Hybrid Variational Autoencoder vs XGBoost for Diabetes Mellitus Prediction: A Latent Space-Based Approach

Abstract

Diabetes mellitus is a leading cause of morbidity and mortality worldwide, making its early prediction a public health priority. This study compares the performance of Extreme Gradient Boosting (XGBoost) and a Hybrid Variational Autoencoder (Hybrid VAE) for diabetes classification, evaluating both their predictive accuracy and clinical interpretability. Using the scikit-learn diabetes dataset (442 samples, 10 clinical variables), two models were implemented: XGBoost with hyperparameter optimization and a Hybrid VAE with an 8-dimensional latent space designed to learn interpretable representations of underlying physiological factors. Accuracy, precision, recall, F1-score, and AUC-ROC were assessed, along with latent space analysis using PCA. The Hybrid VAE outperformed XGBoost in all evaluated metrics: accuracy (73.03% vs. 69.66%), recall (79.55% vs. 70.45%), F1-score (0.7447 vs. 0.6966), and AUC-ROC (0.8045 vs. 0.7702). Latent space analysis revealed a natural separation between diabetic and non-diabetic patients in the principal components, with a cumulative explained variance of 64.0%. The importance of features in XGBoost identified body mass index (BMI) and serum S5 measurement as the most relevant predictors. The Hybrid VAE demonstrates superior performance to XGBoost in diabetes prediction, combining high predictive accuracy with the added advantage of an interpretable latent space that captures the underlying structure of the disease. This hybrid approach represents a promising alternative for clinical applications where both accuracy and understanding of the underlying mechanisms are critical.