Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Establishing Empirical Benchmarks for Lesion-Based Prediction of Language Impairment After Stroke
Poster Session A, Wednesday, September 30, 11:00 am - 1:00 pm, Wangari Maathai
Alessandra Algieri1,2, Matthew Lambon Ralph2, Giacomo Handjaras1, Tammar Truzman2, Ajay Halai2; 1IMT School for Advanced Studies Lucca, 2MRC Cognition and Brain Sciences Unit, University of Cambridge
Introduction: Lesion-symptom mapping has long sought to explain behavioural impairment following stroke from patterns of brain damage. Recently, machine-learning approaches to predicting language outcomes from post-stroke lesions have grown rapidly, spanning classical algorithms, ensemble methods, and deep-learning architectures. However, most studies focus on binary severity classification, single behavioural outcomes, limited model comparison, or lack external validation, making it difficult to establish realistic benchmarks for lesion-based prediction. It remains unclear how much behavioural variance can be explained from lesion anatomy alone, and whether increasingly sophisticated models meaningfully outperform simpler approaches. Here, we aim to establish empirical benchmarks for lesion-based prediction across multiple language tasks in chronic left-hemisphere stroke. We leverage two deeply phenotyped, independent stroke datasets from Manchester (N=97) and Cambridge (N=65), both assessed with identical, extensive language and cognitive batteries. This provides a rare testbed for evaluating within-task, cross-task, and cross-site generalisation, and for estimating the predictive ceiling achievable from lesion information. Methods: Lesion masks were generated on T1-w scans using our pre-trained UNet models, and normalised to MNI space. Lesion-load values were extracted for an atlas informed by vascular territories (N=28). These vectors constitute the input for all predictive models. Here, we outline the methods for two naming tasks—Boston Naming Test (BNT) and Cambridge Naming Test (CNT)—although the framework was applied across all tasks. We developed a lesion-based prediction framework to benchmark multiple machine-learning algorithms packaged in scikit-learn, including Support Vector Regression, Random Forests, Decision Trees, K-Nearest Neighbours, Gradient and Extreme Gradient Boosting, Multilayer Perceptron, and linear regularised regression models (Lasso, ElasticNet, Ridge). Models were evaluated using nested 5-fold stratified cross-validation. The inner loop performed hyperparameter optimisation within model-specific grids, while the outer loop provided unbiased estimates of internal performance. For each outer test fold, we computed MSE, MAE, R², and Pearson’s r, and averaged performance across folds to ensure model comparability. Following internal validation, models were re-trained and optimised on the full dataset, and then used to predict a similar naming task within the Manchester cohort (cross-task validity: ManBNT → ManCNT) as well as to perform external validation on the Cambridge cohort (ManBNT → CamBNT; ManBNT → CamCNT). Results: Internal validation revealed consistent lesion–behaviour mappings, yielding mean correlations of r≈0.4–0.5 (SD≈0.2–0.3). Cross-task prediction showed robust within-cohort generalisation from BNT to CNT: nonlinear models outperformed linear ones (r=0.93 and r=0.59, respectively). Cross-site external validation demonstrated weaker transfer to the Cambridge cohort (r≈0.11–0.35), possibly reflecting its narrower range and milder deficits. Conclusion; These findings provide a proof-of-principle upper bound for lesion-based naming prediction and suggest that higher-capacity models may better capture the predictive structure under well-matched conditions. However, the weaker cross-site transfer indicates that differences in sample characteristics can limit external generalisation, underscoring the need to determine whether these model advantages hold consistently across other tasks and independent datasets. Furthermore, we are refining our models to integrate non-lesion information and applying explainable AI methods (e.g., SHAP) to identify the features driving predictions. All code will be openly accessible.
Topic Areas: Disorders: Acquired, Computational Approaches