Medical-oriented multi-label diagnosis of twelve-lead ECG signals via disease-aware relational modeling and pathology-aware temporal evidence aggregation
- 1Department of Vascular Surgery, The Affiliated Hospital, Southwest Medical University
- 2Department of Intervention Medicine, The Affiliated Hospital, Southwest Medical University
Abstract
Twelve-lead ECG signals are of great value in cardiovascular disease screening and assisted diagnosis, but their multi-label recognition process is simultaneously affected by cross-lead structural dependencies, disease label co-occurrence relationships, and the dispersed distribution of pathological evidence in long-sequence signals. To address these issues, this study proposes a deep learning model for twelve-lead ECG multi-label diagnosis. The model adopts a Transformer-Mamba hybrid structure as the temporal encoding backbone to extract local waveform variations and long-range rhythm features from ECG signals. On this basis, Disease-aware Dual Relational Graph Modeling is designed to establish structured relational representations from both lead relationships and disease label relationships. Furthermore, Pathology-aware Temporal Evidence Aggregation is introduced to adaptively aggregate key pathological evidence from long-sequence ECG segments, thereby improving the discriminative ability for multiple disease labels. This study compares the proposed method with various mainstream deep learning models on the public dataset PTB-XL dataset and the clinical dataset. The results show that the proposed method achieves an Acc of 84.77 % , an F1 of 70.34 % , a Macro-AUC of 91.14 % , and a Macro-AUPRC of 78.10 % on PTB-XL dataset; and an Acc of 89.25 % , an F1 of 76.08 % , a Macro-AUC of 91.48 % , and a Macro-AUPRC of 80.14 % on Clinical dataset. Overall, the proposed method outperforms the compared methods, validating its effectiveness in the twelve-lead ECG multi-label diagnosis task.
Related
- Deploying YOLO for whole-slide mitosis detection: A transparent, auditable, and reproducible pipeline for pathology AI-assisted diagnosis and decision-support tool
- PO034 | ASSESSMENT OF PLATELET AGGREGATION FOR THE DIAGNOSIS OF INHERITED PLATELET DISORDERS USING A FULLY AUTOMATED COAGULATION ANALYSER: COMPARISON WITH LTA AND MEA
- Incidence and pathology-based predictive modeling of lymph node metastasis in early gastric cancer: evidence from a retrospective single-center Chinese surgical cohort with internal validation.
- Glycometabolic burden and myocardial involvement: a relational two-axis hypothesis
- Exploring Sibling Dynamics in the Context of Parentification: A Scoping Review of Relational and Developmental Impacts
BibTeX
@article{chen2026medicaloriented,
title = {{Medical-oriented multi-label diagnosis of twelve-lead ECG signals via disease-aware relational modeling and pathology-aware temporal evidence aggregation}},
author = {Hao Chen and Yi-feng Huang and Lu-pin Yang and Yong Liu and Hu-qiang He},
journal = {Frontiers in Cardiovascular Medicine},
year = {2026},
month = nov,
doi = {10.3389/fcvm.2026.1904647},
url = {https://www.frontiersin.org/articles/10.3389/fcvm.2026.1904647/full},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Medical-oriented multi-label diagnosis of twelve-lead ECG signals via disease-aware relational modeling and pathology-aware temporal evidence aggregation
AU - Hao Chen
AU - Yi-feng Huang
AU - Lu-pin Yang
AU - Yong Liu
AU - Hu-qiang He
PY - 2026
DA - 2026/11/05
JO - Frontiers in Cardiovascular Medicine
DO - 10.3389/fcvm.2026.1904647
UR - https://www.frontiersin.org/articles/10.3389/fcvm.2026.1904647/full
AB - Twelve-lead ECG signals are of great value in cardiovascular disease screening and assisted diagnosis, but their multi-label recognition process is simultaneously affected by cross-lead structural dependencies, disease label co-occurrence relationships, and the dispersed distribution of pathological evidence in long-sequence signals. To address these issues, this study proposes a deep learning model for twelve-lead ECG multi-label diagnosis. The model adopts a Transformer-Mamba hybrid structure as the temporal encoding backbone to extract local waveform variations and long-range rhythm features from ECG signals. On this basis, Disease-aware Dual Relational Graph Modeling is designed to establish structured relational representations from both lead relationships and disease label relationships. Furthermore, Pathology-aware Temporal Evidence Aggregation is introduced to adaptively aggregate key pathological evidence from long-sequence ECG segments, thereby improving the discriminative ability for multiple disease labels. This study compares the proposed method with various mainstream deep learning models on the public dataset PTB-XL dataset and the clinical dataset. The results show that the proposed method achieves an Acc of 84.77 % , an F1 of 70.34 % , a Macro-AUC of 91.14 % , and a Macro-AUPRC of 78.10 % on PTB-XL dataset; and an Acc of 89.25 % , an F1 of 76.08 % , a Macro-AUC of 91.48 % , and a Macro-AUPRC of 80.14 % on Clinical dataset. Overall, the proposed method outperforms the compared methods, validating its effectiveness in the twelve-lead ECG multi-label diagnosis task.
SN - 2297-055X
ER -