NeuroIntegrative Connectivity (NIC) Informatics Tool for Brain Functional Connectivity Network Analysis in Cohort Studies

Abstract

Objective: Brain functional connectivity measures are often used to study interactions between brain regions in various neurological disorders such as epilepsy. In particular, functional connectivity measures derived from high resolution electrophysiological signal data have been used to characterize epileptic networks in epilepsy patients. However, existing signal data formats as well as computational methods are not suitable for complex multi-step methods used for processing and analyzing signal data across multiple seizure events. To address the significant data management challenges associated with signal data, we have developed a new workflow-based tool called NeuroIntegrative Connectivity (NIC) using the Cloudwave Signal Format (CSF) as a common data abstraction model.

Method: The NIC compositional workflow-based tool consists of: (1) Signal data processing component for automated pre- processing and generation of CSF files with semantic annotation using epilepsy domain ontology; and (2) Functional network computation component for deriving functional connectivity metrics from signal data analysis across multiple recording channels. The NIC tool streamlines signal data management using a modular software implementation architecture that supports easy extension with new libraries of signal coupling measures and fast data retrieval using a binary search tree indexing structure called NIC-Index.

Result and Conclusion: We evaluated the NIC tool by processing and analyzing signal data for 28 seizure events in two patients with refractory epilepsy. The result shows that certain brain regions have high local measure of connectivity, such as total degree, as compared to other regions during ictal events in both patients. In addition, global connectivity measures, which characterize transitivity and efficiency, increase in value during the initial period of the seizure followed by decrease towards the end of seizure. The NIC tool allows users to efficiently apply several network analysis metrics to study global and local changes in epileptic networks in patient cohort studies.

1. Introduction

Brain connectivity measures play a significant role in the study of many neurological disorders, including dementia (1), Alzheimer’s Disease (2), and epilepsy (3). In particular, functional connectivity representing temporal coherence between events recorded from different brain regions is an important technique for studying epilepsy seizure network (4-6). Epilepsy is a serious neurological disorder that affects more than 65 million persons worldwide with a significant impact on their quality of life due to repeated seizures with an associated high cost of care (7). More than 30% of epilepsy patients do not respond to anti-epileptic drugs (AED) and they are considered for surgery to resect brain regions that constitute the epileptogenic zone for seizure freedom (8, 9). However, the success rate for epilepsy surgery is not high due several challenges, including difficulty in precise and complete delineation of the epileptogenic zone that results in seizure recurrence for 20% - 50% of the patients who undergo surgery (3). Therefore, there is a critical need to improve the accuracy of existing methods used to delineate epileptogenic zone and improve surgical outcomes in epilepsy patients.

In particular, Stereotactic EEG (SEEG) measures brain electrical activity at a high resolution using multi-contact electrodes implanted directly in the brain structures and records signal data that can be used to derive measures of functional connectivity to characterize the extent of epileptic network (5). The SEEG signal data is often used as “gold standard” during pre-surgical evaluation to characterize the extent of the epileptogenic zone for resection while protecting important cognitive brain functions of the patient (10). SEEG data is recorded as multi-channel signals corresponding to electrical activity from different brain regions and it can be analyzed to: (a) infer whether the EEG signals from different regions are coupled during seizure events; and (b) to characterize the specific type of coupling. In addition, the analysis of EEG data can also provide a measure of directionality regarding the propagation of seizure signals from the region of onset to other brain regions.

The increasing availability of large scale and high-resolution SEEG signal data is an important opportunity to use data-driven techniques to study the characteristics of epileptic network. In particular, multi-institutional collaborative projects are collecting large volumes of high quality neurology data, such as the National Institutes of Health (NIH)-funded multi-center project to study Sudden Unexpected Death in Epilepsy (SUDEP) (11). However, there are several challenges that impede streamlined processing and analysis of these large datasets. For example, the computation of functional connectivity measures from large volume of SEEG data stored using traditional data formats, such as the European Data Format (EDF) (12), is extremely challenging. EDF was originally designed for storage of sleep research related polysomnography data (12-14). In particular, EDF has several limitations that make it difficult to compute functional connectivity measures due to its original design principles and application focus. For example, EDF data storage layout is not suitable for retrieval of channel-oriented data. In addition, EDF files store signal data as a single unit of recording that makes it extremely difficult to extract specific segments of signal data corresponding to individual seizure events (15).

