A Machine Learning Framework for Binary and Multiclass Classification of Cognitive Impairment Using Structural MRI Biomarkers

Abstract

Introduction
Alzheimer’s disease (AD) is the leading cause of dementia worldwide and is characterized by progressive decline in cognitive functions. Mild Cognitive Impairment (MCI) represents an intermediate stage between normal aging and dementia, with measurable cognitive decline but largely preserved functional independence. However, overlapping clinical features and heterogeneous disease patterns may limit the accuracy of conventional classification approaches. Machine learning (ML) methods can support clinical decision-making by identifying complex neuroanatomical patterns in neuroimaging data and improving the characterization of different stages of cognitive impairment.     
Objectives
This study aimed to compare different supervised ML algorithms for binary and multiclass classification of cognitive impairment using structural MRI-derived volumetric biomarkers. Specifically, we evaluated the performance of models trained using neuroanatomical features alone and models integrating MRI biomarkers with demographic variables. In addition, SHAP-based explainability analysis was applied to identify the features contributing most to model predictions and to improve the interpretability of the classification framework.   
Methods
The study included 204 participants, comprising 54 patients with AD, 79 individuals with MCI, and 71 HC. T1-weighted structural MRI scans were processed using automated pipelines to extract cortical and subcortical volumetric measures. Five supervised ML algorithms were evaluated: Logistic Regression, Random Forest, XGBoost, LightGBM, and Support Vector Machine (SVM). All models were implemented within a standardized pipeline including feature scaling, feature selection using SelectKBest with ANOVA F-test, hyperparameter tuning with GridSearchCV, and stratified 5-fold cross-validation. Two experimental settings were compared: models trained only on MRI-derived volumetric features and models trained on a combination of MRI and demographic features. Performance was evaluated using accuracy, F1-score, ROC-AUC, and confusion matrices. Finally, SHAP analysis was applied to the best-performing models to assess the contribution of individual features to classification decisions.
Results
In the binary classification task distinguishing MCI/AD from HC, all models achieved higher performance than in the multiclass setting. Using only MRI-derived volumetric features, Logistic Regression showed the best performance, reaching an accuracy of 0.70 and a ROC-AUC of 0.76. The inclusion of demographic variables further improved classification performance, with SVM achieving the highest results, corresponding to an accuracy of 0.75 and a ROC-AUC of 0.80. In the multiclass classification task involving AD, MCI, and HC, Random Forest achieved the best performance when using only volumetric biomarkers, with an accuracy of 0.53 and a ROC-AUC of 0.69. When MRI and demographic features were combined, SVM achieved the best performance, reaching an accuracy of 0.56 and a ROC-AUC of 0.72. SHAP analysis identified the amygdala, hippocampus, corpus callosum subdivisions, and age as the most influential features driving model predictions.               
Conclusions
The findings support the potential role of structural MRI biomarkers combined with interpretable ML approaches for the classification of cognitive impairment. The integration of demographic variables improved model performance, particularly in binary discrimination tasks, whereas multiclass classification remained more challenging, likely due to the clinical and neuroanatomical overlap between MCI and AD patterns. This framework may represent a promising approach for future clinical decision-support tools aimed at improving diagnostic stratification, although further validation in larger independent cohorts is required.