Optimal channel selection of electroencephalography based on functional network via global graph measurements: application for epilepsy Authors: Seho Lee, Hayom Kim, Jin Hyung Kim, Hyub Huh, Dong-Joo Kim, Jung Bin Kim Published: 2025 ABSTRACT Electroencephalography (EEG) plays a vital role in diagnosing and managing epilepsy, but the traditional technique requires numerous electrodes, making it time-consuming and burdensome for patients. To overcome this, several studies propose reducing the number of channels, often requiring patient-specific selections, which limits their generalizability. This study presents a novel network-based approach for channel selection during the preictal period. We employed 1078 preictal EEG signals collected from 33 epilepsy patients at Korea University Anam Hospital. Using these signals, we established a functional connectivity network and used Pearson’s correlation coefficient to assess the correlation between the selected channels’ network features and the full network’s features. Our findings revealed that a network composed of eight channels, four in each hemisphere located in the midline frontal and parietal regions, exhibited a high correlation with the entire network. Encouragingly, applying the method to the publicly available Siena database yielded similar results. These findings indicate that selecting channels in the midline frontal and parietal regions during the preictal phase effectively preserves the entire network’s features. This suggests that our proposed channel reduction approach could not only enhance our understanding of various brain network-related diseases but also improve patient comfort and efficiency in clinical settings. Introduction Epilepsy, a prevalent neurological condition impacting millions globally, manifests as recurrent spontaneous seizures1. Electroencephalography (EEG) is used to monitor epilepsy and assess patient condition, offering high temporal resolution for tracking seizure spread and speed2. Characteristic biomarkers of epilepsy are evident on EEG, including seizures and interictal epileptiform discharges stemming from excessive, abnormal neuronal firing3,4. However, identifying seizure activity through expert neurologist interpretation of EEG remains challenging due to the multifaceted signals across multiple channels that require deciphering. Therefore, strategic channel selection promises improved utility of EEG across epilepsy diagnosis, monitoring, and management in clinical settings5. Although high-density EEG systems offer detailed spatial resolution, they come with increased setup complexity, higher costs, and computational burdens, making them less practical for real-time monitoring or wearable applications. Furthermore, for various types of epilepsy, including focal seizures, identifying key channels is clinically crucial for applications such as localizing the seizure onset zone (SOZ). Therefore, selecting the number of EEG channels while preserving critical functional network properties can provide a more efficient and patient-friendly alternative, particularly for applications requiring portability and real-time analysis6. By retaining signals with greatest relevance while mitigating complexity, reduced-channel EEG systems could facilitate expert review, automated processing, and even development of wearable seizure detection devices for daily ambulatory monitoring. Epilepsy monitoring benefits from channel reduction, as conventional EEG setups with many electrodes are burdensome for long-term use and emergency care. Fewer electrodes improve practicality while preserving network integrity. Currently conceptualized as a network-level pathology, epilepsy underlies complex interactions between functionally and structurally interconnected cortical brain regions7–10. Identifying and measuring the components that enable activity flow in this network is crucial. However, research has largely focused on identifying the epileptic focus, neglecting the intricate sub-network structure within the larger brain network that precipitates epileptic activity. Understanding network dynamics and synchronized activity is crucial to explain epileptic seizures11. The network architecture plays a pivotal role in demonstrating these dynamics and implies the emergence of synchronized epileptiform patterns12,13. However, a large number of components can obscure synchronized patterns, making it essential to simplify the network and identify core nodes through hierarchical analysis. Mathematical graph analysis allows for the examination of complex networks. Network metrics facilitate description at both nodal and global levels14. Nodal parameters describe specific nodes or connections within the network, while centrality quantifies an element’s local position, importance, or influence15. Global parameters, on the other hand, reflect