Machine Learning–Driven Prediction of Dementia from MRI and Clinical Features: A Comparative Analysis of Ensemble and Baseline Models

Early detection of dementia remains a pressing challenge in clinical neuroscience, as delayed diagnosis limits therapeutic impact and healthcare planning. Leveraging the Open Access Series of Imaging Studies (OASIS) cross-sectional dataset of 436 participants, this study developed a robust machine learning pipeline integrating sociodemographic, clinical, and neuroimaging-derived features. Preprocessing included removal of highly sparse variables (Delay), median imputation of partially missing but clinically essential measures (SES, MMSE, CDR, Educ), Winsorization of extreme values, and skewness correction. The target Clinical Dementia Rating (CDR) was binarized (0 = no dementia, ≥ 0.5 = dementia) to align with clinically actionable screening. Categorical features were numerically encoded, and StandardScaler() was selectively applied to models sensitive to feature magnitude (Logistic Regression, KNN), while ensemble methods (Random Forest, XGBoost, LightGBM) were trained on raw inputs. Stratified shuffle splits of 70:30 and 60:40 were repeated 20 times, with Synthetic Minority Oversampling Technique (SMOTE) applied exclusively to training sets. Performance was assessed using accuracy, precision, recall, F1-score, specificity, and ROC-AUC, complemented by confusion matrices and ROC curves for the best-performing runs. Results indicated superior performance of ensemble learners, with LightGBM yielding the most balanced outcomes (ROC-AUC = 0.954 ± 0.015; accuracy = 0.890 ± 0.022), while XGBoost achieved the highest recall (0.911 ± 0.026), reducing false negatives. These findings demonstrate that gradient-boosting ensembles provide strong and stable performance under repeated internal validation, supporting their use as a comparative methodological baseline for dementia prediction within the studied cohort.