|
چکیده
|
Background. Single-label ECG classification systems fail to capture the multi-morbidity prevalent in cardiac patients, potentially leading to missed comorbidities and suboptimal treatment decisions. This study aimed to apply a Multi label classification (MLC) techniques for joint prediction of cardiac abnormalities using a 12-lead ECG recordings. Methods. We analyzed 10,646, 12-lead ECG recordings from the Chapman-Shaoxing dataset, categorizing diagnoses into eight clinically relevant classes: four rhythm disorders and four pathophysiological conditions (ischemic, electrical, conduction axis, and structural abnormalities). Three MLC algorithms Binary Relevance (BR), Classifier Chains (CC), and Label Powerset (LP) were implemented and evaluated using 10-fold cross-validation. Class imbalance was addressed using SMOTE, To evaluate the performance of the models, precision, recall, F1 score and AUC were used. Results. The Label Powerset model demonstrated superior performance (recall: 0.6880, F1-score: 0.7189, AUC: 0.8202), effectively capturing label dependencies inherent in cardiac multi-morbidity. The model achieved high diagnostic accuracy for rhythm disorders, particularly atrial fibrillation (precision: 87%, recall: 82%) and sinus bradycardia (precision: 84%, recall: 79%). While ischemic and conduction axis disorders proved more challenging, the model maintained clinically acceptable performance. Conclusion. Multi-label classification, particularly the Label Powerset approach, represents a significant advancement in automated ECG interpretation, enabling simultaneous detection of multiple cardiac conditions. This methodology better reflects clinical practice patterns and could enhance diagnostic efficiency and completeness in cardiovascular care.
|