Identification of progressive mild cognitive impairment patients using incomplete longitudinal MRI scans

Area of Science:

• Neuroimaging
• Machine Learning
• Biostatistics

Background:

• Accurate differentiation between progressive mild cognitive impairment (pMCI) and stable mild cognitive impairment (sMCI) is essential for timely intervention in potential Alzheimer's disease (AD) patients.
• Longitudinal data offers valuable insights into disease progression but can be incomplete, posing challenges for analysis.

Purpose of the Study:

• To develop and evaluate a classification framework for pMCI/sMCI using incomplete longitudinal magnetic resonance imaging (MRI) data.
• To enhance diagnostic accuracy by integrating dynamic features derived from serial MRI scans.

Main Methods:

• Extraction of volumetric and dynamic features from longitudinal MRI scans at baseline and 6, 12, and 18 months.
• Feature selection using elastic net sparse learning algorithms and imputation of missing data via low-rank matrix completion.
• Classification of pMCI/sMCI using a multi-kernel support vector machine (mkSVM) on the completed feature matrix.

Main Results:

• The proposed framework achieved a diagnostic accuracy of 78.2% when incorporating longitudinal MRI information.
• This represents a 6.6% improvement in accuracy compared to using only baseline MRI data.
• The study demonstrates the value of longitudinal disease history in improving diagnostic precision.

Conclusions:

• Harnessing incomplete longitudinal MRI data with advanced machine learning techniques can significantly improve the accuracy of distinguishing pMCI from sMCI.
• The developed framework offers a promising approach for early identification of individuals at risk for Alzheimer's disease.
• Integrating dynamic features and robust imputation methods is key to leveraging the full potential of longitudinal neuroimaging data.