Adaptive feature optimization in deep neural architectures for intelligent diabetes prediction

Objective & Problem Statement The proliferation of high-dimensional electronic health records (EHRs) presents a critical computational bottleneck in clinical predictive analytics, as conventional models frequently succumb to the “curse of dimensionality,” resulting in overfitting and diminished generalization capabilities. Proposed Methodology To resolve this deficiency, this study proposes a novel computational architecture that integrates an Adaptive Genetic Algorithm (AGA) with a Hybrid Deep Neural Network (HDNN) for simultaneous dynamic feature selection and classification. Unlike static selection paradigms, the AGA continuously optimizes the feature subspace during training via classification-driven feedback loops. Concurrently, the HDNN enhances representational learning by fusing Convolutional Neural Networks (CNNs) for local spatial extraction, Recurrent Neural Networks (RNNs) for modeling temporal clinical dependencies, and an attention mechanism to explicitly prioritize salient medical indicators. Key Findings (Numerical) Empirical evaluations conducted on the comprehensive UCI Diabetes 130-US dataset, validated across independent runs (reported as Mean±SD), establish the superiority of the proposed framework. The AGA-HDNN architecture achieved a classification accuracy of utilizing an optimized subset of features. Furthermore, the model demonstrated exceptional robustness, maintaining a high predictive accuracy of when restricted to a minimal subset of strictly features. Significance & Impact These findings substantiate the efficacy of the proposed framework as a computationally lean and highly robust paradigm for high-dimensional predictive modeling, offering a scalable methodology for integrating dynamic evolutionary feature optimization with deep learning in complex clinical environments.