Multimodal Ordinal Modeling for Alzheimer's Disease Severity Assessment: An AI Framework Integrating MRI and Clinical Data

The research team proposed a multimodal machine learning framework incorporating attention mechanisms, enabling automatic staging of Alzheimer's disease (AD) severity via ordinal regression. This framework integrates T1-weighted MRI, demographic information, and genetic data, outperforming unimodal methods on the three datasets ADNI, AIBL, and NIFD. It also provides interpretability analysis using Grad CAM++ and SHAP techniques, offering an accurate and reliable AI-assisted tool for AD assessment.

Alzheimer's disease is the most common neurodegenerative disease globally, affecting over 55 million patients. Current clinical staging methods have limitations such as time-intensive (neuropsychological assessments take hours), high variability (subjective differences among evaluators), and resource-intensive (relying on professional doctors). Thus, there is an urgent need for automated and scalable AI assessment tools.

The objective of this study is to develop an automated AD severity assessment system with core features including: 1. Multimodal integration (imaging, demographic, genetic information); 2. Ordinal modeling (respecting the natural order of disease stages); 3. Interpretability (providing clinically understandable decision-making basis); 4. Scalability (handling large-scale data).

Data Modality Integration: T1-weighted MRI (captures AD features such as hippocampal atrophy, cortical thickness, and ventricular enlargement), demographic variables (age, gender, education level), and genetic information (APOE genotype). Ordinal Regression Framework: Adopts ordinal regression instead of standard classification, respecting the ordered structure of the Alzheimer's Disease Clinical Dementia Rating (CDR) scale (CDR0: normal → CDR3: severe dementia), and focuses on adjacent stage accuracy (the ratio of predictions where the difference from the true label is ≤1 stage). Attention Mechanism: Adaptively weights different modalities and features, focusing on AD-related anatomical regions in MRI while enhancing model interpretability.

Experimental design is rigorous to ensure result reliability:

• Multi-dataset Integration: Uses three independent datasets—ADNI (the largest AD cohort), AIBL, and NIFD—to enhance generalization ability.
• Cohort Stratified Division: Training/validation/test sets are representative across all datasets.
• Strict Isolation of Test Set: Test set subjects are completely excluded from training, validation, preprocessing, and hyperparameter tuning; multiple scans of the same subject do not cross sets to avoid data leakage.

Unimodal vs Multimodal:

• Unimodal: T1 MRI model has an adjacent stage accuracy of 0.963 and QWK (consistency with clinical staging) of 0.444; the tabular model (demographics + genetics) has a QWK of 0.433.
• Multimodal: Non-ordinal model has the lowest MAE of 0.340; ordinal model has the highest adjacent stage accuracy of 0.970 and the strongest consistency QWK of 0.549. Value of Ordinal Modeling: Better captures the ordered structure of CDR, predictions are more consistent with clinical staging, and adjacent stage accuracy is superior.

Grad CAM++ Analysis: Visualizes key MRI regions, including the hippocampus (core affected area of AD), medial temporal lobe (related to memory), and cortical regions (brain atrophy), which aligns with known AD pathology. SHAP Analysis: Reveals feature importance—age, hippocampal volume, APOE genotype, and education level are the main predictors, consistent with clinical knowledge, enhancing model credibility.

Clinical Significance: Assists diagnosis (provides second opinions, quantitative indicators, trend analysis); screening tool (preliminary screening, risk stratification, progress monitoring); research applications (clinical trial evaluation, epidemiological analysis, prognosis prediction). Limitations: Based on cross-sectional data, longitudinal studies are needed to verify progress tracking ability; needs validation in more diverse populations (different ethnicities, regions). Future Directions: Longitudinal studies; integration of more modalities such as PET imaging/cerebrospinal fluid biomarkers/cognitive tests; integration into clinical workflows (user interface, uncertainty quantification, anomaly detection).