MRI-Based Multi-Class Relevance Vector Machine Classification of Neurodegenerative Diseases

Area of Science:

• Neuroimaging and computational neuroscience
• Neurology and neurodegenerative disease research
• Artificial intelligence in healthcare

Background:

• Machine learning (ML) shows promise for dementia diagnosis, yet misclassification in MRI-based studies remains a challenge.
• Understanding the drivers of misclassification is critical for advancing automated diagnostic tools.
• Existing ML approaches often struggle with the heterogeneity of neurodegenerative diseases.

Purpose of the Study:

• To implement and evaluate a multi-class classification approach using ML for diagnosing seven neurodegenerative conditions from T1-weighted MRI scans.
• To investigate the reasons behind misclassification by comparing ML predictions with clinical, pathological, genetic, and imaging data.
• To assess the diagnostic performance of ML across different dementia subtypes and controls.

Main Methods:

• Utilized a multi-class classification model employing relevance vector machine and logistic classification.
• Analyzed whole-brain T1-weighted MRI scans from 468 participants across seven diagnostic groups: healthy controls, Alzheimer's disease (AD), behavioral variant frontotemporal dementia (bvFTD), semantic variant primary progressive aphasia (svPPA), non-fluent variant primary progressive aphasia (nfvPPA), corticobasal syndrome (CBS), and progressive supranuclear palsy syndrome (PSPS).
• Validated diagnostic accuracy against clinical, pathological, genetic, and quantitative imaging data.

Main Results:

• The algorithm achieved 71% accuracy in predicting specific neurodegenerative syndromes, 80% for the disease spectrum, and 85% in distinguishing controls from any dementia.
• High performance was observed for healthy controls, moderate for AD, bvFTD, and svPPA, and low for CBS, nfvPPA, and PSPS.
• Misclassified cases often exhibited minimal atrophy and brain volumes similar to controls, particularly early-onset AD and bvFTD with specific genetic mutations or phenocopies.

Conclusions:

• Neurodegenerative disease heterogeneity and the limitations of structural MRI in capturing biological changes contribute to ML misclassification.
• Larger, more inclusive datasets reflecting population heterogeneity are necessary for training robust ML models.
• A degree of uncertainty and margin of error in automated diagnoses is expected and should be considered analogous to clinical diagnostic uncertainty.