A data-driven approach to post-stroke aphasia classification and lesion-based prediction.

Area of Science:

• Neuroscience
• Computational Linguistics
• Neurology

Background:

• Aphasia, a language disorder post-stroke, is traditionally subtyped using the Wernicke-Lichtheim model.
• This 19th-century model distinguishes aphasia based on language production versus comprehension deficits.
• Modern approaches are needed to refine aphasia classification.

Purpose of the Study:

• To apply data-driven methods to re-examine aphasia subtypes.
• To investigate behavioral deficits and lesion correlates in post-stroke aphasia.
• To compare data-driven subtypes with traditional aphasia classifications.

Main Methods:

• Clustering of individuals with aphasia based on behavioral profiles using community detection analysis (CDA).
• Machine learning (random forest classifiers) to predict cluster membership from lesion data.
• Comparison of CDA-derived clusters with traditional aphasia subtypes.

Main Results:

• CDA revealed distinct clusters that did not align with traditional aphasia subtypes (behaviorally or neuroanatomically).
• The primary distinction identified was between phonological and semantic processing, not production and comprehension.
• Lesion-based classification achieved 75% accuracy for CDA-based categories versus 60% for traditional categories.

Conclusions:

• The findings support a data-driven reclassification of aphasia subtypes.
• A new framework distinguishing phonological and semantic processing is proposed.
• This approach offers a more accurate basis for research and clinical classification of aphasia.