 Research
 Open access
 Published:
Individualspecific networks for prediction modelling – A scoping review of methods
BMC Medical Research Methodology volume 22, Article number: 62 (2022)
Abstract
Background
Recent advances in biotechnology enable the acquisition of highdimensional data on individuals, posing challenges for prediction models which traditionally use covariates such as clinical patient characteristics. Alternative forms of covariate representations for the features derived from these modern data modalities should be considered that can utilize their intrinsic interconnection. The connectivity information between these features can be represented as an individualspecific network defined by a set of nodes and edges, the strength of which can vary from individual to individual. Global or local graphtheoretical features describing the network may constitute potential prognostic biomarkers instead of or in addition to traditional covariates and may replace the often unsuccessful search for individual biomarkers in a highdimensional predictor space.
Methods
We conducted a scoping review to identify, collate and critically appraise the stateofart in the use of individualspecific networks for prediction modelling in medicine and applied health research, published during 2000–2020 in the electronic databases PubMed, Scopus and Embase.
Results
Our scoping review revealed the main application areas namely neurology and pathopsychology, followed by cancer research, cardiology and pathology (N = 148). Network construction was mainly based on Pearson correlation coefficients of repeated measurements, but also alternative approaches (e.g. partial correlation, visibility graphs) were found. For covariates measured only once per individual, network construction was mostly based on quantifying an individual’s contribution to the overall grouplevel structure. Despite the multitude of identified methodological approaches for individualspecific network inference, the number of studies that were intended to enable the prediction of clinical outcomes for future individuals was quite limited, and most of the models served as proof of concept that network characteristics can in principle be useful for prediction.
Conclusion
The current body of research clearly demonstrates the value of individualspecific network analysis for prediction modelling, but it has not yet been considered as a general tool outside the current areas of application. More methodological research is still needed on wellfounded strategies for network inference, especially on adequate network sparsification and outcomeguided graphtheoretical feature extraction and selection, and on how networks can be exploited efficiently for prediction modelling.
Introduction
Prediction modelling is essential to an individualized approach to risk assessment, diagnosis, prognosis, and medical decisionmaking. While the conventional approach of model development is mainly based on a small set of patient characteristics, recent developments in biotechnology (e.g. highresolution imaging modalities and highthroughput sequencing methods) have accelerated the generation of individualspecific data at an unprecedented level generally characterized by a highdimensional variable space for each individual and complex correlation structures between the variables. Common methods of prediction modelling are not well suited to deal with such complex data structures in particular in small to moderately sized studies [1,2,3]. Hence the question arises how the abundance of biological information available per individual can be used most efficiently to provide accurate predictions of health outcomes.
Increasingly, it becomes possible to represent individual patient data as individualspecific networks that, apart from individualspecific node measurements, allow to capture connectivity information between variables (nodes) via their edges. Individualspecific networks are exemplified in Fig. 1 for a hypothetical small study cohort. The absence/presence of an edge, or its weight (strength of the connectivity) can be the same for all individuals in the sample (e.g. based on a reference or on sample estimates of a statistical measure of connectivity), or can vary from individual to individual. For example, the individuals depicted in Fig. 1 share the same network structure but are heterogeneous with respect to the edge weights. The network representation is further motivated by subject matter, since complex diseases of the human body are rarely caused by the malfunction of individual molecules but rather by the disruption or dysfunction of the underlying system behaviour or a specific set of biological units [4]. Graphtheoretical features can then capture the heterogeneous variability of system patterns across individuals by describing individualspecific structural and topological network patterns or identify biological modules, a set of nodes acting as key drivers of disease manifestation. For the sake of clarity, variables that originate from individualspecific networks will be referred to as graphtheoretical features throughout this work in order to distinguish them from classic clinical variables. Graphtheoretical features condense a highdimensional predictor space into few quantitative and interpretable descriptors that can be used in prediction models instead of or in addition to classical clinical predictors in order to improve such models. Such an approach may lead to new insights into disease development or progression and may even replace the often unsuccessful search for individual prognostic biomarkers in the highdimensional space.
In the past decades, the call for a complex systembased understanding of human disease mechanisms has led to a general theory and various application frameworks of grouplevel networks, i.e. network inference based on the aggregated study cohort. For example, see Barabási et al. [4] for an extensive review on ‘network medicine’ and Li et al. [5] for an overview of graph representation learning in biology and medicine. However, a “one size fits all” approach for network inference may wash out the individualspecific systems behaviour. Linking individualspecific patterns of connections rather than grouplevel system behaviour across biological components to disease manifestation can not only capture the heterogeneity of biological system behaviour across individuals but also pave the way for the detection of novel biomarkers for more individualized prediction.
To our knowledge, the potential of using graphtheoretical features of individualspecific networks as predictors in clinical prediction models has not yet been systematically explored and hence the stateoftheart in this relatively new field of predictive research remains unknown. Therefore, we conducted a scoping review [6, 7] to systematically examine the scientific literature to identify, collate and critically appraise methodological approaches incorporating individualspecific networks to improve and advance prediction modelling of clinical outcomes in applied health research. We did not consider studies exclusively employing grouplevel networks and those which do not aim at individual outcome prediction.
The remainder of this paper is organized as follows. In Section 2, we discuss the methodology of the scoping review. The general study characteristics and findings of the search strategy are presented in Section 3. Further, we present and collate the identified approaches to individualspecific network inference, graphtheoretical feature extraction and their usage for prediction in Section 4 and conclude with current challenges and future aspects of the identified methodology in Section 5.
Methods
Search strategy
We conducted a scoping literature search in the three electronic databases PubMed, Embase and Scopus to extract peerreviewed articles published between January 1st, 2000 and August 31st, 2020. The search strategy consisted of three sets of terms to cover the research intersection of network analysis and prediction modelling adequately but also to reduce false positive hits by the broad meaning of the term ‘network’. The three sets of terms were: 1) terms associated with network analysis, 2) terms associated with predictive research and 3) exclusion terms (see the Supplementary Material for a detailed overview).
Studies met the inclusion criteria if graphtheoretical features derived from an individualspecific network were considered as candidate predictors in prediction modelling in the medical field. Studies were excluded if they did not focus on prediction modelling of health outcomes and did not consider networks constructed for single individuals. Therefore, we excluded studies that concentrated solely on the descriptive analysis of individualspecific networks without examining their potential association with a clinical outcome or studies considering grouplevel networks computed from aggregated data. Studies that essentially aimed at predicting network behaviour, link prediction or changes in network topology or structure were also discarded.
Selection of studies
All studies identified by the search strategy were initially screened based on the title and then, after inclusion, based on the abstract to determine eligibility. Letters, commentaries and conference abstracts were excluded. Selected articles were then subjected to a fulltext analysis. The first author was responsible for the initial search, application of the exclusion criteria, screening of all the identified articles and the quality evaluation of the included papers. A random subset of studies (consisting of 250, 50 and 25 studies in the title, abstract and fulltext screening phases, respectively) was independently assessed by three additional reviewers (GH, FM, MS) to ensure general validity and reliability of the screening process and data extraction of the first reviewer. Any inconsistencies in selection among the reviewers were discussed and resolved to reach a general consensus., We refer the reader to the Supplementary Material for a more detailed summary of the search strategy and the extraction process. Reporting adhered to the PRISMAScR guidelines [7] to ensure methodological transparency.
Results
Search results and study characteristics
A total of 4988 studies was initially retrieved from the electronic database search together with the manual selection from other sources. After the screening of the titles, 488 articles remained and after reviewing the abstracts, only 227 articles met the eligibility criteria for fulltext analysis of which 79 were excluded due to the following reasons: (1) construction of nonindividualspecific networks (N = 36), (2) the term “network” was used in a different context (N = 17), (3) no association with an outcome of interest was considered (N = 17) or (4) graphtheoretical features were used as dependent variables (N = 9). This left 148 studies (3.0%) out of the initial 4988 studies meeting the eligibility criteria of the review (see Fig. 2).
Through a synthesis of the sources of evidence, four medical domains of application were identified. The majority of the eligible studies (N = 129, 87.2%) covered neurological research, followed by the fields of psychopathology (N = 9, 6.1%), genomics (N = 7, 4.7%), cardiology (N = 2, 1.4%) and pathology (N = 1, 0.7%). The oldest study meeting our inclusion criteria was published in 2009, with the number of studies published annually increasing steadily thereafter (see Supplementary Fig. S1). Most of the studies (N = 100, 67.6%) were published after 2015. Out of the 148 included studies, 135 (91.2%) were identified as quantitative studies and 13 (8.8%) as qualitative studies (e.g. reviews). Besides the process of data acquisition and preparation in applied health research, three main topics emerged in the included articles covering the intersection of network analysis and prediction and were addressed separately in the following subsections: Section 3.2 focuses on datadriven network inference consisting of the individualspecific network construction and network sparsification, Section 3.3 on the extraction of graphtheoretical features and Section 3.4 on predictive analytics using graphtheoretical features. Main modules in a general workflow of considering individualspecific networks in prediction modelling and some aspects of their implementation are illustrated in Fig. 3. However, a detailed presentation of each of the identified analytics for network analysis and prediction modelling as well as their theoretical properties is beyond the scope of this article. Instead, the reader will find useful references throughout this work.
Network construction and sparsification
Notation and concepts
In general, an undirected network (or graph) consists of a pair G = (V, E) where V denotes a finite, nonempty set of p nodes and E is a subset of V × V containing pairs of connected nodes e_{ij} ≔ (v_{i}, v_{j}) referred to as edges. In directed graphs (digraph), each edge has a direction such that e_{ij} ≠ e_{ji}. In weighted networks, each edge e_{ij} is associated with a weight w_{ij} ≔ w(v_{i}, v_{j}) ∈ ℝ. A subnetwork G^{′} = (V^{′}, E^{′}) is a network such that V^{′} ⊆ V and E ′ ⊆ E. The data structure defining a network is the adjacency matrix A = [a_{ij}] in which a_{ij} = 1 indicates the presence of an edge between v_{i} and v_{j}, while a_{ij} = 0 indicates its absence. For weighted networks a_{ij} = w_{ij}, and again a_{ij} = 0 indicates the absence of the respective edge. For individualspecific networks, we assume that for each individual s (s = 1, …, N) a unique network G_{s} = (V_{s}, E_{s}) exists, where N is the number of individuals within the study cohort .
Global graphtheoretical features characterize properties of the entire network, while local features only take the information of a smaller substructure of the network (e.g. node, module) into account. For example, edge density is a global graphtheoretical feature defined as the ratio of the number of actual connections to the number of all possible connections in the network
See Table 1 for an overview of some global and local graphtheoretical features identified in the reviewed studies.
Construction of individualspecific networks
Repeated measurements per variable per individual
Depending on the field of research, the nodes mainly represented regions of interest (ROIs) in the brain, genes or psychotic symptoms (e.g. stress, insomnia). In two individual proofofconcept studies, the nodes corresponded to 5min heart rate variability (HRV) segments [8] or ROI in muscle biopsy images [9]. A variety of methods were identified for defining connectivity between pairs of nodes based on datadriven structural learning, i.e., estimating the graphical structure from the data (see Fig. 4 for a schematic illustration).
The majority of identified studies conducted correlationbased approaches (see Fig. 4A) using repeatedly measured continuous data (e.g. sequential data, time series) to define edges between the prespecified nodes [10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27]. In neurological applications, adjacency matrices most frequently consisted of correlation coefficients (N = 69, P = 48.9%) (e.g., Pearson product moment correlation coefficient) of time series of brain activity (e.g. blood oxygenation level dependent (BOLD) signal) or longitudinal cortical thickness between pairs of ROIs [16]. Fisher’s rtoz transformation was applied to Pearson correlation coefficients to achieve approximate normality [14, 21, 26, 28, 29]. However, correlationbased inference of the network structure can be subject to interfering effects (e.g. outliers, noise) and can only capture pairwise information. Thus, variations of the Pearson correlation such as spatial smoothing have been proposed in which the time series corresponding to a region is obtained by a linear mixture of neighbouring time series or segmentation of the BOLD time series into subdivisions (e.g. snapshot graphs, slidingwindow approach) [10, 30,31,32]. The sliding window or snapshot graph approach only consider small time windows of the full time series yielding a set of graphs G_{s} = {G_{s1}, …, G_{st}} for individual s and t time windows. Then, a presumably more robust final version of the individualspecific network can be obtained by assessing the frequency of appearance of each edge in each G_{sj} for j = 1, . . , t because only a few of these snapshot graphs G_{sj} are influenced by disruptions of the time series by noise artefacts [10]. Some studies also employed several modifications of the standard Pearson correlation coefficient to define connectivity [33,34,35,36,37].
An extension to the bivariate determination of connectivity were partial correlationbased networks [38, 39]. Partial correlation coefficients describe the correlation between two variables that cannot be explained by associations of the variable pair with other variables. In the case of nonnormality of the data, a transformation can be applied prior. Another approach to control for covariates while accounting for temporal information is vector autoregressive (VAR) modelling that regresses the dependent variable measured at time point t on the lagged dependent variable and the predictors evaluated at time point t − 1. VAR was only employed in psychological studies [40, 41]. This way directed individualspecific networks were obtained such that an edge weight w_{ij} associated with a directed link from node v_{i} to node v_{j} corresponds to the respective VAR coefficient in the model. The VAR model and resulting directed networks implied that in general w_{ij} ≠ w_{ji}. Only 2 out of the 135 quantitative studies considered directed networks estimated using VAR models [40, 41].
A recently introduced networkbased representation referred to as a visibility graph allows to derive individualspecific networks from time series data if available per patient in the cohort [42]. A time series with n sequentially ordered data points, {x_{t}}_{t = 1, …, n} is transformed into a network in which each time point t represents a node connected to time point s if the visibility criterion
holds true for an additional data point x_{u} placed between them. Applications of visibility graphs were mostly found in context of EEG data [43,44,45] but also in a cardiologic study dealing with human heart rate variability time series to differentiate between wake/sleep stages [46] and patients of spinal cord injury [47]. A more thorough overview of concepts and algorithms to map time series data into networks can be found in Silva et al. [48].
Alternatively, various distributional similarity approaches were used to evaluate the similarity between two distributions corresponding to a pair of nodes (JensenShannon divergence [49], KullbackLeibler divergence [50], dynamic time warping [8], generalized measure of association [51]). Zhang et al. [49] employed a kernelbased on the JensenShannon divergence to measure the similarity of multivariate time series, whereas Dong et al. [8] assessed similarity between pairs of time series based on dynamic time warping.
Single measurement per variable per individual
The approaches presented so far are only applicable if repeated data points of all variables per individual are available. In the identified studies in which all independent variables were measured only once for an individual, individualspecific network inference was mainly based on quantifying each individual’s contribution to the overall grouplevel structure either by leaveoneout network construction (LOONC) [52, 53] or by a differential perturbation approach [54,55,56,57,58,59]. Strictly speaking, these two approaches do not involve individuallevel parameter inference per se but aim to derive individualspecific networks from grouplevel networks regardless of the grouplevel network inference procedure. More precisely, LOONC removes a single individual from grouplevel network construction and measures the degree of change in all edge weights caused by the removal against the full grouplevel network. Kuijjer et al. [52] proposed to reversely engineer individualspecific networks by linear interpolation between each edge weight \({w}_{ij}^{\left({C}_{int}\right)}\) of the grouplevel network (using all \({N}_{C_{int}}\) observations in the study cohort of interest C_{int}) and the corresponding edge weight \({w}_{ij}^{\left({C}_{int}\backslash q\right)}\) in the network using all observations except an individual of interest q to obtain the edge weight estimate \({w}_{ij}^{(q)}\) for a single individual by
where N can be a generic or individualspecific weight but is usually set to \({N}_{C_{int}}\). In contrast, differential perturbation analysis quantifies the extent to which an edge weight is perturbed by the addition of an individual compared to a reference network constructed independently from the cohort of interest. Edge weights of the individualspecific network G_{q} for individual q are determined by computing the absolute difference in edge weights between the network computed from the reference population augmented by that individual (augmented network) and the reference network as
where \({w}_{ij}^{C_{ref}}\) denotes the edge weights in the reference network and \({w}_{ij}^{\left({C}_{ref}\cup \mathrm{q}\right)}\) in the augmented network.
Neither of the two presented approaches to individualspecific network construction for single measurements of each variable per individual relies on a particular inference method for the grouplevel networks. Equation (2) and (3) require the edge weights of the estimated grouplevel networks but do not specify how these networks have to be inferred. Perturbation analysis in contrast to LOONC, however, depends on an additional reference network to obtain individualspecific edge weights.
In the study by Zhu et al. [54], a grouplevel reference network for the differential perturbation analysis was inferred using Pearson correlation coefficients of gene coexpression data of \({N}_{C_{ref}}\) cancerfree individuals of a reference cohort C_{ref}. An augmented network for an individual with cancer q was constructed by including the expression data of that individual in the reference cohort (C_{ref} ∪ q). The approach was also found with networks constructed using partial correlation [57]. In Kuijjer et al. [52], LOONC was implemented using grouplevel networks derived from Pearson correlation and mutual information, a nonlinear measure of association, while LopesRamos et al. [53] used biologically motivated regulatory networks. Note that since both equations quantify the inferred difference in edge weighting, the obtained individualspecific weights can also assume values outside the boundaries of the underlying statistic of the grouplevel networks. For example, edge weights in the individualspecific networks derived from LOONC or differential perturbation using Pearson correlation can yield values <− 1 and > 1 (see Eq. (2–3)). Both approaches were only found in cancer research [57, 59].
Only one of the identified studies using single measurements per variable did not conduct LOONC or perturbation analysis. Xie et al. [39] proposed conditional Gaussian graphical modelling with mean and covariance matrix depending on an individual’s covariates while assuming homogeneous network structure across individuals in order to infer individualspecific networks. Hence, differences in edge weights between individuals are covariatedependent.
Domainspecific techniques
In neurological application, certain highresolution imaging methods for data acquisition generate domainspecific data structures that require a specifically designed methodology of network inference. For instance, tractography detects fibre pathways linking different anatomical brain regions [60,61,62,63,64,65,66,67,68]. Electroencephalographic (EEG)based connectivity was predominantly assessed by metrics such as phase lag index, coherencebased similarity, or synchronization likelihood index [69,70,71,72]. The interested reader is referred to the references for more details on these applications.
Techniques for network sparsification
Network sparsification removes edges from a network with the intent to optimise the inference of the ‘true’ system by the omission of “spurious” edges, improve interpretability and enable computational feasibility in construction and further processing of the network. In addition, the structural and topological properties of the network might be improved by removing seemingly spurious connections i.e. edges that are erroneously generated during network inference. Given an undirected network G = (V, E), sparsification yields a subnetwork G ’ = (V’, E’) of G with fewer edges such that E > ∣ E^{′}∣.
More than half of the quantitative studies employed network sparsification (N = 74, 54.8%), followed by studies analysing nonsparsified networks (N = 58, 43.0%) and 3 studies that considered both approaches (2.2%). As illustrated in Fig. 4B, the most popular strategy was proportional thresholding (N = 37) in which a common sparsity threshold (e.g. in terms of density) is defined for all individualspecific networks, and edges are removed sequentially in each network in ascending order of their edge weights until the prespecified sparsity threshold is reached [10, 21, 23,24,25, 29, 58, 73]. The second most common approach constituted weightbased thresholding (N = 31) in which one common weight threshold is defined for all individualspecific networks such that all edge weights within an individualspecific network falling below (or above) the threshold are removed (Fig. 5) [8, 11, 67, 68, 74, 75]. The majority of the weightbased approaches employed binarization of all individualspecific networks (N = 21) according to the selected weight threshold τ such that \({w}_{ij}^{\prime }=1\) if w_{ij} > τ and 0 otherwise [23, 27, 29, 75, 76]. Edges with negative weights were often removed due to their questionable interpretability or the inability to compute some graphtheoretical features from them (e.g. clustering coefficient or characteristic path length), omitting potentially meaningful inverse relation between nodes. In N = 6 studies, a sparse representation of a partial correlationbased network was achieved individually by including an L_{q}norm (q ∈ {1, 2}) regularization penalty to the network inference process in order to reinforce strong connections across individuals and drive weak connections towards zero [36, 37, 39, 77, 78]. Due to the individually imposed regularization, intersubject variability might then be inherently induced according to Wee et al. [36]. Hence, they imposed an additional group constraint to encourage a common network topology across the study cohort. Other studies selected a fixed penalty term for all individual networks which is why this approach can also be seen as a special case of weightbased thresholding [34, 37]. Grouplevel edge elimination for each individualspecific network was carried out by univariate testing of the edge weights (N = 17), either by permutation testing to control the probability of including spurious connections at 0.05 or by only including edges with weights significantly different from zero [13, 15,16,17,18,19, 79,80,81]. Then all those edges that were characterized as spurious at the group level were removed from all individualspecific networks.
The implemented strategy of network sparsification across studies varied substantially in terms of: the sparsification method, the number or combination of investigated methods, the selected sparsification threshold(s) and the reasoning behind the chosen strategy. It was not uncommon for studies to combine several sparsification strategies [12, 21, 22, 24] or to examine several strategies separately [33, 34, 37]. Further, all of the identified approaches to network sparsification were dependent on the careful selection of a thresholding parameter by the researcher. To circumvent the arbitrariness associated with the selection of a single threshold, the majority of studies employed multiple thresholding, meaning that a range of cutoff values with small incremental steps was examined such that a series of networks \({\mathcal{G}}_s:= {\left\{{G}_{s,{\tau}_k}\right\}}_{k=1,\dots, \mathrm{T}}\) for each individual s (s = 1, …, N) was obtained with T thresholds τ_{k} where k = 1, …, T [65, 73, 76, 82]. For an illustrated example of multiple thresholding, see Fig. 6. Only one study performed weightbased thresholding with a single, arbitrary threshold [83].
In general, there was considerable heterogeneity in the selection of the optimal threshold value for sparsification in the studies evaluated:(1) a threshold yielding a smallworld index (see Table 1 for definition) above 1 of the networks was chosen [65, 84], (2) an arbitrary fixed threshold was chosen [39], (3) the differently sparsified networks corresponding to a single individual \({\mathcal{G}}_s\) were fused into an average individualspecific network [25], (4) the numerical integral or average of the graphtheoretical feature over the range of network \({\mathcal{G}}_s\) was computed [25, 56, 64, 65, 76], (5) a threshold generating the best results in terms of classification accuracy or association with the outcome was selected [8, 29, 34, 51], and (6) further processing was done including varying network sparsity levels [24, 75, 85].
According to the identified studies conducting network sparsification, the choice of the threshold is crucial to balance between noise removal (spurious edges) and preservation of ‘true’ edges. For instance, proportional thresholding may leave edges with very low edge weights assumed spurious, whereas weightbased thresholding yields different densities across the individualspecific networks which then may affect other graphtheoretical features and in turn, hinder comparability across networks. While the evaluated studies often adequately addressed the selection of the threshold parameter weightbased and proportional thresholding, ‘hidden’ threshold parameters in regularized network inference (i.e. penalty strength) or univariate testing (i.e. significance level) were rarely further investigated.
Graphtheoretical feature extraction
Appropriate features describing the network after its construction reduce the dimensionality of the network, capture aspects of the graph structure and may constitute valuable biomarkers for clinical outcomes in applied health research. The connectivity information within a network can be characterized on different scales: globally or locally. While global features capture properties of the whole network, local features describe their characteristics in defined subareas such as nodes, edges or modules (i.e. clusters of nodes).
The majority of quantitative studies extracted both, local and global graphtheoretical features (N = 56, 41.5%), followed by studies only focusing on local features (N = 44, 32.6%) and studies interested only in global graphtheoretical features (N = 35, 26.0%). The average number of computed global graphtheoretical features describing different aspects of the network across all studies was 3.19 with a standard deviation (SD) of 4.69 and a range of 0 to 44, while the average number of local features was comparatively small with a mean of 1.62 and a SD of 1.72 and a range of 0 to 10. The most frequently examined graphtheoretical features were the clustering coefficient (CC) (N = 84, 62.2%) [86] quantifying the tendency of clustering within the network and the characteristic path length (CPL) (N = 61, 45.2%) defined as the average of all shortest paths over all pairs of nodes in a network. A total of 57 different graphtheoretical features were identified. The most commonly used metrics (examined in more than 10 studies) are briefly explained in Table 1. More detailed descriptions are available elsewhere [87]. In some studies, graphtheoretical features were normalized by dividing them by the same metric computed from a randomly generated network of identical size, density and/or degree distribution to account for differences in network size and density, introducing additional computational complexity. Largely, studies (N = 31, 23.8%) examined the normalized clustering coefficient and characteristic path length obtained by dividing the CC and CPL by the CC and CPL of multiple randomly generated networks [18, 20, 79, 81, 84, 88]. For instance, Imms et al. [89] highlighted the use of normalized CC and CPL as diagnostic biomarkers to differentiate between controls and patients with traumatic brain injury.
Feature selection and prediction modelling
