Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Word-to-Lesion Mapping: Using which words are produced to predict which brain regions are lesioned
Poster Session B, Wednesday, September 30, 4:30 - 6:30 pm, Wangari Maathai
Junhua Ding1, Daniel Mirman2; 1Institute of Psychology, Chinese Academy of Sciences, 2The University of Edinburgh
Introduction: Speech production data are widely used in the diagnosis of neurological disorders, typically relying on coarse summary measures (e.g., words per minute) or clinician-rated fluency (Alyahya et al., 2020; Bryant et al., 2016). These measures have also supported lesion–symptom mapping approaches, which aim to predict behavioral deficits from patterns of brain damage (Bates et al., 2003; Fridriksson et al., 2018). In parallel, research in natural language processing has demonstrated that linguistic features of spoken language can reveal latent characteristics of speakers, including personality traits and early markers of neurodegenerative disease (Cook et al., 2016). However, existing approaches largely focus on mapping from brain lesions to behavioral outcomes, leaving the inverse relationship—whether linguistic output can predict underlying neural damage—relatively underexplored. To address this gap, we developed a novel framework termed word-to-lesion mapping, which leverages the lexical content of connected speech to infer the anatomical location of brain lesions in individuals with post-stroke aphasia. Methods: Using a dataset of connected speech samples from 60 individuals with aphasia following left-hemisphere stroke, each with corresponding neuroimaging data, we evaluated the extent to which word-level information can predict lesion location. We hypothesized that (i) linguistic features embedded in word usage would carry sufficient information to enable above-chance prediction of lesion distribution, and (ii) systematic associations between lexical items and specific brain regions would emerge, reflecting underlying neurocognitive organization. Results: Our results demonstrate that word-based lesion prediction significantly exceeds permutation-based baselines across the lesion territory, achieving accuracies of up to 75%, and enables high-fidelity reconstruction of individual lesion patterns, with peak accuracy approaching 99%. Furthermore, analysis of the similarity structure of prediction weights across words and brain regions reveals meaningful organization linking lexical usage to neuroanatomical substrates. Summary: Together, these findings establish a proof-of-concept for inverting the traditional lesion–symptom mapping framework and highlight the potential of word-based models for clinical applications in diagnosis and prognosis of post-stroke language disorders.
Topic Areas: Disorders: Acquired, Language Production