Key result
Inter-patient ECG evaluation yields an absolute ~18-point lower accuracy than intra-patient beat-level testing.
Why the study?
Heartbeat classification on the MIT-BIH Arrhythmia Database is commonly evaluated with beat-level cross-validation, which places beats from the same subject in both training and test sets, prompting a comparison with record-level separation.
Does record-level (inter-patient) evaluation compared to beat-level (intra-patient) cross-validation alter the classification accuracy and macro-F1 of a 61-dimensional morphological and RR-interval ECG representation for arrhythmia detection?
Does record-level (inter-patient) evaluation compared to beat-level (intra-patient) cross-validation alter the classification accuracy and macro-F1 of a 61-dimensional morphological and RR-interval ECG representation for arrhythmia detection?
Demonstrates a 53.44-percentage-point drop in macro-F1 when moving from intra-patient to inter-patient ECG arrhythmia evaluation, proving that beat-level cross-validation severely overestimates generalization to unseen patients and that class-specific failures (RBBB, LBBB, APC) stem from distinct underlying causes.
Beat-level ECG algorithms should not yet be adopted clinically due to severe inter-patient performance drops; leaves open how to achieve generalizable automated arrhythmia detection.
Heartbeat classification on the MIT-BIH Arrhythmia Database is commonly evaluated with beat-level cross-validation, which places beats from the same subject in both training and test sets. In this work, we evaluate a 61-dimensional heartbeat representation under both that protocol and record-level separation, and compare the two. Each beat is described by 60 signal samples spanning 166.7 ms around the R-peak and one RR interval. Under beat-level five-fold cross-validation, a Random Forest reaches 98.42% accuracy and 95.44% macro-F1 across five classes. Under record-level separation, with the configuration selected entirely within DS1 and DS2 withheld from model selection, the same representation reaches 80.80% accuracy and 42.00% macro-F1, with sensitivity of 0.9864 for Normal, 0.9139 for PVC, 0.2004 for RBBB, 0.0002 for LBBB and 0.0023 for APC. We then separate the causes of these failures experimentally. RBBB sensitivity rises to 0.9178 on one held-out subject when the number of training subjects carrying the class is increased, whereas LBBB sensitivity remains below 0.001 under every configuration and feature subset tested. Across the five classifiers evaluated under DS1/DS2, LBBB remains poorly recognized and RBBB sensitivity remains below 0.40, whereas APC performance varies substantially, showing that the poor transfer is not specific to the Random Forest and that the weak classes are affected by different factors. The comparison illustrates the substantial difference between beat-level and record-level evaluation under the standard DS1/DS2 setting, and shows that per-class failures under record-level separation arise from different causes and require different remedies.
No takes yet. Share an insight, caveat, or question.
Qahtan et al. (2026) studied Cardiac arrhythmias (Normal, PVC, LBBB, RBBB, APC) (n=47). Inter-patient (record-level DS1/DS2 separation) evaluation of a 61-dimensional heartbeat representation (60 morphological samples + 1 RR interval) with Random Forest vs. Intra-patient (beat-level 5-fold cross-validation) evaluation and baseline classifiers (Decision tree, Logistic regression, k-NN, Gaussian naive Bayes, SVM) was evaluated on Classification accuracy and macro-F1 across five heartbeat classes (Normal, PVC, LBBB, RBBB, APC). Record-level inter-patient evaluation of a 61-dimensional ECG representation yielded 80.80% accuracy and 42.00% macro-F1 versus 98.42% accuracy and 95.44% macro-F1 under beat-level cross-validation.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: