An Integrative Multimodal Model for Early Diagnosis of Dementia and Differential Diagnosis of Alzheimer's Disease Using Neuroimaging, Polygenic Risk, and Cognitive Assessments

Abstract

Background: Early diagnosis and etiological classification of dementia remain challenging, as clinicians typically lack tools to integrate cognitive, neuroimaging, and genetic data quantitatively. We developed and validated multimodal risk models to support early diagnosis of dementia and differential diagnosis of Alzheimer's disease (AD) versus non-AD dementias in real-world clinical settings and translated model outputs into individualized risk reports. Methods: Utilizing real-world clinical cohorts (n = 1,100 for early diagnosis of dementia, using clinical diagnoses up to three years after clinical assessment; n = 788 for AD differential diagnosis) from Norwegian Memory Clinics, we trained and validated the Multimodal Hazard Score for Real-World Data (MHS-RWD) model integrating demographics (age, sex), cognitive assessments (MMSE-NR3 or CERAD 10-word delayed recall), the MRI-derived Imaging Hazard Score, and the Polygenic Hazard Score. Discrimination performance was examined using the area under the receiver operating characteristic curve (AUC). Results: In real-world clinical data, the MHS-RWD consistently outperformed any single predictor used alone. For early diagnosis of dementia, the full model achieved an AUC of 0.89 in females and 0.84 in males. For the differential diagnosis of AD from other dementias, the multimodal model yielded an AUC of 0.91 in females and 0.83 in males. A patient-level risk report was designed to present individualized risk estimates. Conclusions: Multimodal integration of cognitive, neuroimaging, and polygenic data in the MHS-RWD tool yields strong discrimination for both early diagnosis of dementia and AD differential diagnosis. The tool relies on data obtainable in clinical care, and genetic information that is becoming increasingly available in routine practice. Delivered through intuitive patient-level risk reports, it could support etiologically informed dementia decisions in real-world settings, with potential utility in primary care.