Boutet Alexandre, Germann Jurgen, Gwun Dave, Loh Aaron, Elias Gavin J B, Neudorfer Clemens, Paff Michelle, Horn Andreas, Kuhn Andrea A, Munhoz Renato P, Kalia Suneil K, Hodaie Mojgan, Kucharczyk Walter, Fasano Alfonso, Lozano Andres M
Joint Department of Medical Imaging, University of Toronto, Toronto, ON, Canada.
University Health Network, Toronto, ON, Canada.
Brain Commun. 2021 Mar 10;3(2):fcab027. doi: 10.1093/braincomms/fcab027. eCollection 2021.
Deep brain stimulation of the subthalamic nucleus has become a standard therapy for Parkinson's disease. Despite extensive experience, however, the precise target of optimal stimulation and the relationship between site of stimulation and alleviation of individual signs remains unclear. We examined whether machine learning could predict the benefits in specific Parkinsonian signs when informed by precise locations of stimulation. We studied 275 Parkinson's disease patients who underwent subthalamic nucleus deep brain stimulation between 2003 and 2018. We selected pre-deep brain stimulation and best available post-deep brain stimulation scores from motor items of the Unified Parkinson's Disease Rating Scale (UPDRS-III) to discern sign-specific changes attributable to deep brain stimulation. Volumes of tissue activated were computed and weighted by (i) tremor, (ii) rigidity, (iii) bradykinesia and (iv) axial signs changes. Then, sign-specific sites of optimal ('hot spots') and suboptimal efficacy ('cold spots') were defined. These areas were subsequently validated using machine learning prediction of sign-specific outcomes with in-sample and out-of-sample data ( = 51 subthalamic nucleus deep brain stimulation patients from another institution). Tremor and rigidity hot spots were largely located outside and dorsolateral to the subthalamic nucleus whereas hot spots for bradykinesia and axial signs had larger overlap with the subthalamic nucleus. Using volume of tissue activated overlap with sign-specific hot and cold spots, support vector machine classified patients into quartiles of efficacy with ≥92% accuracy. The accuracy remained high (68-98%) when only considering volume of tissue activated overlap with hot spots but was markedly lower (41-72%) when only using cold spots. The model also performed poorly (44-48%) when using only stimulation voltage, irrespective of stimulation location. Out-of-sample validation accuracy was ≥96% when using volume of tissue activated overlap with the sign-specific hot and cold spots. In two independent datasets, distinct brain areas could predict sign-specific clinical changes in Parkinson's disease patients with subthalamic nucleus deep brain stimulation. With future prospective validation, these findings could individualize stimulation delivery to optimize quality of life improvement.
丘脑底核的深部脑刺激已成为帕金森病的标准治疗方法。然而，尽管有丰富的经验，但最佳刺激的精确靶点以及刺激部位与个体症状缓解之间的关系仍不清楚。我们研究了机器学习能否在精确的刺激位置信息指导下，预测帕金森病特定症状的改善情况。我们研究了2003年至2018年间接受丘脑底核深部脑刺激的275例帕金森病患者。我们从统一帕金森病评定量表（UPDRS-III）的运动项目中选取深部脑刺激前和最佳的深部脑刺激后评分，以辨别深部脑刺激引起的特定症状变化。计算组织激活体积，并根据（i）震颤、（ii）僵直、（iii）运动迟缓以及（iv）轴性症状变化进行加权。然后，定义特定症状的最佳疗效（“热点”）和次优疗效（“冷点”）部位。随后，使用针对特定症状结果的机器学习预测，对来自样本内和样本外数据（=来自另一家机构的51例丘脑底核深部脑刺激患者）的这些区域进行验证。震颤和僵直热点大多位于丘脑底核外侧和背外侧，而运动迟缓和轴性症状的热点与丘脑底核有更大的重叠。利用组织激活体积与特定症状热点和冷点的重叠情况，支持向量机将患者分为疗效四分位数，准确率≥92%。仅考虑组织激活体积与热点的重叠时，准确率仍然很高（68-98%），但仅使用冷点时，准确率明显较低（41-72%）。仅使用刺激电压时，无论刺激位置如何，模型表现也很差（44-48%）。使用组织激活体积与特定症状热点和冷点的重叠情况时，样本外验证准确率≥96%。在两个独立数据集中，不同的脑区可以预测接受丘脑底核深部脑刺激的帕金森病患者的特定症状临床变化。通过未来的前瞻性验证，这些发现可能使刺激传递个性化，以优化生活质量改善。
