Background: Automated ECG interval and amplitude measurements are widely used as supervision and evaluation targets, yet agreement between major commercial algorithms is imperfect and may depend on waveform shape in ways that aggregate statistics obscure.
Methods: We studied paired Glasgow (UniG) and GE Marquette 12SL outputs on the public PTB-XL+ resource. The analysis unit is a single lead, where each patient-lead pair supplies one median-beat voltage trace to train a self-organising map (SOM), and the corresponding paired UniG and 12SL measurements for agreement assessment. Up to twelve such examples arise per twelve-lead ECG when all leads are available. After fitting the map, we computed Bland-Altman limits of agreement separately within each SOM prototype, pooling all patient-lead pairs assigned to that cell, and visualised prototype-level median waveforms with cell-wise agreement summaries.
Results: Limits of agreement vary markedly across SOM-defined morphology groups for all measures examined, indicating that inter-algorithm discrepancy is strongly shape-dependent rather than homogeneous across the beat manifold. For example, ST Amplitude agreement ranges from approximately 0±20 μV to 10±75 μV between morphology clusters, reflecting substantial morphology‑linked variability.
Conclusions. Morphology-stratified agreement analysis is feasible at scale on PTB-XL+ and suggests that inter-algorithm disagreement in automated ECG measurements is not random but structured by waveform morphology, with implications for their use as training labels or benchmark references. A reproducible pipeline accompanies this work.