Multi-Class Classification of Cardiac Arrhythmias Using Cost-Sensitive Feed-Forward Deep Neural Networks on Imbalanced Tabular Data

Deep Neural Networks (DNNs) have emerged as a powerful tool in biomedical signal processing due to their capacity to capture complex, non-linear relationships within high-dimensional datasets. However, automated multi-class electrocardiogram (ECG) parsing remains heavily constrained by extreme class imbalance and structural data heterogeneity. This study presents a robust feed-forward DNN designed to handle multi-class cardiac arrhythmia classification using the UCI Arrhythmia dataset (n = 452, 280 initial features). A systematic preprocessing architecture was deployed, including sentinel token imputation, zero-variance feature filtering, and isolated training-partition Z-score standardisation to eliminate data leakage. The classification engine consists of a bottleneck MLP architecture (128 and 64 units) regularised with stochastic Dropout (delta = 0.4). Multiple optimisation setups were evaluated, including standalone Early Stopping and hyperparameter grid searching. While a baseline model achieved a raw test accuracy of 65.93%, it exhibited severe overfitting and minority-class neglect. To align the optimisation objective with real-world clinical utility, an inverse-frequency loss weighting framework (compute_class_weight) was integrated. The final optimised model stabilised at a lower aggregate test accuracy of 58.24% but demonstrated significant generalisation improvements across critical rare categories, achieving perfect classification (F1 = 1.00) for Old Anterior Myocardial Infarction and demonstrating full recall (R = 1.00) on single-instance pathologies including Supraventricular Tachycardia and Right Bundle Branch Block. These findings highlight a fundamental precision-recall trade-off under skewed medical data and demonstrate the viability of well-regularised, cost-sensitive DNNs as clinical decision-support tools in high-stakes cardiac screening.