Authors: Dhananjay Yeola, Shyamrao V. Gumaste
Abstract: This paper proposes a lightweight, one-stage deep learning model for the early identification of heart dysfunction (CVD) using electrocardiogram (ECG) images. The framework inte-grates preprocessing, automated data curation, and classification into a single unified pipe-line, thereby reducing redundancy and improving efficiency compared to multistage ap-proaches. The model uses convolutional backbones that have been strengthened with atten-tion techniques to guarantee efficient feature extraction while maintaining important diag-nostic data. The proposed design emphasizes deployability in resource constrained clinical settings, where computational efficiency and interpretability are as important as accuracy. By minimizing computational overhead and improving robustness against noise and varia-bility in ECG data, the framework supports real-time analysis and rapid diagnostic decision-making. A thorough literature survey highlights key gaps in existing research, including challenges in generalization across diverse patient populations, the high computational cost of state-of-the-art models, and the limited interpretability of deep learning predictions. To get around these limitations, this study presents a compact architecture that is optimized for explainability, scalability, and adaptability. Furthermore, we outline a validation plan em-ploying clinically relevant metrics such as sensitivity, specificity, and F1-score to ensure reliability in real-world applications. The proposed approach aims to support cardiologists in timely triage and decision support, ultimately contributing to improved outcomes in cardio-vascular healthcare.
DOI: http://doi.org/10.5281/zenodo.20848033
