Detection and removal of pacing artifacts prior to automated analysis of 12-lead ECG.

BACKGROUND: Pacing artifacts must be excluded from the analysis of paced ECG waveform. This study aimed to develop and validate an algorithm to identify and remove the pacing artifacts on ECG and vectorcardiogram (VCG). METHODS: We developed a semi-automatic algorithm that identifies the onset and offset of a pacing artifact based on the VCG signal slope steepness and designed a graphical user interface that permits quality control and fine-tuning the constraining threshold values. We used 1054 ECGs from the retrospective, multicenter cohort study "Global Electrical Heterogeneity and Clinical Outcomes," including 3825 atrial and 10,031 ventricular pacing artifacts for the algorithm development and 22 ECGs including 108 atrial and 241 ventricular pacing artifacts for validation. Validation was performed per digital sample. We used the kappa-statistic of interrater agreement between manually labeled sample (ground-truth) and automated detection. RESULTS: The constraining parameter values were for onset threshold 13.06 ± 6.21 µV/ms, offset threshold 34.77 ± 17.80 µV/ms, and maximum window size 27.23 ± 3.53 ms. The automated algorithm detected a digital sample belonging to pacing artifact with a sensitivity of 74.5% and specificity of 99.6% and classified correctly 98.8% of digital samples (ROC AUC 0.871; 95%CI 0.853-0.878). The kappa-statistic was 0.785, indicating substantial agreement. The agreement was on 98.81% digital samples, significantly (P < 0.00001) larger than the random agreement on 94.43% of digital samples. CONCLUSIONS: The semi-automated algorithm can detect and remove ECG pacing artifacts with high accuracy and provide a user-friendly interface for quality control.

Eligibility for subcutaneous implantable cardioverter-defibrillator in congenital heart disease.

BACKGROUND: Adult congenital heart disease (ACHD) patients can benefit from a subcutaneous implantable cardioverter-defibrillator (S-ICD). OBJECTIVE: The purpose of this study was to assess left- and right-sided S-ICD eligibility in ACHD patients, use machine learning to predict S-ICD eligibility in ACHD patients, and transform 12-lead electrocardiogram (ECG) to S-ICD 3-lead ECG, and vice versa. METHODS: ACHD outpatients (n = 101; age 42 ± 14 years; 52% female; 85% white; left ventricular ejection fraction [LVEF] 56% ± 9%) were enrolled in a prospective study. Supine and standing 12-lead ECG were recorded simultaneously with a right- and left-sided S-ICD 3-lead ECG. Peak-to-peak QRS and T amplitudes; RR, PR, QT, QTc, and QRS intervals; Tmax, and R/Tmax (31 predictor variables) were tested. Model selection, training, and testing were performed using supine ECG datasets. Validation was performed using standing ECG datasets and an out-of-sample non-ACHD population (n = 68; age 54 ± 16 years; 54% female; 94% white; LVEF 61% ± 8%). RESULTS: Forty percent of participants were ineligible for S-ICD. Tetralogy of Fallot patients passed right-sided screening (57%) more often than left-sided screening (21%; McNemar χ2P = .025). Female participants had greater odds of eligibility (adjusted odds ratio [OR] 5.9; 95% confidence interval [CI] 1.6-21.7; P = .008). Validation of the ridge models was satisfactory for standing left-sided (receiver operating characteristic area under the curve [ROC AUC] 0.687; 95% CI 0.582-0.791) and right-sided (ROC AUC 0.655; 95% CI 0.549-0.762) S-ICD eligibility prediction. Validation of transformation matrices showed satisfactory agreement (<0.1 mV difference). CONCLUSION: Nearly half of the contemporary ACHD population is ineligible for S-ICD. The odds of S-ICD eligibility are greater for female than for male ACHD patients. Machine learning prediction of S-ICD eligibility can be used for screening of S-ICD candidates.