Optimizing deep neural network using genetic algorithm for hypertension risk prediction on mobile apps

Hypertension is a major global health challenge influenced by genetic, environmental, and lifestyle factors. Single Nucleotide Polymorphisms (SNPs) provide an opportunity to support genetic risk assessment by identifying individuals with increased susceptibility before clinical symptoms appear. This study proposes a hybrid Deep Neural Network and Genetic Algorithm (DNN + GA) framework for SNP-based hypertension risk stratification. The Genetic Algorithm was used to optimize the DNN architecture and training hyperparameters, thereby improving predictive performance on high-dimensional genomic data while reducing manual trial-and-error during model design. The cleaned genotype dataset consisted of 2051 samples and 4586 input features. The binary risk label was derived from rs699-based grouping, whereas rs699 itself was excluded from the predictor set to avoid data leakage. After preprocessing, no missing values remained in the dataset. Model performance was evaluated using stratified 5-fold cross-validation. To address class imbalance, SMOTE was applied only within the training folds, preserving unbiased testing and preventing synthetic samples from contaminating validation data. The proposed DNN + GA model achieved the best overall performance by optimizing the DNN architecture and training hyperparameters, with an accuracy of 95.63 ± 1.18%, precision of 95.45 ± 1.15%, sensitivity of 95.20 ± 1.24%, specificity of 96.00 ± 1.08%, F1-score of 95.32 ± 1.17%, and ROC-AUC of 0.982 ± 0.010. Wilcoxon signed-rank testing showed significant improvement over the baseline DNN, with p = 0.043 and a large effect size of r = 0.90. The optimized model was also deployed using TensorFlow Lite for lightweight mobile inference.