A Deep Learning Based Framework Using Multidimensional Feature Fusion for Cardiovascular Disease Classification

Document Type

Conference Proceeding

Publication Date

2026

Abstract

Cardiovascular disease remains the leading cause of global mortality, and while 12-lead Electrocardiogram (ECG) offers a scalable non-invasive diagnostic pathway, most automated classifiers exploit only a single signal representation and largely ignore structured companion clinical measurements. To address this, a Multidimensional fusion framework for multi-label cardiovascular disease classification is proposed. The proposed framework is applied to the PTB-XL and PTB-XL+ datasets targeting five PTB-XL superclasses: Normal ECG (NORM), Myocardial Infarction (MI), ST/T-wave Change (STTC), Conduction Disturbance (CD), and Hypertrophy (HYP). Preprocessing standardizes raw 500 Hz 12-lead ECG waveforms using fold-consistent per-lead normalization and uses structured PTB-XL+ machine-derived measurements as tabular features without validation or test leakage. The framework combines three complementary sources of information: a classwise ensemble of 1D ECG models built from leadwise CNN encoders, BiLSTM temporal modeling, and cross-lead Transformer attention; a classwise ensemble of ResNet-based 2D ECG models; and LightGBM tabular models trained on structured PTB-XL+ measurements. Their outputs are integrated through classwise late fusion so that each modality contributes independently to the final multi-label prediction. Ablation experiments compare unimodal branches, two-modality fusion, trimodal fusion, MIMIC-IV-ECG pretraining, and MIMIC-only evaluation. The proposed Multidimensional fusion framework achieves the best PTB-XL test results with 93.95% macro-AUC, 89.81% accuracy, 79.44% macro-recall, and 77.67% macro-F1 score. Compared with the reviewed PTB-XL studies, the proposed framework achieves the highest macro-F1 score in the reported comparison while maintaining strong macro-AUC performance.

Link to Full Text

Share

COinS