Due to these and other limitations of existing signal data formats (discussed Section 1.1), users often have to copy and transform signal data into different formats based on their requirements. To address this key bottleneck in large-scale research studies, we developed a new model for signal data called Cloudwave Signal Format (CSF) based on the widely used JavaScript Object Notation (JSON) (16). CSF has been designed as a flexible data model for electrophysiological signal data with support for ontology-based event annotation and streamlined storage of signal data as “segments” of fixed time duration for faster retrieval. The key role of CSF as a common abstraction model for signal data enables users to quickly compose new multi-step data processing or analysis pipelines for complex network analysis without cumbersome data transfer and data transformation steps. This is a significant advantage for users who often conduct exploratory data analysis, for example applying both signal amplitude-based correlation functions or frequency based coherence function to accurately characterize the degree of coupling between two signal data streams (15, 17).

In particular, the streamlined processing and analysis of large-scale signal data is critical for gaining deeper insights into the topology of seizure networks during pre-ictal and ictal events using network analysis techniques (18-20). Network analysis techniques have been widely used to characterize complex interactions in a variety of domains, including social media, web information systems, and brain connectivity studies (21, 22). The broad categories of network measures that are used to characterize brain networks are: (1) local measures that characterize the connectivity properties of constituent elements of a network such as node and edge; and (2) global measures that are comprised of individual element measures (21, 23, 24). The primary goals of the NIC tool are: (1) facilitate network analysis of seizure networks using large-scale signal data in patient cohort studies with CSF as a common data abstraction model; and (2) use workflow techniques for efficient and streamlined data processing and analysis. The NIC tool is available for download at (25) and consists of executable files as well as user documentation for deployment and usage.

1.1 Background.

Signal Data Format. There have been several initiatives in the neuroscience community to develop common data formats to facilitate data interoperability, integration, and analysis. For example, the XML-based Clinical and Experimental Data Exchange (XCEDE) schema models the details of the experimental procedure and results (26). Similarly, the NeuroImaging Data Model (NIDM) extends the XCEDE model with support for provenance information describing various metadata elements to enable data reuse and sharing (27). The International Neuroinformatics Coordinating Facility (INCF) has led the Neurodata Without Boundaries (NWB) initiative to standardize neuroscience data models (28). The NWB initiative includes various data formats for storing electrophysiological data, such as Neo (29) and Multiscale Electrophysiology Format (MEF) (30). Although these data formats address some of the limitations of EDF specifications, they have limited or no support for “channel-oriented” storage of partitioned segments of signal data, which is essential for computation of functional connectivity and network analysis. In addition, unlike CSF existing data format for signal data do not include support for semantic annotation of signal data using domain ontologies.

Coupling Measures for Epileptic Network Characterization. Multiple signal analysis techniques have been developed that assume a linear relationship between signal data recorded from different brain regions, for example linear cross- correlation and coherence function. Linear correlation coefficient r between two signals recorded from two sites M (m) and N (n) assumes a model consisting of: (1) an operator L that modifies amplitude of signal m, latency (D), and a noise source (W) (15). The correlation coefficient as a function of time shift between the signals m and n estimates the degree of their interaction and latency value (31). The coherence function uses the signal frequency to compute the correlation between the frequency components of signals m and n (17, 32). The inherent assumption of these measures that the signals are related by a linear function is not always applicable, therefore additional methods, such as average amount of mutual information based on Shannon entropy and non-linear correlation coefficient, have been developed (33, 34). In particular, a method developed by Pijn et al. called nonlinear regression coefficient h2 has been frequently used to characterize the association between two EEG signals during seizure activity (33, 35, 36). In this paper, we describe the implementation of three computational modules to compute linear correlation coefficient 𝜌, nonlinear regression coefficient h2, and mean phase coherence measures in the NIC tool. It is important to note that the modular architecture of the NIC tool allows new coupling measures to be seamlessly added to existing modules in the NIC tool.

