Data-driven neuroanatomical subtypes of primary progressive aphasia

Area of Science:

• Neuroscience
• Radiology
• Computational Biology

Background:

• Primary progressive aphasias (PPA) are language-dominant dementias with distinct variants (semantic, non-fluent/agrammatic, logopenic).
• Neuroimaging, particularly MRI, struggles to clearly differentiate non-fluent/agrammatic and logopenic PPA variants.
• Previous studies relied on phenotype-driven approaches to map PPA neuroanatomy.

Purpose of the Study:

• To utilize a machine learning algorithm (SuStaIn) to discover data-driven neuroanatomical progression profiles in PPA.
• To characterize the heterogeneity of PPA through in-depth subtype-phenotype analysis.
• To identify distinct neuroanatomical patterns independent of clinical diagnosis.

Main Methods:

• Applied the SuStaIn machine learning algorithm to MRI data from 270 PPA participants and validated on an additional 66 patients.
• Segmented MRI scans and analyzed 19 regions of interest to identify neuroanatomical subtypes and their progression stages.
• Assessed the stability of subtype and stage assignments longitudinally and validated findings in an independent dataset.

Main Results:

• Discovered four robust neuroanatomical subtypes of PPA: S1 (left temporal), S2 (insula), S3 (temporoparietal), and S4 (frontoparietal).
• Found strong correlation between S1 and semantic PPA; S2, S3, and S4 showed mixed associations with logopenic and non-fluent/agrammatic variants.
• Subtype assignment remained stable in 84% of patients, and stage assignment in 91.9% over time, with partial validation in the ALLFTD dataset.

Conclusions:

• Identified separable spatiotemporal neuroanatomical phenotypes within the PPA spectrum using machine learning.
• These data-driven subtypes suggest that PPA heterogeneity is complex and does not always align with traditional clinico-anatomical correlations.
• Understanding these multifaceted neuroanatomical profiles can inform clinical decision support and future research in PPA.