Around a third of the 135 quantitative studies (N = 46, 34.1%) conducted implicit techniques for prediction i.e. elementary statistical analytics (e.g. hypothesis testing, correlation analysis) to identify potential biomarkers associated with the outcome [8, 59, 84, 90] while the remaining studies (N = 89, 65.9%) conducted prediction modelling using explicit techniques i.e. methods with the possibility of predicting for an unseen individual. Among the latter, about two thirds aimed at the supervised identification of disease subgroups using graphtheoretical features as independent variables (N = 57, 64.0%) and one third prognosticated a clinical outcome (N = 32, 36.0%). Despite the similarity of the methodological frameworks, the evaluated studies turned out to be very heterogeneous regarding feature selection, sample size, outcome types, the outcome modelling technique and the validation analytics employed.
Feature selection
Feature selection approaches identified in the studies used to define an optimal subset of features can be classified into filter methods (e.g. univariable testing, Pearson correlation) and wrapper methods (iterative optimization of a classification algorithm) but hybrid approaches were also proposed [83]. The majority of studies used filter methods to remove features before training a classifier or a regression model [20, 21, 91]. Wrapper methods (e.g. support vector machine (SVM) with recursive feature elimination (RFE), repeated selection across leaveoneout crossvalidation (LOOCV) runs) either iteratively reduced the features based on a ranking score (e.g. feature importance) of the feature in the prediction algorithm to optimize classification accuracy [25, 63] or selected features with the most discriminative ability out of the full set of features based on repeated least absolute shrinkage and selection operator (LASSO) feature selection across LOOCV runs [77].
Prediction modelling
The median sample size in neurology was 68 (interquartile range (IQR) 41 to 127), in genomics 445 (IQR: 333–761), and in pathopsychology 62 (IQR: 41–97) as illustrated in Fig. 7A. In pathology, only one study with a sample size of 70 was identified and two studies in cardiology with a sample size of 55 and 389 individuals, respectively. The most commonly investigated types of dependent variables across all 135 quantitative studies were binary (N = 86, 63.7%), followed by continuous (N = 37, 27.4%), categorical (N = 14, 10.4%) and lastly, timetoevent (N = 4, 3.0%) with some studies assessing multiple outcome types (N = 6, 4.4%). The considered statistical modelling techniques employed for the task of prediction varied accordingly. Hence, for the sake of clarity, we will distinguish between data modelling and algorithmic modelling approaches according to Breiman [92] (Table 2). The former group contains modelling techniques that connect covariates to the outcome variable by a stochastic model (e.g. linear or logistic regression), while approaches belonging to the latter group use an algorithm to predict the outcome from the covariates (e.g. SVM, random forest). For a complete list of the grouping of the identified methods, we refer to the supplementary material. In general, SVMs were the most popular approach (N = 40, 45.0%) [27], followed by linear regression (N = 25, 28.0%) [19, 22, 28, 64, 67, 81, 93, 94], random forests (N = 12, 13.5%) [15, 41, 58, 72, 95, 96] and logistic regression (N = 9, 10.1%) [41, 61, 65, 97, 98].
The majority of quantitative studies built discriminative classifiers for discrete outcome labels mainly using only local graphtheoretical features (see Table 2 and Fig. 7B). The most common approach extracted local graphtheoretical features with filterbased feature selection and then trained a linear SVM for outcome classification [16, 30, 37, 62, 77, 83]. A popular choice for feature selection before SVM training and classification was univariable feature screening of statistical significance [10, 13, 14, 34, 36, 74, 78]. However, this method is known to suffer from conceptual shortcomings [99]. In the case of multiple thresholding [74], the use of multiple data modalities [14] or the extraction of multiple local features [32, 100] for classification, a supervised multiplekernel learning approach was adopted to fuse layers of features to predict disease subgroups. A subset of studies even trained multiple classifiers for the binary classification of the presence of disease (e.g. SVM, knearest neighbour, decision tree, random forests, Naïve Bayes, Adaptive Boosting) [11, 26, 35, 69, 96, 101]. For instance, [26] stated to have built 67 classifiers to differentiate between children with autism spectrum disorder and agematched controls and 23 advanced regression models for phenotypic prediction.
In terms of model validation, nearly all diagnostic studies reported model discrimination and performance of the classifier in terms of crossvalidated accuracy, sensitivity, specificity and the area under the receiver operating characteristic curve (AUC) [12, 23, 25, 27, 70, 77]. Some studies reported estimated predictive accuracy solely utilizing correlation between the predicted and the true values [93, 96]. In a few studies, performance was claimed to exceed previous approaches, but such a claim was often unsubstantiated and solely based on the comparison between their overall accuracy or AUC and that of the existing studies without consideration of other study factors (study design, set of variables, sample size, heterogeneity between study cohorts) [26, 102, 103]. Imbalanced distribution of class labels was stated as an issue for training accurate classifiers but was accounted for in some studies by assessing the balanced accuracy (i.e. the arithmetic mean of sensitivity and specificity) [58, 70, 74, 100]. Further, calibration analysis as a reliability assessment of the predicted probability of the event actually occurring with the observed relative frequency by the means of a calibration curve was rarely reported in any of the evaluated studies.
Regressionbased modelling was mainly performed by linear regression (N = 28, 60.7%), followed by logistic regression (N = 8, 17.4%), Cox proportional hazards regression (N = 2, 4.3%) and mixedeffects modelling (N = 2, 4.3%). In case global graphtheoretical features were extracted from each individualspecific network or a continuous outcome type was of interest, regressionbased analytics were the preferred course of action (see Table 2). The majority of the studies used only graphtheoretical features as independent variables and rarely adjusted for clinical characteristics (e.g. in [15, 18, 21, 22, 67, 73, 80, 97]). Even fewer studies adjusted for network size and density within the regression model to account for their intersubject differences affecting the extracted graphtheoretical features [79, 81]. Model adjustment by clinical information was more common in studies investigating the association of global graphtheoretical features with an outcome of interest rather than studies interested in local features. Batalle et al. [61] stated that the graphtheoretical features included in addition to the clinicalepidemiological covariates even proved to yield a higher contribution in terms of statistical significance after separate (blockwise) stepwise backward elimination of variables. However, clinical variables were not only used to adjust the model. For example, Xie et al. [39] proposed a twostage approach consisting of first, deriving individualspecific network models from conditional Gaussian graphical models dependent on an individual’s clinical features and second, the clinical outcome model to estimate the graphtheoretical parameters’ effects on an outcome of interest adjusted by the same clinical features.
The sample size of studies conducting regressionbased methods was slightly higher and showed a significantly greater variance in comparison with the sample size studies employing classificationbased methods (Table 2). Furthermore, possible overfitting of predictive data modelling was rarely directly addressed as in the study by Batalle et al. [61] who reduced preselected graphtheoretical features into a single summary index or in Anderson et al. [98] who restricted the variables which entered the model to 4 to have at least 5 events per variables [104].
Only three of the 130 quantitative studies conducted Cox proportional hazards model to model a timetoevent outcome [66, 77, 79]. Tuladhar et al. [66] identified global efficiency instead of conventional MRI biomarkers as a predictor of allcause dementia with lower global efficiency associated with a higher risk of dementia onset while adjusting for a set of clinical features. Similarly, Liu et al. [77] stated that none of the traditional clinical features was consistently selected in the majority of the LOOCV runs for overall survival time of highgrade glioma patients and out of the three most important selected features, two were graphtheoretical features.
Discussion
In this scoping review, we identified the stateoftheart statistical methodology currently employed when using individualspecific networks for prediction in medicine and applied health research. We found a wide range of applications and methodological concepts in our review. We collated the key concepts identified across the 148 included studies considering three main aspects of modelling with individualspecific networks: (1) individualspecific network inference, (2) extraction of graphtheoretical features, and (3) prediction modelling. Within each of these aspects, there is considerable methodological heterogeneity in the implementation, use, areas of application, and reporting. However, all approaches outlined in this work are in principle generalizable to any field of research and may be suitable to answer various prognostic or diagnostic research questions in medicine.
Individualspecific networks were frequently constructed by evaluating correlations between repeated measurements of pairs of variables (e.g. timeseries data of two brain regions). Here we identified two main approaches based on bivariate and partial correlation analysis, some variants thereof, and some further approaches that were often tied to the process of data acquisition itself. Furthermore, the recently proposed visibility graphs offer a flexible approach to individualspecific network inference in relation to time series data. However, sometimes several time series are available per individual so that the individualspecific network is ambiguous, which leads to a further layer of complexity. In the absence of repeated measurements per variable, two novel and promising approaches were LOONC and differential perturbation. Despite their flexibility and ease of application to continuous independent variables, a big challenge in the latter two approaches remains the considerable computational burden, in particular in highdimensional data settings, due to the repeated computation of the augmented network for each individual [57].