overall network properties. Global metrics such as efficiency and clustering can be used for comparative assessments16,17. However, variations in basic network components can distort results, so metrics involving weak inter-regional interactions must be used cautiously to avoid noisy outputs. To ensure accuracy, analysis should focus on the densely interconnected network core of central nodes18. If the selected network reliably captures the core elements that constitute the essential network, the subsequent sub-network global parameters should remain relatively stable19,20. Therefore, solid selection methods that yield a representative core sub-network could demonstrate consistent global parameters before and after node reduction. This study aims to identify optimal EEG channels using global network parameters, preserving features regardless of seizure onset zone. Focusing on metrics, this approach offers a generalized framework for studying epilepsy-related networks while preserving functional connectivity (Fig. 1). Therefore, the hypothesis is that network analysis using global parameters can maintain network characteristics even with reduced spatial resolution. Selecting nodes based on their network information capture ability ensures accurate network metrics from reduced channels. We analyzed one minute of preictal EEG data from epilepsy patients using 19 channels and developed a channel selection algorithm based on global parameters. The selected channels were validated using external data from a different population. A schematic overview of the analytical process is shown in (Fig. 1a).Fig. 1Schematic of a process for reconstructing the brain network from a minimum number of channels while preserving information from the entire network during the preictal period. (a) The flow diagram for optimal channel reduction for preictal network. (b) Process for construction of whole network and reduced network. (c) Evaluation of correlation between the whole network and the reduced network using graph parameters. (d) Selection of the reduced network that well preserves the properties of the whole network. Schematic of a process for reconstructing the brain network from a minimum number of channels while preserving information from the entire network during the preictal period. (a) The flow diagram for optimal channel reduction for preictal network. (b) Process for construction of whole network and reduced network. (c) Evaluation of correlation between the whole network and the reduced network using graph parameters. (d) Selection of the reduced network that well preserves the properties of the whole network. Materials and methods Subject and data acquisition This study investigated inpatients from the long-term video-EEG monitoring database at Korea University Anam Hospital (KU database) between January 2019 and December 2020. Patients with EEG data containing ictal discharges were screened and selected after a comprehensive evaluation by board-certified epileptologists (HK, JBK), including neurological examination, electroclinical diagnosis, and neuroimaging studies. This study included patients whose EEG data showed at least two ictal discharges. Specifically, patients with repeated seizure activity in a short period (less than 10 min), making the timing of ictal onset and termination uncertain, were excluded. Therefore, this study included 84 patients diagnosed with focal epilepsy, selected from the long-term video-EEG monitoring database (Fig. 2). Patients were required to have at least two recorded ictal EEG events for inclusion. Seizure onset zones were classified based on seizure semiology and EEG findings, including frontal (n = 12), temporal (n = 9), parietal (n = 3), occipital (n = 1), and multifocal epilepsy (n = 10). Patients with one or no recorded ictal EEG events were excluded from the primary analysis (n = 49). EEG was recorded at a sampling rate of 200 Hz (Comet-PLUS, Grass Technologies Inc., West Warwick, RI, USA) with 19 electrodes (Fp1, Fp2, F3, F4, Fz, F7, F8, C3, C4, Cz, P3, P4, Pz, O1, O2, T3, T4, T5, and T6) according to the International 10-20 system. This study investigated optimal reduced electrodes in 33 epilepsy patients with various seizure foci using 1078 preictal EEG signals. This study adhered to the ethical guidelines of the Declaration of Helsinki and received approval from the Institutional Review Board of Korea University Anam Hospital (No. 2020AN0435). To validate the channel selection algorithm based on network graph analysis, we used the publicly available Siena scalp EEG database. This database, collected from 14 epilepsy patients at the University of Siena (Italy), comprises 19 EEG channels (Fp1, Fp2, F3, F4, Fz, F7, F8, C3, C4, Cz, P3, P4, Pz, O1, O2, T3, T4, T5, T6) recorded at 512 Hz using the International 10-20 system. Similar to our in-house database, preictal