To establish a prediction model for the conversion risk from mild cognitive impairment (MCI) to Alzheimer's disease (AD) based on structural MRI features, and employ interpretable algorithms to quantify the contribution of critical brain region features to model discrimination.

A total of 356 sets of T1-weighted structural MRI data were enrolled from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. All samples were stratified and randomly divided into a training set containing 237 subjects [119 mild cognitive impairment converters (MCIc) and 118 mild cognitive impairment non-converters (MCInc)] and an independent test set consisting of 119 subjects (60 MCIc and 59 MCInc). Five-fold cross-validation was conducted on the training set to evaluate model stability. FreeSurfer v7.4.1 software was used to extract 1,481 brain structural features for constructing the feature matrix. Boruta nonlinear wrapper feature selection was performed, and 17 core features were retained ultimately. Four models, including support vector machine (SVM), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM) and tabular prior-data fitted network (TabPFN), were constructed using the original and filtered feature sets separately. Logistic regression combined with LASSO regression (LR+LASSO) was introduced as a classic linear baseline for comparative analysis. SHapley Additive exPlanations (SHAP) was further applied to interpret brain region features associated with high conversion risk.

After Boruta feature screening, the separability of samples in the principal component space was improved substantially. The 17 retained features mainly involve brain regions such as the hippocampus, amygdala, temporal lobe and parietal lobe. On the independent test set, the Boruta+TabPFN pipeline achieved an accuracy of 0.765, an area under curve (AUC) of 0.823 and a false negative rate (FNR) of 0.200. Compared with LR+LASSO, Boruta+TabPFN achieved better performance in accuracy, precision, specificity and AUC. However, LR+LASSO yielded slightly higher recall at the target recall threshold, indicating that the linear baseline model prioritizes high sensitivity at the expense of specificity.

Effective nonlinear feature screening is required to build robust discriminative models with high-dimensional structural MRI features. Although Boruta+TabPFN did not obtain the optimal value for every single evaluation metric, it delivered steady performance in AUC, recall and false negative control, which better satisfies the clinical requirement of identifying high-risk individuals in the early screening of MCI-to-AD conversion. SHAP interpretability analysis demonstrated that the risk of MCI progression to AD is driven by coordinated morphological changes across multiple brain regions including the hippocampus, amygdala, temporal lobe, parietal lobe and limbic system, providing interpretable imaging evidence for early disease risk evaluation.