Julius Fridriksson

Researchers in artificial intelligence increasingly intervene on language models to study how their functions are organized, silencing attention components in ways reminiscent of the brain lesions long used to map language in post-stroke aphasia. A common move is to ablate an attention head and read the resulting drop in a behavior as evidence that the head implements it. How much that establishes remains unclear, because a head whose removal disrupts a behavior is necessary for it but need not be the site where it is computed. We examined this question for picture naming, taking the loss of naming (anomia) that defines aphasia as the behavior of interest. Across six vision-language models spanning three language backbones, we ablated each attention head in turn during single-word picture naming and measured accuracy before and after. Localization varied widely. In LLaVA-1.6-Vicuna-13B, a single early head was necessary: removing it alone reduced naming accuracy from 0.99 to 0.006. It was not sufficient, because retaining it while ablating the other 1,599 heads also produced 0% accuracy. Two Mistral-backbone models had no critical head. Early-layer dependence appeared in every model but varied in strength, whereas dependence on any single head ranged from dominant to absent. Within Qwen2.5-VL, a dominant head appeared in the 7B but not the 3B model, indicating that this concentration emerged with scale. Ablating the head whose removal disrupts naming does not establish that the head computes the behavior, or that naming localizes to a head at all.

Complex datasets used for lesion‐symptom mapping (LSM) present serious challenges for interpretable statistical analysis. We introduce Critical Network Lesion‐Symptom Mapping (CN‐LSM), a neuroanatomical atlas‐based method that uses low‐dimensional regression models and out‐of‐sample prediction to identify brain regions that contribute significantly to the prediction of clinical behavioral scores. Rather than focusing on a single most‐predictive model, CN‐LSM identifies a set of plausible models analyzed as an ensemble. CN‐LSM offers three primary advantages over existing approaches: (1) greater robustness to analytical decisions that should be (but often are not) inconsequential to results; (2) holistic quantification of uncertainty across the entire statistical map rather than at individual voxels or regions; and (3) explicit accommodation of multiple plausible explanations, reducing false negatives while preserving stringent control over false positive rates.

Purpose: While aphasia treatment studies commonly use picture-naming performance as an outcome measure, narrative discourse better reflects functional language use. Discourse variables may also hold prognostic value for naming treatment response, but their predictive role remains underexplored. Method: We analyzed baseline and posttreatment narrative discourse samples from 95 chronic stroke survivors with aphasia enrolled in a lexical retrieval intervention study. Participants received 3 weeks each of phonological and semantic naming therapy in a crossover design. The Cinderella story retells were analyzed for a range of discourse features: mean length of utterance, words per minute, verbs per utterance, propositional density, type–token ratio, core lexicon, main concepts analysis, and error ratios. We used univariate (generalized) binomial and linear mixed-effects modeling with multiple predictors to assess whether baseline discourse variables predicted naming gains on the Philadelphia Naming Test and whether discourse variables themselves changed following treatment. Results: Higher aphasia severity and lower baseline propositional density predicted greater naming gains across time points and treatment types. Gains after phonological therapy were also predicted by higher baseline core lexicon production. Posttreatment discourse showed gains in mean length of utterance, words per minute, and core lexicon, which were maintained at 6 months, while phonological errors declined only at follow-up. Phonological treatment led to increases in words per minute. Conclusions: Discourse variables reflecting propositional efficiency and lexical appropriateness uniquely predict treatment response beyond general aphasia severity. Lexical retrieval therapy generalizes to improvements in narrative discourse, underscoring the clinical value of incorporating discourse-level measures in aphasia assessment and treatment outcome evaluation. Supplemental Material: https://doi.org/10.23641/asha.32568822

Large language models (LLMs) have achieved remarkable capabilities, yet methods to verify which model components are truly necessary for language function remain limited. Current interpretability approaches rely on internal metrics and lack external validation. Here we present the Brain-LLM Unified Model (BLUM), a framework that leverages lesion-symptom mapping—the gold standard for establishing causal brain-behavior relationships for over a century—as an external reference structure for evaluating LLM perturbation effects. Using data from individuals with chronic post-stroke aphasia (N = 410), we trained symptom-to-lesion models that predict brain damage location from behavioral error profiles, applied systematic perturbations to transformer layers, administered identical clinical assessments to perturbed LLMs and human patients, and projected LLM error profiles into human lesion space. LLM error profiles were sufficiently similar to human error profiles that predicted lesions corresponded to actual lesions in error-matched humans above chance in 67% of picture naming conditions and 68.3% of sentence completion conditions, with semantic-dominant errors mapping onto ventral-stream lesion patterns and phonemic-dominant errors onto dorsal-stream patterns. These findings open a new methodological avenue for LLM interpretability in which clinical neuroscience provides external validation, establishing human lesion-symptom mapping as a reference framework for evaluating artificial language systems and motivating direct investigation of whether behavioral alignment reflects shared computational principles.

Background and Objectives: Advanced brain age relative to chronological age may predict cognitive decline in healthy aging, but prospective evidence remains limited. We tested whether the brain age gap (BAG) predicts cognition and 5-year cognitive decline in the AGES-Reykjavik Study. Methods: Participants with structural MRI and neuropsychological data were included cross-sectionally (n=4,606); mean age 76.4y; 58% female) and longitudinally (n=2,561; mean follow-up 5.2y). BAG was estimated from T1-weighted MRI. Mixed-effects models examined BAG effects on Memory (M), processing speed (PS), and executive function (EF), adjusting for age, sex, and education. Results: Higher BAG was associated with worse cognition across all domains (b=-0.037 to -0.047, all p<0.001). BAG significantly predicted greater decline in M (b=-0.007, p=0.046); subgroup analyses revealed steeper decline in M and EF with advanced brain age. Discussion: BAG predicts longitudinal cognitive decline beyond chronological age, supporting its utility as a biomarker in preventative care of healthy aging.

