Sperber Christoph, Gallucci Laura, Kasties Vanessa, Arnold Marcel, Umarova Roza M
Department of Neurology, Inselspital, University Hospital Bern, University of Bern, Bern 3010, Switzerland.
Child Development Center, University Children's Hospital Zurich, University of Zürich, Zürich 8008, Switzerland.
Brain Commun. 2025 Jul 28;7(4):fcaf268. doi: 10.1093/braincomms/fcaf268. eCollection 2025.
The prediction of stroke outcome from imaging markers could be used to guide individualized therapeutic approaches. We aimed to find the best imaging marker to predict the global functional impact of a stroke lesion among low- to high-level connectomic measures-indirect estimations of structural connectivity, graph representations, or brain modes-as well as spatial lesion features. This observational study retrospectively analysed clinical routine data from patients with acute first-ever ischaemic stroke. We traced lesions in diffusion-weighted MRI and computed 21 topographic or connectomic measures, including (i) tract-wise, voxel-wise and interregional white matter disconnection that were indirectly estimated by reference to healthy connectome data; (ii) interregional network structure by graph measures; and (iii) brain modes, which represent elementary interactions between grey matter regions. We used all features to predict stroke severity [National Institutes of Health Stroke Scale (NIHSS) 24 h] or classify poor functional outcome (mRS 3 months ≥ 2) in a nested cross-validation with high-dimensional machine-learning models. For comparison to specific, granular post-stroke cognitive deficits, we replicated the modelling procedures in another sample for selective attention and phonemic word fluency. The study included 755 patients [mean age = 66.9 ± 15.3 years; NIHSS 24 h median (IQR) = 2 (1; 5); mRS 3 months = 1 (0; 2)]. For both measures, simple spatial lesion features (NIHSS 24 h: ² = 0.395 ± 0.059; mRS: accuracy = 65.62% ± 3.45, positive predictive value = 0.72 ± 0.13; negative predictive value = 0.64 ± 0.04) outperformed connectomic measures (all < 0.0007), even though the predictions of the best measures in each category were numerically close. Control analyses on specific cognitive deficits in a sample of 182 patients found connectomic measures to be equal or even superior to spatial lesion features. Connectomic stroke imaging markers provide no benefit in the prediction of acute stroke severity and functional outcome at 3 months. Spatial lesion imaging features seem to effectively capture the global neurological perturbation caused by a stroke lesion and could provide a basis for personalized prediction algorithms. On the other hand, connectomic stroke imaging markers may be warranted when modelling specific post-stroke cognitive deficits.
利用影像学标志物预测卒中预后可用于指导个体化治疗方案。我们旨在从低水平到高水平的连接组学测量（结构连接性的间接估计、图形表示或脑模式）以及空间病变特征中找到最佳的影像学标志物，以预测卒中病变对整体功能的影响。这项观察性研究回顾性分析了首次发生急性缺血性卒中患者的临床常规数据。我们在扩散加权磁共振成像中追踪病变，并计算了21种地形学或连接组学测量指标，包括：（i）通过参考健康连接组数据间接估计的逐束、逐体素和区域间白质断开；（ii）通过图形测量得到的区域间网络结构；以及（iii）代表灰质区域间基本相互作用的脑模式。我们使用所有特征，通过高维机器学习模型进行嵌套交叉验证，以预测卒中严重程度[美国国立卫生研究院卒中量表（NIHSS）24小时]或对不良功能预后（3个月改良Rankin量表≥2）进行分类。为了与特定的、详细的卒中后认知缺陷进行比较，我们在另一个样本中重复了建模过程，以评估选择性注意力和音素词流畅性。该研究纳入了755例患者[平均年龄 = 66.9 ± 15.3岁；NIHSS 24小时中位数（四分位间距）= 2（1；5）；3个月改良Rankin量表 = 1（0；2）]。对于这两项测量指标，简单的空间病变特征（NIHSS 24小时： = 0.395 ± 0.059；改良Rankin量表：准确率 = 65.62% ± 3.45，阳性预测值 = 0.72 ± 0.13；阴性预测值 = 0.64 ± 0.04）优于连接组学测量指标（所有 < 0.0007），尽管每类中最佳测量指标的预测值在数值上很接近。对182例患者样本中的特定认知缺陷进行的对照分析发现，连接组学测量指标与空间病变特征相当甚至更优。连接组学卒中影像学标志物在预测急性卒中严重程度和3个月时的功能预后方面并无益处。空间病变影像学特征似乎能有效捕捉卒中病变引起的整体神经功能紊乱，并可为个性化预测算法提供依据。另一方面，在对特定的卒中后认知缺陷进行建模时，连接组学卒中影像学标志物可能是必要的。
