Modality Imbalance? Dynamic Multi-Modal Knowledge Distillation in Automatic Alzheimer's Disease Recognition.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Medical Informatics

Background:

• Alzheimer's disease (AD) is the most common dementia, requiring early detection for better patient outcomes.
• Machine learning models using speech and text show promise for automated AD detection.
• Existing multi-modal approaches struggle with modality imbalance, leading to suboptimal performance.

Purpose of the Study:

• To propose and evaluate a novel Dynamic Multi-Modal Knowledge Distillation (DMMKD) approach.
• To address the modality imbalance challenge in multi-modal AD detection.
• To improve the accuracy and robustness of automated Alzheimer's disease identification.

Main Methods:

• Developed a Dynamic Multi-Modal Knowledge Distillation (DMMKD) algorithm.
• DMMKD dynamically identifies dominant and weak modalities for targeted knowledge distillation.
• Employed inter(cross)-modal and intra-modal distillation to balance learning speeds.

Main Results:

• DMMKD demonstrated substantial performance improvements over standard multi-modal approaches.
• Relative accuracy gains of 15.4% on the ADReSSo dataset and 10.9% on the ADReSS-M dataset were observed.
• Achieved state-of-the-art accuracies of 91.5% (ADReSSo) and 87.0% (ADReSS-M).

Conclusions:

• The proposed DMMKD effectively mitigates modality imbalance in multi-modal learning for AD detection.
• DMMKD significantly enhances diagnostic accuracy compared to existing methods.
• This approach offers a promising solution for improving early Alzheimer's disease identification.