High-dimensional routine healthcare records provide an opportunity to identify individuals at elevated risk of Alzheimer’s disease before formal diagnosis, but the value of thousands of available variables cannot be judged from predictive discrimination alone. This study re-examines an existing machine-learning experiment to determine how aggressive feature-space reduction affects discrimination, class-balanced performance, and model behavior when Alzheimer’s disease is predicted 3–4 years before clinical diagnosis. Integrated primary-healthcare and social-service records were used to construct a 4,311-feature modelling space. Analysis of variance, mutual information, and SHapley Additive exPlanations were used to generate 25- and 50-feature subsets, which were evaluated against the complete feature space across Logistic Regression, Random Forest, Gaussian Naïve Bayes, Multi-Layer Perceptron, XGBoost, and LightGBM. Feature compression produced strongly model-dependent effects. The highest AUC was 0.755 for XGBoost using 50 SHAP-selected variables, whereas the highest balanced accuracy was 0.668 for Logistic Regression using the same feature count and selection strategy. Several compact configurations approached or exceeded the discrimination achieved with all 4,311 variables, but substantial deterioration occurred in other model–selection combinations, particularly Gaussian Naïve Bayes with SHAP-selected features. These findings indicate that dimensionality reduction cannot be evaluated as a single universally beneficial preprocessing step. Compact Alzheimer’s prediction models should instead be judged according to how much discrimination they retain, whether balanced performance improves or deteriorates, how strongly results depend on the learning algorithm, and whether reduced feature spaces support more stable and interpretable modelling.
