基于脑网络组图谱和机器学习的卒中后失语症转归预测模型的构建及康复机制研究

Aphasia is a common advanced cerebral dysfunction after cerebral stroke. Precise prediction of language function is beneficial for relief of the social-psychological burden and decision-making of individualized rehabilitation, which is a difficult and popular issue in neurorehabilitation. Recent research has revealed that indices of multi-modal magnetic resonance imaging have the potential to improve the accuracy of traditional predictive models based on demographic parameters and clinical scores, but the efficiency is still not good enough for clinical application. Additionally, the neural mechanisms for language recovery are also unclarified. Our preliminary findings show that neuroimaging indices including white matter integrity of specific fiber segment are associated with scores of multiple language subtests, which are of potential predictive value. This project is based on the theory of contemporary language pathway and brain network processing model. We plan to perform subtle cortical parcellation and subcortical segmentation using the Human Brainnetome Atlas, and select critical imaging features for the predictive model from multiple levels including white matter pathway, rest-state functional network and complicated structural network. Multiple machine learning algorithms will be chosen to train and test the predictive model, and determine the variables and parameters with the optimal discrimination and accuracy. Finally, we will investigate the dynamic changes of neural pathways and networks by follow-up. The research on neuroimaging mechanisms of language recovery will provide novel insights for individualized rehabilitation strategies in patients with post-stroke aphasia.

失语症是脑卒中常见的脑高级功能障碍。卒中后失语症患者语言功能预后的个体化精准预测可以减轻社会-心理负担并协助康复方案的决策，已经成为神经康复领域的难点和热点。近年研究发现，多模态磁共振影像指标可以提高行为学评分和人口学特征所构建的传统预测模型的准确性，但综合效能未达到理想的临床适用性，且语言康复的神经机制尚未完全明确。本项目前期工作发现局部白质纤维完整性等神经影像学指标与多项语言评分存在关联，具有潜在的预测价值。本项目拟在现代语言通路和脑网络加工模型的基础上，利用人类脑网络组图谱细化分割皮层及皮层下核团亚区结构，从白质纤维通路、静息态功能网络、复杂脑结构网络等多个层面筛选关键的预测变量纳入模型。借助多种机器学习算法在预测模型上反复训练、验证，确定最优综合效能的参数变量组合。最后，通过随访对比神经通路和脑网络的动态变化，为研究卒中后失语症康复的脑影像学机制和个体化治疗策略提供新思路。

失语症是脑卒中常见的脑高级功能障碍。卒中后失语症患者语言功能预后的个体化精准预测可以减轻社会-心理负担并协助康复方案的决策，已经成为神经康复领域的难点和热点。在本项目的资助下，本研究团队组建了多模态失语症队列数据库，利用多模态磁共振技术实现了失语症早期快速筛查。在现代语言通路和脑网络加工模型的理论基础上，项目组利用人类脑网络组图谱细化分割皮层及皮层下核团亚区结构，从白质纤维通路、静息态功能网络、复杂脑结构网络等多个层面筛选关键的预测变量。本研究发现，腹背侧双通路中交叉纤维的轴索内体积下降，特异性纤维群体的的微结构损伤特征是敏感的失语症生物标记物。右侧皮层下镜像脑区的灌注水平和皮层-皮层下功能连接强度的改变可以反映命名能力，而扣带回的神经血管耦联状态与理解分数存在关联。基于以上多模态影像关键特征，本研究建立了结合机器学习的语言转归预测模型，提升了模型的预测准确性。此外，本研究还分析了失语症患者中神经通路和脑网络的动态变化，阐明了非优势侧半球皮质-皮质下环路参与卒中后命名功能重组的潜在神经生理学机制，为今后失语症患者康复策略提供了基础。