Department of Electrical Engineering, Eindhoven University of Technology, P.O. Box 513, 5600MB, Eindhoven, The Netherlands; Department of Behavioral Sciences, Epilepsy Center Kempenhaeghe, P.O. Box 61, 5590 VE, Heeze, The Netherlands.
Department of Electrical Engineering, Eindhoven University of Technology, P.O. Box 513, 5600MB, Eindhoven, The Netherlands; Department of Neurology, Maastricht University Medical Center, P.O. Box 5800, 6202 AZ, Maastricht, The Netherlands; Department of Behavioral Sciences, Epilepsy Center Kempenhaeghe, P.O. Box 61, 5590 VE, Heeze, The Netherlands.
Comput Methods Programs Biomed. 2018 Feb;154:143-151. doi: 10.1016/j.cmpb.2017.11.017. Epub 2017 Nov 16.
The autism spectrum disorder (ASD) diagnosis requires a long and elaborate procedure. Due to the lack of a biomarker, the procedure is subjective and is restricted to evaluating behavior. Several attempts to use functional MRI as an assisting tool (as classifier) have been reported, but they barely reach an accuracy of 80%, and have not usually been replicated or validated with independent datasets. Those attempts have used functional connectivity and structural measurements. There is, nevertheless, evidence that not the topology of networks, but their temporal dynamics is a key feature in ASD. We therefore propose a novel MRI-based ASD biomarker by analyzing temporal brain dynamics in resting-state fMRI.
We investigate resting-state fMRI data from 2 independent datasets of adolescents: our in-house data (12 ADS, 12 controls), and the Leuven dataset (12 ASD, 18 controls, from Leuven university). Using independent component analysis we obtain relevant socio-executive resting-state networks (RSNs) and their associated time series. Upon these time series we extract wavelet coherence maps. Using these maps, we calculate our dynamics metric: time of in-phase coherence. This novel metric is then used to train classifiers for autism diagnosis. Leave-one-out cross validation is applied for performance evaluation. To assess inter-site robustness, we also train our classifiers on the in-house data, and test them on the Leuven dataset.
We distinguished ASD from non-ASD adolescents at 86.7% accuracy (91.7% sensitivity, 83.3% specificity). In the second experiment, using Leuven dataset, we also obtained the classification performance at 86.7% (83.3% sensitivity, and 88.9% specificity). Finally we classified the Leuven dataset, with classifiers trained with our in-house data, resulting in 80% accuracy (100% sensitivity, 66.7% specificity).
This study shows that change in the coherence of temporal neurodynamics is a biomarker of ASD, and wavelet coherence-based classifiers lead to robust and replicable results and could be used as an objective diagnostic tool for ASD.
自闭症谱系障碍（ASD）的诊断需要一个冗长而精心的过程。由于缺乏生物标志物，该过程具有主观性，仅限于评估行为。已经有一些使用功能磁共振成像（fMRI）作为辅助工具（作为分类器）的尝试，但它们的准确性仅勉强达到 80%左右，并且通常没有使用独立数据集进行复制或验证。这些尝试使用了功能连接和结构测量。然而，有证据表明，不是网络的拓扑结构，而是其时间动态是 ASD 的一个关键特征。因此，我们通过分析静息态 fMRI 中的大脑时间动态提出了一种新的基于 MRI 的 ASD 生物标志物。
我们研究了来自两个独立青少年数据集的静息态 fMRI 数据：我们的内部数据集（12 名 ASD，12 名对照）和鲁汶数据集（12 名 ASD，18 名对照，来自鲁汶大学）。使用独立成分分析，我们获得了相关的社会执行静息态网络（RSN）及其相关的时间序列。在这些时间序列上，我们提取了小波相干图。使用这些图谱，我们计算了我们的动力学指标：同相相干的时间。然后，使用这个新指标来训练自闭症诊断的分类器。使用留一法交叉验证来评估性能。为了评估站点间的稳健性，我们还在内部数据集上训练我们的分类器，并在鲁汶数据集上进行测试。
我们以 86.7%的准确率（91.7%的敏感性，83.3%的特异性）区分了 ASD 和非 ASD 青少年。在第二个实验中，我们还使用鲁汶数据集获得了 86.7%的分类性能（83.3%的敏感性和 88.9%的特异性）。最后，我们使用内部数据训练的分类器对鲁汶数据集进行分类，准确率为 80%（100%的敏感性，66.7%的特异性）。
本研究表明，时间神经动力学的相干性变化是 ASD 的生物标志物，基于小波相干的分类器可产生稳健且可复制的结果，可作为 ASD 的客观诊断工具。
