Centrality-based nearest-neighbor projected-distance regression (C-NPDR) feature selection for correlation predictors with application to resting-state fMRI of major depressive disorder

Centrality-based nearest-neighbor projected-distance regression (C-NPDR) feature selection for correlation predictors with application to resting-state fMRI of major depressive disorder | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Centrality-based nearest-neighbor projected-distance regression (C-NPDR) feature selection for correlation predictors with application to resting-state fMRI of major depressive disorder Elizabeth Kresock, Henry Luttbeg, Jamie Li, Rayus Kuplicki, B. A. McKinney, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4193488/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background. Nearest-neighbor projected-distance regression (NPDR) is a metric-based machine learning feature selection algorithm that uses distances between samples and projected differences between variables to identify variables or features that may interact to affect the prediction of complex outcomes. Typical bioinformatics data consist of separate variables of interest like genes or proteins. In contrast, resting-state functional MRI (rs-fMRI) data is composed of time-series for brain Regions of Interest (ROIs) for each subject, and these within-brain time-series are typically transformed into correlations between pairs of ROIs. These pairs of variables of interest can then be used as input for feature selection or other machine learning. Straightforward feature selection would return the most significant pairs of ROIs; however, it would also be beneficial to know the importance of individual ROIs. Results. We extend NPDR to compute the importance of individual ROIs from correlation-based features. We present correlation-difference and centrality-based versions of NPDR. The centrality-based NPDR can be coupled with any centrality method and can be coupled with importance scores other than NPDR, such as random forest importance. We develop a new simulation method using random network theory to generate artificial correlation data predictors with variation in correlation that affects class prediction. Conclusions. We compare feature selection methods based on detecting functional simulated ROIs, and we apply the new centrality NPDR approach to a resting-state fMRI study of major depressive disorder (MDD) and healthy controls. We determine that the areas of the brain that are the most interactive in MDD patients include the middle temporal gyrus, the inferior temporal gyrus, and the dorsal entorhinal cortex. The resulting feature selection and simulation approaches can be applied to other domains that use correlation-based features. Resting-state fMRI statistical interactions machine learning feature selection Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Background Resting-state functional MRI (rs-fMRI) measures the blood-oxygen-level-dependent (BOLD) signal in regions of interest (ROIs) throughout the entire brain of a subject while at rest (i.e., not during a cognitive task) [ 1 , 2 , 3 , 4 ]. The BOLD signal for each ROI is a time-series whose low frequency fluctuations show correlation between other ROIs. These correlations can be used to represent the functional connectivity (FC) between ROIs in a weighted brain network [ 5 , 6 , 7 , 8 , 9 ]. Differential changes, or rewiring, of correlation FC between subjects with mood disorders and healthy controls may reveal neural mechanisms of disease. Feature selection and machine learning methods that use FC measures as predictors have potential as fMRI-based biomarkers and disease classification [ 23 ]. Machine learning classification and feature selection have been used widely for rs-fMRI data to detect and better understand the mechanisms of mood disorders, including major depressive disorder (MDD) [ 24 ]. A wide variety of algorithms have been used, including Support Vector Machines (SVM), XGBoost, Random Forests, and deep learning. Various biomarkers and measures of FC between brain regions have been used as fMRI-based ML predictors, including correlation, mutual information, amplitude of low frequency fluctuation (ALFF) and regional homogeneity (ReHo). ML classification of MDD with rs-fMRI has been promising, but using ML for diagnosis is likely premature [ 24 ]. Feature selection, which is the focus of the current study, while not directly diagnostic, may provide valuable insights into the biological mechanisms of MDD. Multiple approaches have been used to define the fMRI input predictors for ML. Some approaches for ROI feature engineering have been more data-driven, for example, by using independent component analysis (ICA) to define brain networks and ROIs [ 25 ]. Similarly, multi-voxel pattern analysis (MVPA) [ 12 ] uses SVM to build classifiers without assumptions about the organization of the brain. But these can be difficult to generalize between data sets and studies. We use preexisting ROIs based on brain atlases constructed by experts using anatomical and functional information. (Fig. 1 A). In this study, we use Pearson's correlation across the full time series to quantify FC between brain regions as a representation of temporal synchrony of their activity. For each atlas ROI (Fig. 1 A), we average the voxel time series in the ROI (Fig. 1 B) and then compute the correlation between all pairs of n ROI time series (Fig. 1 D). For each subject, the upper triangle of the correlation matrices are stretched to create datasets where the n(n-1)/2 predictor variables are pairwise correlations between ROIs (Fig. 1 D). Standard feature selection methods can then be used for this data to give the importance of ROI pairs. One of the goals of the current study is to disentangle the many important pairs to identify important individual ROIs. A common way to identify important ROIs is to perform a seed-based analysis, where the global correlation between a given seed region and all other brain regions is computed [ 26 ], and then this centrality quantity can be tested for all ROIs for association with an outcome like MDD. In the current study, we use centrality in a different way with our nearest-neighbor projected distance regression (NPDR) approach [ 15 ]. We apply NPDR to correlation-based predictors (Fig. 1 D) and then use network theory to determine the cumulative effect of the differential correlations for each ROI. In addition to this centrality-NPDR (c-NPDR), we also use a correlation-based projected difference to directly compute the importance of ROIs. We also apply this centrality approach with random forest (c-rf). We compare the rs-fMRI feature selection methods using a new simulation tool, and we apply the methods to a previous rs-fMRI study of MDD. Methods We describe a new centrality-based algorithm and projected correlation distance algorithm for ranking the importance of ROIs for predicting a class variable. Both of these methods are based on nearest-neighbor projected-distance regression (NPDR), a machine learning algorithm that is able to detect statistical interactions using nearest neighbors in a high-dimensional space [ 15 ]. NPDR minimizes a contrastive loss function for pairs of samples \((i,j)\) . The contrastive loss \({\delta }_{ij}\left(y\right)\) is an indicator of whether samples \((i,j)\) are in the same class or a different class based on class variable y . The contrastive loss can be penalized by LASSO or Ridge, or it can be unpenalized and P-values can be computed. Rather than using the predictor/attribute values directly in the regression, NPDR uses the difference \({\stackrel{⃑}{d}}_{ij}\left(X\right)\) (or projected distance onto the attributes X) between subjects \((i,j)\) . The vector denotes the projected distance for all attributes in the set X. In the current application, the attributes are Pearson correlations between pairs of ROIs. For centrality NPDR (c-NPDR), the projected distance or diff, \({d}_{ij}\left(p\right)\) , is the absolute difference between subjects \((i,j)\) for one correlation attribute p (correlation between a pair of ROIs): $${d}_{ij}\left(p\right)=\left|{A}_{p}^{\left(i\right)}-{A}_{p}^{\left(j\right)}\right|$$ 1 , where \({A}_{p}^{\left(i\right)}\) is the correlation for subject i between a pair of ROIs, represented by p . Thus, if there are n ROIs, the NPDR design matrix consists of n(n-1)/2 attribute columns; one for each pair of ROIs. The NPDR-selected ROI pairs can then be used in any number of centrality algorithms to rank the importance of individual ROIs. For comparison, we use the following centralities: degree, betweenness, eigenvector and integrated value of influence (IVI) [ 30 ]. The other NPDR-based method (correlation-diff-NPDR) for ranking the importance of ROIs from ROI-pair correlation data uses a more complex projected distance, \({d}_{ij}^{CD}\) , but directly gives importance of individual ROIs without centrality calculations [ 27 ]. The correlation-diff (CD) or correlation projected distance for ROI r is given by $${d}_{ij}^{CD}\left(r\right)={\sum }_{k\ne r}\left|{A}_{rk}^{\left(i\right)}-{A}_{rk}^{\left(j\right)}\right|$$ 2 , where \({A}_{rk}^{\left(i\right)}\) is the correlation between ROIs r and k for subject i . Thus, the correlation-diff for ROI r is the absolute sum of differences between r and all other ROIs. If there