Interpretable machine learning model based on multimodal MRI radiomics for Alzheimer's disease diagnosis

Abstract

Objective Alzheimer's disease (AD), the most common neurodegenerative disorder, is a leading cause of cognitive impairment and dementia in older adults. This study aimed to develop an interpretable machine learning model using multimodal MRI radiomics for the diagnosis of Alzheimer's disease. Materials and methods A total of 110 subjects (48 AD, 62 healthy control subjects) underwent 3D T1WI, DWI, and T2WI scans. Radiomics features were extracted from eight AD-related brain regions and selected using a three-step approach: variance thresholding, independent t-test, and LASSO–all conducted strictly within the training cohort. Logistic regression (LR) and random forest (RF) models were constructed for single-sequence and combined-sequence data. Model performance was evaluated using ROC analysis, calibration curves, and decision curve analysis. SHapley Additive exPlanations (SHAP) was applied for interpretability. Results Sixteen core radiomics features were retained. Combined-sequence models outperformed single-sequence models, achieving test AUCs of 0.989 and 0.970 for LR and RF, respectively. The LR combined-sequence model achieved an accuracy of 0.882, sensitivity of 0.800, and specificity of 0.947. SHAP analysis identified texture features from the parietal lobe as key contributors. A nomogram integrating radiomics and clinical factors (homocysteine, triglycerides) demonstrated excellent calibration and clinical net benefit. Conclusion Multimodal MRI radiomics combined with interpretable machine learning provides an accurate and explainable tool for AD diagnosis, with the combined LR model exhibiting superior performance.