Automated detection of non-physiological artifacts on ECG signal: UK Biobank and CRIC.
An electrocardiogram (ECG) is commonly used in clinical practice. Poor data quality, artifacts, and misplacement of electrodes have to be identified before the clinical interpretation of ECG. We aimed to develop an algorithm to automatically identify ECG artifacts and lead misplacement.
We utilized 42,743 ECGs from UK Biobank (UKB; n = 42,743 participants; age 55±8 y; cardiovascular disease 1.2 %; diabetes 0.9 %; chronic kidney disease 0.5 %; ventricular pacing 0 %) for the algorithm development and 41,495 ECGs from the Chronic Renal Insufficiency Cohort (CRIC; n = 3912 participants; age 63 ± 11 y; cardiovascular disease 78 %; diabetes 56 %; chronic kidney disease 100 %; ventricular pacing 3.5 %) for external validation. We developed a fully automated algorithm to detect non-physiological ECG artifacts, such as high or low peak-to-peak amplitude, frequency-based outliers, and misplaced electrodes. In UKB, the algorithm demonstrated a sensitivity of 84.9 %, a specificity of 100 %, an ROC AUC of 0.924, and a Kappa statistic of 0.91. We observed 98.81 % agreement between ground truth and algorithm-identified non-physiological ECG artifacts, significantly (p < 0.00001) larger than the random agreement of 86.91 % expected at the observed 7.6 % prevalence. The misplacement of limb lead electrodes in UKB affected the Wilson Central Terminal. In CRIC, we observed an agreement of 94.90 %, which was significantly (p < 0.00001) better than by chance (93.27 % at the observed 5.3 % prevalence, including pacing artifacts), 16.8 % sensitivity, 99.3 % specificity, and an ROC AUC of 0.580.
The fully automated algorithm can accurately detect ECG artifacts and potential lead misplacement, thus permitting automated quality control of ECG analysis. The code is provided at https://github.com/Tereshchenkolab/ECG-quality-control.
We utilized 42,743 ECGs from UK Biobank (UKB; n = 42,743 participants; age 55±8 y; cardiovascular disease 1.2 %; diabetes 0.9 %; chronic kidney disease 0.5 %; ventricular pacing 0 %) for the algorithm development and 41,495 ECGs from the Chronic Renal Insufficiency Cohort (CRIC; n = 3912 participants; age 63 ± 11 y; cardiovascular disease 78 %; diabetes 56 %; chronic kidney disease 100 %; ventricular pacing 3.5 %) for external validation. We developed a fully automated algorithm to detect non-physiological ECG artifacts, such as high or low peak-to-peak amplitude, frequency-based outliers, and misplaced electrodes. In UKB, the algorithm demonstrated a sensitivity of 84.9 %, a specificity of 100 %, an ROC AUC of 0.924, and a Kappa statistic of 0.91. We observed 98.81 % agreement between ground truth and algorithm-identified non-physiological ECG artifacts, significantly (p < 0.00001) larger than the random agreement of 86.91 % expected at the observed 7.6 % prevalence. The misplacement of limb lead electrodes in UKB affected the Wilson Central Terminal. In CRIC, we observed an agreement of 94.90 %, which was significantly (p < 0.00001) better than by chance (93.27 % at the observed 5.3 % prevalence, including pacing artifacts), 16.8 % sensitivity, 99.3 % specificity, and an ROC AUC of 0.580.
The fully automated algorithm can accurately detect ECG artifacts and potential lead misplacement, thus permitting automated quality control of ECG analysis. The code is provided at https://github.com/Tereshchenkolab/ECG-quality-control.
Authors
Bukhari Bukhari, Kewalramani Kewalramani, Witzigreuter Witzigreuter, Pourbemany Pourbemany, Barbato Barbato, Daw Daw, Dhar Dhar, Rincon-Choles Rincon-Choles, Rao Rao, Bhat Bhat, Soliman Soliman, Tereshchenko Tereshchenko,
View on Pubmed