Authors: Hongling Zhu, M.D., Ph.D., Yinuo Jiang, B.Eng., Cheng Cheng, Ph.D., Jingyi Wang, B.S., Linghao Zhu, B.Eng., Xia Chen, B.S., Kuan Feng, Yujian Liu, M.D., Ph.D., Longjianjie Zhang, B.Eng., Qiushi Luo, M.S., Xin He, M.A., Chen Ruan, B.S., Heng Zhang, M.S., Jurong Tian, M.S., Chunyan Wang, B.S., Yawen Zou, B.S., Hao Du, M.D., Ph.D., Zhouping Tang, M.D., Ph.D., Haitao Song, Ph.D., Guohua Wan, Ph.D., Hesong Zeng, M.D., Ph.D., Ye Yuan, Ph.D., and Xiaoyun Yang, M.D., Ph.D.
Background: The electrocardiogram (ECG) remains the most commonly used screening tool for cardiac diseases. Although cardiac hypertrophy, dilation, and enlargement are important causes of heart failure and sudden death, they are mainly diagnosed via echocardiography after symptom onset, due to the low sensitivity of human ECG interpretation. This study challenges the mainstream diagnostic methodology by implementing a reduced-channel deep learning–based model that utilizes an ECG (or a four-channel ECG) as a single data source for early diagnosis.
Methods: We constructed a large-scale database comprising 90,895 ECGs from 74,562 patients taken from a total of 2,386,886 ECGs and 988,257 echocardiograms from January 1, 2012 to July 17, 2021, from Tongji Hospital, Wuhan, China. A multi-label deep learning–based model using ECG as a single input was created, with echocardiography as the gold standard at the model training stage. Four distinct datasets were used for testing. Furthermore, we applied an aggregated attribution score for each lead, based on the expected gradient of the model, to investigate the representative lead of the model.
Results: The sensitivity value increased from 0.270 (as reported by six participating ECG physicians with 6 to 24 years of experience) to 0.586 after using the proposed model, demonstrating a twofold increase in average sensitivity. Therefore, in over half of the patients with cardiac hypertrophy, dilation, and enlargement, cases can potentially be detected during routine ECG monitoring. The calculated attribution score identified the four highest-performing leads: I, aVR, V1, and V5. The performance of the reduced-channel model, trained with I, aVR, V1, and V5 leads, is equivalent to that of the 12-channel model, which supports the feasibility of wearable devices as an alternative to echocardiography.
Conclusions: ECGs can serve as a viable method for early diagnosis of cardiac hypertrophy, dilation, and enlargement through routine monitoring. The four representative leads can assist in human ECG annotation and inform portable device design using fewer embedded channels. Using a large-scale cardiac hypertrophy, dilation, and enlargement database comprising 90,895 ECGs from 74,562 patients who underwent ECG and echocardiography during a single visit or within a short time frame, we more than doubled the average sensitivity across all cardiovascular regions, suggesting that, in over half of the patients with these conditions, cases can potentially be detected during routine ECG monitoring. The four highest-performing leads were identified (I, aVR, V1, and V5), supporting the potential for efficient diagnosis with fewer embedded channels. (Funded by the National Natural Science Foundation of China and others.)