are n ROIs, the NPDR design matrix for Eq. ( 2 ) will have n columns as opposed to n(n-1)/2 for Eq. ( 1 ). Thus, NPDR with Eq. ( 2 ) yields importance scores for ROIs, while NPDR with Eq. ( 1 ) yields the importance of ROI pairs. In both cases (Eqs. 1 and 2 ), NPDR importance can be computed in terms of individual P-values or in a multivariate model with LASSO or Ridge. In order to threshold the results of correlation-diff-NPDR and cNPDR, we use regularization and p-values. We use the LASSO penalty, also known as the L1 penalty, which is a regularization technique used in regression models to prevent overfitting and to enhance the model's prediction accuracy and interpretability. For non-penalized methods, we use a p-value adjusted cutoff, where ROI pairs that had an adjusted p-value > 0.05 are removed from the network. To threshold the random forest results, we use a cutoff of the top 100 pairs of ROIs (Fig. 2 ). Correlation-diff-NPDR directly yields a list of significant individual ROIs. However, the centrality methods need an additional step to map pair importance to individual importance. The c-NPDR and c-rf methods yield lists of important pairs of ROIs, so we apply centralities to the resulting edge lists to obtain a list of important individual ROIs (Fig. 2 ). The significant pairs of ROIs are graphed as a network, where the nodes are ROIs and edges are defined when the ROI pairs have a correlation that affects the outcome variable (e.g., MDD). This interaction network is a way to visualize the importance of MDD nodes based on their connections and visualize local structure. We quantify the importance of individual ROIs using common centralities, degree, eigenvector, betweenness, and IVI [ 30 ]. IVI combines multiple centrality measures. We compare the NPDR methods to a centrality version of random forest (c-rf). We use the correlation predictor data (Fig. 1 d) with random forest permutation importance with 5000 trees, filter the correlation pairs to the top 200 to create a network, and then computed ROI centralities. Simulation Method and Real Data. Simulation Approach. We develop a random network approach to simulate correlation-based features, a fraction of which are functional or associated with the case-control status (Fig. 3 ). The application we have in mind is correlation between brain ROIs in resting-state fMRI studies, where correlation is calculated from the BOLD signal time-series. We do not simulate the time series, but rather directly simulate the correlations and their differences between groups. Features or predictors are correlations between pairs of ROIs rather than ROIs themselves. We note that these simulations and feature selection methods would also apply to other types of correlation-based data in other research domains. The user specifies the number of ROIs, the number of cases and controls, the number of functional ROIs (i.e., that are associated with the outcome), the effect size and the type of underlying random network for the brain. The user can specify their own network, for example from real data, or they can generate any network from the igraph library. Initial correlation matrices are generated for each sample based on the network, where connected ROIs have higher random correlations than unconnected ROIs. Functional nodes are chosen from the largest connected component (i.e. group of nodes such that there is a path between any pair of nodes in the group). Edges between the functional nodes (green edges in Fig. 3 ) are then used to create differential correlation between cases and controls (black dots in Fig. 3 heatmaps). We use a parameter called “multiway” that controls how many edges we randomly select to generate differential correlation. For example, a multiway of 2 will use only a subset of the possible edges between functional nodes (a subset of possible edges will be green). If we set multiway to the maximum, then all possible edges between functional nodes will have differential correlation (green). We use multiway = 5 in this application. We generate replicate simulations to compare feature selection methods based on the ability to detect the ground truth functional ROIs. We use the F1 Score to test whether the top ROI features selected by a method overlap with the top functional features. Real Data . We compare feature selection methods on data from the Tulsa 1000 (T1000), a longitudinal study at the Laureate Institute for Brain Research following 1000 individuals, including healthy individuals and those with mood and other disorders [ 28 ]. We use rs-fMRI time series for 188 MDD subjects and 47 healthy controls (HC) from T1000 (163 female and 72 male). We use the Automated Anatomical Labelling Atlas (AAL Atlas) with 87 ROIs and the Brainnetome Atlas with 246 ROIs to define consistent and interpretable mappings for selected features [ 29 , 10 ]. The Brainnetome Atlas parcellates the brain based on structural and connectivity features. It is based on neuroimaging data, particularly rs-fMRI and diffusion tensor imaging (DTI) data, which reveal both the functional and structural connectivity patterns in the brain. For each atlas, we detrended the signals and averaged the time-series for the voxels within an atlas ROI. Results We simulate 50 replicate datasets each with 100 cases, 100 controls, and 100 ROIs. We select 10 functional ROIs, but their effects are detected through their correlations with each other in correlation-predictor datasets. The underlying correlation networks are based on Erdos-Renyi random networks with p = 0.1 connection probability. We used a medium effect size of 0.5 Cohen’s d. We apply six feature selection methods to the replicate simulations (Fig. 4 ) and compare them based on their average ability to detect the 10 functional ROIs. Centrality-Random Forest (c-rf) with degree centrality (red, Fig. 4 ) has a similar mean F1 score to correlation-diff NPDR (corr-diff) using Ridge regression (green CD Ridge, Fig. 4 ), and they are both similar to centrality-NPDR (c-npdr) with degree centrality (left blue). The npdr-based methods corr-diff and c-npdr with degree have slightly less variation than c-rf. The F1 scores for centrality-based NPDR methods (all blue, Fig. 4 ) depend on the centrality method used. C-NPDR works best with degree, whereas IVI, betweenness, and eigenvector centralities are noticeably worse. The close similarity between corr-diff and c-npdr (Eq. 1 ) with degree suggests that the corr-diff NPDR difference metric (Eq. 2 ) works mathematically like degree. We apply four feature selection methods to the real rs-fMRI study to compare the selected important ROIs for MDD (Table 1). The correlation-diff NPDR with LASSO selects the fewest features because it has a tendency to eliminate correlated features (first column, Table 1). The other correlation-diff NPDR (second column, Table 1) uses an adjusted p-value cutoff, and it results in more selected features in part due to inclusion of more correlated features. The centrality methods with degree, NPDR (column 3) and random forest (column four), use a manual threshold because degree centrality does not have a statistical threshold. For the Brainnetome atlas (top, Table 1), NPDR correlation-diff LASSO (column 1) selects the most parsimonious list of ROIs, and the these features are included in the longer lists of NPDR methods (columns 2 and 3). The random forest method has a slightly different set of selected ROIs because it is not metric based and it tends to find more main effects than interactions compared to NPDR methods. We highlight the selected ROIs involving MTG (middle temporal gyrus) because it is the top ROI found by NPDR correlation-diff and has been associated with MDD [ 20 ], [ 31 ]. For the AAL atlas (bottom section, Table 1), the three NPDR methods have a consensus of ROIs selected by the LASSO method. These regions include the dorsal and ventral default mode networks (DMN) and the right executive control network (ECN). The random forest centrality method includes multiple blocks of correlated variables (bottom, Table 1) in regions like Anterior Salience, Auditory, and dorsal DMN. The NPDR methods have reduced correlation, and the LASSO version automatically selects a parsimonious set of ROIs. The highest NPDR degree nodes for AAL (Dorsal_DMN_ 03 and Ventral_DMN_07) are in the same Louvain graph cluster and the third highest degree node (Right_ECN_04) is in a different cluster (Fig. 5 ). Discussion The middle temporal gyrus (MTG) was found by all feature selection methods to be important for MDD (Table 1). The centrality random forest method only identified the left MTG while the LASSO correlation-diff NPDR identified only the right MTG. The other NPDR methods, including the centrality based method, identified both left and right MTG. MTG is critical for semantic memory processing, visual perception, and language processing [ 17 ], and studies have found associations with MDD. For example, studies using structural and functional MRI identified significant gray matter abnormalities in the right MTG among patients with treatment-resistant depression (TRD) and treatment-responsive depression (TSD) compared to healthy controls [ 20 ]. The reduced gray matter volume in the bilateral MTG is indicative of structural changes associated with MDD. Similarly, a previous study found fractional amplitude of low-frequency