Search Abstracts | Symposia | Slide Sessions | Poster Sessions
A universal bidirectional model of aphasic behavior and left-hemisphere lesion in 296 patients
Poster Session A, Wednesday, September 30, 11:00 am - 1:00 pm, Wangari Maathai
Roger Newman-Norlund1, Kalil Warren2, Saeed Ahmadi1, Nadra Salman2, Xiang Guan3, Yong Yang3, Regan Willis3, Leondardo Bonilha4, Julius Fridriksson1; 1University of South Carolina, Department of Communication Sciences and Disorders, 2University of South Carolina, Linguistics Department, 3University of South Carolina, Department of Computer Science and Engineering, 4USC School of Medicine, Department of Neurology
Background. In 1861, Paul Broca inferred from a behavioral observation a lesion he would not see until autopsy. The inference ran in two directions at once: the deficit implicated a location, and the location, once confirmed, accounted for the deficit. This two-way reasoning is now distributed across the translational team: the speech-language pathologist characterizes behavior, the neurologist and neuroradiologist characterize the lesion, and the cognitive neuroscientist moves between the two. Yet statistical aphasiology has not matched this fluidity. The forward map (lesion-to-behavior, or lesion–symptom mapping, LSM) has a 60-year history; the inverse (behavior-to-lesion, or symptom–lesion mapping, SLM) was, until recently, only a theoretical goal. Our companion paper (Newman-Norlund et al., 2026) introduced human SLM and showed it is tractable in the CSTAR cohort analyzed here, reaching cross-validated R² up to 0.46 in classical perisylvian regions. That work was not, however, designed to put the two directions in one model. We argue they are two readings of a single object, and show that one model recovers both. Participants and measures. We analyzed the same 296 chronic post-stroke participants from the Center for the Study of Aphasia Recovery (CSTAR) cohort used in the companion SLM paper. Behavior was summarized as ten routine clinical measures spanning syndrome composites, overall severity, naming, fluency, comprehension, and repetition. Lesion was summarized as proportional damage to the fifteen left-hemisphere regions the companion paper identified as most predictable, spanning classical perisylvian cortex and the arcuate fasciculus. Modeling. A linear latent-variable model (partial least squares regression, five components, no neural network) was trained on 266 patients and tested on 30, stratified by WAB aphasia type. The entire procedure was repeated across 20 random train/test splits, and the same fitted parameters were evaluated in both directions. Results. The five components the model identifies, without being told to, correspond to overall severity plus the four classical aphasic syndromes. Component 1 indexes overall severity. Component 2 is the Broca–Wernicke axis, pairing frontal damage with reduced fluency and posterior-temporal damage with impaired comprehension. Component 3 captures the conduction profile. Component 4 captures temporal-lobe anomia. Component 5 captures anterior motor-speech involvement. Across 20 held-out splits, the behavior-to-lesion direction correctly localized damage (mean 11.8 of 15 regions predicted above cohort baseline; R² = 0.21 ± 0.12), and the lesion-to-behavior direction predicted clinical scores (mean 8.3 of 10; R² = 0.26 ± 0.12). A round-trip check, predict lesion from behavior, then predict behavior back from that predicted lesion, reached high fidelity in 41 ± 7% of held-out patients per split, and moderate fidelity in more than 60% of patients in every split. Conclusions. The forward and inverse maps of aphasia are the same map. Overall severity and the four classical aphasic syndromes, rather than being imposed on the data, emerge as its latent structure. For the translational team already distributing this two-way reasoning across its members, a single set of parameters now quantifies both directions on the same scale. Here we show that forward (LSM) and inverse (SLM) maps can be one.
Topic Areas: Disorders: Acquired, Computational Approaches