Uncovering spatiotemporal patterns of atrophy in progressive supranuclear palsy using unsupervised machine learning

Area of Science:

• Neuroimaging
• Machine Learning
• Neurology

Background:

• Progressive supranuclear palsy (PSP) exhibits significant pathological and phenotypic heterogeneity.
• Understanding the links between PSP subtypes and clinical presentation is crucial for diagnosis and treatment.
• Current diagnostic methods may not fully capture the spectrum of PSP variations.

Purpose of the Study:

• To apply a novel unsupervised machine learning algorithm, Subtype and Stage Inference (SSI), to a large MRI dataset of PSP patients.
• To identify distinct subtypes of PSP based on spatiotemporal patterns of brain atrophy.
• To investigate the relationship between these identified subtypes, clinical phenotypes, and disease progression.

Main Methods:

• Applied SSI to volumetric MRI features from 426 PSP cases and 290 controls.
• Utilized structural T1-weighted MRI scans for baseline and follow-up analyses.
• Compared clinical phenotypes across identified subtypes and validated subtype stability longitudinally.

Main Results:

• Identified two distinct subtypes: 'subcortical' and 'cortical', each with unique atrophy progression patterns.
• Showed strong association between clinical diagnosis and SSI-derived subtypes (e.g., 82% of PSP-cortical cases assigned to cortical subtype).
• Found that the 'subcortical' subtype was associated with worse clinical severity scores compared to the 'cortical' subtype.

Conclusions:

• Empirically identified two distinct image-based subtypes of spatiotemporal atrophy in PSP.
• These subtypes correlate with specific clinical syndromes and exhibit different clinical characteristics.
• Accurate subtyping and staging at baseline have implications for clinical trial screening and disease progression tracking.