Directional discordance between a deployed commercial 12-lead ECG interpretation algorithm and expert cardiologists, and its distribution by sex and age: a post-deployment audit using PTB-XL.

Automated 12-lead ECG algorithms are widely used, yet their disagreement with cardiologists across patient subgroups is rarely audited on public data. We characterized the directional discordance of the GE Marquette 12SL algorithm against expert labels.
Methods: In this retrospective audit of 14,885 ECGs from 13,480 patients (PTB-XL/PTB-XL+ dataset), we compared 12SL (v23) statements with standardized cardiologist labels for myocardial infarction (MI), conduction disturbance (CD), and left ventricular hypertrophy (LVH). We calculated observed agreement and paired marginal differences with 95% confidence intervals. Discordance was modeled using generalized estimating equations incorporating age, sex, and their interaction.
Results: Observed agreement was high (0.842-0.896) but directionally divergent. 12SL produced fewer positive statements than cardiologists for CD (paired marginal difference 0.092, 95% CI 0.085-0.098) and more for LVH (-0.070, 95% CI -0.075 to -0.064); MI showed no net difference. The false-negative proportion for CD decreased with age, while false-positive proportions rose with age for CD and MI. For LVH false positives, the age association was strongly modified by sex (interaction 1.60, 95% CI 1.47-1.73), being absent in women but steep in men. The LVH direction was specific to primary 12SL statements and reversed if supportive statements were included.
Conclusion: A widely deployed commercial ECG algorithm exhibits diagnosis-specific, directional discordance with expert reading that varies by age and sex, underscoring the value of reproducible post-deployment auditing. Because no clinical gold standard was available, these findings describe agreement with experts rather than absolute diagnostic accuracy.