fluctuation (fALFF) was higher in patients with MDD in the right and left MTG [ 31 ]. All tested methods also identified STG (superior temporal gyrus) and ITG (inferior temporal gyrus) as important for MDD. In patients with anxious depression versus healthy controls, a previous study found increased fALFF values in the left STG [ 32 ]. While not previously linked to MDD, ITG showed gray matter volume reductions in the MTG and ITG in chronic schizophrenia patients [ 21 ]. The link to another psychiatric condition could indicate a broader role for ITG in mood disorders. Conclusions The application of machine learning and feature selection algorithms to fMRI data is increasingly critical toward understanding biological mechanisms of disorders. New methods are needed that can account for interactions between variables and regions of interest. NPDR has been shown previously to be able to detect interactions, and we extended it to handle data where the predictors are correlations between pairs of ROIs. We also developed a new simulation approach to compare these methods. We found that c-NPDR with degree is similar to the correlation-diff NPDR. This suggests a mathematical connection between the correlation-diff metric and degree centrality, and it suggests ways that correlation-diff might be improved by using other centralities. We applied these NPDR methods and a random forest approach to correlation data from a real rs-fMRI dataset for MDD, and the consensus between these methods found MTG, ITG, and STG to be important MDD ROIs. Abbreviations rs-fMRI resting-state functional magnetic resonance imaging MDD major depressive disorder HC healthy control ROI region of interest Declarations Author Contribution EK and BM wrote the main manuscript. EK implemented the centrality approach and performed data analysis. BD developed the corr-diff approach and simulation method. HL performed simulation analysis. JL and RK contributed to analysis design. All authors reviewed the manuscript. BM led the overall design. References Van Den Heuvel MP, Hulshoff Pol HE. Exploring the brain network: A review on resting-state fMRI functional connectivity, European Neuropsychopharmacology, vol. 20, no. 8, pp. 519–534, Aug. 2010, 10.1016/j.euroneuro.2010.03.008 . Salvador R, Suckling J, Coleman MR, Pickard JD, Menon D, Bullmore E. Neurophysiological Architecture of Functional Magnetic Resonance Images of Human Brain, Cerebral Cortex, vol. 15, no. 9, pp. 1332–1342, Sep. 2005, 10.1093/cercor/bhi016 . Iraji A, et al. The connectivity domain: Analyzing resting state fMRI data using feature-based data-driven and model-based methods. NeuroImage. Jul. 2016;134:494–507. 10.1016/j.neuroimage.2016.04.006 . Preti MG, Bolton TA, Van De Ville D. The dynamic functional connectome: State-of-the-art and perspectives, NeuroImage, vol. 160, pp. 41–54, Oct. 2017, 10.1016/j.neuroimage.2016.12.061 . Biswal BB et al. Mar., Toward discovery science of human brain function, Proc. Natl. Acad. Sci. U.S.A., vol. 107, no. 10, pp. 4734–4739, 2010, 10.1073/pnas.0911855107 . Smith SM et al. Jan., Network modelling methods for FMRI, NeuroImage, vol. 54, no. 2, pp. 875–891, 2011, 10.1016/j.neuroimage.2010.08.063 . Van Dijk KRA, Sabuncu MR, Buckner RL. The influence of head motion on intrinsic functional connectivity MRI. NeuroImage. Jan. 2012;59(1):431–8. 10.1016/j.neuroimage.2011.07.044 . Zuo X-N, Xing X-X. Test-retest reliabilities of resting-state FMRI measurements in human brain functional connectomics: A systems neuroscience perspective, Neuroscience & Biobehavioral Reviews, vol. 45, pp. 100–118, Sep. 2014, 10.1016/j.neubiorev.2014.05.009 . Fornito A, Zalesky A, Breakspear M. Graph analysis of the human connectome: Promise, progress, and pitfalls, NeuroImage, vol. 80, pp. 426–444, Oct. 2013, 10.1016/j.neuroimage.2013.04.087 . Fan L, et al. The Human Brainnetome Atlas: A New Brain Atlas Based on Connectional Architecture. Cereb Cortex. Aug. 2016;26(8):3508–26. 10.1093/cercor/bhw157 . Zhuo J, Fan L, Liu Y, Zhang Y, Yu C, Jiang T. Connectivity Profiles Reveal a Transition Subarea in the Parahippocampal Region That Integrates the Anterior Temporal–Posterior Medial Systems, J. Neurosci., vol. 36, no. 9, pp. 2782–2795, Mar. 2016, 10.1523/JNEUROSCI.1975-15.2016 . Al-Zubaidi A, Mertins A, Heldmann M, Jauch-Chara K, Münte TF. Machine Learning Based Classification of Resting-State fMRI Features Exemplified by Metabolic State (Hunger/Satiety), Frontiers in Human Neuroscience, vol. 13, 2019, Accessed: Feb. 05, 2024. [Online]. Available: https://www.frontiersin.org/articles/ 10.3389/fnhum.2019.00164 . Álvarez JD, Matias-Guiu JA, Cabrera-Martín MN, Risco-Martín JL, Ayala JL. An application of machine learning with feature selection to improve diagnosis and classification of neurodegenerative disorders, BMC Bioinformatics, vol. 20, no. 1, p. 491, Oct. 2019, 10.1186/s12859-019-3027-7 . Shi C, Zhang J, Wu X. An fMRI Feature Selection Method Based on a Minimum Spanning Tree for Identifying Patients with Autism, Symmetry, vol. 12, no. 12, p. 1995, Dec. 2020, 10.3390/sym12121995 . Le TT, Dawkins BA, McKinney BA. Nearest-neighbor Projected-Distance Regression (NPDR) for detecting network interactions with adjustments for multiple tests and confounding. Bioinformatics. May 2020;36(9):2770–7. 10.1093/bioinformatics/btaa024 . Breiman L. Random Forests Mach Learn. 2001;45(1):5–32. 10.1023/A:1010933404324 . Loh D. Middle temporal gyrus | Radiology Reference Article | Radiopaedia.org, Radiopaedia. Accessed: Feb. 05, 2024. [Online]. Available: https://radiopaedia.org/articles/middle-temporal-gyrus . Loh D. Inferior temporal gyrus | Radiology Reference Article | Radiopaedia.org, Radiopaedia. Accessed: Feb. 05, 2024. [Online]. Available: https://radiopaedia.org/articles/inferior-temporal-gyrus . Gaillard F. Entorhinal cortex | Radiology Reference Article | Radiopaedia.org, Radiopaedia. Accessed: Feb. 05, 2024. [Online]. Available: https://radiopaedia.org/articles/entorhinal-cortex . Ma C, et al. Resting-State Functional Connectivity Bias of Middle Temporal Gyrus and Caudate with Altered Gray Matter Volume in Major Depression. PLoS ONE. Sep. 2012;7(9):e45263. 10.1371/journal.pone.0045263 . Onitsuka T, et al. Middle and Inferior Temporal Gyrus Gray Matter Volume Abnormalities in Chronic Schizophrenia: An MRI Study. AJP. Sep. 2004;161(9):1603–11. 10.1176/appi.ajp.161.9.1603 . Kim IB, Park S-C. Neural Circuitry–Neurogenesis Coupling Model of Depression. IJMS. Feb. 2021;22(5):2468. 10.3390/ijms22052468 . Du Y, Fu Z, Calhoun VD. Classification and Prediction of Brain Disorders Using Functional Connectivity: Promising but Challenging. Front Neurosci. Aug. 2018;12:525. 10.3389/fnins.2018.00525 . Chen Y, Zhao W, Yi S, Liu J. The diagnostic performance of machine learning based on resting-state functional magnetic resonance imaging data for major depressive disorders: a systematic review and meta-analysis. Front Neurosci. Sep. 2023;17:1174080. 10.3389/fnins.2023.1174080 . Calhoun VD, de Lacy N. Neuroimaging Clin N Am. Nov. 2017;27(4):561. 10.1016/j.nic.2017.06.012 . Ten Key Observations on the Analysis of Resting-state Functional MR Imaging Data Using Independent Component Analysis. Joel SE, Caffo BS, van Zijl PC, Pekar JJ. On the relationship between seed-based and ICA-based measures of functional connectivity, Magn Reson Med, vol. 66, no. 3, pp. 644–657, Sep. 2011, 10.1002/mrm.22818 . Dawkins BA, Le TT, McKinney BA. Theoretical properties of distance distributions and novel metrics for nearest-neighbor feature selection. PLoS ONE. Feb. 2021;16(2):e0246761. 10.1371/journal.pone.0246761 . Victor TA, et al. Tulsa 1000: a naturalistic study protocol for multilevel assessment and outcome prediction in a large psychiatric sample. BMJ Open. Jan. 2018;8(1):e016620. 10.1136/bmjopen-2017-016620 . Rolls ET, Huang C-C, Lin C-P, Feng J, Joliot M. Automated anatomical labelling atlas 3. NeuroImage. Feb. 2020;206:116189. 10.1016/j.neuroimage.2019.116189 . Salavaty A, Ramialison M, Currie PD. Integrated Value of Influence: An Integrative Method for the Identification of the Most Influential Nodes within Networks. Patterns. Aug. 2020;1(5):100052. 10.1016/j.patter.2020.100052 . Zhang Q, et al. Associations between abnormal spontaneous neural activity and clinical variables, eye movements, and event-related potential indicators in major depressive disorder. Front Neurosci. Jan. 2023;16:1056868. 10.3389/fnins.2022.1056868 . Zhao P, et al. Altered fractional amplitude of low-frequency fluctuations in the superior temporal gyrus: a resting-state fMRI study in anxious depression. BMC Psychiatry. Nov. 2023;23(1):847. 10.1186/s12888-023-05364-w . Tables Table 1 is available in the Supplementary Files section. Additional Declarations No competing interests reported. Supplementary Files Table1.