Laureate Institute for Brain Research (LIBR), Tulsa, OK, USA.
Sci Rep. 2019 Jan 23;9(1):419. doi: 10.1038/s41598-018-36699-0.
Head motion (HM) during fMRI acquisition can significantly affect measures of brain activity or connectivity even after correction with preprocessing methods. Moreover, any systematic relationship between HM and variables of interest can introduce systematic bias. There is a large and growing interest in identifying neural biomarkers for psychiatric disorders using resting state fMRI (rsfMRI). However, the relationship between HM and different psychiatric symptoms domains is not well understood. The aim of this investigation was to determine whether psychiatric symptoms and other characteristics of the individual predict HM during rsfMRI. A sample of n = 464 participants (174 male) from the Tulsa1000, a naturalistic longitudinal study recruiting subjects with different levels of severity in mood/anxiety/substance use disorders based on the dimensional NIMH Research Domain Criteria framework was used for this study. Based on a machine learning (ML) pipeline with nested cross-validation to avoid overfitting, the stacked model with 15 anthropometric (like body mass index, BMI) and demographic (age and sex) variables identifies BMI and weight as the most important variables and explained 10.9 percent of the HM variance (95% CI: 9.9-11.8). In comparison ML models with 105 self-report measures for state and trait psychological characteristics identified nicotine and alcohol use variables as well as impulsivity inhibitory control variables but explain only 5 percent of HM variance (95% CI: 3.5-6.4). A combined ML model using all 120 variables did not perform significantly better than the model using only 15 physical variables (combined model 95% confidence interval: 10.2-12.4). Taken together, after considering physical variables, state or trait psychological characteristics do not provide additional power to predict motion during rsfMRI.
头部运动（HM）在 fMRI 采集过程中即使经过预处理方法校正，也会显著影响脑活动或连通性的测量值。此外，HM 与感兴趣变量之间的任何系统关系都会引入系统偏差。使用静息状态 fMRI（rsfMRI）识别精神疾病的神经生物标志物引起了广泛而日益增长的兴趣。然而，HM 与不同精神症状领域之间的关系尚不清楚。本研究旨在确定精神症状和个体的其他特征是否可以预测 rsfMRI 期间的 HM。本研究使用了来自塔尔萨 1000 号（Tulsa1000）的样本，这是一项自然主义纵向研究，根据 NIMH 研究领域标准框架，基于维度招募了不同严重程度的情绪/焦虑/物质使用障碍的受试者。该研究使用具有嵌套交叉验证的机器学习（ML）管道来避免过度拟合，堆叠模型包含 15 个人体测量学（如体重指数，BMI）和人口统计学（年龄和性别）变量，确定 BMI 和体重是最重要的变量，并解释了 10.9%的 HM 方差（95%CI：9.9-11.8）。相比之下，使用 105 种用于状态和特质心理特征的自我报告措施的 ML 模型确定了尼古丁和酒精使用变量以及冲动性抑制控制变量，但仅解释了 5%的 HM 方差（95%CI：3.5-6.4）。使用所有 120 个变量的组合 ML 模型的性能并不明显优于仅使用 15 个物理变量的模型（组合模型 95%置信区间：10.2-12.4）。综合考虑，在考虑了物理变量后，状态或特质心理特征并不能为预测 rsfMRI 期间的运动提供额外的动力。
