based multi-scale dual-branch CNN-transformer network for 12-lead ECG signal classification[J]. Knowl Based Syst, 2023, 280: 111006. doi：10.1016/j.knosys.2023.111006
[12]	Zheng JW, Zhang JM, Danioko S, et al. A 12-lead electrocardiogram database for arrhythmia research covering more than 10, 000 patients[J]. Sci Data, 2020, 7(1): 48. doi：10.1038/s41597-020-0386-x
[13]	Petch J, Di S, Nelson W. Opening the black box: the promise and limitations of explainable machine learning in cardiology[J]. Can J Cardiol, 2022, 38(2): 204-13. doi：10.1016/j.cjca.2021.09.004
[14]	Zhang DD, Yang S, Yuan XH, et al. Interpretable deep learning for automatic diagnosis of 12-lead electrocardiogram[J]. iScience, 2021, 24(4): 102373. doi：10.1016/j.isci.2021.102373
[15]	Lundberg SM, Lee SI. A unified approach to interpreting model predictions[C]//Neural Information Processing Systems., 2017
[16]	Reddy L, Talwar V, Alle S, et al. IMLE-net: an interpretable multi-level multi-channel model for ECG classification[C]//2021 IEEE International Conference on Systems, Man, and Cybernetics (SMC). October 17-20, 2021. Melbourne, Australia. IEEE, 2021: 1068-1074. doi：10.1109/smc52423.2021.9658706
[17]	Liu FF, Liu CY, Zhao LN, et al. An open access database for evaluating the algorithms of electrocardiogram rhythm and morphology abnormality detection[J]. J Med Imaging Hlth Inform, 2018, 8(7): 1368-73. doi：10.1166/jmihi.2018.2442
[18]	Zhang HZ, Zhao W, Liu S. SE-ECGNet: a multi-scale deep residual network with squeeze-and-excitation module for ECG signal classification[C]//2020 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). December 16-19, 2020. Seoul, Korea. IEEE, 2020: 2685-2691. doi：10.1109/bibm49941.2020.9313548
[19]	He KM, Zhang XY, Ren SQ, et al. Deep residual learning for image recognition[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). June 27-30, 2016, Las Vegas, NV, USA. IEEE, 2016: 770-8. doi：10.1109/cvpr.2016.90
[20]	Poli M, Massaroli S, Nguyen E, et al. Hyena hierarchy: Towards larger convolutional language models[C]//International Conference on Machine Learning. PMLR, 2023: 28043-78.
[21]	Yang W, Deo R, Guo WS. Functional feature extraction and validation from twelve-lead electrocardiograms to identify atrial fibrillation[J]. Commun Med (Lond), 2025, 5(1): 32. doi：10.1038/s43856-025-00749-2
[22]	Hochreiter S, Schmidhuber J. Long short-term memory[J]. Neural Comput, 1997, 9(8): 1735-80. doi：10.1162/neco.1997.9.8.1735
[23]	Kingma D P, Ba J. Adam: A method for stochastic optimization[J]. arxiv preprint arxiv:, 2014.
[24]	Barredo Arrieta A, Díaz-Rodríguez N, Del Ser J, et al. Explainable artificial intelligence (XAI): concepts, taxonomies, opportunities and challenges toward responsible AI[J]. Inf Fusion, 2020, 58: 82-115. doi：10.1016/j.inffus.2019.12.012
[25]	Sundararajan M, Taly A, Yan Q. Gradients of counterfactuals[J]. arxiv preprint arxiv:, 2016.
[26]	Sokolova M, Lapalme G. A systematic analysis of performance measures for classification tasks[J]. Inf Process Manag, 2009, 45(4): 427-37. doi：10.1016/j.ipm.2009.03.002
[27]	Hwang S, Cha J, Heo J, et al. Multi-label abnormality classification from 12-lead ECG using a 2D residual U-net[C]//ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). April 14-19, 2024, Seoul, Korea, Republic of. IEEE, 2024: 2265-9. doi：10.1109/icassp48485.2024.10448259
[28]	Strodthoff N, Wagner P, Schaeffter T, et al. Deep learning for ECG analysis: benchmarks and insights from PTB-XL[J]. IEEE J Biomed Health Inform, 2021, 25(5): 1519-28. doi：10.1109/jbhi.2020.3022989
[29]	Conen D, Adam M, Roche F, et al. Premature atrial contractions in the general population: frequency and risk factors[J]. Circulation, 2012, 126(19): 2302-8. doi：10.1161/circulationaha.112.112300
[30]	Joglar JA, Chung MK, Armbruster AL, et al. 2023 ACC/AHA/ACCP/HRS guideline for the diagnosis and management of atrial fibrillation: a report of the American college of cardiology/American heart association joint committee on clinical practice guidelines[J]. Circulation, 2024, 149(1): e1-156. doi：10.1161/cir.0000000000001207
[31]	Ikeda T. Right bundle branch block: current considerations[J]. Curr Cardiol Rev, 2021, 17(1): 24-30. doi：10.2174/1573403x16666200708111553

