Deep Learning for Type 2 Diabetes Prediction: A Benchmarking Study Using the PIMA Indians Diabetes Dataset

Abstract

Type 2 diabetes mellitus (T2DM) is a chronic metabolic disorder with a growing global burden, making early and accurate risk prediction a public-health priority. This study develops and evaluates a deep neural network (DNN) for T2DM prediction using the PIMA Indian Diabetes dataset (768 records and 8 clinical features). Physiologically implausible zero values in glucose, blood pressure, skin thickness, insulin, and body-mass index were treated as missing and replaced using median imputation. Inputs were then standardized before modelling. The proposed regularized multilayer DNN used batch normalization, dropout, early stopping, and stratified 5-fold cross-validation, followed by evaluation on a held-out 20% test set. We evaluated both models using the same data splits on logistic regression, random forest, and support vector machine models. The DNN achieved an accuracy of 77.27%, 67.27% precision, 68.52% recall, 67.89% F1-score, and an AUC of 83.61% on the independent test set. DNN was also the most discriminative on both datasets. Glucose, BMI, diabetes pedigree function, and age were the dominant predictors. The limitations of this work consist of small size of the dataset used and noisy labels. The next step in future research should be aimed at improving the model's explainability, increasing the number of samples in the cohort, and federated learning.