In the pursuit to separate ‘real’ from ‘spurious’ connections, in addition to the need for a reduced computational burden, network sparsification has become an important aspect of network analysis. However, sparsification not only depends on the selected technique but, more importantly, also on the chosen threshold, and hence, often multiple thresholds were employed to reduce the impact of a possibly flawed choice, yielding a sequence of networks per individual with varying edge weights. Various approaches were then applied to deal with the sequence of networks; numerical integration or averaging were the most popular approaches together with a threshold selection yielding the best AUROC. Although the majority of studies refrained from using a single, arbitrary threshold value, multiple edge weighting schemes and sparsification strategies were seldom guided by model fit. In addition, sensitivity analyses evaluating the impact of threshold choice on predictive performance were expected but hardly found.
For the extraction of graphtheoretical features, we found a set of global and local features (e.g. see Table 2) that were used in many studies across research fields. For the most part, the clustering coefficient, the characteristic path length and the edge weights were examined in the search for potential biomarkers of the outcome of interest. Furthermore, the extraction of graphtheoretical features did not follow a deliberate process but often consisted of a greedy collection of network characteristics i.e. the computation and outcomeassociated investigation of as many graphtheoretical variables as possible.
Despite the multiplicity of identified methodological approaches for individualspecific network inference across fields of application, the number of studies that proposed models actually intended to provide clinical outcome prediction for future individuals was quite limited, and most models were estimated to provide a proofofconcept that graphtheoretical features may in principle be useful for outcome prediction. The lack of deployable clinical prediction models could be either a consequence of the fundamental challenges in network inference methods shared by all areas of application, in particular concerning the lack of a gold standard for network construction and sparsification, or of the general unawareness in how individualspecific networks can be exploited for prediction.
Through the systematic collation of the identified analytical approaches, we found some areas interesting for future research and which may help to reduce some of the arbitrariness of some analytical choices.
First, more research is needed to reduce the computational burden related to the construction and analysis of relatively large and dense networks and the inherent computational complexity of graph metric computation. Large omics studies generate substantial amounts of data which can lead to major computational difficulties in network inference and further analysis, if each node in the network represents a single variable. In particular, LOONC and differential perturbation approaches would suffer in such a data setting due to the computational burden caused by the repeated network computation for each individual in the study cohort. One possible strategy for reducing network complexity and facilitating network analysis and feature extraction could be node aggregation over groups of connected, nonindependent nodes (modules). In this sense, variables could be combined as modules either through unsupervised clustering algorithms or through biological background knowledge, so that each node represents a group of independent variables and no longer a single variable.
Second, network sparsification and multivariable model estimation could be linked more closely by guiding the search for a suitable threshold or the integration over several thresholds by the model fit. Multiple thresholding yields a set of sequential graphtheoretical estimates of the networks that are computed over a fine grid on a continuous domain using incremental steps between threshold values (e.g. grid searching). Ideas from functional data analysis could be transferred to the area of modelling with individualspecific networks. Briefly, instead of choosing a threshold that provides univariably optimal prediction performance, or integration over multiple prespecified thresholds with equal weights of each threshold, one could interpret the individualspecific sequence of graphtheoretical features corresponding to the set of sparsified networks \({\mathcal{G}}_s:= {\left\{{G}_{s,{\tau}_k}\right\}}_{k=1,\dots, \mathrm{T}}\) as a functional data predictor. Then, one may define a flexible weighting function of τ_{k}, f(τ_{k}), through outcomeguided calibration to optimize clinical prediction. Consequently, such an analysis would also yield an estimate of the relative importance of different thresholds for network sparsification. Alternatively, carefully conducted sensitivity analyses would allow evaluating to what extent the reported results depended on the choice of the selected threshold value in particular for studies continuing the search for the optimal threshold parameter by univariate analysis.
Further, the comparison of networks of varying sizes and edge densities can impose issues for prediction modelling since some graphtheoretical features are confounded by them [80]. The inclusion of these two features regardless of ‘significance’ may reduce the magnitude of bias, and improve prediction performance and explainability of such models [105]. Omission of these confounding variables could mask the actual effects of interest in model explanation. Generally, graphtheoretic features are inherently associated with each other, and more research is needed to better understand these associations.
In contrast, a reoccurring problem of multivariable model building was found in the evaluated studies: “univariable prefiltering” of variables in which only variables with a statistically significant association with the outcome are included in the model [106, 107]. However, a pvalue above the statistical significance threshold of 5% is not sufficient evidence for the lack of an effect of the independent variable [108]. The popularity of prefiltering across the evaluated studies can presumably be traced back to the greedy collection of graphtheoretical features or to a disproportional number of local graph features that were obtained relative to sample size. Since the actual goal of univariable preselection was often a considerable reduction of the number of independent variables proportional to the sample size to seemingly avoid overfit, defining a minimum basic set of features (MBSF) to investigate may be beneficial when network analysis is employed for prediction. The identification of such an MBSF, however, is not an easy task and may require investigations across a range of applications embedded in the respective research fields.
Lastly, future studies should investigate the added benefit of graphtheoretical features in addition to clinical variables so as not to examine their clinical utility separately from traditional variables and hence, improve existing clinical prediction models. We have seen that graphtheoretical variables were mostly examined independently, which was partly due to the relatively low samples across the evaluated studies but also due to the dominant preference of algorithmic classification approaches with local graphtheoretical features, where clinical information was largely ignored. In some studies, in which graphtheoretical features were examined together with clinical variables, these even turned out to be stronger predictors than the traditional set of clinical variables. It remains elusive if this could be explained by publication bias or demonstrates the clinical relevance of graphtheoretical features.
Research on the aforementioned points is essential to establish a stateoftheart and to provide more evidencebased guidance in using individualspecific networks for the prediction of clinical outcomes to applied researchers. Despite the identified aspects for future research, our scoping review is subject to some limitations. First, we may have missed some relevant applications due to the lack of standardized terminology to describe the intersection of network analysis (in particular approaches to construct individualspecific networks) and prediction modelling but also because of the ambiguity of the term ‘network’. Secondly, by limiting our study to applications in medicine and health research, we may not have captured studies that employed individualspecific networks in which the individuals do not represent patients but other individual entities. Thirdly, since the screening of 4988 studies and extracting data for the 148 articles included in this review was laborious, some time passed between identification of studies and completion of data extraction. Nevertheless, we decided not to update our search to include more recent articles (i.e. published from August 2020 onwards) because we do not believe that there were substantial changes in practice in the intervening time. Last but not least, this review focused on the use of individualspecific network analysis for prediction which is why existing refinements of the presented methods for network construction and extraction of graphtheoretic ones might not have passed the inclusion criteria of our search strategy.
At this juncture, it is important to emphasize that we have not assessed the quality of included studies and did not perform a risk of bias assessment. This was done in agreement with the general guidelines on conducting a scoping review. Consequently, our review is unable to make definitive recommendations for practice but rather describes the current methodological practice and possible areas of future research [7, 109].
Conclusion
Network analysis offers a flexible tool for personalized medicine, hence prediction with individualspecific networks is an emerging field full of potential for future research. The application in clinical research is still in its infancy but our findings can strengthen the methodological conduct to incorporate individualspecific network analysis in predictive tasks. The framework still requires further refinement, and research must cover statistical, computational and applicationspecific aspects. In addition to methodological advances, comparative studies of proposed methodologies are needed to understand how methods compare and which method works best in a specific setting. This may eventually lead to establishing a stateoftheart in this novel and fascinating scientific arena located at the crosssection of statistics, computer science and medicine.
Availability of data and materials
Not applicable.
Abbreviations
 AUC:

Area under the curve
 A:

Assortativity
 BC:

Betweenness centrality
 BOLD:

Blood oxygenation leveldependent
 CC:

Clustering coefficient
 CPL:

Characteristic path length
 DAG:

Directed acyclic graph;
 Dg:

Degree
 Ds:

Density
 EEG:

Electroencephalography
 EW:

Edge weight
 GE:

Global efficiency
 HRV:

Heart rate variability
 IQR:

Interquartile range
 LASSO:

Least absolute shrinkage and selection operator
 LE:

Local efficiency
 LOONC:

Leaveoneout network construction
 LOOCV:

Leaveoneout crossvalidation
 M:

Modularity
 MBSF:

Minimum basic set of features
 RFE:

Recursive feature elimination
 SD:

Standard deviation
 SVM:

Support vector machine
 SWI:

Small world index
 ROI:

Region of interest
 VAR:

Vector autoregressive
References
Hastie T, Tibshirani R, Friedman J. The elements of statistical learning. New York: Springer Series in Statistics; 2001.
Van Calster B, van Smeden M, De Cock B, Steyerberg EW. Regression shrinkage methods for clinical prediction models do not guarantee improved performance: simulation study. Stat Methods Med Res. 2020;29(11):3166–78.
Šinkovec H, Heinze G, Blagus R, Geroldinger A. To tune or not to tune, a case study of ridge logistic regression in small or sparse datasets. BMC Med Res Methodol. 2021;21(1):199.
Barabási AL, Gulbahce N, Loscalzo J. Network medicine: a networkbased approach to human disease. Nat Rev Genet. 2011;12(1):56–68.
Li MM, Huang K, Zitnik M. Graph Representation Learning in Biomedicine. arXiv. 2021;210404883.v2. https://arxiv.org/abs/2104.04883. Accessed 10 Nov 2021.
Peters MD, Marnie C, Tricco AC, Pollock D, Munn Z, Alexander L, et al. Updated methodological guidance for the conduct of scoping reviews. JBI Evid Synth. 2020;18(10):2119–26.
Tricco AC, Lillie E, Zarin W, O'Brien KK, Colquhoun H, Levac D, et al. PRISMA extension for scoping reviews (PRISMAScR): checklist and explanation. Ann Intern Med. 2018;169(7):467–73.
Dong Z, Li X, Chen W. Frequency network analysis of heart rate variability for obstructive apnea patient detection. IEEE J Biomed Health Inform. 2017;22(6):1895–905.
Sáez A, Rivas E, MonteroSánchez A, Paradas C, Acha B, Pascual A, et al. Quantifiable diagnosis of muscular dystrophies and neurogenic atrophies through network analysis. BMC Med. 2013;11(1):1–11.
Bian J, Xie M, Topaloglu U, Cisler JM. A Probabilistic Model of Functional Brain Connectivity Network for Discovering Novel Biomarkers. AMIA Summ Transl Sci Proc. 2013;2013:21.
Saghayi M, Greenberg J, O’Grady C, Varno F, Hashmi MA, Bracken B, et al. Brain network topology predicts participant adherence to mental training programs. Netw Neurosci. 2020;4(3):528–55.
Xu X, Li W, Mei J, Tao M, Wang X, Zhao Q, et al. Feature selection and combination of information in the functional brain connectome for discrimination of mild cognitive impairment and analyses of altered brain patterns. Front Aging Neurosci. 2020;12:28.
Khazaee A, Ebrahimzadeh A, BabajaniFeremi A. Identifying patients with Alzheimer’s disease using restingstate fMRI and graph theory. Clin Neurophysiol. 2015;126(11):2132–41.
Wee CY, Yap PT, Zhang D, Denny K, Browndyke JN, Potter GG, et al. Identification of MCI individuals using structural and functional connectivity networks. NeuroImage. 2012;59(3):2045–56.
Paldino MJ, Zhang W, Chu ZD, Golriz F. Metrics of brain network architecture capture the impact of disease in children with epilepsy. NeuroImage: Clin. 2017;13:201–8.
Li Y, Wang Y, Wu G, Shi F, Zhou L, Lin W, et al. Discriminant analysis of longitudinal cortical thickness changes in Alzheimer's disease using dynamic and network features. Neurobiol Aging. 2012;33(2):427 e15. e30.
Paldino MJ, Golriz F, Zhang W, Chu ZD. Normalization enhances brain network features that predict individual intelligence in children with epilepsy. PLoS One. 2019;14(3):e0212901.
Rimkus CM, Schoonheim MM, Steenwijk MD, Vrenken H, Eijlers AJ, Killestein J, et al. Gray matter networks and cognitive impairment in multiple sclerosis. Mult Scler J. 2019;25(3):382–91.
Finn ES, Shen X, Scheinost D, Rosenberg MD, Huang J, Chun MM, et al. Functional connectome fingerprinting: identifying individuals using patterns of brain connectivity. Nat Neurosci. 2015;18(11):1664–71.
van Duinkerken E, Ijzerman RG, Klein M, Moll AC, Snoek FJ, Scheltens P, et al. Disrupted subjectspecific gray matter network properties and cognitive dysfunction in type 1 diabetes patients with and without proliferative retinopathy. Hum Brain Mapp. 2016;37(3):1194–208.
De Baene W, Rutten GJM, Sitskoorn MM. Cognitive functioning in glioma patients is related to functional connectivity measures of the nontumoural hemisphere. Eur J Neurosci. 2019;50(12):3921–33.
Wang Z, Zhang D, Liang B, Chang S, Pan J, Huang R, et al. Prediction of biological motion perception performance from intrinsic brain network regional efficiency. Front Hum Neurosci. 2016;10:552.
Sen B, Bernstein GA, Mueller BA, Cullen KR, Parhi KK. Subgraph entropy based network approaches for classifying adolescent obsessivecompulsive disorder from restingstate functional MRI. NeuroImage: Clin. 2020;26:102208.
Cheng H, Newman S, Goñi J, Kent JS, Howell J, Bolbecker A, et al. Nodal centrality of functional network in the differentiation of schizophrenia. Schizophr Res. 2015;168(1–2):345–52.
Wen H, Liu Y, Rekik I, Wang S, Chen Z, Zhang J, et al. Combining disrupted and discriminative topological properties of functional connectivity networks as neuroimaging biomarkers for accurate diagnosis of early tourette syndrome children. Mol Neurobiol. 2018;55(4):3251–69.
Zhou Y, Yu F, Duong T. Multiparametric MRI characterization and prediction in autism spectrum disorder using graph theory and machine learning. PLoS One. 2014;9(6):e90405.
Hojjati SH, Ebrahimzadeh A, BabajaniFeremi A. Identification of the early stage of Alzheimer's disease using structural MRI and restingstate fMRI. Front Neurol. 2019;10:904.
Yamashita M, Kawato M, Imamizu H. Predicting learning plateau of working memory from wholebrain intrinsic network connectivity patterns. Sci Rep. 2015;5(1):1–8.
Zhang T, Zhao Z, Zhang C, Zhang J, Jin Z, Li L. Classification of early and late mild cognitive impairment using functional brain network of restingstate fMRI. Front Psych. 2019;10:572.
Zhou L, Wang Y, Li Y, Yap PT, Shen D. Initiative AsDN. Hierarchical anatomical brain networks for MCI prediction: revisiting volumetric measures. PLoS One. 2011;6(7):e21935.
Richiardi J, Achard S, Bunke H, Van De Ville D. Machine learning with brain graphs: predictive modeling approaches for functional imaging in systems neuroscience. IEEE Signal Process Mag. 2013;30(3):58–70.
Zhang Y, Zhang H, Chen X, Lee SW, Shen D. Hybrid highorder functional connectivity networks using restingstate functional MRI for mild cognitive impairment diagnosis. Sci Rep. 2017;7(1):1–15.
Damaraju E, Allen EA, Belger A, Ford JM, McEwen S, Mathalon D, et al. Dynamic functional connectivity analysis reveals transient states of dysconnectivity in schizophrenia. NeuroImage: Clin. 2014;5:298–308.
Qiao L, Zhang H, Kim M, Teng S, Zhang L, Shen D. Estimating functional brain networks by incorporating a modularity prior. NeuroImage. 2016;141:399–407.
Cecchi GA, Rish I, Thyreau B, Thirion B, Plaze M, PaillereMartinot ML, et al. Discriminative Network Models of SchizophreniaNIPS; 2009.
Wee CY, Yap PT, Zhang D, Wang L, Shen D. Groupconstrained sparse fMRI connectivity modeling for mild cognitive impairment identification. Brain Struct Funct. 2014;219(2):641–56.
Bohland JW, Saperstein S, Pereira F, Rapin J, Grady L. Network, anatomical, and nonimaging measures for the prediction of ADHD diagnosis in individual subjects. Front Syst Neurosci. 2012;6:78.