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-4193488","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":285721681,"identity":"7057c1fd-35e7-4b2c-91e2-513e5e7b7c1f","order_by":0,"name":"Elizabeth Kresock","email":"","orcid":"","institution":"The University of Tulsa","correspondingAuthor":false,"prefix":"","firstName":"Elizabeth","middleName":"","lastName":"Kresock","suffix":""},{"id":285721682,"identity":"b12bb478-7627-4155-a3a3-1a43958fb049","order_by":1,"name":"Henry Luttbeg","email":"","orcid":"","institution":"The University of Tulsa","correspondingAuthor":false,"prefix":"","firstName":"Henry","middleName":"","lastName":"Luttbeg","suffix":""},{"id":285721683,"identity":"8a5e22e6-670f-4ccb-838c-7579c8a7b3d0","order_by":2,"name":"Jamie Li","email":"","orcid":"","institution":"The University of Tulsa","correspondingAuthor":false,"prefix":"","firstName":"Jamie","middleName":"","lastName":"Li","suffix":""},{"id":285721684,"identity":"2740c902-3be2-4dd4-8562-155d6bf1f9f4","order_by":3,"name":"Rayus Kuplicki","email":"","orcid":"","institution":"Laureate Institute for Brain Research","correspondingAuthor":false,"prefix":"","firstName":"Rayus","middleName":"","lastName":"Kuplicki","suffix":""},{"id":285721685,"identity":"53a0aa96-0bca-495b-83bb-1b108aa73336","order_by":4,"name":"B. A. McKinney","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABAklEQVRIiWNgGAWjYBACAxhDghlM2QAxcwOIxdiAWwtjwwGEljQGHqhiIrRA+IcJazGXPnz88QeGbfKS7bzHHnxsO5+4XyKx+cMPBhvZDQewa7HsS0sE2nLbcDYzX7rhzLbbiT0SiW2SPQxpxri0GJzhMWw4+O92ghwzj5k0L1QL0HWHE3Fr4f8IsgWi5W/bOZCW5o9/GP7j0cLDCNYiDdLC2HYApKVBmofhAE4tlj1shjPOAP0ys5nHTLLnXLJxz5mHbdIyBsnGM3FoMedhfvChguG2vMT5M2YSP8rsZNvbkw9/fFNhJ9uHQwsqYGSDO5gY5WDwh2iVo2AUjIJRMIIAANPRYByHFQ7vAAAAAElFTkSuQmCC","orcid":"","institution":"The University of Tulsa","correspondingAuthor":true,"prefix":"","firstName":"B.","middleName":"A.","lastName":"McKinney","suffix":""},{"id":285721686,"identity":"347ebe00-81cb-47f8-9f80-127e7538162d","order_by":5,"name":"Brett McKinney","email":"","orcid":"","institution":"The University of Tulsa","correspondingAuthor":false,"prefix":"","firstName":"Brett","middleName":"","lastName":"McKinney","suffix":""},{"id":285721687,"identity":"f6451ca4-44c2-47a1-8ab2-3a232932c523","order_by":6,"name":"Bryan Dawkins","email":"","orcid":"","institution":"SomaLogic, Inc","correspondingAuthor":false,"prefix":"","firstName":"Bryan","middleName":"","lastName":"Dawkins","suffix":""}],"badges":[],"createdAt":"2024-03-30 20:29:13","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-4193488/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4193488/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":53879514,"identity":"a02b7e3e-af2f-46ea-ad36-e68a4379ad8b","added_by":"auto","created_at":"2024-04-01 17:11:37","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":367315,"visible":true,"origin":"","legend":"\u003cp\u003eIllustration of resting-state fMRI data used for machine learning feature selection. Regions of interest (ROIs) are made up of groups of voxels within the brain. Three ROIs (A) are used for illustration (green, blue, and red cubes/voxels), but the number of ROIs is typically on the order of 200. Each voxel has an associated time series, which are averaged within ROIs to create the green, red and blue time series (B).\u0026nbsp; From these time-series, pairwise ROI correlations are calculated and stored in a matrix for each subject (C). The upper triangle of each subject’s correlation matrix can be stretched into a sample vector, s\u003csub\u003ei\u003c/sub\u003e, to form rows of a dataset (D), where the predictors (columns) are ROI-ROI correlations.\u0026nbsp;\u003c/p\u003e","description":"","filename":"image1.png","url":"https://assets-eu.researchsquare.com/files/rs-4193488/v1/7c2cd3b9907d9d0692a62803.png"},{"id":53879515,"identity":"75cdb212-ad81-4361-bc05-b77a57b1f75c","added_by":"auto","created_at":"2024-04-01 17:11:37","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":126907,"visible":true,"origin":"","legend":"\u003cp\u003eAnalysis methods for rs-fMRI data with correlation-based features and a class variable. On the left, correlation-diff-NPDR (Eq. 2) can directly rank the importance of ROIs using P-values or penalized regression coefficients. On the right, centrality-NPDR (C-NPDR, Eq. 1) and centrality random forest (c-rf) rank the importance of pairs of ROIs, and then centralities of the resulting ROI-ROI networks are used to rank the importance of individual ROIs.\u003c/p\u003e","description":"","filename":"image2.png","url":"https://assets-eu.researchsquare.com/files/rs-4193488/v1/7bc6dcde1046f6456ff3fee7.png"},{"id":53879513,"identity":"8914b5b7-9ed2-4b12-862b-0d0f801f51ec","added_by":"auto","created_at":"2024-04-01 17:11:37","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":788127,"visible":true,"origin":"","legend":"\u003cp\u003eSimulation method for correlation-based features. A random network is generated (Erdos-Renyi in the example) between the number of regions of interest or ROIs (10 circles in the example). For each sample, random correlations are mapped to connected regions of the network and lower random correlations to unconnected regions. Pairs of regions are selected to be functional (associated with the outcome variable), which are indicated by green edges in the brain network and black dots in the heat maps. Each heatmap is a different sample. For the cases (left heat maps), the selected functional pairs are perturbed to have higher correlation, and for the controls (right heat maps), the functional pairs are perturbed lower. The final heat maps represent case-control datasets with correlation-based features containing noise and functional ROI-pairs.\u003c/p\u003e","description":"","filename":"image3.png","url":"https://assets-eu.researchsquare.com/files/rs-4193488/v1/dc6d68458b237454230e9713.png"},{"id":53879518,"identity":"0e63575f-6499-45ea-abcc-b77cda9c6909","added_by":"auto","created_at":"2024-04-01 17:11:37","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":51432,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of feature selection methods for identifying simulated top 10 functional ROIs (out of 100) that are associated with the class variable. F1 score is used to quantify whether the top 10 ranking by a given method is enriched for the top 10 functional ROIs. Each violin plot represents 50 replicate simulations. Simulations are created using random network theory (Fig. 3). The same cutoff was chosen for all methods because random forest does not have a statistical threshold for the importance score. The centrality-Random Forest method (c-rf, red) uses degree to compute ROI importance. The NPDR method that uses the correlation diff (corr-diff, green) uses NPDR-Ridge to rank ROIs directly as opposed to using cenrality. Centrality-NPDR methods (right three blues) use degree, IVI, and eigenvector centrality to compute ROI importance from the NPDR network.\u003c/p\u003e","description":"","filename":"image4.png","url":"https://assets-eu.researchsquare.com/files/rs-4193488/v1/4e1e80ff80ee7400f8d38db1.png"},{"id":53879516,"identity":"fb1693c1-0763-41a3-88f2-70a5fd8b0e14","added_by":"auto","created_at":"2024-04-01 17:11:37","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":169513,"visible":true,"origin":"","legend":"\u003cp\u003eNPDR network of brain regions of interest (ROIs) for major depressive disorder (MDD) using the AAL atlas (Table 1, bottom section). Nodes are sized by degree (Table 1, bottom section, column 3) and colored by Louvain clustering.\u003c/p\u003e","description":"","filename":"image5.png","url":"https://assets-eu.researchsquare.com/files/rs-4193488/v1/be3531b3ee1b144e0234b79d.png"},{"id":54868074,"identity":"d9511b10-4c98-43f5-9621-5a20a688cea4","added_by":"auto","created_at":"2024-04-17 22:37:35","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1818581,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4193488/v1/0abc3faa-bd62-411e-b6fb-e0cf33b6c024.pdf"},{"id":53879517,"identity":"80afd82f-2d4c-4dcd-b592-9ea6589f66b0","added_by":"auto","created_at":"2024-04-01 17:11:37","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":75860,"visible":true,"origin":"","legend":"","description":"","filename":"Table1.docx","url":"https://assets-eu.researchsquare.com/files/rs-4193488/v1/a06b7ca5ef798da71a3edc1b.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Centrality-based nearest-neighbor projected-distance regression (C-NPDR) feature selection for correlation predictors with application to resting-state fMRI of major depressive disorder","fulltext":[{"header":"Background","content":"\u003cp\u003eResting-state functional MRI (rs-fMRI) measures the blood-oxygen-level-dependent (BOLD) signal in regions of interest (ROIs) throughout the entire brain of a subject while at rest (i.e., not during a cognitive task) [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. The BOLD signal for each ROI is a time-series whose low frequency fluctuations show correlation between other ROIs. These correlations can be used to represent the functional connectivity (FC) between ROIs in a weighted brain network [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Differential changes, or rewiring, of correlation FC between subjects with mood disorders and healthy controls may reveal neural mechanisms of disease. Feature selection and machine learning methods that use FC measures as predictors have potential as fMRI-based biomarkers and disease classification [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eMachine learning classification and feature selection have been used widely for rs-fMRI data to detect and better understand the mechanisms of mood disorders, including major depressive disorder (MDD) [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. A wide variety of algorithms have been used, including Support Vector Machines (SVM), XGBoost, Random Forests, and deep learning. Various biomarkers and measures of FC between brain regions have been used as fMRI-based ML predictors, including correlation, mutual information, amplitude of low frequency fluctuation (ALFF) and regional homogeneity (ReHo). ML classification of MDD with rs-fMRI has been promising, but using ML for diagnosis is likely premature [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. Feature selection, which is the focus of the current study, while not directly diagnostic, may provide valuable insights into the biological mechanisms of MDD.