Network Analysis of Functional Connectivity Network in Epilepsy. Several brain network analysis studies have identified “small world” connectivity as an important feature of brain functional network, which consists of efficient local and global connections (21, 22, 37). Many studies have focused on the organization of specific brain regions or modules during events between seizures (interictal), during seizure (ictal), and after seizure (postictal) using scalp EEG, electrocorticographic (ECoG), and SEEG data (18, 19, 38). Using undirected graph models, many of these studies have identified increased organization in the network topology during ictal events with dissolution of these subgraph motifs towards the end of seizure (19, 37, 39). Studies focused on the analysis of network topology during interictal period have not detected consistent trends. In this paper, we describe the use of a directed graph model for analysis of network topology during different periods of an ictal event. In addition, unlike previous studies that often use an ad hoc approach for data processing as well analysis, we describe the advantages of using the NIC tool to streamline the data processing and analysis steps for large-scale signal datasets.

2. Methods

Computation of functional connectivity measures from signal data corresponding to a clinical event involves multiple data processing and analysis steps, including extraction of one or more signal data segments corresponding to specific time duration, recording channels, and computation of coupling measure. Existing approaches use computational steps that are difficult to share, and require significant effort to maintain over a period of time. Scientific workflow systems are widely used to automate multi-step data processing and analysis pipelines in a wide range of scientific application domains, including biomedical research (40, 41). In this section, we describe the components and features of the NIC tool that uses a workflow approach to process and analyze signal data. The components of the NIC tool can be divided into two categories based on their functionalities: (1) component I: to generate CSF files from EDF files with semantic annotations; (2) component II: to compute signal coupling measures to derive functional connectivity measures for network analysis.

2.1 Component I: Generation of CSF files with Semantic Annotations. The first component of the NIC tool has been developed to support automated conversion of EDF files into CSF files. This signal data processing and transformation component consists of five modules (Figure 1 shows a detailed overview of the workflow with input and output values together with control flow between different modules). This NIC component takes four user-defined values as input parameters: (1) the duration of a signal fragment (epoch duration) in a CSF file (the default duration of a fragment is 30 seconds); (2) the number of fragments to be stored in a single CSF file; (3) the location of the source EDF files and the associated annotation files (with clinical events); and (4) the storage location for output CSF files. The first two modules of this component extract description of the research study and metadata information for each recording electrode from the EDF file.

Figure 1. The Component I of the NIC Platform (Generation of CSF files with Semantic Annotations) with input/output parameters and computational modules.

In addition, the clinical annotation values are mapped to terms defined in a domain ontology called the Epilepsy and Seizure Ontology (EpSO) for semantic annotation of CSF files (42). EpSO has been designed as a domain ontology using the well-known four- dimensional classification of epileptic seizures and epilepsies: (1) Seizures that represent signs or symptoms due to abnormal electrophysiological activity in brain; (2) Anatomical brain locations corresponding to seizure activities; (3) Etiology to describe the cause of epilepsy; and (4) Medical conditions that are related to epilepsy (42). The current version of EpSO with more than 1350 ontology classes models a variety of terms describing clinical events that occur during EEG recording and these terms are modeled as ontology classes using the Web Ontology Language (OWL2) properties (42). The third module extracts and transforms data from EDF file records into channel-oriented data records using a “key-value” structure. The third module also partitions the signal data into smaller fragments corresponding to epoch duration (a user-defined parameter). The output of the third module are fragments of signal data with channel-oriented layout that are optimized for use in functional network analysis applications. We note that each CSF file is self-descriptive with study metadata, channel-specific metadata, clinical event annotations, and fragments of signal data. Therefore, a CSF file can be stored in a distributed file system, such as the Hadoop Distributed File Systems (HDFS) for large-scale signal data analysis (16).

