Restoration of missing or low-quality 12-lead ECG signals using ensemble deep-learning model with optimal combination. (May 2023)

• Record Type: Journal Article
• Authors: Yoo, Hakje; Yum, Yunjin; Kim, Yoojoong; Kim, Jong-Ho; Park, Hyun-Joon; Joo, Hyung Joon
• Is Part Of: Biomedical signal processing and control. Volume 83(2023)
• Publication Date: 2023-05
• Subjects: 12-lead Electrocardiogram -- Missing signal -- Restoration model -- Linear regression -- Ensemble model -- Bidirectional long short-term memory -- Convolution natural network; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2023.104690
• Languages: English
• ISSNs: 1746-8094

Abstract

Graphical abstract: Highlights: For this study, a high-quality 12-lead electrocardiogram dataset was constructed based on a large database from the general hospital. The accuracy of the restored signals through 165 candidate combinations appears linear, and there is a threshold combination in which the accuracy rapidly decreases. The optimal combination that best reflects the missing signal could improve the restoration accuracy of electrocardiogram signal. The results of restoring electrocardiogram signal in clinical case using the ensemble model potentially demonstrate clinical usability. Abstract: Background and Objective: In a 12-lead electrocardiogram (ECG) examination, the ECG signals often have low-quality data problems due to high-frequency noise caused by muscles and low-frequency noise caused by body movement, breathing. These problems cause delays in examination results and increase medical costs. For this reason, solving low-quality data and missing ECG data problems can provide patients with improved medical services, reducing the work-loss and medical costs. The purpose of this study is to develop a signal restoration model for each of the 12 signals to solve the low-quality and missing data problems caused by mechanical and operator errors during 12-lead ECG examinations. Methods: For this study, 13, 862 high-quality 12-lead ECG recordings for multiple diseases were obtained from the 12-lead ECG database of a general hospital from 2016 to 2020. Two strategies were adopted to develop an accurate restoration model. First, to obtain the optimal input parameters for the ECG regeneration model for each ECG signal, linear regression (LR) models were developed for all 165 three-signal combinations of 11 signals. Second, the restoration models were constructed in a parallel architecture combining bidirectional long short-term memory (Bi-LSTM) with a convolutional neural network (CNN) to learn the temporal and spatial features of optimal combinations. Results: The performances of the 165 candidate combinations for restoring missing signal were analyzed through the LR model to find the optimal input parameter for all ECG signals. The average root mean square error of the optimal combinations was 0.082 μ V . The average RMSE of the signal restoration model made using the optimal combinations and deep-learning model (Bi-LSTM&CNN) was 0.037 μ V, and the cosine simplicity was 0.991. Conclusions: This ECG restoration technology obtained optimal input parameters through the LR model and developed ECG restoration model through the Bi-LSTM&CNN combined model to restore ECG signals for multiple diseases. The 12-lead ECG signal restoration model developed through this study offers high accuracy for the magnitude and direction components of all 12 signals. This technology can be used in emergency medical systems and remote ECG measurement situations, as well as in synthetic ECG generation technologies for constructing research datasets.