EEG signals were extracted from Siena for external validation, utilizing data from 47 preictal segments of the 14 subjects.Fig. 2Patient distribution and inclusion criteria. Patient distribution and inclusion criteria. Data processing The EEG signal from the KU database was extracted 1 min before the ictal period (Fig. 1b). EEG signals contaminated by motion artifacts, apparatus noise, etc., were manually rejected. For algorithm verification, the EEG signal from the Siena database was extracted 1 min pre-ictal using the same process as the KU database. This database, with differing sampling rates, were resampled to 200 Hz. All EEG signals were epoched into 2-s segments. Additionally, artifact signals were manually removed before functional connectivity analysis. To analyze functional connectivity using the artifact-free signal, coherence was employed. Coherence captures frequency-specific synchronization between EEG channels and serves as a useful measure of functional connectivity patterns in both full and reduced networks, despite its susceptibility to volume conduction. The coherence was estimated using the Hermes toolbox in MATLAB (MathWorks, Natick, MA, USA)21. It was calculated in four distinct frequency bands: delta (0.5–4 Hz), theta (4–8 Hz), alpha (8–12 Hz), and beta (12–30 Hz): where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{i,j}(f)$$\end{document} is the cross spectrum of the signals between channel i and j. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{i,i}(f)$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{j,j}(f)$$\end{document} are the spectral of each channel i and j. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$COH_{i,j}(f)$$\end{document} was represented that 0 corresponds to no synchrony and 1 corresponds to strong synchrony at the frequency f. Graph analysis Four key graph metrics were extracted to analyze network properties: Strength (ST), Clustering Coefficient (CC), Eigenvector Centrality (EVC), and Global Efficiency (GE). These graph metrics provide insights into both nodal and global network characteristics, ensuring that the selected subnetwork retains essential functional properties. An index of strength measures the strength of the connection between two channels (Fig. 1b). The strength of a node (ST) is calculated by the following formula:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} ST_{i} = \sum _{j\ne i}w_{ij} \end{aligned}$$\end{document}, which \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\omega _{ij}$$\end{document} is represented the weights between node i and j. The higher a node’s ST, the stronger its overall connections are with other nodes. The global efficiency (GE) of a network is a measure of the efficiency of information transmission between remote nodes. GE is calculated with the average path length between node i and j. The average path length (PL) is the mean of the weight between node i and j. It quantifies the efficiency of information transfer across the network, where lower values indicate more direct connectivity between nodes. Then, path length is represented below equation.2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} PL=\frac{1}{n\left( n-1\right) }\sum _{i,j\in N,i\ne j}\ w_{ij} \end{aligned}$$\end{document}GE is described to the inverse of the average path length between all nodes in the network, and is calculated as follows:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} GE = \frac{1}{n(n-1)}\sum _{j.k\in G}\frac{1}{PL_{j,k}} \end{aligned}$$\end{document}where n is the number of nodes and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$w_{ij}$$\end{document} is the weight of the edge between nodes i and j, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L_{j,k}$$\end{document} is the average path length between nodes j and k. A high GE indicates a more efficient network structure, allowing for rapid and effective integration of information between distant brain regions. A clustering coefficient (CC) indicates the degree to which a region is interconnected with other regions. CC is calculated by calculating the clustering degree between two different channels, including a specific channel. In a weighted network, CC can be calculated by the following formula:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} & CC_{i} = \frac{2\sum _{j=1}^{n-1}\sum _{k=j+1}^{n}(\widetilde{w_{ij}}\widetilde{w_{ik}}\widetilde{w_{jk}})^{1/3}}{deg_{i}(deg_{i}-1)} \end{aligned}$$\end{document}5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} & \widetilde{w_{i,j}} = \frac{w_{i,j}}{max\left\{ w_{ij}|i,j \in Z, 1 \le i,j\le n \right\} } \end{aligned}$$\end{document}Where, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$CC_{i}$$\end{document} is a measure of the clustering coefficient of node i and n is the number of nodes. It is calculated as the 3-root of the product of the weights of the edges between node i and its neighbors. