Diagnostic Preservation After ECG Digitization: From Risk Stratification to Reliability Assurance

As Electrocardiogram (ECG) digitization becomes increasingly accurate, waveform similarity alone does not establish preservation of diagnostically relevant information. This study investigates whether post-digitization diagnostic instability can be identified without access to the native digital ECG and whether such reference-free observability is sufficient to support reliability assurance. PTB-XL waveforms were rendered as 3×4 ECG pages at 25 mm/s, 10 mm/mV, and 4 pixels/mm and re-digitized with a fixed image-based trace-recovery pipeline. Analyses on PTB-XL showed that unobserved lead-time regions contributed only limited incremental diagnostic information under the tested setting, indicating that missing temporal support alone could not explain the observed preservation loss. A reference-free instability score derived solely from digitized ECGs retained above-chance record-level discrimination on a patient-disjoint cohort (AUROC 0.614, unitless, for distinguishing ECGs with versus without at least one probe-defined new diagnostic error), but calibration deteriorated after population transfer. Selective acceptance reduced the occurrence of probe-defined new diagnostic errors without satisfying the multimetric preservation requirements, while class-wise analysis revealed strongly directional and diagnosis-dependent failures. Application to 99 real-world ECGs from Tam Duc Hospital further showed that the representation remained technically applicable despite structured feature-domain shifts. These findings establish a distinction between risk observability and preservation assurance: diagnostic instability can be partially stratified without an oracle reference, but successful ranking alone is insufficient to guarantee stable diagnostic decisions or calibrated reliability