Multimodal multitask learning for predicting MCI to AD conversion using stacked polynomial attention network and...

Area of Science:

• Neuroscience
• Medical Imaging
• Machine Learning

Background:

• Early identification and treatment of mild cognitive impairment (MCI) are vital for delaying Alzheimer's disease (AD) progression and preserving cognitive function.
• Precise prediction of MCI stages and conversion to AD is essential for timely diagnosis and intervention strategies.

Purpose of the Study:

• To develop and evaluate a multimodal framework for differentiating early MCI (eMCI) from late MCI (lMCI).
• To predict the conversion timeline of MCI patients to Alzheimer's disease (AD).
• To enhance the representation learning from small multimodal datasets using an attention-based module.

Main Methods:

• A multimodal framework utilizing multitask learning was employed, integrating clinical data and radiomics features from MRI.
• An attention-based module, Stack Polynomial Attention Network (SPAN), was proposed for robust feature encoding from limited data.
• Adaptive Exponential Decay (AED) was used to compute a potent factor for improving multimodal data learning.
• Experiments were conducted on the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort, including 249 eMCI and 427 lMCI participants.

Main Results:

• The proposed multimodal strategy achieved a c-index score of 0.85 for predicting MCI to AD conversion time.
• The framework demonstrated high accuracy in MCI-stage categorization.
• The performance of the developed model was comparable to contemporary research in the field.

Conclusions:

• The multimodal framework effectively differentiates MCI stages and predicts Alzheimer's disease conversion.
• The integration of clinical data, radiomics, and advanced deep learning techniques (SPAN, AED) shows promise for early AD detection.
• This approach offers a valuable tool for clinical decision-making and the development of personalized treatment strategies for MCI and AD.