2.2 Component II: Computation of Functional Connectivity Measures from Signal Data. The second component of the NIC tool implements modular libraries for computing various measures of functional connectivity and allows users to select one or more coupling measures. Figure 2 illustrates the various modules of the second component of the NIC tool together with the control flow between the different modules. The input parameters for this component includes the location of the CSF files, the start and end time of the event (T1, T2), and the set of channels (Cn) involved in recording signal data corresponding to an event. The workflow uses the NIC-Index traversal algorithm to locate and access relevant segments of signal data from a CSF file (43). The NIC-Index implements the red-black binary search tree with a well-defined search time complexity, which supports fast access to segments of signal data in a CSF file (the details of the NIC-Index are discussed in our previous work (43)). The modular design of the NIC tool makes it easy to add new coupling measures (for example, Granger causality and transfer entropy) and modification of existing coupling measure modules with minimal impact on users. The second component of the NIC tool generates a correlation or coherence value matrix for each pair of channels listed in the user input.

Figure 2. The Component II of the NIC platform (Computation of Functional Connectivity Measures from Signal Data) with input/output parameters and computational modules.

The matrix values corresponding to coupling measures are analyzed to characterize the degree of correlation or coherence between signal channels. As described earlier, the nonlinear correlation coefficient (h2) values are asymmetric, that is it is possible that h2(x|y) ¹ h2(y|x), which can be used to infer directionality of coupling between signals (33). The directionality computed from EEG signal data can be used for source localization in epileptic zone, for example Wendling et al. have proposed the “direction index” to predict a signal that “causally drives” another signal (5, 35). The non-linear measure is maximized by taking into account all possible time lag values τ between the two signals in the direction X to Y and using the maximum value of the measure, that is,

$$h_{XY}^2={\max }_{{\tau }_{\min }}<\tau <{\tau }_{\max }{h}_{XY}^2$$

. The NIC component generates the output coupling measures as comma separated values (CSV) and stores them in a file, which can be directly used for network analysis.

2.2 Network Analysis of Seizure Networks. Network analysis of graph models is a widely-used technique to interpret complex interactions between different brain regions as part of functional network analysis. For example, degree distribution of a directed graph network is used to characterize the resilience of the network in terms of nodes that are important to maintain the network and the density of edges representing the state of network (21, 24, 44). The degree of a node based on the number of edges incident on the node, either as outbound edges or inbound edges in a directed graph (Figure 3), is a common measure used to characterize the local properties of a graph (24). The aggregate of outbound edges and inbound edges of a node is called the total degree of a node and the distribution of node degree values for a graph is called the degree distribution of the graph (24). Nodes with high degree values often have a key role in a network and their removal can lead to decrease in the connectivity of the network, potentially isolating specific nodes, and ultimately fragmenting the network (24, 44). Therefore, this local network measure is of interest in analyzing epileptic networks.

Figure 3. The degree distribution of electrode contacts (represented as nodes in seizure network) corresponding to six events for Seizure 1 and Seizure 2 for patient 1.

In addition to local network measures, there are multiple global measures of network analysis that can be used to characterize the transitivity, efficiency, and presence of modular structures in seizure networks (21). The characteristic path length (CPL) measure is computed as the average of shortest paths between all pairs of nodes in the graph and represents the separation between vertices in a graph (45). In contrast to CPL, the global efficiency is influenced more by short paths and this measure is computed as the average of inverse shortest path lengths (46). The global efficiency measure is used to characterize the efficiency of communications across the network and small world networks have high global efficiency reflecting low cost of network communications (46). The clustering coefficient measure represents the connectivity between nodes in the neighborhood of a specific node and is interpreted as the occurrence of “cliques” in a network (45). The clustering coefficient measure is computed as the ratio of the total number of triangles formed among the neighbors of a node and the total number of potential triangles that can form between the neighbors of a node (45). In the next section, we describe the results of using the NIC workflow with the NIC-Index to process signal data from 28 seizure-related events followed by both local and global network analysis using SEEG signal data recorded from two patients with refractory epilepsy.

