ACS-DNN: A hybrid Adam-Cuckoo Search deep neural network with dynamic learning rate for multi-class disease risk stratification in electronic health records

Electronic health records (EHRs) contain rich clinical data that can enable automated disease risk stratification, yet existing deep learning approaches for EHR-based risk assessment suffer from slow convergence, overfitting, and suboptimal parameter optimization in high-dimensional feature spaces. This paper presents ACS-DNN, a novel Adam-Cuckoo Search-based Deep Neural Network that addresses these limitations through two key innovations: (1) Intra-training Lévy-flight exploration is integrated directly into the parameter update process to enhance parameter-space exploration and avoid premature convergence; and (2) a gradient-variance-driven dynamic global learning rate that adapts the base learning rate based on second moment estimates. The ACS-DNN classifies EHR instances into three clinically meaningful risk levels: Normal, Less Critical, and Very Critical. Evaluated on the Cleveland Heart Disease dataset and the Pima Indian Diabetes dataset using stratified 10-fold cross-validation, ACS-DNN achieves classification accuracies of 95.4% and 94.3%, respectively, outperforming seven comparative methods, including Random Forest, XGBoost, LSTM, PSO-DNN, WOA-DNN, SGD-DNN, and Adam-DNN, across the evaluated performance metrics. Furthermore, ACS-DNN converges in approximately 30 epochs compared to 40 (Adam) and 50+ (SGD) epochs, demonstrating a 25–40% improvement in convergence speed. These results demonstrate the effectiveness of the proposed hybrid optimization strategy for automated multi-class EHR risk stratification.