Department of Psychological and Brain Sciences, Washington University in St. Louis.
Department of Psychology and Neuroscience, Duke University.
J Psychopathol Clin Sci. 2024 Nov;133(8):656-666. doi: 10.1037/abn0000919.
Brain structure correlates of obsessive-compulsive personality disorder (OCPD) remain poorly understood as limited OCPD assessment has precluded well-powered studies. Here, we tested whether machine learning (ML; elastic net regression, gradient boosting machines, support vector regression with linear and radial kernels) could estimate OCPD scores from personality data and whether ML-predicted scores are associated with indices of brain structure (cortical thickness and surface area and subcortical volumes). Among older adults (ns = 898-1,606) who completed multiple OCPD assessments, ML elastic net regression with Revised NEO Personality Inventory personality items as features best predicted Five-Factor Obsessive-Compulsive Inventory-Short Form (FFOCI-SF) scores, root-mean-squared error (RMSE)/SD = 0.66; performance generalized to a sample of college students (n = 175; RMSE/SD = 0.51). Items from all five-factor model personality traits contributed to predicted FFOCI-SF (p-FFOCI-SF) scores; conscientiousness and openness items were the most influential. In college students (n = 1,253), univariate analyses of cortical thickness, surface area, and subcortical volumes revealed only a positive association between p-FFOCI-SF and right superior frontal gyrus cortical thickness after adjusting for multiple testing (b = 2.21, p = .0014; all other |b|s < 1.04; all other ps > .009). Multivariate ML models of brain features predicting FFOCI, conscientiousness, and neuroticism performed poorly (RMSE/SDs > 1.00). These data reveal that all five-factor model traits contribute to maladaptive OCPD traits and identify greater right superior frontal gyrus cortical thickness as a promising correlate of OCPD for future study. Broadly, this study highlights the utility of ML to estimate unmeasured psychopathology phenotypes in neuroimaging data sets but that our application of ML to neuroimaging may not resolve unreliable associations and small effects characteristic of univariate psychiatric neuroimaging research. (PsycInfo Database Record (c) 2024 APA, all rights reserved).
脑结构与强迫症人格障碍（OCPD）的相关性仍知之甚少，因为有限的 OCPD 评估使得有力的研究无法进行。在这里，我们测试了机器学习（ML；弹性网络回归、梯度提升机、具有线性和径向核的支持向量回归）是否可以从人格数据中估计 OCPD 分数，以及 ML 预测分数是否与大脑结构指标（皮质厚度和表面积以及皮质下体积）相关。在完成多次 OCPD 评估的老年人（n=898-1606）中，使用具有修订的 NEO 人格量表人格项目的 ML 弹性网络回归对五因素强迫症检查表-短表（FFOCI-SF）得分进行了最佳预测，均方根误差（RMSE）/SD=0.66；性能推广到大学生样本（n=175；RMSE/SD=0.51）。来自五因素模型人格特质的所有项目都有助于预测 FFOCI-SF（p-FFOCI-SF）得分；尽责性和开放性项目的影响力最大。在大学生（n=1253）中，在调整了多次检验后，皮质厚度、表面积和皮质下体积的单变量分析仅显示 p-FFOCI-SF 与右侧额上回皮质厚度之间存在正相关（b=2.21，p=0.0014；所有其他|b|s<1.04；所有其他 p>0.009）。预测 FFOCI、尽责性和神经质的大脑特征的多元 ML 模型表现不佳（RMSE/SDs>1.00）。这些数据表明，所有五因素模型特质都有助于适应不良的 OCPD 特质，并确定右侧额上回皮质厚度增加是未来研究中 OCPD 的一个有前途的相关因素。总的来说，这项研究强调了 ML 用于估计神经影像学数据集中未测量的精神病理学表型的效用，但我们对神经影像学的 ML 应用可能无法解决不可靠的关联和单变量精神神经影像学研究的小效应特征。（PsycInfo 数据库记录（c）2024 APA，保留所有权利）。
