ECG software showed high QRS detection (98.74-99.57% sensitivity) and rhythm classification accuracy across ethnic groups, with some variation in VEB and AF sensitivity.
Does an automated ECG analysis software accurately detect beats and rhythms across diverse racial and ethnic populations compared to technician annotations?
Automated ECG analysis software maintains high diagnostic accuracy for arrhythmia detection across diverse racial and ethnic populations, mitigating potential algorithmic bias.
Absolute Event Rate: 0% vs 0%
Abstract Background Ambulatory ECG monitoring coupled with robust analysis software enables long-term arrhythmia assessment outside hospital settings. AI models trained on homogeneous populations may lack generalisability, leading to inaccurate ECG classifications and potential healthcare disparities. Limited research exists on racial and ethnic disparities in ambulatory ECG analysis software performance, underscoring the need for validation across diverse populations. Purpose To evaluate the performance of an ECG analysis software in detecting beats and rhythms across ethnically diverse ambulatory cohorts. Methods A dataset was collected comprising 471 20-minute ECG recordings from 387 patients under a standardised protocol. Patients self-identified within the following racial and ethnic groups: White (38.0%, n=147), Asian (27.1%, n=105), Black/African-American (20.9%, n=81), and Hispanic (14.0%, n=54). ECG recordings were processed through the software for automated QRS detection followed by beat and rhythm analysis. The software's performance was evaluated against technician-annotated labels for QRS detection, Sinus Rhythm (SR), Atrial Fibrillation (AF), and Ventricular Ectopic Beats (VEB) classification using sensitivity (Se), specificity (Sp), and PPV evaluation metrics. Differences in QRS amplitude were analysed using technician-annotated labels averaged per recording. Results Despite standardised ECG collections, QRS amplitudes differed among ethnic groups (Kruskal-Wallis, p 0.001), with low-amplitude signals (0.5mV) in 39.16% of Caucasian, 11.43% of Asian, 19.75% of Black, and 16.66% of Hispanic patients. Although amplitude differences were observed, overall QRS detection remained consistent with sensitivity between 98.74-99.57% and PPV between 97.31–99.55% across all groups. VEB sensitivity was 91.22% (95% CI: 90.25–92.18%) in Caucasian and 97.91% (95% CI: 97.22–98.60%) in Asian cohorts, compared to 85.50% (95% CI: 82.47-88.52%) in Black and 87.27% (95% CI: 84.14-90.40%) in Hispanic cohorts. Specificity exceeded 98.71% in all cohorts. The software classified AF episodes with 100% sensitivity in the Asian (95% CI: 96.90-100.00%) and Hispanic cohorts (95% CI: 71.11-100.00%), with specificity above 98.82%. Sensitivity in Caucasian and Black cohorts was 90.57% (95% CI: 87.96-93.19%) and 91.07% (95% CI: 83.12-99.00%), respectively, with specificity of 93.34% and 97.11%. SR sensitivity was 99.45% (95% CI: 97.96-100.00%) in Asian and 99.40% (95% CI: 97.77-100.00%) in Black cohorts, compared to 86.60% (95% CI: 81.94-91.27%) in Caucasian and 89.53% (84.88-94.19%) in Hispanic cohorts. Conclusion(s) The software consistently yielded high levels of QRS detection accuracy, VEB performance and rhythm performance across diverse ethnic groups, demonstrating generalisability and minimal performance bias. These findings underscore the importance of validating AI-driven ECG analysis tools in diverse populations to ensure equitable clinical outcomes.
Wiseman et al. (Sat,) reported a other. ECG software showed high QRS detection (98.74-99.57% sensitivity) and rhythm classification accuracy across ethnic groups, with some variation in VEB and AF sensitivity.