Feature transferability and the lead configuration barrier in cross database ventricular ectopy detection

Machine-learning models for heartbeat classification are usually validated within a single database, leaving the roles of sampling-rate and lead-configuration mismatch in cross-database failure unresolved. We ask how much of the gap is sampling, how much is imbalance, and how much is morphology. SARA-ML is a 34-feature framework in which every temporal descriptor is defined in milliseconds and every spectral descriptor as a dimensionless fraction, so feature vectors are comparable across sampling rates without resampling. A Random Forest and gradient-boosted trees were tested under three imbalance strategies (none; minority SMOTE to 20%; balanced downsampling/oversampling) using patient-independent grouped 5-fold cross-validation on MIT-BIH (47 subjects, 109,468 beats, 360 Hz) and applied unchanged to INCART (22 patients, 118,369 beats, 257 Hz). Ventricular sensitivity was 92 to 95% within MIT-BIH and 92.9% on the standard 44-record subset. Across databases, balanced rebalancing raised ventricular sensitivity from 70.5 to 82.3%. Transfer generalized to a third database (European ST-T, 250 Hz): ventricular sensitivity 90.6%, reverse transfer 94.8%. For burden quantification, cross-database positive predictive value was 0.96 (burden correlation 0.95). Removing morphology raised transfer sensitivity to 0.841, localizing the residual gap. Sample-count-independent features remove sampling rate as a degradation source, rebalancing removes much of the remainder, and the residual gap is morphological. The framework provides an interpretable, reproducible cross-database ECG baseline.