Multimodal manifold-regularized transfer learning for MCI conversion prediction.

Area of Science:

• Neuroscience
• Medical Imaging
• Machine Learning

Background:

• Mild cognitive impairment (MCI) is an early stage of Alzheimer's disease (AD) with a high conversion rate.
• Accurate prediction of MCI to AD conversion is crucial for early diagnosis and pre-symptomatic risk assessment.
• Existing methods often lack robustness by relying solely on target domain samples.

Purpose of the Study:

• To develop a novel multimodal manifold-regularized transfer learning (M2TL) method for enhanced MCI conversion prediction.
• To leverage auxiliary domain data (AD vs. normal controls) and unlabeled samples to improve prediction performance.
• To enhance classifier robustness by incorporating sample selection mechanisms.

Main Methods:

• Utilized a kernel-based maximum mean discrepancy criterion to mitigate domain distribution differences.
• Employed a semi-supervised multimodal manifold-regularized least squares classification approach integrating multiple data sources.
• Incorporated a group sparsity constraint for informative sample selection.

Main Results:

• The M2TL method achieved a classification accuracy of 80.1% for MCI conversion prediction on the ADNI database.
• Demonstrated significant improvement over existing state-of-the-art methods.
• Validated the effectiveness of jointly utilizing target, auxiliary, and unlabeled domain samples.

Conclusions:

• The proposed M2TL method offers a robust and effective approach for predicting Alzheimer's disease progression from MCI.
• Transfer learning combined with semi-supervised and manifold regularization techniques enhances predictive accuracy.
• This method holds promise for early diagnosis and risk stratification in Alzheimer's disease.