Answers are grounded in this paper and cite their sources.
Background Structural magnetic resonance imaging (MRI)-based machine learning has shown potential for distinguishing cognitively normal (CN) individuals from those with mild cognitive impairment (MCI). However, previous classification studies have often emphasized discrimination, while calibration, comparative evaluation across different algorithms, and anatomically interpretable model contributions have been less consistently integrated. Moreover, the identification of important MRI features does not necessarily establish their independent biological significance. This study aimed to develop and internally evaluate a structural MRI-based machine-learning framework for CN–MCI classification using data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). Methods This retrospective cross-sectional study included 2,649 participants aged over 65 years from ADNI, comprising 1,285 CN individuals and 1,364 participants with MCI. Eighteen MRI-derived morphometric features selected using least absolute shrinkage and selection operator (LASSO) regression in the training set were combined with age, sex, and education. Five algorithms—logistic regression, support vector machine (SVM), random forest, Light Gradient Boosting Machine (LightGBM), and extreme gradient boosting (XGBoost)—were compared using an internally held-out test set. Model performance was assessed using discrimination, classification metrics, calibration, and decision-curve analysis. Shapley Additive Explanations (SHAP) were used for post hoc model interpretation. Results The XGBoost-based final model achieved an area under the receiver operating characteristic curve (AUC) of 0.917 (95% confidence interval [CI], 0.889–0.945), compared with 0.865 (95% CI, 0.836–0.894) for logistic regression (paired DeLong P = 0.048). The final model achieved an accuracy of 0.847, sensitivity of 0.854, specificity of 0.839, and Brier score of 0.143. The most influential MRI features included bilateral inferior lateral ventricular and hippocampal volumes and left middle temporal cortical thickness. These rankings describe contributions to model predictions rather than independently established biomarkers or disease mechanisms. Conclusions A structural MRI-based machine-learning framework incorporating demographic information demonstrated good internal discrimination between CN and MCI participants in the ADNI cohort. Calibration assessment and post hoc feature interpretation provided complementary information beyond discrimination alone. The findings should be interpreted as cohort-specific classification evidence, and independent external validation is required before clinical application. Clinical trial number Not applicable.
No takes yet. Share an insight, caveat, or question.
Wu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: