Diagnosis of Alzheimer's Disease Based on Multimodal MRI and Amyloid/Tau PET Fusion: Quantification of Modality Complementarity and Interpretability Analysis

Abstract Background: Alzheimer's disease (AD) diagnosis relies on multimodal information, yet the optimal fusion strategy and the true complementarity among modalities remain unclear. This study aimed to systematically compare different modality configurations for AD diagnosis and to quantify modality complementarity. Methods: A total of 874 participants from the ADNI cohort (CN = 608, AD = 266) were included. A total of 39 features were extracted, comprising structural MRI (20 features), amyloid PET (5), Tau PET (10), and clinical variables (4). We employed a three-step feature selection process using variance thresholding, t-tests, and LASSO to evaluate 7 modality configurations and 5 classifiers, and quantified modality complementarity using centered kernel alignment (CKA), Pearson correlation coefficients, and the Jaccard index. Results: The multimodal fusion model achieved an AUC of 0.9558 in the full dataset and 0.9910 in the complete PET subset, outperforming all single-modality and dual-modality combinations. The Jaccard index for MRI and PET was 0.0000, indicating no overlap in misclassification patterns. SHAP analysis indicated that the hippocampal volume ratio (HIPPO_RATIO) was the most important feature. Conclusions: Multimodal fusion integrating structural MRI, amyloid/Tau PET, and clinical variables provides a robust and interpretable tool for AD diagnosis. More importantly, it addresses a key technical gap in multimodal research---the lack of a quantitative framework for assessing modality complementarity---by demonstrating that MRI and PET exhibit completely non-overlapping error patterns, which constitutes the core mechanism driving the improved performance of multimodal fusion.