3. Results

The objective of our evaluation is to demonstrate and validate the use of NIC tool for computing functional connectivity measures followed by network analysis of the epileptic network. We use data for 28 seizure events across 4 seizures in two epilepsy patients to demonstrate the practical utility of the NIC platform in research studies. The two patients were selected as both of them were refractory to medication and were considered for surgery.

Patient 1 was diagnosed with intractable focal epilepsy of 5-years duration (47). The patient was stereotactically implanted with intracranial electrodes with 10-12 contacts of length in 31cms in insulae, mesial temporal, and opercular regions based on the results of surface EEG for seizure lateralization (47). Patient 2 was diagnosed with generalized tonic clonic (GTC) and automotor seizure of 5-years duration. The patient was implanted with intracranial electrodes with 10 intracranial electrodes in amygdala, hippocampal head and body, posterior cingulate, basal temporo-occipital, temporal pole, orbitofrontal, frontal track in area of encephalomalacia, anterior insula, and posterior insula. The two seizures (out of multiple seizures) for each patient were selected as they had the maximum number of clinically differentiated periods during a seizure, which are labeled as ictal periods and correspond to spread of discharge at specific time (e.g., location of discharge onset is labeled as seizure onset, location of discharge spread in 1st 200 milliseconds labeled as ictal period 1 etc.). The patient 1 has three ictal periods and patient 2 had five ictal periods.

In the first step, the component I of the NIC platform was used to process and convert 12 EDF files into 204 CSF files for patient 1 and 25 EDF files into 2858 CSF files for patient 2 using default parameters. The resulting CSF files together with time as well as channels information associated with ictal periods were used as input to NIC component II to compute three functional connectivity measures. The non-linear correlation coefficient measure (h2) was used to derive directed graph network models, which were subsequently analyzed using both global and local graph network analysis metrics (24, 44).

3.1 Local and Global Measures of Seizure Networks. It is extremely difficult to manually quantify the changes in network topology during various ictal period due to the large number of nodes and complex interactions between these nodes during ictal period. Therefore, network analysis metrics significantly improve our ability to systematically quantify the network connectivity during ictal period. Figure 3 shows the degree distribution of nodes during Seizure 1 and Seizure 2 events for patient 1. Figure 3 shows that a specific set of nodes have high total degree across both Seizure 1 and Seizure 2 ictal period as compared to others, for example contacts on electrodes RF, LK, and LF. In contrast, contacts on electrodes LJ, LI have low total degree for both Seizure 1 and Seizure 2 with contacts on electrode RJ also showing low total degree for Seizure 2 related ictal periods. These nodes can be studied further in the context of their role in the formation and maintenance of seizure network structure. Similarly, Figure 4 shows the degree distribution during Seizure 1 and Seizure 2 ictal periods for patient 2. Figure 4 shows that contacts on electrodes implanted in amygdala, temporal pole, and to a lesser extent hippocampal head as well as basal temporal occipital have consistently high degree during both Seizures 1 and 2. In contrast, electrode contacts in hippocampal body have low degree across both the seizures and electrode contacts in posterior cingulate have low degree during Seizure 2.

Figure 4. The degree distribution of electrode contacts (represented as nodes in seizure network) corresponding to six events for Seizure 1 and Seizure 2 for patient 2.

