Sex-specific infarct volume associations and early prediction of language impairment progression following stroke surgery: a network approach.

Background: Post-stroke language impairment affects nearly one-third of acute stroke patients or 30% of ischemic stroke patients, yet predicting its progression remains challenging due to multifactorial recovery processes. Methods: Predictors were selected using Least Absolute Shrinkage and Selection Operator (LASSO) regression. The interrelationships among these predictors and their mechanistic links to gender and post-stroke language impairment were subsequently explored employing Graphical LASSO (GLASSO) and Bayesian network analysis. Furthermore, Network Outcome Analysis (NOA) was applied to investigate the associations between preoperative predictors and the severity of postoperative language impairment deterioration. Results: LASSO identified eight predictors for inclusion in the model. The ROC analysis demonstrated favorable predictive efficacy (AUC = 0.80 [95% CI: 0.719-0.876], accuracy = 0.73, sensitivity = 0.78, specificity = 0.69, PPV = 0.71, NPV = 0.76). The DCA results (probability range: 0-0.81, 0.91) further indicated good clinical utility of the model. Preoperative GCS emerged as the primary direct predictor. Although male patients exhibited larger infarct volumes (22.72 mL vs. 15.19 mL in females), this difference was not directly associated with poorer language outcomes. Discussion: This multimodal model, enhanced by network analysis, accurately predicts language impairment progression and highlights preoperative consciousness as a key mediator, supporting precision stroke rehabilitation by capturing complex predictor interrelationships.