Introduction: Reliable rhythm diagnosis is paramount for atrial fibrillation (AF) screening with single-lead ECG devices. 

Aim: To identify patient characteristics prone to incorrect rhythm diagnosis employing a machine learning algorithm for ECG signal quality estimation (Q_EST).

Methods: One-minute single-lead ECG tracings (MyDiagnostick, Applied Biomedical Systems, Maastricht, Netherlands) recorded for AF screening in community pharmacies were analyzed for signal quality by blinded human overread and a machine learning approach rated from excellent to uninterpretable (Figure 1). ECG recordings with less than 30 seconds of consecutive interpretable ECG tracing were excluded from analysis. Validity of every single recording was automatically assessed based on the occurrence of motion artifacts and other disturbances. Twenty-six individual signal features were extracted from the ECG recordings to train a Gaussian process regression model for automated signal quality assessment. 

Results: In total, 6918 out of 7031 ECGs were eligible for analysis. In the cohort, 357 (5.2%) subjects were identified with AF. Reasons for incorrect automated heart rhythm diagnosis by the ECG device were analyzed in a binary logistic regression with forward selection considering sex, body mass index (continuous), age (continuous), congestive heart failure, hypertension, diabetes mellitus, stroke/transient ischemia, and vascular disease. Male sex (odds ratio (OR) 1.79, 95% confidence interval (CI) 1.35-2.39, P<0.001), lower body mass index (per step) (OR 0.9, 95% CI 0.87-0.94, P<0.001), and higher age (per year) (OR 1.06, 95% CI 1.03-1.08, P<0.001) were significantly associated with incorrect rhythm diagnosis. Individual Q_EST score was calculated by automated ECG signal quality assessment. The Q_EST score correlated with incorrect measurement (rho 0.35, P<0.001) and showed higher scores for male sex, age older than 75 years, and body mass index below 25 kg/m² (Table 1).

Conclusion: Automated signal quality (Q_EST score) assessment by a machine learning approach correlated significantly with incorrect measurement of an automated AF screening performed with a single-lead ECG device. The Q_EST score indicates groups of patients with a higher chance for lower ECG signal quality which is associated with higher probability of incorrect measurements. Automated AF screening is recommended for elderly, and male sex has a higher incidence of AF. Our data shows that these groups tend to have lower ECG signal quality, which may contradict successful automated screening. 

Table 1 Estimated ECG signal quality (Q_EST score) correlates with incorrect heart rhythm diagnosis for Male sex, age older than 75 years and BMI <25 significantly. Lower Q_EST score indicates better signal quality with higher chance of correct measurement. *Pearson Chi-Square test; #One-way ANOVA.