Introduction: Little is known about the effect of ECG signal quality on correct rhythm diagnosis in single-lead ECG atrial fibrillation (AF) screening. 

Aim: Single-lead ECG recordings were rated upon ECG signal quality by human overread and a machine learning algorithm to determine ECG signal quality (Q_EST) and support rhythm diagnosis.

Methods: A total of 7031 one-minute single-lead ECG tracings recorded for AF screening in community pharmacies were analyzed for rhythm diagnosis and signal quality by blinded human overread (Figure 1A). In addition, a multi-step machine learning approach was implemented to automatically assess signal quality. In the first step, the validity of every single recording was evaluated based on the occurrence of motion artifacts and other disturbances. Inclusion criterium was a minimum of 30 seconds of consecutive interpretable ECG tracing according to current guideline recommendations. Subsequently, a Gaussian process regression model was trained to assess signal quality using 26 individual signal features extracted from the recording. Like human overreading, signal quality was rated with continuous numbers from 0 (excellent quality) to 3 (uninterpretable).

Results: In total, 6918 ECGs were eligible for analysis and initial automated ECG device screening identified 397 subjects with AF (5.7%). By human overread 357 (5.2%) were identified with AF. In 209 subjects (3%), automated rhythm analysis was incorrect with 113 false positives (1.6%) and 96 false-negative (1.4%) results. Incorrect measurements were significantly associated with ECG signal quality according to human and automated signal quality (Q_EST score) assessment via the machine learning approach (Table 1). In receiver operating analysis (Figure 1B), human signal quality assessment performed well in identifying good signal quality for correct heart rhythm diagnosis with an area under the curve (AUC) of 0.78 (sensitivity 81%, specificity 63%). Automated ECG signal assessment further improved correct heart rhythm identification with an AUC of 0.86 (sensitivity 71%, specificity 80%).

Conclusion: Human overread is recommended in automated single-lead ECG AF screening. However, it is time-consuming, investigator dependent and prone to investigator experience. Automated single-lead AF detection is highly dependent on ECG signal quality. Adding an ECG signal quality score may improve reliability of automated ECG rhythm diagnosis and support physician overread for final rhythm diagnosis.

Table 1 ECG signal quality in groups according to human rating. Incorrect automated diagnosis by the single-lead ECG device was higher for worse ECG signal quality groups. Q_EST score increases with worse ECG signal quality. *Pearson Chi-Square test; #One-way ANOVA.