Lord LD, Allen P, Expert P, Howes O, Broome M, Lambiotte R, et al. Functional brain networks before the onset of psychosis: a prospective fMRI study with graph theoretical analysis. NeuroImage: Clin. 2012;1(1):91–8.
Xie S, Li X, McColgan P, Scahill RI, Zeng D, Wang Y. Identifying diseaseassociated biomarker network features through conditional graphical model. Biometrics. 2020;76(3):995–1006.
Booij SH, Wichers M, De Jonge P, Sytema S, Van Os J, Wunderink L, et al. Study protocol for a prospective cohort study examining the predictive potential of dynamic symptom networks for the onset and progression of psychosis: the Mapping Individual Routes of Risk and Resilience (Mirorr) study. BMJ Open. 2018;8(1):e019059.
Lutz W, Schwartz B, Hofmann SG, Fisher AJ, Husen K, Rubel JA. Using network analysis for the prediction of treatment dropout in patients with mood and anxiety disorders: A methodological proofofconcept study. Sci Rep. 2018;8(1):1–9.
Lacasa L, Luque B, Ballesteros F, Luque J, Nuno JC. From time series to complex networks: The visibility graph. Proc Natl Acad Sci. 2008;105(13):4972–5.
Ahmadlou M, Adeli H, Adeli A. New diagnostic EEG markers of the Alzheimer’s disease using visibility graph. J Neural Transm. 2010;117(9):1099–109.
Bajestani GS, Behrooz M, Khani AG, NouriBaygi M, Mollaei A. Diagnosis of autism spectrum disorder based on complex network features. Comput Methods Prog Biomed. 2019;177:277–83.
Grobelny BT, London D, Hill TC, North E, Dugan P, Doyle WK. Betweenness centrality of intracranial electroencephalography networks and surgical epilepsy outcome. Clin Neurophysiol. 2018;129(9):1804–12.
Hou F, Li F, Wang J, Yan F. Visibility graph analysis of very shortterm heart rate variability during sleep. Phys A Stat Mech Appl. 2016;458:140–5.
Chen S, Gallagher MJ, Hogg F, Papadopoulos MC, Saadoun S. Visibility graph analysis of intraspinal pressure signal predicts functional outcome in spinal cord injured patients. J Neurotrauma. 2018;35(24):2947–56.
Silva VF, Silva ME, Ribeiro P, Silva F. Time series analysis via network science: Concepts and algorithms. Wiley Interdiscip Rev Data Min Knowl Discov. 2021;11(3):e1404.
Zhang Z, Ding J, Xu J, Tang J, Guo F. Multiscale Timeseries Kernelbased Learning Method for Brain Disease Diagnosis. IEEE J Biomed Health Inform. 2021;25(1):209–17.
Homan P, Argyelan M, DeRosse P, Szeszko PR, Gallego JA, Hanna L, et al. Structural similarity networks predict clinical outcome in earlyphase psychosis. Neuropsychopharmacology. 2019;44(5):915–22.
Philips GR, Daly JJ, Príncipe JC. Topographical measures of functional connectivity as biomarkers for poststroke motor recovery. J Neuroeng Rehabil. 2017;14(1):1–16.
Kuijjer ML, Tung MG, Yuan G, Quackenbush J, Glass K. Estimating samplespecific regulatory networks. Iscience. 2019;14:226–40.
LopesRamos CM, Kuijjer ML, Ogino S, Fuchs CS, DeMeo DL, Glass K, et al. Gene regulatory network analysis identifies sexlinked differences in colon cancer drug metabolism. Cancer Res. 2018;78(19):5538–47.
Zhu K, Pian C, Xiang Q, Liu X, Chen Y. Personalized analysis of breast cancer using samplespecific networks. PeerJ. 2020;8:e9161.
Liu X, Wang Y, Ji H, Aihara K, Chen L. Personalized characterization of diseases using samplespecific networks. Nucleic Acids Res. 2016;44(22):e164e.
Audrain S, Barnett AJ, McAndrews MP. Language network measures at rest indicate individual differences in naming decline after anterior temporal lobe resection. Hum Brain Mapp. 2018;39(11):4404–19.
Huang Y, Chang X, Zhang Y, Chen L, Liu X. Disease characterization using a partial correlationbased samplespecific network. Brief Bioinform. 2021;22(3):bbaa062.
Das T, Borgwardt S, Hauke DJ, Harrisberger F, Lang UE, RiecherRössler A, et al. Disorganized gyrification network properties during the transition to psychosis. JAMA Psychiatry. 2018;75(6):613–22.
Park B, Lee W, Park I, Han K. Finding prognostic gene pairs for cancer from patientspecific gene networks. BMC Med Genet. 2019;12(8):1–14.
Boot EM, van Leijsen EM, Bergkamp MI, Kessels RP, Norris DG, de Leeuw FE, et al. Structural network efficiency predicts cognitive decline in cerebral small vessel disease. NeuroImage: Clin. 2020;27:102325.
Batalle D, Eixarch E, Figueras F, MuñozMoreno E, Bargallo N, Illa M, et al. Altered smallworld topology of structural brain networks in infants with intrauterine growth restriction and its association with later neurodevelopmental outcome. NeuroImage. 2012;60(2):1352–66.
Sun Y, Bi Q, Wang X, Hu X, Li H, Li X, et al. Prediction of conversion from amnestic mild cognitive impairment to Alzheimer's disease based on the brain structural connectome. Front Neurol. 2019;9:1178.
Wee CY, Yap PT, Li W, Denny K, Browndyke JN, Potter GG, et al. Enriched white matter connectivity networks for accurate identification of MCI patients. NeuroImage. 2011;54(3):1812–22.
Welton T, Constantinescu CS, Auer DP, Dineen RA. Graph theoretic analysis of brain Connectomics in multiple sclerosis: Reliability and relationship with cognition. Brain Connectivity. 2020;10(2):95–104.
Du J, Wang Y, Zhi N, Geng J, Cao W, Yu L, et al. Structural brain network measures are superior to vascular burden scores in predicting early cognitive impairment in post stroke patients with small vessel disease. NeuroImage: Clin. 2019;22:101712.
Tuladhar AM, van Uden IW, RuttenJacobs LC, Lawrence A, van der Holst H, van Norden A, et al. Structural network efficiency predicts conversion to dementia. Neurology. 2016;86(12):1112–9.
Yeo RA, Ryman SG, Van Den Heuvel MP, De Reus MA, Jung RE, Pommy J, et al. Graph metrics of structural brain networks in individuals with schizophrenia and healthy controls: group differences, relationships with intelligence, and genetics. J Int Neuropsychol Soc. 2016;22(2):240.
Gou L, Zhang W, Li C, Shi X, Zhou Z, Zhong W, et al. Structural brain network alteration and its correlation with structural impairments in patients with depression in de novo and drugnaive Parkinson's disease. Front Neurol. 2018;9:608.
Liu W, Zhang C, Wang X, Xu J, Chang Y, Ristaniemi T, et al. Functional connectivity of major depression disorder using ongoing EEG during music perception. Clin Neurophysiol. 2020;131(10):2413–22.
BabajaniFeremi A, Noorizadeh N, Mudigoudar B, Wheless JW. Predicting seizure outcome of vagus nerve stimulation using MEGbased network topology. NeuroImage: Clin. 2018;19:990–9.
GomezPilar J, de LuisGarcía R, Lubeiro A, de la Red H, Poza J, Núñez P, et al. Relations between structural and EEGbased graph metrics in healthy controls and schizophrenia patients. Hum Brain Mapp. 2018;39(8):3152–65.
Van Diessen E, Otte WM, Braun KP, Stam CJ, Jansen FE. Improved diagnosis in children with partial epilepsy using a multivariable prediction model based on EEG network characteristics. PLoS One. 2013;8(4):e59764.
Wen W, Zhu W, He Y, Kochan NA, Reppermund S, Slavin MJ, et al. Discrete neuroanatomical networks are associated with specific cognitive abilities in old age. J Neurosci. 2011;31(4):1204–12.
Jie B, Zhang D, Wee CY, Shen D. Topological graph kernel on multiple thresholded functional connectivity networks for mild cognitive impairment classification. Hum Brain Mapp. 2014;35(7):2876–97.
dos Santos SA, Biazoli Junior CE, Comfort WE, Rohde LA, Sato JR. Abnormal functional restingstate networks in ADHD: graph theory and pattern recognition analysis of fMRI data. Biomed Res Int. 2014;2014:380531.
Hou Z, Wang Z, Jiang W, Yin Y, Yue Y, Zhang Y, et al. Divergent topological architecture of the default mode network as a pretreatment predictor of early antidepressant response in major depressive disorder. Sci Rep. 2016;6(1):1–9.
Liu L, Zhang H, Wu J, Yu Z, Chen X, Rekik I, et al. Overall survival time prediction for highgrade glioma patients based on largescale brain functional networks. Brain Imaging Behaviour. 2019;13(5):1333–51.
