Subspace corrected relevance learning with application in neuroimaging.

Area of Science:

• Machine Learning
• Neuroimaging Analysis
• Medical Data Science

Background:

• Combining data from multiple sources in machine learning can introduce unwanted variation.
• Multi-center neuroimaging data, such as FDG-PET scans for neurodegenerative diseases, often exhibits center-dependent variations due to differences in equipment and protocols.

Purpose of the Study:

• To develop a method to mitigate center-dependent variations in multi-center datasets.
• To improve the generalization and interpretability of machine learning models for classifying neurodegenerative diseases, specifically Parkinson's disease stages.

Main Methods:

• A two-step approach using Generalized Matrix Learning Vector Quantization (GMLVQ) was proposed.
• Step 1: Train GMLVQ on healthy control data to identify a 'relevance space' distinguishing between data collection centers.
• Step 2: Construct a correction matrix from the relevance space to refine a second GMLVQ system for the diagnostic task.

Main Results:

• The proposed method successfully reduced bias in machine learning systems by eliminating center-specific information during training.
• The approach yielded more informative relevance profiles interpretable by medical experts.
• Effectiveness was validated on both real-world multi-center datasets and simulated data.

Conclusions:

• The developed method effectively addresses extraneous variation in multi-center data for improved machine learning model performance.
• This technique enhances specificity and interpretability in neurodegenerative disease classification.
• The approach is adaptable to other domains requiring the identification of a relevant space to construct a correction matrix.