Taming arbitrary modality missingness and imbalance: A unified graph-MoE framework for Alzheimer's disease diagnosis
Guangqian Yang1, Ye Du1, Xiaowei Hu2
1Department of Biomedical Engineering, The Hong Kong Polytechnic University, Hong Kong Special Administrative Region of China.
Medical Image Analysis
|July 23, 2026
Summary
This study introduces UniMIX-AD, a novel framework for Alzheimer's disease diagnosis using multimodal biomarkers. It effectively handles missing data and imbalanced combinations, improving accuracy for rare but critical diagnostic cases.
Area of Science:
- Biomedical Informatics
- Machine Learning for Healthcare
- Neurodegenerative Disease Diagnostics
Background:
- Multimodal biomarkers show promise for Alzheimer's disease (AD) diagnosis.
- Current multimodal learning methods fail in real-world clinical settings due to missing data and imbalanced modality combinations.
- Existing models are biased towards common data patterns, neglecting rare but crucial combinations for early AD detection.
Purpose of the Study:
- To develop a robust framework, UniMIX-AD, for accurate AD diagnosis under realistic, imperfect multimodal data conditions.
- To address challenges of arbitrary modality missingness and long-tailed modality-combination imbalance in clinical datasets.
- To improve diagnostic accuracy, especially for rare combinations critical for early AD identification.
Main Methods:
- UniMIX-AD integrates a Unified Missing-modality Imputation (UMI) module to reconstruct missing data by leveraging cross-modal dependencies.
- A Graph-coordinated Sparse Mixture-of-Experts (G-Sparse-MoE) adaptively routes patients and transfers knowledge from frequent to rare modality combinations.
- The framework ensures comprehensive patient representations and mitigates the long-tailed imbalance.
Main Results:
- UniMIX-AD demonstrated superior performance compared to state-of-the-art methods on ADNI and OASIS-3 datasets.
- Achieved a 7.1% relative average accuracy improvement across datasets.
- Provided a significant accuracy boost of up to 14.3% for rare (tail) modality combinations in the ADNI dataset.
Conclusions:
- UniMIX-AD offers a robust solution for Alzheimer's disease diagnosis in practical, multimodal clinical scenarios.
- The framework successfully handles missing data and modality-combination imbalance, enhancing diagnostic performance.
- UniMIX-AD shows significant potential for improving early and accurate detection of Alzheimer's disease.
Related Concept Videos
01:22Dementia l: Introduction
Dementia is an acquired, progressive syndrome characterized by a decline in multiple cognitive domains severe enough to impair daily functioning and reduce independence. Although memory loss is a central feature, the diagnosis requires additional deficits involving language, executive function, visuospatial skills, judgment, calculation, or abstract reasoning. These cognitive impairments reflect underlying neurodegenerative or vascular processes that gradually disrupt neuronal networks...
01:26Alzheimer's Disease: Overview
Alzheimer's Disease (AD) is a continually advancing neurodegenerative disorder, distinguished by escalating memory loss, cognitive dysfunction, and dementia. The disease unfolds in three stages: preclinical, mild cognitive impairment (MCI), and dementia. Its onset is insidious, and the progression gradual, with the cause not well explained by other disorders.
The clinical diagnosis of AD hinges on the presence of memory and other cognitive impairments. Biomarkers, such as changes in Aβ and tau...
The clinical diagnosis of AD hinges on the presence of memory and other cognitive impairments. Biomarkers, such as changes in Aβ and tau...
01:29Alzheimer Disease l: Introduction
Alzheimer disease is a chronic, progressive, and irreversible neurodegenerative disorder and the most common cause of dementia in older adults. It leads to gradual neuronal loss, causing cognitive decline, behavioral changes, and loss of functional independence.Risk Factors and EtiologyThe disease is multifactorial. Age is the strongest risk factor, with prevalence doubling every 5 years after age 65. Genetic factors include mutations in genes such as APP, PSEN1, and PSEN2, which are associated...