Poststroke cognitive impairment (PSCI) is highly prevalent in stroke survivors and correlated with unfavorable clinical outcomes. This study aimed to identify the neural substrate of PSCI using atlas-based disconnectome analysis and assess the value of disconnection score, a baseline measure for stroke-induced structural disconnection, in PSCI prediction.A multicenter prospective cohort of 676 first-ever patients with acute ischemic stroke was enrolled from 3 independent hospitals in China. Sociodemographic, clinical, and neuroimaging data were collected at acute stage of stroke. Cognitive assessment was performed at 3 months after stroke. Voxel-wise and tract-wise disconnectome analysis were performed to uncover the strategic structural disconnection pattern for global PSCI. Disconnection score was calculated for each participant in leave-one-dataset-out cross-validation. Multivariable logistic regression was performed for the association between disconnection score and PSCI. Prediction models with and without disconnection score were developed, cross-validated, and compared in terms of discrimination and goodness-of-fit.Compared with lesions of non-PSCI, those of PSCI were more likely to have fiber connections with left prefrontal cortex and left deep structures (thalamus and basal ganglia). Disconnection score could predict the risk and severity of PSCI during cross-validation, and was independently associated with PSCI after controlling for all baseline covariates (odds ratio, 1.38 [95% CI, 1.17-1.64]; P<0.001). Incorporating disconnection score into a reference model with 6 known predictors resulted in significant improvement in both discrimination and goodness-of-fit throughout cross-validation.A strategic structural disconnection pattern centered on left prefrontal cortex, thalamus, and basal ganglia is identified for global PSCI using indirect disconnectome analysis. The baseline disconnection score is independently predictive of PSCI and has significant incremental value to preexisting sociodemographic, clinical, and neuroimaging predictors.URL: http://www.chictr.org.cn/enIndex.aspx; Unique identifier: ChiCTR-ROC-17013993.