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$CC_{i}=1$$\end{document} means that all of the neighbors of node i are connected to each other in a triangle. The CC of a node quantifies how densely connected its neighbors are to one another. In graph theory, eigenvector centrality (EVC) is one of the metrics used to calculate the influence of a channel in a functional connectivity (FC) network. EVC is calculated as follows:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} Ax = \lambda x, \lambda x = \sum _{j=1}^{n}a_{ij}x_{i}, i = 1,\cdots ,n \end{aligned}$$\end{document}An EVC measures how importance a node is in a network. It is calculated as the sum of the centralities of the nodes that are connected to the node. The largest eigenvalue of the adjacency matrix of the network, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda$$\end{document}, is proportional to the sum of the similarity scores of all the nodes that are connected to the node by the corresponding eigenvector. A high EC score indicates that a node is connected to many other highly influential nodes. Channel selection method To identify the optimal subset of channels, an exhaustive enumeration approach was employed (Fig. 1d). Specifically, we generated all possible combinations of 3, 4, and 5 channels from the set of eight left-hemisphere electrodes (Fp1, F3, F7, C3, P3, O1, T7, and P7). All corresponding paired electrodes from the right hemisphere (e.g., Fp2, F4, F8, and C4) were automatically included. However, as epileptic discharges during the preictal period typically do not propagate to both hemispheres. In addition, the midline channels (Fz, Cz, Pz) were excluded as they lack direct hemispheric counterparts and, therefore, could not be incorporated into our symmetric pairing framework. Statistical analysis A one-way analysis of variance (ANOVA) test was conducted to evaluate whether the number of selected channels had a statistically significant impact on the sum of correlation coefficients for each frequency band. ANOVA was chosen as it allows for comparing mean differences across multiple groups to determine if variations in channel selection influence network preservation. If a significant effect was found, Tukey’s post-hoc analysis was performed to identify specific group differences while controlling for multiple comparisons. To perform network-based channel selection, correlation analysis was conducted on four graph parameters: ST, CC, EVC, and GE. Channel combinations with the highest sum of the correlation coefficients for all graph metric were considered the most promising candidates for channel reduction. The implementation of equal weighting was executed to ensure a balanced interaction between nodal (ST, EVC) and global (CC, GE) network properties, thereby preventing the predominance of one aspect over the other. Nodal measures reflect local connectivity and centrality within the network, while global measures capture overall network efficiency and integration. Given the critical role of both local and global characteristics in maintaining functional connectivity, equal weighting was applied to avoid the analysis being biased toward either aspect. Additionally, Pearson correlation coefficients were calculated to evaluate the similarity of network metrics between the reduced and full networks (Fig. 1c). Only results with p-values below 0.05 were reported, ensuring statistical significance. This criterion applies to all Pearson correlation values reported in both the main text and Supplementary Materials. Results Korea University Hospital database This study proposes a method for optimal channel selection in epileptic seizure prediction, aiming to reduce the number of channels while preserving the network parameter characteristics of the epileptic brain network (Table. 1). We calculated four graph parameters: ST, GE, CC, and EVC for all available channels and for subsets of 6, 8, and 10 selected channels. Notably, the sum of these parameters for the 8- and 10-channel combinations showed no statistically significant differences across the delta, theta, and alpha frequency bands (Fig. 3). To assess how well the selected channels capture network information compared to all channels, we calculated correlation coefficients for each network indicator. In the 6-channel case, channels Fp1, C3, and P3 (with their corresponding paired channels Fp2, C4, and P4) showed the highest correlation across all frequency bands. Similarly, for the 8-channel case, the optimal combination across delta, theta, alpha, and beta bands included Fp1, C3, P3, P7 and their paired counterparts (Table 2, Fig. 4). Lastly, Fp1, F3, C3, P3, and P7 (paired channels: Fp2, F4, C4, P4, and P8) among 10 channels combinations was selected in delta, theta, and alpha band. However, this