Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Severity scores blur, error types reveal: behavioral choice shapes lesion–language mapping
Poster Session B, Wednesday, September 30, 4:30 - 6:30 pm, Wangari Maathai
Roger Newman-Norlund1, Kalil Warren2, Saeed Ahmadi1, Nadra Salman2, Yong Yang3, Regan Willis3, Xiang Guan3, Leonardo Bonilha4, Julius Fridriksson1; 1University of South Carolina, Department of Communication Sciences and Disorders, 2University of South Carolina, Department of Linguistics, 3University of South Carolina, Department of Computer Science and Engineering, 4USC School of Medicine, Department of Neurology
Introduction. When researchers relate language deficits to brain damage in stroke patients, they must choose both a behavioral test and a way of summarizing it. The standard choice is the Western Aphasia Battery (WAB) reported as severity subscores reflecting fluency, comprehension, and naming. These scores collapse heterogeneous behavior into a small number of severity ratings. An alternative is to record the types of errors a patient makes using a defined scoring scheme. Both the WAB sentence completion and responsive speech subtests (10 items combined) and the Philadelphia Naming Test (PNT, 175 items) can be scored this way. We asked whether the choice between severity aggregated and error typed behavior changes the brain anatomy that can be recovered, and whether the number of test items matters. Participants. 296 chronic post stroke participants from the Center for the Study of Aphasia Recovery (CSTAR), all with unilateral left hemisphere stroke and lesion segmentations on the JHU atlas (189 regions). WAB subscores were available for N=288, WAB error proportions for N=259, and PNT error proportions for N=214. All three lenses were present in N=183 patients used for paired contrasts. Methods. Using non negative matrix factorization, we decomposed each patient’s lesion pattern into five spatially coherent atoms: anterior temporal or temporal pole, posterior occipitotemporal, dorsal stream or SLF or inferior parietal, striatal or internal capsule, and frontal or IFG. The 5 atom solution was reproducible across split halves (mean atom correlation 0.81 ± 0.08). For each behavioral score we computed a five number anatomical fingerprint, one partial Spearman correlation per atom, controlling for the other four atoms and total lesion volume. The main statistic was the mean off diagonal correlation between behavioral fingerprints within a lens. Results. WAB subscore fingerprints were highly similar (mean off diagonal r = +0.45, bootstrap 95% CI [+0.08, +0.73]), indicating that the five severity subscores point to largely the same anatomy. WAB error fingerprints were not similar (r = −0.03 [−0.14, +0.31]). PNT error fingerprints were also not similar (r = −0.05 [−0.16, +0.43]). In paired contrasts (N=183), WAB subscores minus PNT errors = +0.48 [+0.05, +0.81], and WAB subscores minus WAB errors = +0.49 [+0.07, +0.81], both P(>0) = 0.98. WAB errors and PNT errors did not differ (Δ = −0.01 [−0.37, +0.32], P(>0) = 0.49) despite a 17 fold difference in item count. Individual atom by error correlations were modest (|ρ| ≤ 0.27) and did not survive correction. Conclusions. What you measure determines what anatomy you can see. The critical dimension is not the test or its length, but whether behavior is summarized as severity or as error type. WAB severity subscores converge on a general impairment signal, whereas both 10 item WAB and 175 item PNT error proportions preserve distinct anatomical signals at comparable resolution. Severity scores remain useful clinically, but when the goal is to relate language deficits to specific brain anatomy, error typed behavior should be the default.
Topic Areas: Computational Approaches, Disorders: Acquired