Two-step paretial least square regression classifiers in brain-state decoding using functional magnetic resonance imaging

RESEARCH ARTICLE Two-step paretial least square regression classifiers in brain-state decoding using functional magnetic resonance imaging Zhiying Long ID1, Yubao Wang1, Xuanping Liu1, Li Yao1,2* 1 State Key Laboratory of Cognitive Neuroscience and Learning, Beijing Normal University, Beijing, China, 2 School of Information Science and Technology, Beijing Normal University, Beijing, China a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 OPEN ACCESS Citation: Long Z, Wang Y, Liu X, Yao L (2019) Two-step paretial least square regression classifiers in brain-state decoding using functional magnetic resonance imaging. PLoS ONE 14(4): e0214937. https://doi.org/10.1371/journal. pone.0214937 Editor: Yong Fan, University of Pennsylvania Perelman School of Medicine, UNITED STATES Received: October 16, 2018 Accepted: March 24, 2019 Published: April 10, 2019 Copyright: © 2019 Long et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability Statement: Data can be downloaded from the website https://figshare.com/ s/03832c6ecadd85120b20 (DOI: 10.6084/m9. figshare.7914308). Funding: Our work is supported by the National Key Research and Development Program of China under grant 2017YFB1002502, the National Natural Science Foundation of China (61671067), the Key Program of National Natural Science Foundation of China (61731003), the Interdiscipline Research Funds of Beijing Normal University and the * Abstract Multivariate analysis methods have been widely applied to decode brain states from functional magnetic resonance imaging (fMRI) data. Among various multivariate analysis methods, partial least squares regression (PLSR) is often used to select relevant features for decoding brain states. However, PLSR is seldom directly used as a classifier to decode brain states from fMRI data. It is unclear how PLSR classifiers perform in brain-state decoding using fMRI. In this study, we propose two types of two-step PLSR classifiers that use PLSR/sparse PLSR (SPLSR) to select features and PLSR for classification to improve the performance of the PLSR classifier. The results of simulated and real fMRI data demonstrated that the PLSR classifier using PLSR/SPLSR to select features outperformed both the PLSR classifier using a general linear model (GLM) and the support vector machine (SVM) using PLSR/SPLSR/GLM in most cases. Moreover, PLSR using SPLSR to select features showed the best performance among all of the methods. Compared to GLM, PLSR is more sensitive in selecting the voxels that are specific to each task. The results suggest that the performance of the PLSR classifier can be largely improved when the PLSR classifier is combined with the feature selection methods of SPLSR and PLSR. Introduction One of the key challenges in cognitive neuroscience is to map the human brain activities to different brain states. Multi-voxel pattern analysis (MVPA) using machine learning models has been widely applied to functional magnetic resonance imaging (fMRI) datasets to address this question [1]. The models that are trained on stimulus-evoked brain activity can be used to discriminate multiple cognitive processes [2–4]. Two critical steps that include feature selection and classification are involved in decoding brain states. Among the various MVPA methods, the partial least squares regression (PLSR) is a powerful method for multivariate data analysis that can be used in either feature/variable PLOS ONE | https://doi.org/10.1371/journal.pone.0214937 April 10, 2019 1 / 16 Two-step PLSR classifiers in decoding using fMRI Fundamental Research Funds for the Central Universities (2017XTCX04). Competing interests: The authors have declared that no competing interests exist. selection or classification. The variable selection methods based on PLSR can be categorized into three categories: filter, wrapper, and embedded methods [5]. In contrast to wrapper and embedded methods, filter methods that use the output of PLSR to identify a subset of important variables are fast and easy to compute. So far, PLSR filter methods have been applied to select features from fMRI data. Chou et al. (2014) used the magnitude of the PLS regression coefficient as an importance index to select features [6]. Swathi et al. (2015) proposed an effective feature (voxel) selection strategy that used PLSR to form an index for the informative voxels to prioritize the voxel selection based on the degree of association with the stimulus conditions [7]. Tu et al. (2016) applied PLSR to extract neural features that were correlated with the intensity of laser-evoked nociceptive pain from electroencephalography and fMRI data [8]. Except for feature/variable selection, PLSR can also be used for classification. After the PLS regression coefficients are estimated, some discriminate approaches on PLS-predicted response or scores are applied to fMRI-based decoding. Rodriguez (2010) applied the partial least squares (PLS) regression combined with a maximum output threshold to predict locations in the spatial navigation of humans with fMRI [9]. Moreover, PLSR using linear discriminant analysis (PLSLDA) was applied to decode emotional states from fMRI data [10] and to discriminate individuals with disease from those at high risk [11]. Although PLS has been widely applied to variable selection, it does not automatically select relevant variables. In recent years, sparse PLSR (SPLSR) methods that add sparse constraint to the PLSR have been proposed and applied to automatic variable selection [12] or classification [13]. However, SPLSR has seldom been applied to fMRI-based decoding. For fMRI data, the number of features is much larger than the number of samples. Because PLSR and SPLSR can manage a large number of features (p) and a small sample size (n), it is essential to investigate how PLSR and SPLSR should be applied effectively to fMRI-based decoding. The previous studies applied PLSR/SPLSR to either feature selection or classification separately. It is unclear whether PLSR/SPLSR can be effectively used simultaneously for both feature selection and classification in fMRI-based decoding. In this study, we proposed a two-step PLSR framework that first used PLSR or SPLSR for feature selection and then used PLSR for classification. The method that uses PLSR for both feature selection and classification is named P_PLSR. The method that uses SPLSR for feature selection and PLSR for classification is named SP_PLSR. We investigated the robustness and feasibility of P_PLSR and SP_PLSR in brain-state decoding based on fMRI. Moreover, we compared P_PLSR and SP_PLSR using a general linear model (GLM) for feature selection and a support vector machine (SVM) for classification (G_SVM), PLSR for feature selection and SVM for classification (P (...truncated)