Purpose: Verb argument structure (VAS) is often impaired in poststroke aphasia. Previous studies investigated VAS impairments using constrained experiments or less often narrative speech, which relies on time-consuming manual annotations of VAS in speech transcriptions. Here, we aimed to develop and validate an automated approach to quantify VAS use in narrative speech using natural language processing (a dependency parser) and to apply it to a new data set. Method: First, we validated this approach by comparing manually annotated measures of VAS from a previous study, with automatically annotated VAS measures developed here. We then applied this approach to a new data set of participants with aphasia ( n = 106), whose narrative discourse samples were compared to a control group from AphasiaBank to (a) replicate previous findings and (b) assess VAS impairments in agrammatism specifically. Results: The analysis of the new data set using automatically annotated VAS revealed that participants with Broca's aphasia and, according to grammatical distinctions, agrammatic participants were more likely to select verbs used with fewer arguments, on average, and to produce fewer arguments than controls. Conclusions: This study reveals that dependency parsers are suitable to characterize VAS use in spontaneous speech with virtually absent manual labor, confirming the feasibility of developing clinically applicable tools for complex language analysis. The results of the study additionally show that participants with aphasia differ from controls in their VAS sensitivity and abilities.

Effective rehabilitation tools are essential for improving language outcomes in chronic aphasia. Speech entrainment is a behavioral treatment that has shown promise in enhancing speech output in nonfluent aphasia, potentially by acting as an external mechanism to synchronize anterior and posterior language regions in the left hemisphere. Transcranial alternating current stimulation has been hypothesized to enhance functional connectivity between brain regions by amplifying endogenous oscillations. This proof-of-concept study explored whether high-definition tACS (HD-tACS) could improve speech fluency in nonfluent aphasia when paired with speech entrainment. In a double-blind, pseudorandomized study, 1 mA of HD-tACS at 7 Hz was applied to anterior and posterior left-hemisphere regions of individuals with nonfluent aphasia (N = 13). Stimulation was applied under three conditions: in-phase, anti-phase, and sham, and paired speech entrainment. Three outcome measures were examined: (1) number of words produced; (2) number of errors, and (3) ‘entrainment’ to the speech entrainment model. Group-level analyses for two of the three outcome measures reveal statistically significant differences between the experimental conditions. In-phase alternating current stimulation yielded more words and better entrainment to the audiovisual model than the sham condition. This study provides promising evidence that HD-tACS could improve speech production in individuals with nonfluent aphasia. These results contribute to growing evidence supporting the therapeutic potential of non-invasive brain stimulation approaches as an adjuvant to traditional behavioral speech-language therapy in stroke survivors.

Background: Recovery from chronic post-stroke aphasia is highly heterogeneous and shaped by lesion characteristics, brain integrity, and systemic health. Traditional group-level models struggle to capture this multidimensional, dynamic variability. Digital twin approaches - patient-specific, continually updating models - may enable individualized prediction and counterfactual evaluation of modifiable risk factors. Therefore, the aim was to develop and validate a proof-of-concept digital twin that predicts individual naming outcomes during language treatment and quantifies the estimated impact of modifiable health factors on naming. This study represents the first application of digital twin modeling to aphasia recovery, and we hypothesize that this could constitute a critical first step toward dynamically adaptive, personalized models for aphasia rehabilitation. Methods: We analyzed longitudinal data from 106 chronic stroke survivors with aphasia enrolled in the POLAR randomized clinical trial. For each participant we combined baseline demographic/health variables (age, sex, education, days post-stroke, hypertension, diabetes, BMI), lesion load in left-hemisphere language ROIs (JHU atlas), ROI-level white-matter microstructure (FA), and resting-state functional connectivity restricted to language regions. A continual-learning linear model (River framework; Adam optimizer) was pretrained on baseline data and updated across timepoints. Model performance was assessed by R² at the final timepoint. Counterfactual simulations systematically altered hypertension, diabetes, and BMI to estimate isolated and combined effects on predicted Philadelphia Naming Test (PNT) scores. Results: The digital twin predicted final PNT scores with R² = 0.5848 (explaining approximately 58% of variance). The largest contributors were prior naming performance, age, lesion load in language regions, and white-matter integrity in temporal regions (notably right MTG and STG pole). Counterfactual results estimated modest but consistent effects of health factors, with them collectively accounting for approximately 25% of the variance in treatment gains. The average change in PNT score with counterfactual changes was 7.92 (SD = 16.11). Therefore, diabetic status explained 2% of the variance in treatment gains, hypertensive status explained 4.75%, and increasing BMI explained 18.5%. Conclusions: This study demonstrate the feasibility and clinical potential of applying a digital twin framework to chronic post-stroke aphasia, with the model successfully predicting more than half the variance in naming performance during language treatment. Through counterfactual simulation, we demonstrated that modifiable health factors exert measurable, bidirectional influences on predicted treatment outcomes, underscoring the role of systemic health in shaping language recovery. Although the individual effects of these factors were modest in magnitude, their cumulative influence on treatment gains illustrates how multiple small biological contributors can add up to shape meaningful differences in language outcomes. More broadly, these findings illustrate the potential value of digital twin models for aphasia treatment, particularly as a tool to integrate diverse biological factors and generate individualized, dynamically updated predictions. Graphical