Early diagnosis of acute myocardial infarction using ECG signals with explainable artificial intelligence

Abstract

Acute myocardial infarction (AMI)—a life-threatening reduction or interruption of coronary blood flow that encompasses ST-elevation, non-ST-elevation, and occlusion subtypes—remains a leading cause of death worldwide. Although dozens of AI-based ECG studies exist, most highlight isolated technical advances; an end-to-end, interpretable workflow suitable for routine clinical use is still lacking. Rather than proposing a single "black-box" diagnostic device, this thesis presents a practical blueprint that spans automated signal extraction from PDF ECG reports, model training, counterfactual explanation generation, and clinical validation. The workflow begins with an automated pipeline that converts PDF ECGs into lead-specific time-series signals. A deep-learning model, pretrained on PTB-XL and fine-tuned on an in-house cohort, first distinguishes myocardial-infarction traces from normal ones. To complement this model, we introduce a fully interpretable machine-learning framework that maps salient ECG features to clinically recognised patterns and generates counterfactual explanations—offering clinicians actionable, signal-level cues and reinforcing trust. Two experienced cardiologists oversaw every stage of development: they validated the preprocessing pipeline, confirmed the diagnostic logic, and conducted a qualitative review of model outputs and counterfactuals, ensuring alignment with everyday cardiology practice. The deep-learning model achieved 96.4% accuracy on the PTB-XL test set but dropped to 70.9% on the in-house cohort, revealing a marked cross-dataset gap. The decline was sharper for the four-class task ("MI subtypes + healthy"), where accuracy fell to 46.4%. Fine-tuning with PTB-XL weights partially recovered performance, raising accuracy to 74.2% (binary) and 51.6% (four-class). Among traditional learners, XGBoost remained strongest, posting 89.4% accuracy on PTB-XL. The cardiologists' review judged 31 of 46 counterfactual reports clinically meaningful; high-quality reports averaged an alignment score of 0.83 ± 0.12. By detailing both system architecture and practical hurdles—such as extracting raw signals from PDFs and showing performance gaps between public and real-world data—this thesis underscores that high benchmark scores alone do not guarantee clinical utility. Future work should emphasise rigorous local validation, thoughtful incorporation of metadata, and multi-centre collaborations to narrow the remaining gulf between AI research and day-to-day cardiology practice.