Bio-Inspired Feature Selection and Deep Neural Networks for Accurate Diabetes Prediction

Abstract

Accurate and early prediction of diabetes is critical for effective disease management and timely intervention. This study proposes a hybrid deep learning framework that integrates a bio-inspired feature selection technique with a deep neural network (DNN) classifier to enhance predictive performance. An improved Shuffled Frog Leaping Algorithm (SFLA) is employed to select the most informative features using a multi-objective fitness function combining classification accuracy and feature subset compactness. The optimised features are then input into a regularised DNN equipped with dropout layers and a softmax-based attention mechanism to improve generalisation and interpretability. Bayesian hyperparameter tuning and early stopping further refine the training process. The model is evaluated on the PIMA Indians Diabetes Dataset using stratified 10-fold cross-validation. The proposed method achieves 86.4% accuracy, 82.1% F1-score, and 0.91 AUC, outperforming traditional classifiers such as SVM, Random Forest, and standard DNNs. This approach provides a robust, scalable, and explainable framework for effective diabetes prediction in clinical environments.