Hybrid Deep Learning Approach for Heart Disease Prediction using Multi-Head Attention Network and Stacked Long Short-Term Memory

Heart disease (HD) is a specific abnormal condition that includes disorders occurring in the heart and affecting the bloodstream, while insufficient blood vessel function impacts heart rhythm. Diagnosis of HD from both patient history and medical data such as blood sugar, blood pressure, age, genetics, and cholesterol is necessary due to the increasing number of deaths and the need to minimize the risk of HD. However, irregularities in data patterns such as patient variability, pressure rate, and ECG signals make it difficult to recognize the hidden patterns of HD, which complicates the detection of its signs. To overcome these challenges, the proposed Multi-Head Attention Network–Stacked Long Short-Term Memory (MHAN-SLSTM) is introduced for accurate HD prediction. The proposed MHAN-SLSTM improves HD prediction by integrating multi-head attention for capturing complex feature dependencies with SLSTM for deep hierarchical learning. Missing values and min-max normalization techniques are applied during the preprocessing stage. Cheetah Optimization (CO) is utilized in the feature selection stage to enhance prediction accuracy. The proposed MHAN-SLSTM achieves 99.36% accuracy when compared with the Convolutional Neural Network–Optimized Gated Recurrent Unit (CNN+OGRU).