Multimodal prediction of MCI-to-dementia conversion over a two-year window.
Authors
Affiliations (8)
Affiliations (8)
- Center for Advanced Technology and Education, Department of Electrical and Computer Engineering, Florida International University, Miami, FL, USA. [email protected].
- Center for Advanced Technology and Education, Department of Electrical and Computer Engineering, Florida International University, Miami, FL, USA.
- Institute for Human Genetics, Clemson University, Greenwood, SC, USA.
- 1Florida Alzheimer's Disease Research Center, University of Florida, Gainesville, FL, USA.
- Departments of Psychiatry and Behavioral Sciences, University of Miami Miller School of Medicine, Miami, FL, USA.
- Robert Stempel College of Public Health and Social Work, School of Social Work, Florida International University, Miami, FL, USA.
- Knight Foundation School of Computing and Information Sciences, Florida International University, Miami, FL, USA.
- Center for Cognitive Neuroscience and Aging, University of Miami Miller School of Medicine, Miami, FL, USA.
Abstract
Mild cognitive impairment (MCI) represents a heterogeneous clinical state where some individuals remain stable while others progress to dementia. Identification of patients who are at high risk for near-term conversion remains a critical challenge for timely intervention. We developed an explainable multimodal machine-learning framework to predict progression from MCI to dementia within a clinically meaningful two-year period. We analyzed 2008 samples from 828 unique MCI subjects from the Alzheimer's Disease Neuroimaging Initiative (ADNI), including cognitive/functional assessments, demographic/genetic variables, and structural MRI features. Subjects were classified as stable MCI (sMCI, n = 1306) or progressive MCI (pMCI, n = 702) based on two-year outcomes. Machine learning algorithms were evaluated on unimodal and multimodal using nested cross-validation, and explainability was assessed using SHapley Additive exPlanations (SHAP) and permutation feature importance. The combination of cognitive assessments, demographics/risk factors and structural MRI features outperformed individual modalities, highlighting the complementary relationship between cognitive-functional decline and neuroanatomical degeneration. The best-performing model, a calibrated XGBoost, achieved a balanced accuracy of 81.28% ± 2.56% during cross-validation and 81.99% (bootstrap mean 81.97%, 95% CI: [78.22%-85.51%]) on an independent held-out test set. Explainable AI analyses identified functional impairment measures (Functional Activities Questionnaire, Clinical Dementia Rating - Sum of Boxes), cognitive performance scores (Alzheimer's Disease Assessment Scale - Cognitive Subscale 11-item version, Alzheimer's Disease Assessment Scale - Cognitive Subscale 13-item version, Mental State Examination), APOE ε4 status, and structural abnormalities within the hippocampus, lateral ventricle, parietal and temporal regions, and amygdala, as the most influential predictors of conversion. Short-term progression from MCI to dementia can be predicted with high accuracy using multimodal information obtained from a single clinical visit. The convergence of cognitive impairment and region-specific neurodegeneration emerged as the strongest indicator of conversion risk.