Comparative Analysis of Neural Network Architectures for Cardiovascular Disease Diagnosis Using Electrocardiogram Data

Tamirlan Kulzhanov, Alua Myrzakerimova

2025

Closed access Metadata (CSL-JSON)
DOI 10.1109/sist61657.2025.11139351
Data confidence 70 medium
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

Cardiovascular diseases persist as the leading cause of global mortality, necessitating advancements in early diagnosis via electrocardiogram analysis. While deep learning architectures—including convolutional neural networks, recurrent neural networks, multi-layer perceptrons, and transformers—have automated electrocardiogram interpretation, the optimal balance of accuracy, efficiency, and interpretability remains contested. This study conducts a comprehensive comparative analysis of these architectures, integrating a systematic literature review with empirical validation on the MIT-BIH arrhythmia dataset. Contrary to prior reports favoring multi-layer perceptrons (99.3% accuracy on small datasets), our experimental results demonstrate that hybrid architectures like CRT-Net achieve superior diagnostic precision (99.24% accuracy), closely approximating cardiologist-level performance. However, multi-layer perceptrons exhibit unmatched computational efficiency (98.93% accuracy in 42 seconds), challenging assumptions about complexity-accuracy trade-offs. Transformers underperform in isolation (Vision Transformer: 94.56%) but excel when integrated into hybrids, while pure recurrent neural networks lag in accuracy (Long Short-Term Memory networks: 96.77%) and scalability. Our findings reveal that convolutional neural networks remain robust for feature extraction (97.80% accuracy), but CRT-Net’s fusion of convolutional neural networks, recurrent neural networks, and self-attention mechanisms sets a new benchmark for clinical-grade diagnostics. Conversely, multi-layer perceptrons emerge as ideal candidates for wearable devices due to their real-time processing capabilities. This work advocates a context-driven paradigm: CRT-Net for high-stakes hospital settings and multi-layer perceptrons for resource-constrained environments. The study concludes with a roadmap for optimizing lightweight hybrids and electrocardiogram-specific transformers, emphasizing the need to harmonize accuracy, efficiency, and interpretability for equitable artificial intelligence-driven cardiology.