Deep Learning Algorithm of 12-Lead Electrocardiogram for Parkinson Disease Screening

Background: Although idiopathic Parkinson's disease (IPD) is increasing with the aging population, there is no adequate screening test for early diagnosis of IPD. Cardiac autonomic dysfunction begins in the early stages of IPD, and an electrocardiogram (ECG) contains precise information on the heart. Objective: This study is to develop an ECG deep learning algorithm that can efficiently screen for IPD. Methods: Data were collected from 751 IPD patients (2,138 ECGs), 751 age and sex-matched non-IPD patients (2,673 ECGs) as a control group, and 297 drug-induced Parkinsonism (DPD) patients (875 ECGs) as a disease control group. ECG data were randomly divided into training set, validation set, and test set at a ratio of 6:2:2. We developed a deep-convolutional neural network (CNN) consisting of 16 layers with Bayesian optimization that classified IPD patients by ECG data. The robustness of the deep learning model was verified through 5-fold cross-validation. Results: The AUROC of the model for detection of IPD was 0.924 (95% CI, 0.913-0.936) in the test set. That for detecting DPD was 0.473 (95% CI, 0.453-0.504). The sensitivities of the model according to Unified Parkinson's Disease Rating Scale III and Hoehn & Yahr scale were also similar. Conclusion: In conclusion, the CNN-based deep learning model using ECG data showed quite good performance in identifying IPD patients. Standardized 12-lead ECG test could be one of the clinically feasible candidate methods for early screening of IPD in the future.