A hybrid DMO-CNN-LSTM framework for feature selection and diabetes prediction: a deep learning perspective

Abstract

The early and accurate prediction of diabetes mellitus remains a significant challenge in clinical decision-making due to the high dimensionality, noise, and heterogeneity of medical data. This study proposes a novel hybrid classification framework that integrates the dwarf mongoose optimization (DMO) algorithm for feature selection with a convolutional neural network–long short-term memory (CNN-LSTM) deep learning architecture for predictive modeling. The DMO algorithm is employed to intelligently select the most informative subset of features from a large-scale diabetes dataset collected from 130 U.S. hospitals over a 10-year period. These optimized features are then processed by the CNN-LSTM model, which combines spatial pattern recognition and temporal sequence learning to enhance predictive accuracy. Extensive experiments were conducted and compared against traditional machine learning models (logistic regression, random forest, XGBoost), baseline deep learning models (MLP, standalone CNN, standalone LSTM), and state-of-the-art hybrid classifiers. The proposed DMO-CNN-LSTM model achieved the highest classification performance with an accuracy of 96.1%, F1-score of 94.6%, and ROC-AUC of 0.96, significantly outperforming other models. Additional analyses, including confusion matrix, ROC curves, training convergence plots, and statistical evaluations confirm the robustness and generalizability of the approach. These findings suggest that the DMO-CNN-LSTM framework offers a powerful and interpretable tool for intelligent diabetes prediction, with strong potential for integration into real-world clinical decision-support systems.