MedXpert‐AI: Context‐Aware and Explainable Deep Learning Framework for Intelligent ECG Interpretation
Complexity · 2026-01 · 2026 (1)
DOI: 10.1155/cplx/8694851 ↗Kurz gefasst
Fachartikel in der Zeitschrift Complexity, online erschienen am 01.11.2026. Verfasst von Anto Lourdu Xavier Raj Arockia Selvarathinam und 6 weiteren Autor:innen. Die Arbeit nennt 46 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Taif University.
Abstract (Originalsprache)
Cardiovascular diseases remain a major cause of morbidity and mortality worldwide, creating a need for accurate and interpretable automated electrocardiogram (ECG) analysis. Although deep learning has demonstrated strong performance in ECG interpretation, many existing approaches primarily rely on data‐driven representations and provide limited integration of structured clinical knowledge and model‐level interpretability. This study presents MedXpert‐AI, a context‐aware and explainable hybrid framework for 27‐label multilabel ECG classification. The framework integrates CNN‐, BiLSTM‐, and transformer‐based multilevel feature learning with clinical knowledge representation comprising a medical knowledge graph and rule‐based clinical reasoning. A graph neural network (GNN) is employed to learn knowledge‐guided representations, while adaptive attention feature fusion (AAFF) adaptively integrates ECG‐derived and clinical representations for final diagnostic prediction. SHAP, Grad‐CAM, and transformer attention heatmaps are incorporated to provide complementary model‐level explanations. The framework was evaluated using 43,101 12‐lead ECG recordings from the PhysioNet/Computing in Cardiology (CinC) 2020 dataset, comprising 27 scored diagnostic labels. For the primary multilabel evaluation, MedXpert‐AI achieved 98.85% AUROC, 97.64% AUPRC, 97.20% Macro F1, 97.46% Micro F1, and 0.94 MCC. Controlled ablation experiments showed performance reductions following removal of the GNN, transformer, and clinical knowledge representation modules, as well as replacement of AAFF with simple feature concatenation. Computational analysis further demonstrated a relatively low resource requirement under the evaluated experimental conditions, with approximately 2.3 billion FLOPs and 1.5 GB GPU memory usage. The explainability analysis provided complementary visualizations of feature attribution, activation localization, and temporal attention, although formal clinician‐based validation of explanation quality was not performed. These findings support MedXpert‐AI as a knowledge‐aware and explainable framework for automated multilabel ECG analysis, while external multicenter and prospective clinical validation remain necessary to establish generalizability and clinical utility.
Förderung
- Taif University
Zitieren
APA
Selvarathinam, A. L. X. R. A., Simaiya, S., Baqasah, A. M., Alroobaea, R., Alsafyani, M., Khalid, A., & Tekale, T. H. (2026). MedXpert‐AI: Context‐Aware and Explainable Deep Learning Framework for Intelligent ECG Interpretation. Complexity, 2026(1). https://doi.org/10.1155/cplx/8694851
Vancouver
Selvarathinam ALXRA, Simaiya S, Baqasah AM, Alroobaea R, Alsafyani M, Khalid A, et al. MedXpert‐AI: Context‐Aware and Explainable Deep Learning Framework for Intelligent ECG Interpretation. Complexity. 2026;2026(1). doi:10.1155/cplx/8694851
BibTeX
@article{Selvarathinam2026,
author = {Selvarathinam, Anto Lourdu Xavier Raj Arockia and Simaiya, Sarita and Baqasah, Abdullah M. and Alroobaea, Roobaea and Alsafyani, Majed and Khalid, Alharbi and Tekale, T. Hiyab},
title = {MedXpert‐AI: Context‐Aware and Explainable Deep Learning Framework for Intelligent ECG Interpretation},
journal = {Complexity},
year = {2026},
volume = {2026},
number = {1},
doi = {10.1155/cplx/8694851}
}
EUVIMED