Demidenko Michael I, Mumford Jeanette A, Poldrack Russell A
Department of Psychology, Stanford University, Stanford, CA, United States.
Imaging Neurosci (Camb). 2024 Sep 10;2. doi: 10.1162/imag_a_00262. eCollection 2024.
Empirical studies reporting low test-retest reliability of individualblood oxygen-level dependent (BOLD) signal estimates in functional magneticresonance imaging (fMRI) data have resurrected interest among cognitiveneuroscientists in methods that may improve reliability in fMRI. Over the lastdecade, several individual studies have reported that modeling decisions, suchas smoothing, motion correction, and contrast selection, may improve estimatesof test-retest reliability of BOLD signal estimates. However, it remainsan empirical question whether certain analytic decisionsimprove individual- and group-levelreliability estimates in an fMRI task across multiple large, independentsamples. This study used three independent samples (s: 60, 81,119) that collected the same task (Monetary Incentive Delay task) across tworuns and two sessions to evaluate the effects of analytic decisions on theindividual (intraclass correlation coefficient [ICC(3,1)]) and group(Jaccard/Spearman) reliability estimates of BOLD activityof task fMRI data. The analytic decisions in this study vary across fourcategories: smoothing kernel (five options), motion correction (four options),task parameterizing (three options), and task contrasts (four options), totaling240 different pipeline permutations. Across all 240 pipelines, the median ICCestimates are consistently low, with a maximum median ICC estimate of .43- .55 across the 3 samples. The analytic decisions with the greatestimpact on the median ICC and group similarity estimates are thecontrast, Cue Model parameterization, and a largersmoothing kernel. Using anin a contrastcondition meaningfully increased group similarity and ICC estimates as comparedwith using thecue. This effect was largest for the CueModel parameterization; however, improvements in reliability came at the cost ofinterpretability. This study illustrates that estimates of reliability in theMID task are consistently low and variable at small samples, and a highertest-retest reliability may not always improve interpretability of theestimated BOLD signal.
实证研究表明，功能磁共振成像（fMRI）数据中个体血氧水平依赖（BOLD）信号估计的重测信度较低，这重新唤起了认知神经科学家对可能提高fMRI信度的方法的兴趣。在过去十年中，几项个体研究报告称，诸如平滑处理、运动校正和对比度选择等建模决策，可能会提高BOLD信号估计的重测信度。然而，在多个大型独立样本的fMRI任务中，某些分析决策是否能提高个体和组水平的信度估计，这仍然是一个实证问题。本研究使用了三个独立样本（样本量分别为60、81、119），在两次扫描和两个时间段内收集相同的任务（金钱激励延迟任务），以评估分析决策对任务fMRI数据中BOLD活动的个体（组内相关系数[ICC(3,1)]）和组（杰卡德/斯皮尔曼）信度估计的影响。本研究中的分析决策分为四类：平滑核（五种选项）、运动校正（四种选项）、任务参数化（三种选项）和任务对比度（四种选项），总共240种不同的流程排列。在所有240种流程中，ICC估计值的中位数一直很低，在这三个样本中，ICC估计值的最大中位数为0.43 - 0.55。对中位数ICC和组相似性估计影响最大的分析决策是对比度、线索模型参数化和更大的平滑核。与使用线索相比，在对比度条件下使用有意义地提高了组相似性和ICC估计值。这种效应在线索模型参数化中最为明显；然而，信度的提高是以可解释性为代价的。本研究表明，在小样本情况下，金钱激励延迟任务中的信度估计一直很低且变化不定，较高的重测信度并不总是能提高估计的BOLD信号的可解释性。
