Machine Learning Models for Identifying Neural Predictors of TMS Treatment Response in MDD

经颅磁刺激（TMS）是治疗抑郁症的一种有效且易耐受的方法 疾病（MDD）。经颅磁刺激是昂贵和耗时的，因此确定反应标志将减少财政和 心理负担。此外，治疗反应是高度可变的。临床和诊断异质性 抑郁症导致反应的神经标志物的显著变化。关于神经标记物的文献 TMS强度和靶点的可变性使其复杂化，这可能进一步改变反应。电场 模型通过考虑强度和结构来估计目标受到刺激的程度， 每个参与者的信息，但在这个时候，没有研究已经调查了协会 脑电场和治疗反应之间的联系此外，神经生物学相关的 背外侧（dlPFC）TMS治疗反应还没有得到很好的理解。机器学习也许能帮助我们 了解这些复杂的特征及其与治疗反应的关系。因此， 个性化治疗，我将开发一个数据驱动的机器学习模型，使用以下内容：（1）预 反映电路失调的治疗静息状态连接;（2）电场建模，以估计 个体患者大脑上的电场或电压，作为刺激充分性的标志;和（3） 预期的目标网络连接作为目标参与的标记。之前我们已经证实 机器学习预测MDD抗抑郁反应的可行性。我们将优化和扩展一个模型 基于预测dlPFC TMS响应的多伦多大学存档数据开发。我们会证实这一点 外部基于三组数据：我们在匹兹堡大学收集的数据，布朗大学的档案数据， 和假TMS数据。作为一个探索性的目标，我们将确定我们的模型，预测dlPFC TMS 治疗反应能够预测对dmPFC TMS刺激的反应。在我读博士的时候， 生物工程，我开发了基于内核的机器学习模型，以个性化神经网络标记， 抗抑郁反应鉴于抑郁症的临床和神经异质性，我将利用我的机器 通过接受高级优化方法的培训获得学习和神经成像经验， 抑郁症神经生物学，以确定TMS治疗反应的稳定，可重复的神经预测因子， 临床上可转化的个性化治疗。这将使我能够开发出最佳的治疗模式， 响应和促进我的长期职业目标，以开发个性化的治疗算法，使用大规模的 数据我以前在机器学习、生物工程、神经成像以及 对抑郁症的初步了解使我能够最大限度地发挥培训目标的作用 在这份提案中。 1

Transcranial magnetic stimulation (TMS) is an effective and easy-to-tolerate treatment for major depressive disorder (MDD). TMS is costly and time-intensive so identifying markers of response would reduce financial and psychological burden. Further, treatment response is highly variable. Clinical and diagnostic heterogeneity of depression contributes to significant variability in neural markers of response. The literature on neural markers is complicated by variability in TMS intensity and targets, which may further modify response. Electrical field models estimate the degree to which a target is stimulated by considering both the intensity and structural information of each participant but at this time there are no studies that have investigated the association between brain electrical fields and treatment response. Moreover, the neurobiological correlates of dorsolateral (dlPFC) TMS treatment response are not well understood. Machine learning may be able to help us understand these complex set of features and their association to treatment response. Thus to appropriately personalize treatments, I will develop a data-driven machine learning model that uses the following: (1) pre- treatment resting state connectivity that reflects circuit dysregulation; (2) electrical field modeling to estimate the electrical field or voltage on individual patient’s brain, as a marker of sufficiency of stimulation; and (3) expected target network connectivity as a marker of target engagement. We have previously demonstrated feasibility of machine learning to predict antidepressant response in MDD. We will optimize and expand a model developed on archival University of Toronto data that predicted dlPFC TMS response. We will validate this externally on three sets of data: data we collect at University of Pittsburgh, archival data from Brown University, and sham TMS data. As an exploratory aim, we will identify whether our model that predicts dlPFC TMS treatment response is capable of predicting response to dmPFC TMS stimulation. During my PhD in Bioengineering, I developed kernel-based machine learning models to personalize neural networks markers of antidepressant response. Given the clinical and neural heterogeneity of depression, I will leverage my machine learning and neuroimaging experience by receiving training in advanced optimization approaches and depression neurobiology to identify stable, reproducible neural predictors of TMS treatment response to achieve clinically translatable personalized treatments. This will allow me to develop optimized models of treatment response and facilitate my long-term career goal to develop personalized treatment algorithms using large-scale data. My previous experiences in machine learning, bioengineering, neuroimaging, as well as the preliminary understanding of depression uniquely position me to maximize the benefits of training aims outlined in this proposal. 1