JMIR Publications

ABSTRACT

Automated arrhythmia detection is among the most clinically consequential applications of electrocardiography (ECG). Substantial evidence confirms that ECG morphology, including QRS duration, QRS voltage, early repolarization patterns, and bundle branch blocks varies across racial and ethnic groups. These variations are real, but race and ethnicity are best understood as proxies for an underlying bundle of genetic, anthropometric, environmental, and social factors that existing studies rarely disentangle; which of these is the dominant driver of any clinically significant change in ECG morphology is, at present, unknown. We argue that the field should stop treating this variation as a settled racial effect and instead make a deliberate effort to identify its dominant drivers. These considerations are particularly relevant for machine-learned algorithms, whose performance is shaped by the composition of the training distribution and therefore requires development and validation across populations that reflect known sources of ECG variation. Rule-based algorithms may be less susceptible to these effects for gross arrhythmia detection, but this assumption has rarely been independently evaluated, and early prospective evidence suggests residual, classification-specific differences in performance. We conducted a systematic search (PROSPERO Database: CRD42023495643) across MEDLINE, Cochrane Trial Register, and EMBASE, retrieving 503,573 patients from all populated continents; not one of the 12 studies meeting inclusion criteria had race- or ethnicity-stratified detection performance as its primary objective. As a subordinate but practical point, we contend that routinely collecting race and ethnicity in diagnostic-performance studies is hypothesis-generating. While demographic variables such as age and sex are already incorporated into some diagnostic algorithms and influence established interpretation thresholds, consistent reporting of race and ethnicity may help identify previously unrecognized variations in diagnostic performance and guide subsequent investigations into their underlying causes. We therefore suggest that race and ethnicity be routinely reported alongside other demographic and anthropometric characteristics, including age, sex, height, and weight.