Towards Accurate Diabetes Prediction: A Synergistic Approach Using Adaptive Deep Learning Techniques

Diabetes is a burgeoning international health issue, requiring effective and precise predictive models. Here, we compare Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) for predicting diabetes using the Pima Indians Diabetes Dataset (containing 768 samples with 8 clinical features). For improving model performance, we utilized data normalization, removal of outliers, and feature scaling in preprocessing. A 5-fold cross-validation approach was used to maintain robustness. The best CNN model had an accuracy of 85.20%, which was better than the RNN model (83.50%) and current methods like CNN-RNN hybrids, Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) models. Hyper parameters like learning rate, batch size, and neurons were optimized to prevent over fitting and gain maximum performance. Comparative analysis depicts that deep learning models, i.e., CNNs, can effectively distinguish diabetic and non-diabetic cases. In future work, emphasis will also be given to blending other datasets and optimization techniques for greater generalizability.