Benchmarking Regression Models for Predicting Diabetes Progression: A Reproducible Evaluation

Abstract Predicting future disease progression from baseline measurements is a useful research task for studying clinical risk modelling, but predictive performance must be interpreted in light of dataset size and evaluation design. This paper compares Linear Regression, Ridge Regression, Random Forest, and Gradient Boosting on the diabetes progression dataset distributed with scikit-learn. The dataset contains 442 observations and 10 baseline variables. A fixed 80/20 train–test split (random_state=42) was used, and five-fold cross-validation was conducted on the training portion. Models were assessed using mean absolute error (MAE), root mean squared error (RMSE), and coefficient of determination (R²). On this run, the best held-out R² was 0.454 for Ridge Regression, while test MAE ranged from 42.794 to 44.603. The experiment provides a transparent baseline comparison, not a clinical prediction tool. The small dataset, single split, and lack of external validation limit generalization. The code and software environment should be archived alongside any public submission.