combination was ranked in second in beta band. The rank of combinations of 6 channels and 10 channels combinations was detailed in Supplementary table 1. and Supplementary table 2. All parameters were highly correlated between the first combination in paired 4 channels of Fp1, C3, and P3 (paired channels: Fp2, C4, and P4) (Fig. 5). The analysis identified channel combinations with high weights in the global network analysis from the KU database. In the frontal region, (F3, F4) was most common, followed by (Fp1, Fp2) and (F7, F8). Similarly, in the parietal region, (P3, P4) was most common, followed by P7 and P8. Interestingly, while (C3, C4) in the central region was not frequently selected, it ranked first. Consequently, the KU database generated 8-channel combinations (Fp1, F3, F7, P3, P7, C3, with their corresponding paired channels: Fp2, F4, F8, P4, P8, and C4) exhibiting high correlation with the global network (Fig. 5).Fig. 3The results of the ANOVA on the sum of the correlation coefficient, which are generated in each band, according to the combination of channels. In the delta band, theta band and alpha band, there were no significant differences between the results of 8 channels and 10 channels combinations. The bar is represented the standard deviation. Comparison of network metrics for each channel combination using a Tukey post-hoc test. N.S. not significant.Table 1Results of graph parameters in KU database. The value indicates the mean (standard deviation) of graph parameters.STGECCEVCDelta band (1–4 Hz)0.361(0.084)0.361(0.084)0.312(0.088)0.219(0.004)Theta band (4–8 Hz)0.317(0.067)0.317(0.066)0.266(0.067)0.218(0.005)Alpha band (8–12 Hz)0.308(0.070)0.308(0.070)0.260(0.070)0.216(0.004)Beta band (12–30 Hz)0.285(0.070)0.284(0.070)0.246(0.071)0.214(0.006)1 ST strength, GE global efficiency, CC clustering coeffient, EVC eigenvector centrality.Table 2Pearson’s correlation of graph parameters comparison with 8 channels and whole channels in KU database.Delta band (1–4 Hz)STGECCEVCSumFp1, C3, P3, P70.9760.9790.9760.7643.695F3, F7, P3, P70.9720.9640.9670.7473.649Fp1, F3, P3, P70.9570.9440.9560.8163.673Fp1, F3, C3, P70.9490.9330.9510.7963.629F3, F7, P3, O10.9450.9390.9420.8043.630Theta band (4–8 Hz)STGECCEVCSumFp1, C3, P3, P70.9750.9770.9730.8193.745F3, F7, P3, P70.9540.9460.9460.7003.546Fp1, F3, P3, P70.9330.9120.9300.8103.585F3, F7, P3, O10.9370.9260.9310.7503.544Fp1, F3, C3, P70.9120.9000.9130.8413.566Alpha band (8–12 Hz)STGECCEVCSumFp1, C3, P3, P70.9680.9730.9700.7953.707F3, F7, P3, O10.9650.9610.9580.7103.594F3, F7, P3, P70.9650.9520.9610.6803.558Fp1, F3, P3, P70.9500.9320.9510.7493.581Fp1, F3, C3, P70.9440.9320.9460.7683.590Beta band (12–30 Hz)STGECCEVCSumFp1, C3, P3, P70.9450.9570.9540.8873.744F3, F7, P3, P70.9620.9320.9630.8423.698Fp1, C3, P3, T70.9530.9340.9560.7593.602Fp1, C3, P3, O10.9400.9420.9490.7083.538F3, F7, T7, O10.9310.9320.9350.8143.6131 ST strength, GE global efficiency, CC clustering coeffient, EVC eigenvector centrality. The results of the ANOVA on the sum of the correlation coefficient, which are generated in each band, according to the combination of channels. In the delta band, theta band and alpha band, there were no significant differences between the results of 8 channels and 10 channels combinations. The bar is represented the standard deviation. Comparison of network metrics for each channel combination using a Tukey post-hoc test. N.S. not significant. Table 1Results of graph parameters in KU database. The value indicates the mean (standard deviation) of graph parameters.STGECCEVCDelta band (1–4 Hz)0.361(0.084)0.361(0.084)0.312(0.088)0.219(0.004)Theta band (4–8 Hz)0.317(0.067)0.317(0.066)0.266(0.067)0.218(0.005)Alpha band (8–12 Hz)0.308(0.070)0.308(0.070)0.260(0.070)0.216(0.004)Beta band (12–30 Hz)0.285(0.070)0.284(0.070)0.246(0.071)0.214(0.006)1 ST strength, GE global efficiency, CC clustering coeffient, EVC eigenvector centrality. Results of graph parameters in KU database. The value indicates the mean (standard deviation) of graph parameters. 1 ST strength, GE global efficiency, CC clustering coeffient, EVC eigenvector centrality. Table 2Pearson’s correlation of graph parameters comparison with 8 channels and whole channels in KU database.Delta band (1–4 Hz)STGECCEVCSumFp1, C3, P3, P70.9760.9790.9760.7643.695F3, F7, P3, P70.9720.9640.9670.7473.649Fp1, F3, P3, P70.9570.9440.9560.8163.673Fp1, F3, C3, P70.9490.9330.9510.7963.629F3, F7, P3, O10.9450.9390.9420.8043.630Theta band (4–8 Hz)STGECCEVCSumFp1, C3, P3, P70.9750.9770.9730.8193.745F3, F7, P3, P70.9540.9460.9460.7003.546Fp1, F3, P3, P70.9330.9120.9300.8103.585F3, F7, P3, O10.9370.9260.9310.7503.544Fp1, F3, C3, P70.9120.9000.9130.8413.566Alpha band (8–12 Hz)STGECCEVCSumFp1, C3, P3, P70.9680.9730.9700.7953.707F3, F7, P3, O10.9650.9610.9580.7103.594F3, F7, P3, P70.9650.9520.9610.6803.558Fp1, F3, P3, P70.9500.9320.9510.7493.581Fp1, F3, C3, P70.9440.9320.9460.7683.590Beta band (12–30 Hz)STGECCEVCSumFp1, C3, P3, P70.9450.9570.9540.8873.744F3, F7, P3, P70.9620.9320.9630.8423.698Fp1, C3, P3, T70.9530.9340.9560.7593.602Fp1, C3, P3, O10.9400.9420.9490.7083.538F3, F7, T7, O10.9310.9320.9350.8143.6131 ST strength, GE global efficiency, CC clustering coeffient, EVC eigenvector centrality. Pearson’s correlation of graph parameters comparison with 8 channels and whole channels in KU database. 