\u003c/p\u003e \u003cp\u003eMultiple approaches have been used to define the fMRI input predictors for ML. Some approaches for ROI feature engineering have been more data-driven, for example, by using independent component analysis (ICA) to define brain networks and ROIs [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. Similarly, multi-voxel pattern analysis (MVPA) [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e] uses SVM to build classifiers without assumptions about the organization of the brain. But these can be difficult to generalize between data sets and studies. We use preexisting ROIs based on brain atlases constructed by experts using anatomical and functional information. (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eA).\u003c/p\u003e \u003cp\u003eIn this study, we use Pearson's correlation across the full time series to quantify FC between brain regions as a representation of temporal synchrony of their activity. For each atlas ROI (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eA), we average the voxel time series in the ROI (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eB) and then compute the correlation between all pairs of n ROI time series (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eD). For each subject, the upper triangle of the correlation matrices are stretched to create datasets where the n(n-1)/2 predictor variables are pairwise correlations between ROIs (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eD). Standard feature selection methods can then be used for this data to give the importance of ROI pairs. One of the goals of the current study is to disentangle the many important pairs to identify important individual ROIs.\u003c/p\u003e \u003cp\u003eA common way to identify important ROIs is to perform a seed-based analysis, where the global correlation between a given seed region and all other brain regions is computed [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e], and then this centrality quantity can be tested for all ROIs for association with an outcome like MDD. In the current study, we use centrality in a different way with our nearest-neighbor projected distance regression (NPDR) approach [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. We apply NPDR to correlation-based predictors (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eD) and then use network theory to determine the cumulative effect of the differential correlations for each ROI. In addition to this centrality-NPDR (c-NPDR), we also use a correlation-based projected difference to directly compute the importance of ROIs. We also apply this centrality approach with random forest (c-rf). We compare the rs-fMRI feature selection methods using a new simulation tool, and we apply the methods to a previous rs-fMRI study of MDD.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003eWe describe a new centrality-based algorithm and projected correlation distance algorithm for ranking the importance of ROIs for predicting a class variable. Both of these methods are based on nearest-neighbor projected-distance regression (NPDR), a machine learning algorithm that is able to detect statistical interactions using nearest neighbors in a high-dimensional space [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. NPDR minimizes a contrastive loss function for pairs of samples \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\((i,j)\\)\u003c/span\u003e\u003c/span\u003e. The contrastive loss \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\delta }_{ij}\\left(y\\right)\\)\u003c/span\u003e\u003c/span\u003e is an indicator of whether samples \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\((i,j)\\)\u003c/span\u003e\u003c/span\u003e are in the same class or a different class based on class variable \u003cem\u003ey\u003c/em\u003e. The contrastive loss can be penalized by LASSO or Ridge, or it can be unpenalized and P-values can be computed. Rather than using the predictor/attribute values directly in the regression, NPDR uses the difference \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\stackrel{⃑}{d}}_{ij}\\left(X\\right)\\)\u003c/span\u003e\u003c/span\u003e (or projected distance onto the attributes X) between subjects \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\((i,j)\\)\u003c/span\u003e\u003c/span\u003e. The vector denotes the projected distance for all attributes in the set X. In the current application, the attributes are Pearson correlations between pairs of ROIs. For centrality NPDR (c-NPDR), the projected distance or diff, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({d}_{ij}\\left(p\\right)\\)\u003c/span\u003e\u003c/span\u003e, is the absolute difference between subjects \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\((i,j)\\)\u003c/span\u003e\u003c/span\u003e for one correlation attribute \u003cem\u003ep\u003c/em\u003e (correlation between a pair of ROIs):\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$${d}_{ij}\\left(p\\right)=\\left|{A}_{p}^{\\left(i\\right)}-{A}_{p}^{\\left(j\\right)}\\right|$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e,\u003c/p\u003e \u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({A}_{p}^{\\left(i\\right)}\\)\u003c/span\u003e\u003c/span\u003eis the correlation for subject \u003cem\u003ei\u003c/em\u003e between a pair of ROIs, represented by \u003cem\u003ep\u003c/em\u003e. Thus, if there are \u003cem\u003en\u003c/em\u003e ROIs, the NPDR design matrix consists of n(n-1)/2 attribute columns; one for each pair of ROIs.\u003c/p\u003e \u003cp\u003eThe NPDR-selected ROI pairs can then be used in any number of centrality algorithms to rank the importance of individual ROIs. For comparison, we use the following centralities: degree, betweenness, eigenvector and integrated value of influence (IVI) [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe other NPDR-based method (correlation-diff-NPDR) for ranking the importance of ROIs from ROI-pair correlation data uses a more complex projected distance, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({d}_{ij}^{CD}\\)\u003c/span\u003e\u003c/span\u003e, but directly gives importance of individual ROIs without centrality calculations [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. The correlation-diff (CD) or correlation projected distance for ROI \u003cem\u003er\u003c/em\u003e is given by\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$${d}_{ij}^{CD}\\left(r\\right)={\\sum }_{k\\ne r}\\left|{A}_{rk}^{\\left(i\\right)}-{A}_{rk}^{\\left(j\\right)}\\right|$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e,\u003c/p\u003e \u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({A}_{rk}^{\\left(i\\right)}\\)\u003c/span\u003e\u003c/span\u003e is the correlation between ROIs \u003cem\u003er\u003c/em\u003e and \u003cem\u003ek\u003c/em\u003e for subject \u003cem\u003ei\u003c/em\u003e. Thus, the correlation-diff for ROI \u003cem\u003er\u003c/em\u003e is the absolute sum of differences between \u003cem\u003er\u003c/em\u003e and all other ROIs. If there are \u003cem\u003en\u003c/em\u003e ROIs, the NPDR design matrix for Eq.\u0026nbsp;(\u003cspan refid=\"Equ2\" class=\"InternalRef\"\u003e2\u003c/span\u003e) will have \u003cem\u003en\u003c/em\u003e columns as opposed to \u003cem\u003en(n-1)/2\u003c/em\u003e for Eq.\u0026nbsp;(\u003cspan refid=\"Equ1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Thus, NPDR with Eq.\u0026nbsp;(\u003cspan refid=\"Equ2\" class=\"InternalRef\"\u003e2\u003c/span\u003e) yields importance scores for ROIs, while NPDR with Eq.\u0026nbsp;(\u003cspan refid=\"Equ1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) yields the importance of ROI pairs. In both cases (Eqs.\u0026nbsp;\u003cspan refid=\"Equ1\" class=\"InternalRef\"\u003e1\u003c/span\u003e and \u003cspan refid=\"Equ2\" class=\"InternalRef\"\u003e2\u003c/span\u003e), NPDR importance can be computed in terms of individual P-values or in a multivariate model with LASSO or Ridge.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn order to threshold the results of correlation-diff-NPDR and cNPDR, we use regularization and p-values. We use the LASSO penalty, also known as the L1 penalty, which is a regularization technique used in regression models to prevent overfitting and to enhance the model's prediction accuracy and interpretability. For non-penalized methods, we use a p-value adjusted cutoff, where ROI pairs that had an adjusted p-value\u0026thinsp;\u0026gt;\u0026thinsp;0.05 are removed from the network. To threshold the random forest results, we use a cutoff of the top 100 pairs of ROIs (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eCorrelation-diff-NPDR directly yields a list of significant individual ROIs. However, the centrality methods need an additional step to map pair importance to individual importance. The c-NPDR and c-rf methods yield lists of important pairs of ROIs, so we apply centralities to the resulting edge lists to obtain a list of important individual ROIs (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). The significant pairs of ROIs are graphed as a network, where the nodes are ROIs and edges are defined when the ROI pairs have a correlation that affects the outcome variable (e.g., MDD). This interaction network is a way to visualize the importance of MDD nodes based on their connections and visualize local structure. We quantify the importance of individual ROIs using common centralities, degree, eigenvector, betweenness, and IVI [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. IVI combines multiple centrality measures.