A Novel Autonomous GA-LSTM Framework for Cost-Sensitive ECG Arrhythmia Classification

Cardiovascular disease kills more people worldwide than any other cause, which puts real weight behind the push for automated, reliable arrhythmia detection from electrocardiogram (ECG) signals. Long Short-Term Memory (LSTM) networks are a natural fit for this problem, but two things routinely undermine them in practice: hyperparameters still get tuned by hand, and real ECG data is badly imbalanced. We built GA-LSTM, a framework in which a Genetic Algorithm (GA) performs an autonomous, budget-constrained search of the LSTM hyperparameter space (hidden units, learning rate, dropout), guided by a macro-F1 fitness function that couples the search itself to class-imbalance robustness, combined with cost-sensitive learning to handle severe class imbalance. On the MIT-BIH Arrhythmia database, the GA-optimized network reached 98.34% accuracy, a weighted F1 of 98.31%, and a Macro-F1 of 90.43%, with the largest gains on rare classes such as Fusion beats. On the PTB Diagnostic ECG Database, the model reached 99.42% accuracy and 99.41% F1, catching 98.39% of true Abnormal cases. The GA-selected configuration outperforms a manually-tuned baseline by a statistically significant margin (McNemar test). GA-LSTM is an accurate, lightweight model whose inference latency is compatible with real-time deployment; we discuss inter-patient validation, noise robustness, and component-wise ablation as open questions for establishing full clinical readiness.