1 ST strength, GE global efficiency, CC clustering coeffient, EVC eigenvector centrality. Fig. 4The scatter plot shows the correlation in the Korea university hospital database between the reduced network (Y-axis) and the whole network (X-axis). All graph parameters from the reduced network exhibited a strong correlation with those from the whole network, and there was across all frequency bands. Fig. 5The numerical distribution of channels that have the highest selection rate in the Top 5 candidates to reduce channels in the Korea university hospital database. Left panel: high distribution in the prefrontal and Parietal lobes across Delta, Theta, and Alpha bands. Concentration of channels in the Centroparietal region for the Beta band. Right panel: overall the reduced channels located in the prefrontal and parietal lobes in all frequency bands. The scatter plot shows the correlation in the Korea university hospital database between the reduced network (Y-axis) and the whole network (X-axis). All graph parameters from the reduced network exhibited a strong correlation with those from the whole network, and there was across all frequency bands. The numerical distribution of channels that have the highest selection rate in the Top 5 candidates to reduce channels in the Korea university hospital database. Left panel: high distribution in the prefrontal and Parietal lobes across Delta, Theta, and Alpha bands. Concentration of channels in the Centroparietal region for the Beta band. Right panel: overall the reduced channels located in the prefrontal and parietal lobes in all frequency bands. Siena database The optimal channel selection method from the KU database was applied to the Siena database . Similar to the KU analysis, we calculated graph parameters and correlation coefficients for different channel combinations (6, 8, and 10) (Tables 3 and 4). While the specific top-ranked combinations differed in Siena, a key finding emerged: Fp1, C3, P3, and P7 (with paired channels) consistently appeared in the top 5 across all bands and showed the highest correlation with the network parameters using all channels. Notably, this combination matched the optimal combination identified in the KU database. Additionally, F3, F7, P3, and P7 (with paired channels) also ranked highly in most frequency bands of Siena. Detailed information about specific channel rankings can be found in Supplementary Tables 3 and 4. The results of the preictal EEG obtained from Siena database are as follows, similar to the KU database. As shown in the KU database, the combination of channels Fp1, Fp2, P3, P4, P7, P8, C3, and C4 ranked at the top in all frequency bands. In addition, for each region, (F3, F4), (Fp1, Fp2), and (F7, F8) account for a large proportion in the frontal region, and (P3, P4) and (P7, P8) account for the highest proportion in the parietal region. In addition, the channel (C3, C4) also has a high proportion. It was confirmed that relatively fewer channels were selected in the temporal and occipital regions compared to other regions. Therefore, in the Siena database, it was confirmed that 8 combinations of Fp1, F3, F7, P3, P7, C3 (paired channels: Fp2, F4, F8, P4, P8, C4) have a high correlation with graph analysis using 19 channels, similar to the results of the KU database (Table 5).Table 3Results of graph parameters in Siena database. The value indicates the mean (standard deviation) of graph parameters.STGECCEVCDelta band (1–4 Hz)0.306(0.113)0.306(0.113)0.258(0.115)0.217(0.006)Theta band (4–8 Hz)0.268(0.092)0.268(0.092)0.219(0.091)0.217(0.006)Alpha band (8–12 Hz)0.242(0.069)0.242(0.069)0.196(0.067)0.218(0.008)Beta band (12–30 Hz)0.217(0.084)0.217(0.084)0.186(0.081)0.215(0.008)1 ST strength, GE global efficiency, CC clustering coeffient, EVC eigenvector centrality.Table 4Pearson’s correlation of Graph Parameters comparison with 8 channels and whole channels in Siena database.Delta band (1–4 Hz)STGECCEVCSumFp1, C3, P3, P70.9700.9690.9650.6083.512F7, C3, P3, P70.9740.9620.9730.5283.437F3, F7, P3, P70.9510.9490.9390.6513.489F3, C3, T7, O10.9620.9550.9600.5153.391Fp1, F3, P3, P70.9440.9440.9360.6283.451Theta band (4–8 Hz)STGECCEVCSumFp1, C3, P3, P70.9630.9590.9480.7683.638F7, C3, P3, P70.9690.9500.9580.5973.474F3, F7, P3, P70.9420.9340.9290.8003.604F3, C3, T7, O10.9510.9390.9400.6783.507Fp1, F3, C3, T70.9620.9320.9530.5813.427Alpha band (8–12 Hz)STGECCEVCSumFp1, C3, P3, P70.9580.9390.9460.7453.588Fp1, F3, C3, P70.9560.9530.9340.7203.564F3, C3, T7, O10.9420.9280.9340.7753.578Fp1, F3, C3, T70.9540.9230.9390.6893.506F3, F7, C3, P70.9660.9420.9500.5003.358Beta band (12–30 Hz)STGECCEVCSumFp1, C3, P3, P70.9650.9500.9650.7503.630F7, C3, P3, P70.9690.9500.9650.6993.583Fp1, F3, C3, T70.9590.9420.9540.7253.581F3, C3, T7, O10.9360.9430.9290.7873.596F3, C3, F7, O10.9590.9530.9590.5343.404ST strength, GE global efficiency, CC clustering coefficient, EVC eigenvector centrality.Table 5The selected channels in all combinations in Siena database.DeltaThetaAlphaBetaAll FrequencyFp1/Fp222329F3/F4334313F7/F822127C3/C4345517P3/P4431210P7/P8433212T7/T812227O1/O211125Significant values are in bold. Table 3Results of graph parameters in Siena database. The value indicates the mean (standard deviation) of graph parameters.STGECCEVCDelta band (1–4 Hz)0.306(0.113)0.306(0.113)0.258(0.115)0.217(0.006)Theta band (4–8 Hz)0.268(0.092)0.268(0.092)0.219(0.091)0.217(0.006)Alpha band (8–12 Hz)0.242(0.069)0.242(0.069)0.196(0.067)0.218(0.008)Beta band (12–30 Hz)0.217(0.084)0.217(0.084)0.186(0.081)0.215(0.008)1 ST strength, GE global efficiency, CC clustering coeffient, EVC eigenvector centrality. Results of graph parameters in Siena database. The value indicates the mean (standard deviation) of graph parameters. 1 ST strength, GE global efficiency, CC clustering coeffient, EVC eigenvector centrality. Table 4Pearson’s correlation of Graph Parameters comparison with 8 channels and whole channels in Siena database.Delta band (1–4 Hz)STGECCEVCSumFp1, C3, P3, P70.9700.9690.9650.6083.512F7, C3, P3, P70.9740.9620.9730.5283.437F3, F7, P3, P70.9510.9490.9390.6513.489F3, C3, T7, O10.9620.9550.9600.5153.391Fp1, F3, P3, P70.9440.9440.9360.6283.451Theta band (4–8 Hz)STGECCEVCSumFp1, C3, P3, P70.9630.9590.9480.7683.638F7, C3, P3, P70.9690.9500.9580.5973.474F3, F7, P3, P70.9420.9340.9290.8003.604F3, C3, T7, O10.9510.9390.9400.6783.507Fp1, F3, C3, T70.9620.9320.9530.5813.427Alpha band (8–12 Hz)STGECCEVCSumFp1, C3, P3, P70.9580.9390.9460.7453.588Fp1, F3, C3, P70.9560.9530.9340.7203.564F3, C3, T7, O10.9420.9280.9340.7753.578Fp1, F3, C3, T70.9540.9230.9390.6893.506F3, F7, C3, P70.9660.9420.9500.5003.358Beta band (12–30 Hz)STGECCEVCSumFp1, C3, P3, P70.9650.9500.9650.7503.630F7, C3, P3, P70.9690.9500.9650.6993.583Fp1, F3, C3, T70.9590.9420.9540.7253.581F3, C3, T7, O10.9360.9430.9290.7873.596F3, C3, F7, O10.9590.9530.9590.5343.404ST strength, GE global efficiency, CC clustering coefficient, EVC eigenvector centrality. Pearson’s correlation of Graph Parameters comparison with 8 channels and whole channels in Siena database. ST strength, GE global efficiency, CC clustering coefficient, EVC eigenvector centrality. Table 5The selected channels in all combinations in Siena database.DeltaThetaAlphaBetaAll FrequencyFp1/Fp222329F3/F4334313F7/F822127C3/C4345517P3/P4431210P7/P8433212T7/T812227O1/O211125Significant values are in bold. The selected channels in all combinations in Siena database. Significant values are in bold. Discussion This study investigated the hypothesis that the global characteristics of a network can be maintained with a reduced number of channels by selecting optimal channels based on inferred networks derived from multichannel EEG signals. This study focused on epilepsy patients to explore the feasibility of reducing EEG channels while preserving global network characteristics associated with epilepsy mechanisms. Patients included in the study exhibited diverse seizure onset zones; however, the primary goal was not to localize these zones but rather to assess the network-level changes commonly observed in epilepsy. our method employs graph theoretical metrics to ensure that the reduced set of channels retains the essential functional properties of the entire network, rather than merely preserving local connectivity. By utilizing network-based correlation analysis, the proposed approach optimizes channel selection to ensure that the chosen subset effectively preserves the global efficiency and clustering properties of the full EEG network. The results showed that the global network characteristics of the network obtained by reducing the channels can be well preserved when compared with the network measured using the standard 19 channels through graph measurements of the network. Considering the user convenience of measuring brain signals while maintaining the global network properties, selecting four pairs of channels symmetrically in both hemispheres seem to be the optimal choice. In addition, the optimally selected channels were confirmed to be located in the midline frontal region and the parietal region. Furthermore, the results of the Siena database showed that the channels of the four pairs of channels in both hemispheres were selected in the midline frontal and the