Module	In/Out & K and Other Parameters
Input Layer	In: 12×T
Multi-scale feature extraction frontend	In: 12×T; Out: 64×T/2; K: 7, 15, 31; Stride: 2
ResNet-34 Layer1	In: 64; Out: 64; K: 7; BasicBlock1d: 3; Downsample: No
ResNet-34 Layer2	In: 64; Out: 128; K: 7; BasicBlock1d: 4; Downsample: Yes
ResNet-34 Layer3	In: 128; Out: 256; K: 7; BasicBlock1d: 6; Downsample: Yes
ResNet-34 Layer4	In: 256; Out: 512; K: 7; BasicBlock1d: 3; Downsample: Yes
Global feature mixing module	In: 512; Out: 512; K: 31
Temporal analysis module (LSTM)	In: 512; Hidden state: 64
Temporal analysis module (Hyena)	In: 64; Out: 64; K: 15
Feature fusion and classifier	Concatenated in: 512×2+64; Fully connected layer Output: Classification results

Module	In/Out & K and Other Parameters
Input Layer	In: 12×T
Multi-scale feature extraction frontend	In: 12×T; Out: 64×T/2; K: 7, 15, 31; Stride: 2
ResNet-34 Layer1	In: 64; Out: 64; K: 7; BasicBlock1d: 3; Downsample: No
ResNet-34 Layer2	In: 64; Out: 128; K: 7; BasicBlock1d: 4; Downsample: Yes
ResNet-34 Layer3	In: 128; Out: 256; K: 7; BasicBlock1d: 6; Downsample: Yes
ResNet-34 Layer4	In: 256; Out: 512; K: 7; BasicBlock1d: 3; Downsample: Yes
Global feature mixing module	In: 512; Out: 512; K: 31
Temporal analysis module (LSTM)	In: 512; Hidden state: 64
Temporal analysis module (Hyena)	In: 64; Out: 64; K: 15
Feature fusion and classifier	Concatenated in: 512×2+64; Fully connected layer Output: Classification results

Item	Configuration
Operation system	Windows 11
CPU	Intel Core i5-13490F
GPU	NVIDIA GeForce RTX 4060Ti
RAM	16GB
Cuda version	12.7
Pytorch version	2.3.1
Anaconda version	24.9.2

Item	Configuration
Operation system	Windows 11
CPU	Intel Core i5-13490F
GPU	NVIDIA GeForce RTX 4060Ti
RAM	16GB
Cuda version	12.7
Pytorch version	2.3.1
Anaconda version	24.9.2

Method	Accuracy	AUC	F1s	Precision	Recall
Zhang D et al.^[14]	0.963	0.961	0.823	0.826	0.821
Reddy L et al.^[16] (IMLE-Net)	0.895	0.913	0.810	0.812	0.809
Hwang S et al.^[27] (ResU2D-LC)	0.905	0.928	0.797	0.808	0.787
Strodthoff N et al.^[28] (inception1d)	0.901	0.941	0.813	0.836	0.792
MRHL-ECGNet	0.972	0.983	0.864	0.873	0.857