Date of Award

Spring 2026

Document Type

Thesis - SCU Access Only

Publisher

Santa Clara : Santa Clara University, 2026

Department

Bioengineering

First Advisor

Hamed Akbari

Abstract

Myocardial infarction (MI) remains a leading cause of death worldwide, with outcomes heavily dependent on rapid and accurate diagnosis. While the 12-lead electrocardiogram (ECG) is the primary diagnostic tool, other methods such as biomarkers, echocardiography, and angiography can be invasive, time-dependent, expensive, or require specialized expertise. Existing artificial intelligence ECG systems improve MI classification but often lack interpretability, generalizability, and the ability to localize ischemia or quantify severity. To address these limitations, HEARTNet (High-Resolution ECG-Based Activation Reconstruction Transformer Network) is proposed as a physiology-informed deep learning framework for real-time MI detection and 3D ischemia localization from standard 12-lead ECG signals. HEARTNet uses a temporal transformer architecture with multi-head self-attention to generate both a continuous ischemia severity score and lead-attention heatmaps projected onto a 3D probabilistic cardiac atlas. The model was trained and validated on 14,981 ECG records from the PTB-XL dataset and achieved an AUC of 0.9647, sensitivity of 85.63%, and specificity of 93.90%. Qualitative validation demonstrated anatomically consistent ischemia localization. HEARTNet provides real-time, interpretable, and spatially resolved ischemia detection without requiring patient-specific imaging, offering a portable and accessible tool for improving MI triage across clinical and pre-hospital settings.

Download

Share

COinS