parietal region. Based on the results of this study, the generalizability of the optimal channel selection method was verified by applying it to other populations with the same pathology. In this study, generalizability is defined not by identical channel selections across datasets, but by the preservation of key network properties regardless of dataset variations. Specifically, generalizability is assessed based on the consistency of global network characteristics, including the correlation structure of graph metrics and the inclusion of core functional regions. Despite variations in the exact electrode selections, the retention of these essential network features supports the applicability of the method beyond the original dataset, demonstrating its potential for broader use in epilepsy-related electroencephalogram (EEG) analysis. From a network perspective, channel selection can be made independently of subject-specific characteristics if network characteristics are preserved after reducing the number of channels. The preservation of key network characteristics does not imply that the reduced network is identical to the full-channel network; rather, it retains essential connectivity patterns. While two networks may exhibit similar efficiency and clustering values, our analysis ensures that the selected subnetwork maintains the structural integrity of the original network by capturing key nodal and global properties. Figure 5 demonstrates that the reduced montage preserves the spatial organization of network connectivity, supporting the validity of our approach. The characteristics of the network can be measured by evaluating the efficiency of information exchange between regions in the global network regardless of nodal features, as the brain needs to function together in multiple regions22. In this sense, GE is an indicator of the efficiency of segregation and integration of functions in the brain network23. GE can be interpreted as a reflection of the efficiency of the network as a whole, even though the number of channels has been reduced from the perspective of global information transfer. The CC represents the degree of interconnection between regions through connections with neighboring nodes. Alternatively, it may reflect the density of neighboring nodes related to the characteristic path length24. Thus, the characteristics of CC would be maintained whether nodes are located with similar neighbors12. Therefore, even if channels with similar neighbors are selected, the characteristics of the global network are expected to be maintained through the CC, which indicates the degree of interconnectedness. EVC, a measure of the influence of a node on a network, can be used to identify the most important nodes25. Therefore, a subnetwork composed of these essential nodes is likely to reflect the characteristics of the nodes that make up the global network. The selected subchannels were chosen to preserve key network characteristics, including nodal (Strength, Eigenvector Centrality) and global (Clustering Coefficient, Global Efficiency) parameters. Notably, EVC demonstrates heightened sensitivity to local variations, thereby serving as a pivotal element in confirming that the reduced montage preserves significant functional connectivity patterns, as opposed to merely mirroring global signal trends. Collectively, these metrics represent local connectivity and overall network efficiency. While no single metric fully captures the complexity of functional networks, the strong correlation between parameters in full and reduced networks supports the reduced montage as a representative model. This model minimizes functional information loss while optimizing electrode placement. Consequently, it is thought that a technique that uses a variety of graph measurements, such as GE, CC, and EVC, to construct a sub-network with the minimum number of channels can preserve the characteristics of the global network while reducing the number of channels. The findings of this study indicate that the channel position of the sub-network with the capacity to reflect the global network may be situated within the key components of the default mode network (DMN) (Fig. 6). This observation is consistent with prior reports suggesting that the DMN plays a substantial role in the propagation of seizures in individuals diagnosed with epilepsy. Recent research has shown that epilepsy can be better diagnosed by examining it at the network level. During seizure progression, brain networks undergo dynamic modulations, resulting in increased regularity and decreased randomness in the ictal versus interictal states. In the preictal phase, the network rapidly evolves, with epileptic discharges propagating through specific subnetworks. Using preictal EEG signals, we identified core subnetworks in epilepsy patients prior to ictal onset. