Automated training data generation for multiple DNNs in ECG analysis

Abstract Electrocardiograms (ECGs) exhibit diverse waveforms depending on the type of disease, patient age, and electrode positions, making automated analysis challenging. Although deep neural networks have been applied to ECG interpretation, their performance is highly dependent on extensive, high‐quality training data, which are difficult to obtain for rare or novel patterns. This study proposes an automatic retraining algorithm that enhances the accuracy of ECG waveform boundary detection by utilizing recognition scores (RS) and quantifies the consistency of predictions across six independently trained models within a multiple deep neural network (mDNN) framework. The initial mDNN was trained using ECG data from five healthy individuals. Using ECG data from 30 elderly patients, the algorithm identifies low‐RS cases, corrects errors, and generates refined training data for subsequent mDNN retraining. Evaluation with ECG data from 20 previously unseen patients showed that RS values for P and T waves nearly doubled. Furthermore, the similarity between mDNN‐derived ECG parameters and expert annotations increased by 29%–98%, depending on the specific parameter. The retrained mDNN also exhibited more consistent results than human experts when analyzing single‐patient data.