\u003c/p\u003e \u003cp\u003eWe compare the NPDR methods to a centrality version of random forest (c-rf). We use the correlation predictor data (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ed) with random forest permutation importance with 5000 trees, filter the correlation pairs to the top 200 to create a network, and then computed ROI centralities.\u003c/p\u003e \u003cp\u003e \u003cb\u003eSimulation Method and Real Data.\u003c/b\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eSimulation Approach.\u003c/b\u003e We develop a random network approach to simulate correlation-based features, a fraction of which are functional or associated with the case-control status (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). The application we have in mind is correlation between brain ROIs in resting-state fMRI studies, where correlation is calculated from the BOLD signal time-series. We do not simulate the time series, but rather directly simulate the correlations and their differences between groups. Features or predictors are correlations between pairs of ROIs rather than ROIs themselves. We note that these simulations and feature selection methods would also apply to other types of correlation-based data in other research domains.\u003c/p\u003e \u003cp\u003eThe user specifies the number of ROIs, the number of cases and controls, the number of functional ROIs (i.e., that are associated with the outcome), the effect size and the type of underlying random network for the brain. The user can specify their own network, for example from real data, or they can generate any network from the igraph library. Initial correlation matrices are generated for each sample based on the network, where connected ROIs have higher random correlations than unconnected ROIs.\u003c/p\u003e \u003cp\u003eFunctional nodes are chosen from the largest connected component (i.e. group of nodes such that there is a path between any pair of nodes in the group). Edges between the functional nodes (green edges in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e) are then used to create differential correlation between cases and controls (black dots in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e heatmaps). We use a parameter called \u0026ldquo;multiway\u0026rdquo; that controls how many edges we randomly select to generate differential correlation. For example, a multiway of 2 will use only a subset of the possible edges between functional nodes (a subset of possible edges will be green). If we set multiway to the maximum, then all possible edges between functional nodes will have differential correlation (green). We use multiway\u0026thinsp;=\u0026thinsp;5 in this application.\u003c/p\u003e \u003cp\u003eWe generate replicate simulations to compare feature selection methods based on the ability to detect the ground truth functional ROIs. We use the F1 Score to test whether the top ROI features selected by a method overlap with the top functional features.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eReal Data\u003c/b\u003e. We compare feature selection methods on data from the Tulsa 1000 (T1000), a longitudinal study at the Laureate Institute for Brain Research following 1000 individuals, including healthy individuals and those with mood and other disorders [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. We use rs-fMRI time series for 188 MDD subjects and 47 healthy controls (HC) from T1000 (163 female and 72 male). We use the Automated Anatomical Labelling Atlas (AAL Atlas) with 87 ROIs and the Brainnetome Atlas with 246 ROIs to define consistent and interpretable mappings for selected features [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. The Brainnetome Atlas parcellates the brain based on structural and connectivity features. It is based on neuroimaging data, particularly rs-fMRI and diffusion tensor imaging (DTI) data, which reveal both the functional and structural connectivity patterns in the brain. For each atlas, we detrended the signals and averaged the time-series for the voxels within an atlas ROI.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eWe simulate 50 replicate datasets each with 100 cases, 100 controls, and 100 ROIs. We select 10 functional ROIs, but their effects are detected through their correlations with each other in correlation-predictor datasets. The underlying correlation networks are based on Erdos-Renyi random networks with p\u0026thinsp;=\u0026thinsp;0.1 connection probability. We used a medium effect size of 0.5 Cohen\u0026rsquo;s d.\u003c/p\u003e \u003cp\u003eWe apply six feature selection methods to the replicate simulations (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e) and compare them based on their average ability to detect the 10 functional ROIs. Centrality-Random Forest (c-rf) with degree centrality (red, Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e) has a similar mean F1 score to correlation-diff NPDR (corr-diff) using Ridge regression (green CD Ridge, Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e), and they are both similar to centrality-NPDR (c-npdr) with degree centrality (left blue). The npdr-based methods corr-diff and c-npdr with degree have slightly less variation than c-rf. The F1 scores for centrality-based NPDR methods (all blue, Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e) depend on the centrality method used. C-NPDR works best with degree, whereas IVI, betweenness, and eigenvector centralities are noticeably worse. The close similarity between corr-diff and c-npdr (Eq.\u0026nbsp;\u003cspan refid=\"Equ1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) with degree suggests that the corr-diff NPDR difference metric (Eq.\u0026nbsp;\u003cspan refid=\"Equ2\" class=\"InternalRef\"\u003e2\u003c/span\u003e) works mathematically like degree.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eWe apply four feature selection methods to the real rs-fMRI study to compare the selected important ROIs for MDD (Table\u0026nbsp;1). The correlation-diff NPDR with LASSO selects the fewest features because it has a tendency to eliminate correlated features (first column, Table\u0026nbsp;1). The other correlation-diff NPDR (second column, Table\u0026nbsp;1) uses an adjusted p-value cutoff, and it results in more selected features in part due to inclusion of more correlated features. The centrality methods with degree, NPDR (column 3) and random forest (column four), use a manual threshold because degree centrality does not have a statistical threshold.\u003c/p\u003e \u003cp\u003eFor the Brainnetome atlas (top, Table\u0026nbsp;1), NPDR correlation-diff LASSO (column 1) selects the most parsimonious list of ROIs, and the these features are included in the longer lists of NPDR methods (columns 2 and 3). The random forest method has a slightly different set of selected ROIs because it is not metric based and it tends to find more main effects than interactions compared to NPDR methods. We highlight the selected ROIs involving MTG (middle temporal gyrus) because it is the top ROI found by NPDR correlation-diff and has been associated with MDD [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e], [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eFor the AAL atlas (bottom section, Table\u0026nbsp;1), the three NPDR methods have a consensus of ROIs selected by the LASSO method. These regions include the dorsal and ventral default mode networks (DMN) and the right executive control network (ECN). The random forest centrality method includes multiple blocks of correlated variables (bottom, Table\u0026nbsp;1) in regions like Anterior Salience, Auditory, and dorsal DMN. The NPDR methods have reduced correlation, and the LASSO version automatically selects a parsimonious set of ROIs. The highest NPDR degree nodes for AAL (Dorsal_DMN_ 03 and Ventral_DMN_07) are in the same Louvain graph cluster and the third highest degree node (Right_ECN_04) is in a different cluster (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e).