Classification of Mild Cognitive Impairment With Multimodal Data Using Both Labeled and Unlabeled Samples

Area of Science:

• Neuroscience
• Medical Imaging
• Genetics

Background:

• Mild Cognitive Impairment (MCI) represents a preclinical stage of Alzheimer's Disease (AD).
• Accurate MCI classification is vital for timely AD diagnosis and intervention.
• MCI exhibits significant clinical heterogeneity, complicating diagnosis.

Purpose of the Study:

• To investigate the efficacy of a multimodal co-training method for MCI classification.
• To leverage both labeled and unlabeled data for improved diagnostic performance.
• To integrate structural magnetic resonance imaging (sMRI) and genotype data for enhanced MCI detection.

Main Methods:

• Utilized 364 MCI samples (228 labeled, 136 unlabeled) from the Alzheimer's Disease Neuroimaging Initiative (ADNI-1) cohort.
• Extracted quantitative trait (QT) features from sMRI and SNP features from genotype data.
• Implemented a multimodal co-training framework with random forest classification on an independent ADNI-2 dataset.

Main Results:

• Achieved an accuracy of 85.50% for MCI classification.
• Obtained an Area Under the Curve (AUC) of 0.825.
• Demonstrated that combined sMRI and SNP data significantly improve MCI classification performance.

Conclusions:

• The multimodal co-training approach effectively utilizes both labeled and unlabeled data.
• Integration of sMRI and genotype data offers a powerful strategy for MCI classification.
• This framework shows significant potential for early and accurate diagnosis of MCI.