Yu R, Zhang H, An L, Chen X, Wei Z, Shen D. Connectivity strengthweighted sparse group representationbased brain network construction for M CI classification. Hum Brain Mapp. 2017;38(5):2370–83.
Tijms BM, Ten Kate M, Gouw AA, Borta A, Verfaillie S, Teunissen CE, et al. Gray matter networks and clinical progression in subjects with predementia Alzheimer's disease. Neurobiol Aging. 2018;61:75–81.
Tijms BM, Yeung HM, Sikkes SA, Möller C, Smits LL, Stam CJ, et al. Singlesubject gray matter graph properties and their relationship with cognitive impairment in earlyand lateonset Alzheimer's disease. Brain Connectivity. 2014;4(5):337–46.
Hawkins R, Shatil A, Lee L, Sengupta A, Zhang L, Morrow S, et al. Reduced global efficiency and random network features in patients with relapsingremitting multiple sclerosis with cognitive impairment. Am J Neuroradiol. 2020;41(3):449–55.
Jie B, Liu M, Zhang D, Shen D. Subnetwork kernels for measuring similarity of brain connectivity networks in disease diagnosis. IEEE Trans Image Process. 2018;27(5):2340–53.
Dai D, He H, Vogelstein JT, Hou Z. Accurate prediction of AD patients using cortical thickness networks. Mach Vis Appl. 2013;24(7):1445–57.
Langer N, Pedroni A, Gianotti LR, Hänggi J, Knoch D, Jäncke L. Functional brain network efficiency predicts intelligence. Hum Brain Mapp. 2012;33(6):1393–406.
Hashmi JA, Kong J, Spaeth R, Khan S, Kaptchuk TJ, Gollub RL. Functional network architecture predicts psychologically mediated analgesia related to treatment in chronic knee pain patients. J Neurosci. 2014;34(11):3924–36.
Watts DJ, Strogatz SH. Collective dynamics of ‘smallworld’networks. nature. 1998;393(6684):440–2.
Rubinov M, Sporns O. Complex network measures of brain connectivity: uses and interpretations. NeuroImage. 2010;52(3):1059–69.
Tijms BM, Möller C, Vrenken H, Wink AM, de Haan W, van der Flier WM, et al. Singlesubject grey matter graphs in Alzheimer's disease. PLoS One. 2013;8(3):e58921.
Imms P, Clemente A, Cook M, D’Souza W, Wilson PH, Jones DK, et al. The structural connectome in traumatic brain injury: A metaanalysis of graph metrics. Neurosci Biobehav Rev. 2019;99:128–37.
Lee J, Lee M, Kim DS, Kim YH. Functional reorganization and prediction of motor recovery after a stroke: a graph theoretical analysis of functional networks. Restor Neurol Neurosci. 2015;33(6):785–93.
Raamana PR, Weiner MW, Wang L, Beg MF. Initiative AsDN. Thickness network features for prognostic applications in dementia. Neurobiol Aging. 2015;36:S91–S102.
Breiman L. Statistical modeling: The two cultures (with comments and a rejoinder by the author). Stat Sci. 2001;16(3):199–231.
ChristovMoore L, Reggente N, Douglas PK, Feusner JD, Iacoboni M. Predicting empathy from resting state brain connectivity: A multivariate approach. Front Integr Neurosci. 2020;14:3.
Doucet GE, Rider R, Taylor N, Skidmore C, Sharan A, Sperling M, et al. Presurgery restingstate local graphtheory measures predict neurocognitive outcomes after brain surgery in temporal lobe epilepsy. Epilepsia. 2015;56(4):517–26.
Corps J, Rekik I. Morphological brain age prediction using multiview brain networks derived from cortical morphology in healthy and disordered participants. Sci Rep. 2019;9(1):1–10.
Gheiratmand M, Rish I, Cecchi GA, Brown MR, Greiner R, Polosecki PI, et al. Learning stable and predictive networkbased patterns of schizophrenia and its clinical symptoms. NPJ Schizophr. 2017;3(1):1–12.
Dicks E, Tijms BM, Ten Kate M, Gouw AA, Benedictus MR, Teunissen CE, et al. Gray matter network measures are associated with cognitive decline in mild cognitive impairment. Neurobiol Aging. 2018;61:198–206.
Anderson ED, Giudice JS, Wu T, Panzer MB, Meaney DF. Predicting concussion outcome by integrating finite element modeling and network analysis. Front Bioeng Biotechnol. 2020;8:309.
Sun GW, Shook TL, Kay GL. Inappropriate use of bivariable analysis to screen risk factors for use in multivariable analysis. J Clin Epidemiol. 1996;49(8):907–16.
Jie B, Wee CY, Shen D, Zhang D. Hyperconnectivity of functional networks for brain disease diagnosis. Med Image Anal. 2016;32:84–100.
Xu M, Sanz DL, Garces P, Maestu F, Li Q, Pantazis D. A Graph Gaussian Embedding Method for Predicting Alzheimer’s Disease Progression with MEG Brain Networks. IEEE Trans Biomed Eng. 2021;68(5):1579–88.
Rashid B, Calhoun V. Towards a brainbased predictome of mental illness. Hum Brain Mapp. 2020;41(12):3468–535.
Wei R, Li C, Fogelson N, Li L. Prediction of conversion from mild cognitive impairment to Alzheimer's Disease using MRI and structural network features. Front Aging Neurosci. 2016;8:76.
Vittinghoff E, McCulloch CE. Relaxing the rule of ten events per variable in logistic and Cox regression. Am J Epidemiol. 2007;165(6):710–8.
Shrier I, Platt RW. Reducing bias through directed acyclic graphs. BMC Med Res Methodol. 2008;8(1):1–15.
Heinze G, Wallisch C, Dunkler D. Variable selection–a review and recommendations for the practicing statistician. Biom J. 2018;60(3):431–49.
Heinze G, Dunkler D. Five myths about variable selection. Transpl Int. 2017;30(1):6–10.
Steyerberg EW. Clinical prediction models: a practical approach to development, validation, and updating. New York: Springer; 2019.
Khalil H, Peters M, Tricco A, Pollock D, Alexander L, McInerney P, et al. Conducting high quality scoping reviewschallenges and solutions. J Clin Epidemiol. 2021;130:156–60.
Acknowledgements
KVS and FM acknowledge the European Union’s Horizon 2020 research and innovation programme under grant agreement No 860895 (TranSYSh2020transys.eu). KVS also acknowledges the European Union’s Horizon 2020 research and innovation programme under the Marie SklodowskaCurie grant agreement No 813533 (MLFPMmlfpm.eu).
Funding
This project has received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie SkłodowskaCurie grant agreement No 860895 (FM, KVS).
Author information
Authors and Affiliations
Contributions
MG was responsible for the conception, design of the study, search, screening, inclusion of relevant articles, data extraction of eligible studies and wrote the initial draft of the manuscript. GH was responsible for the conception, design and carried out the screening and data extraction of a fraction of articles. FM and MS conducted screening and data extraction of a fraction of articles. GH, KVS and SM revised the draft critically for important intellectual content. All authors agreed to the final submission.
Corresponding author
Ethics declarations
Ethics approval and consent to participate
Not applicable.
Consent for publication
Not applicable.
Competing interests
All authors declare that they have no competing interests.
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary Information
Additional file 1.
Preferred Reporting Items for Systematic reviews and MetaAnalyses extension for Scoping Reviews (PRISMAScR) Checklist
Additional file 2: Supplementary Material S1.
Detailed Methods. Table S1. Keyword sets used in the search strategy across the databases for article extraction. Figure S1. Year of publication of the identified studies. Table S2. Scale of the graphtheoretical features used as candidate predictors stratified by the area of application.
Additional file 3. Supplementary Material S2.
Questionnaire for fulltext screening
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.
About this article
Cite this article
Gregorich, M., Melograna, F., Sunqvist, M. et al. Individualspecific networks for prediction modelling – A scoping review of methods. BMC Med Res Methodol 22, 62 (2022). https://doi.org/10.1186/s12874022015446
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s12874022015446