\u003c/p\u003e "},{"header":"Discussion","content":"\u003cp\u003eThe middle temporal gyrus (MTG) was found by all feature selection methods to be important for MDD (Table\u0026nbsp;1). The centrality random forest method only identified the left MTG while the LASSO correlation-diff NPDR identified only the right MTG. The other NPDR methods, including the centrality based method, identified both left and right MTG. MTG is critical for semantic memory processing, visual perception, and language processing [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e], and studies have found associations with MDD. For example, studies using structural and functional MRI identified significant gray matter abnormalities in the right MTG among patients with treatment-resistant depression (TRD) and treatment-responsive depression (TSD) compared to healthy controls [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. The reduced gray matter volume in the bilateral MTG is indicative of structural changes associated with MDD. Similarly, a previous study found fractional amplitude of low-frequency fluctuation (fALFF) was higher in patients with MDD in the right and left MTG [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAll tested methods also identified STG (superior temporal gyrus) and ITG (inferior temporal gyrus) as important for MDD. In patients with anxious depression versus healthy controls, a previous study found increased fALFF values in the left STG [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]. While not previously linked to MDD, ITG showed gray matter volume reductions in the MTG and ITG in chronic schizophrenia patients [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. The link to another psychiatric condition could indicate a broader role for ITG in mood disorders.\u003c/p\u003e"},{"header":"Conclusions","content":"\u003cp\u003eThe application of machine learning and feature selection algorithms to fMRI data is increasingly critical toward understanding biological mechanisms of disorders. New methods are needed that can account for interactions between variables and regions of interest. NPDR has been shown previously to be able to detect interactions, and we extended it to handle data where the predictors are correlations between pairs of ROIs. We also developed a new simulation approach to compare these methods. We found that c-NPDR with degree is similar to the correlation-diff NPDR. This suggests a mathematical connection between the correlation-diff metric and degree centrality, and it suggests ways that correlation-diff might be improved by using other centralities. We applied these NPDR methods and a random forest approach to correlation data from a real rs-fMRI dataset for MDD, and the consensus between these methods found MTG, ITG, and STG to be important MDD ROIs.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cdiv class=\"DefinitionList\"\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003ers-fMRI\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eresting-state functional magnetic resonance imaging\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eMDD\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003emajor depressive disorder\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eHC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003ehealthy control\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eROI\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eregion of interest\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eEK and BM wrote the main manuscript. EK implemented the centrality approach and performed data analysis. BD developed the corr-diff approach and simulation method. HL performed simulation analysis. JL and RK contributed to analysis design. All authors reviewed the manuscript. BM led the overall design.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eVan Den Heuvel MP, Hulshoff Pol HE. Exploring the brain network: A review on resting-state fMRI functional connectivity, European Neuropsychopharmacology, vol. 20, no. 8, pp. 519\u0026ndash;534, Aug. 2010, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.euroneuro.2010.03.008\u003c/span\u003e\u003cspan address=\"10.1016/j.euroneuro.2010.03.008\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSalvador R, Suckling J, Coleman MR, Pickard JD, Menon D, Bullmore E. Neurophysiological Architecture of Functional Magnetic Resonance Images of Human Brain, Cerebral Cortex, vol. 15, no. 9, pp. 1332\u0026ndash;1342, Sep. 2005, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/cercor/bhi016\u003c/span\u003e\u003cspan address=\"10.1093/cercor/bhi016\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIraji A, et al. The connectivity domain: Analyzing resting state fMRI data using feature-based data-driven and model-based methods. NeuroImage. Jul. 2016;134:494\u0026ndash;507. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.neuroimage.2016.04.006\u003c/span\u003e\u003cspan address=\"10.1016/j.neuroimage.2016.04.006\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePreti MG, Bolton TA, Van De Ville D. The dynamic functional connectome: State-of-the-art and perspectives, NeuroImage, vol. 160, pp. 41\u0026ndash;54, Oct. 2017, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.neuroimage.2016.12.061\u003c/span\u003e\u003cspan address=\"10.1016/j.neuroimage.2016.12.061\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBiswal BB et al. Mar., Toward discovery science of human brain function, Proc. Natl. Acad. Sci. U.S.A., vol. 107, no. 10, pp. 4734\u0026ndash;4739, 2010, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1073/pnas.0911855107\u003c/span\u003e\u003cspan address=\"10.1073/pnas.0911855107\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSmith SM et al. Jan., Network modelling methods for FMRI, NeuroImage, vol. 54, no. 2, pp. 875\u0026ndash;891, 2011, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.neuroimage.2010.08.063\u003c/span\u003e\u003cspan address=\"10.1016/j.neuroimage.2010.08.063\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVan Dijk KRA, Sabuncu MR, Buckner RL. The influence of head motion on intrinsic functional connectivity MRI. NeuroImage. Jan. 2012;59(1):431\u0026ndash;8. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.neuroimage.2011.07.044\u003c/span\u003e\u003cspan address=\"10.1016/j.neuroimage.2011.07.044\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZuo X-N, Xing X-X. Test-retest reliabilities of resting-state FMRI measurements in human brain functional connectomics: A systems neuroscience perspective, Neuroscience \u0026amp; Biobehavioral Reviews, vol. 45, pp. 100\u0026ndash;118, Sep. 2014, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.neubiorev.2014.05.009\u003c/span\u003e\u003cspan address=\"10.1016/j.neubiorev.2014.05.009\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFornito A, Zalesky A, Breakspear M. Graph analysis of the human connectome: Promise, progress, and pitfalls, NeuroImage, vol. 80, pp. 426\u0026ndash;444, Oct. 2013, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.neuroimage.2013.04.087\u003c/span\u003e\u003cspan address=\"10.1016/j.neuroimage.2013.04.087\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFan L, et al. The Human Brainnetome Atlas: A New Brain Atlas Based on Connectional Architecture. Cereb Cortex. Aug. 2016;26(8):3508\u0026ndash;26. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/cercor/bhw157\u003c/span\u003e\u003cspan address=\"10.1093/cercor/bhw157\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhuo J, Fan L, Liu Y, Zhang Y, Yu C, Jiang T. Connectivity Profiles Reveal a Transition Subarea in the Parahippocampal Region That Integrates the Anterior Temporal\u0026ndash;Posterior Medial Systems, J. Neurosci., vol. 36, no. 9, pp. 2782\u0026ndash;2795, Mar. 2016, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1523/JNEUROSCI.1975-15.2016\u003c/span\u003e\u003cspan address=\"10.1523/JNEUROSCI.1975-15.2016\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAl-Zubaidi A, Mertins A, Heldmann M, Jauch-Chara K, M\u0026uuml;nte TF. Machine Learning Based Classification of Resting-State fMRI Features Exemplified by Metabolic State (Hunger/Satiety), Frontiers in Human Neuroscience, vol. 13, 2019, Accessed: Feb. 05, 2024. [Online]. Available: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.frontiersin.org/articles/\u003c/span\u003e\u003cspan address=\"https://www.frontiersin.org/articles/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fnhum.2019.00164\u003c/span\u003e\u003cspan address=\"10.3389/fnhum.2019.00164\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e\u0026Aacute;lvarez JD, Matias-Guiu JA, Cabrera-Mart\u0026iacute;n MN, Risco-Mart\u0026iacute;n JL, Ayala JL. An application of machine learning with feature selection to improve diagnosis and classification of neurodegenerative disorders, BMC Bioinformatics, vol. 20, no. 1, p. 491, Oct. 2019, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s12859-019-3027-7\u003c/span\u003e\u003cspan address=\"10.1186/s12859-019-3027-7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShi C, Zhang J, Wu X. An fMRI Feature Selection Method Based on a Minimum Spanning Tree for Identifying Patients with Autism, Symmetry, vol. 12, no. 12, p. 1995, Dec. 2020, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/sym12121995\u003c/span\u003e\u003cspan address=\"10.3390/sym12121995\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLe TT, Dawkins BA, McKinney BA. Nearest-neighbor Projected-Distance Regression (NPDR) for detecting network interactions with adjustments for multiple tests and confounding. Bioinformatics. May 2020;36(9):2770\u0026ndash;7. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/bioinformatics/btaa024\u003c/span\u003e\u003cspan address=\"10.1093/bioinformatics/btaa024\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBreiman L. Random Forests Mach Learn. 2001;45(1):5\u0026ndash;32. