Biomarkers

Area of Science:

• Neuroscience
• Computational Biology
• Machine Learning

Background:

• Frontotemporal dementia (FTD) involves cognitive decline affecting behavior, judgment, and language.
• FTD subtypes (bvFTD, SD, nfvPPA) are often misdiagnosed, highlighting the need for better diagnostic tools.
• Current FTD models lack continuous metrics and longitudinal data analysis.

Purpose of the Study:

• To develop and validate a trajectory modeling approach for enhanced characterization of FTD cognitive progression.
• To adapt and implement the Generalized Matrix Learning Vector Quantization (GMLVQ) algorithm for FTD.
• To improve the accuracy of FTD diagnosis and subtyping using machine learning.

Main Methods:

• Utilized the Neuroimaging in Frontotemporal Dementia (NIFD) dataset with longitudinal clinical, cognitive, and MRI data.
• Applied Freesurfer to extract approximately 180 neuroimaging features, including cortical thickness and volume.
• Extended the GMLVQ algorithm for both binary and multi-class classification of FTD subtypes.

Main Results:

• Achieved 94.4% accuracy in binary classification distinguishing semantic dementia (SD) from other FTD subtypes.
• Attained 79.9% accuracy in multi-class classification of all three FTD subtypes (bvFTD, SD, nfvPPA).
• Demonstrated 52.2% accuracy in a six-class classifier including FTD subtypes, Alzheimer's disease, MCI, and controls.

Conclusions:

• GMLVQ trajectory modeling shows potential for advancing FTD diagnosis and assessment.
• Further tuning of the multi-class model is recommended for improved performance.
• The approach offers a promising direction for analyzing longitudinal neurodegenerative disease data.