Table 1 shows the results of computing four global network measures for all six ictal periods of Seizure 1 and Seizure 2 in patient 1. The average degree measure is used to characterize the network density property of a graph and the values in Table 1 show that there is a notable increase in the average degree of the seizure networks during ictal periods for both Seizure 1 and Seizure 2 as compared to the period before seizure onset with ictal period 1 associated with higher average degree for both seizures. We note that almost all CPL values for Seizure 1 ictal periods are higher than Seizure 2 (except ictal period 2) and the CPL values increase during ictal events for both Seizure 1 and Seizure 2 in contrast to seizure onset and ictal periods before seizure onset, which corresponds to existence of longer paths during ictal periods (CPL is known to be influenced by long paths (21)). In Table 1, we see that although the global efficiency measures for both Seizure 1 and Seizure 2 are not high (9% to 7%), they increase during ictal periods (ictal period 2 in particular with 16% and 12% for Seizure 1 and 2 respectively), which may correspond to more efficient transmission of signals across the seizure network during ictal periods.

Table 1: Global measures of seizure networks corresponding to six events across two seizures in Patient 1.

	5 minutes Preictal		10 seconds Preictal		Seizure Onset		Ictal Period 1		Ictal Period 2		Ictal Period 3
	Seizure 1	Seizure 2	Seizure 1	Seizure 2	Seizure 1	Seizure 2	Seizure 1	Seizure 2	Seizure 1	Seizure 2	Seizure 1	Seizure 2
Average Degree	2.27	1.6	2.53	1.87	2.27	1.87	3.6	2.4	2.27	2.8	2.67	1.73
Characteristic Path Length	0.14	0.01	0.11	0.10	0.1	0.09	0.38	0.13	0.14	0.18	0.13	0.11
Global Efficiency	0.09	0.07	0.09	0.08	0.09	0.07	0.09	0.09	0.16	0.12	0.10	0.07
Clustering Coefficient	0.12	0	0.15	0	0.15	0.15	0.19	0.27	0.12	0.23	0.26	0

It is interesting to note that the clustering coefficient measure for Seizure 2 network increases significantly during two of the three ictal periods from a value of 0 during pre-ictal period before returning to a value of 0 during ictal period 3. These rapid changes in Seizure 2 represent the formation and dissolution of clustered subnetworks during ictal periods that may represent high rate of signal transmission across specific parts of the seizure network. In contrast to Seizure 2, the clustering coefficient values for Seizure 1 related ictal periods remain almost unchanged across five of the six periods with approximately two-fold increase during ictal period 3 (from 0.12 to 0.26).

Table 2 shows the results of applying the same global network measures for Seizure 1 and Seizure 2 in patient 2. Similar to patient 1, the average degree value increases during both Seizure 1 and Seizure 2 for patient 2 with significantly greater increase during ictal period 1 followed by gradual decrease during ictal periods 2, 3, 4, and 5. Similarly, CPL and global efficiency measures increase during Seizure 1 and Seizure 2 during the initial phase of the seizure with subsequent decrease during later ictal periods. However, the clustering coefficient values are relatively stable during seizure onset and all five ictal periods, which is unlike the increase and decrease of clustering coefficient values for patient 1. Overall, the global network measures for patient 1 and patient 2 show increased connectivity during the initial phases of both the seizures followed by gradual decrease during later part of the seizures. These results are consistent with previous studies that have used ECoG and scalp electrodes.

Table 2: Global measures of seizure networks corresponding to six events across two seizures in Patient 2.

	Threshold Period		10 seconds Pre-ictal		Seizure Onset		Ictal Period 1		Ictal Period 2		Ictal Period 3		Ictal Period 4		Ictal Period 5
	Seizure 1	Seizure 2	Seizure 1	Seizure 2	Seizure 1	Seizure 2	Seizure 1	Seizure 2	Seizure 1	Seizure 2	Seizure 1	Seizure 2	Seizure 1	Seizure 2	Seizure 1	Seizure 2
Average Degree	2.82	2.52	8.45	4.78	2.18	4.26	7.73	14.5	5.55	7.22	1.73	3.04	1.27	4.17	2.27	4
CPL	0.09	0.1	1.62	0.44	0.58	0.41	1.62	1.59	0.57	0.82	0.26	0.27	0.16	0.27	0.48	0.42
Global Efficiency	0.07	0.07	0.38	0.17	0.16	0.15	0.36	0.57	0.21	0.28	0.11	0.1	0.08	0.13	0.17	0.14
Clustering Coefficient	0.82	0.82	0.63	0.62	0.62	0.71	0.58	0.58	0.63	0.6	0.6	0.7	0.6	0.8	0.68	0.66

3.2 User Evaluation Using Survey. We performed a user evaluation of the two components of the NIC tool using a questionnaire that recorded the satisfaction of the user with respect to the usability and functionality. The survey questionnaire was prepared following guidelines defined by the US Department of Health and Human Services Usability.gov site (48). The survey involved a group of five persons who are representative of the expected users of the NIC tool, including a physician, an epidemiology researcher, a postdoctoral scholar, and two graduate students participating in research studies focused on neurological disorders. Overall, the users were highly satisfied or satisfied with the tool, they found the tool easy to use, and they were likely to refer the tool to other users. Further, three of the five participants had used similar tools earlier and they found the NIC tool to be better or somewhat better. The users also provided descriptive feedback in terms of new functionalities and improvements in the NIC tool, for example specific changes to the user manual for improving usability, and adding support for new versions of the Java execution environment. Based on the user feedback, the next release of the NIC platform will incorporate the suggested modifications to further improve the usability and functionality.

4. Discussions and Limitations

The evaluation results demonstrate that the NIC tool is a practical tool for processing and analyzing signal data from multiple events across multiple seizures. The output of the NIC platform in form of directed graph models can be directly used to perform systematic network analysis, which provides insight regarding the pattern of interactions between different brain regions during seizure events. For example, the degree distribution analysis show that electrode contacts in certain brain regions (in both the patients) have significantly higher degree of interactions as compared to other brain regions, which may correlate to their role in the formation and persistence of network structure during seizures. In addition, the global network analysis measures show that there is an increase in the level of organization in brain network during the initial phase of seizures and this characteristic persists through multiple ictal periods. It is interesting to note that there are notable variations in clustering coefficient values across events in patient 1, while they change minimally across events in patient 2. This may reflect consistent formation of fully connected subgraphs in patient 2 that have a significant role in seizure activity and need further investigation.

In this paper, we did not use specific signal frequency bands, for example beta or gamma band, during network analysis, which a potential direction for future studies in the context of network analysis metrics. The current version of the NIC tool has not been evaluated with respect to scalability. In particular, as part of our ongoing work, different components of the NIC tool are being implemented using Apache Spark (49) to support faster and scalable analysis of large volume of data. Finally, an important limitation of the version of NIC tool described in this paper is the lack of capability to capture provenance metadata to support reproducibility of the data analysis performed using the NIC tool. Scientific reproducibility is a critical aspect of research studies and provenance metadata plays a central role in supporting reproducibility (50). Therefore, we are developing a NIC module to collect provenance metadata during the processing and analysis of signal data, which will be available in the next version of the NIC tool.

5. Conclusions

Large-scale epilepsy cohort studies are important for developing measures that can accurately lateralize the epileptogenic zone and improve outcome of epilepsy surgery. In this paper, we introduced a workflow-based tool that uses CSF as a common abstraction model for signal data for processing and analyzing signal data across multiple seizure-related events and patients. The NIC tool is a practical and efficient tool for cohort studies as it does not require users to perform cumbersome data transformation or data processing steps required for performing network analysis of epileptic networks. The network analysis of complex interactions during seizure-related events provides important insights into the topology of brain network and the NIC tool aims to enable users to perform this task in a streamlined manner.