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1023/A:1010933404324\u003c/span\u003e\u003cspan address=\"10.1023/A:1010933404324\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLoh D. Middle temporal gyrus | Radiology Reference Article | Radiopaedia.org, Radiopaedia. Accessed: Feb. 05, 2024. [Online]. Available: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://radiopaedia.org/articles/middle-temporal-gyrus\u003c/span\u003e\u003cspan address=\"https://radiopaedia.org/articles/middle-temporal-gyrus\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLoh D. Inferior temporal gyrus | Radiology Reference Article | Radiopaedia.org, Radiopaedia. Accessed: Feb. 05, 2024. [Online]. Available: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://radiopaedia.org/articles/inferior-temporal-gyrus\u003c/span\u003e\u003cspan address=\"https://radiopaedia.org/articles/inferior-temporal-gyrus\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGaillard F. Entorhinal cortex | Radiology Reference Article | Radiopaedia.org, Radiopaedia. Accessed: Feb. 05, 2024. [Online]. Available: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://radiopaedia.org/articles/entorhinal-cortex\u003c/span\u003e\u003cspan address=\"https://radiopaedia.org/articles/entorhinal-cortex\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMa C, et al. Resting-State Functional Connectivity Bias of Middle Temporal Gyrus and Caudate with Altered Gray Matter Volume in Major Depression. PLoS ONE. Sep. 2012;7(9):e45263. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1371/journal.pone.0045263\u003c/span\u003e\u003cspan address=\"10.1371/journal.pone.0045263\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOnitsuka T, et al. Middle and Inferior Temporal Gyrus Gray Matter Volume Abnormalities in Chronic Schizophrenia: An MRI Study. AJP. Sep. 2004;161(9):1603\u0026ndash;11. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1176/appi.ajp.161.9.1603\u003c/span\u003e\u003cspan address=\"10.1176/appi.ajp.161.9.1603\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim IB, Park S-C. Neural Circuitry\u0026ndash;Neurogenesis Coupling Model of Depression. IJMS. Feb. 2021;22(5):2468. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/ijms22052468\u003c/span\u003e\u003cspan address=\"10.3390/ijms22052468\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDu Y, Fu Z, Calhoun VD. Classification and Prediction of Brain Disorders Using Functional Connectivity: Promising but Challenging. Front Neurosci. Aug. 2018;12:525. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fnins.2018.00525\u003c/span\u003e\u003cspan address=\"10.3389/fnins.2018.00525\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen Y, Zhao W, Yi S, Liu J. The diagnostic performance of machine learning based on resting-state functional magnetic resonance imaging data for major depressive disorders: a systematic review and meta-analysis. Front Neurosci. Sep. 2023;17:1174080. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fnins.2023.1174080\u003c/span\u003e\u003cspan address=\"10.3389/fnins.2023.1174080\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCalhoun VD, de Lacy N. Neuroimaging Clin N Am. Nov. 2017;27(4):561. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.nic.2017.06.012\u003c/span\u003e\u003cspan address=\"10.1016/j.nic.2017.06.012\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Ten Key Observations on the Analysis of Resting-state Functional MR Imaging Data Using Independent Component Analysis.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJoel SE, Caffo BS, van Zijl PC, Pekar JJ. On the relationship between seed-based and ICA-based measures of functional connectivity, Magn Reson Med, vol. 66, no. 3, pp. 644\u0026ndash;657, Sep. 2011, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1002/mrm.22818\u003c/span\u003e\u003cspan address=\"10.1002/mrm.22818\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDawkins BA, Le TT, McKinney BA. Theoretical properties of distance distributions and novel metrics for nearest-neighbor feature selection. PLoS ONE. Feb. 2021;16(2):e0246761. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1371/journal.pone.0246761\u003c/span\u003e\u003cspan address=\"10.1371/journal.pone.0246761\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVictor TA, et al. Tulsa 1000: a naturalistic study protocol for multilevel assessment and outcome prediction in a large psychiatric sample. BMJ Open. Jan. 2018;8(1):e016620. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1136/bmjopen-2017-016620\u003c/span\u003e\u003cspan address=\"10.1136/bmjopen-2017-016620\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRolls ET, Huang C-C, Lin C-P, Feng J, Joliot M. Automated anatomical labelling atlas 3. NeuroImage. Feb. 2020;206:116189. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.neuroimage.2019.116189\u003c/span\u003e\u003cspan address=\"10.1016/j.neuroimage.2019.116189\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSalavaty A, Ramialison M, Currie PD. Integrated Value of Influence: An Integrative Method for the Identification of the Most Influential Nodes within Networks. Patterns. Aug. 2020;1(5):100052. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.patter.2020.100052\u003c/span\u003e\u003cspan address=\"10.1016/j.patter.2020.100052\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang Q, et al. Associations between abnormal spontaneous neural activity and clinical variables, eye movements, and event-related potential indicators in major depressive disorder. Front Neurosci. Jan. 2023;16:1056868. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fnins.2022.1056868\u003c/span\u003e\u003cspan address=\"10.3389/fnins.2022.1056868\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhao P, et al. Altered fractional amplitude of low-frequency fluctuations in the superior temporal gyrus: a resting-state fMRI study in anxious depression. BMC Psychiatry. Nov. 2023;23(1):847. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s12888-023-05364-w\u003c/span\u003e\u003cspan address=\"10.1186/s12888-023-05364-w\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003eTable 1 is available in the Supplementary Files section.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":true,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Resting-state fMRI, statistical interactions, machine learning feature selection","lastPublishedDoi":"10.21203/rs.3.rs-4193488/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4193488/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground.\u003c/h2\u003e \u003cp\u003eNearest-neighbor projected-distance regression (NPDR) is a metric-based machine learning feature selection algorithm that uses distances between samples and projected differences between variables to identify variables or features that may interact to affect the prediction of complex outcomes. Typical bioinformatics data consist of separate variables of interest like genes or proteins. In contrast, resting-state functional MRI (rs-fMRI) data is composed of time-series for brain Regions of Interest (ROIs) for each subject, and these within-brain time-series are typically transformed into correlations between pairs of ROIs. These pairs of variables of interest can then be used as input for feature selection or other machine learning. Straightforward feature selection would return the most significant pairs of ROIs; however, it would also be beneficial to know the importance of individual ROIs.\u003c/p\u003e\u003ch2\u003eResults.\u003c/h2\u003e \u003cp\u003eWe extend NPDR to compute the importance of individual ROIs from correlation-based features. We present correlation-difference and centrality-based versions of NPDR. The centrality-based NPDR can be coupled with any centrality method and can be coupled with importance scores other than NPDR, such as random forest importance. We develop a new simulation method using random network theory to generate artificial correlation data predictors with variation in correlation that affects class prediction.\u003c/p\u003e\u003ch2\u003eConclusions.\u003c/h2\u003e \u003cp\u003eWe compare feature selection methods based on detecting functional simulated ROIs, and we apply the new centrality NPDR approach to a resting-state fMRI study of major depressive disorder (MDD) and healthy controls. We determine that the areas of the brain that are the most interactive in MDD patients include the middle temporal gyrus, the inferior temporal gyrus, and the dorsal entorhinal cortex. The resulting feature selection and simulation approaches can be applied to other domains that use correlation-based features.\u003c/p\u003e","manuscriptTitle":"Centrality-based nearest-neighbor projected-distance regression (C-NPDR) feature selection for correlation predictors with application to resting-state fMRI of major depressive disorder","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-04-01 17:11:32","doi":"10.21203/rs.3.rs-4193488/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"b2f79e32-573b-4a43-aa11-b0196245ae58","owner":[],"postedDate":"April 1st, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2024-04-17T22:29:27+00:00","versionOfRecord":[],"versionCreatedAt":"2024-04-01 17:11:32","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-4193488","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4193488","identity":"rs-4193488","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}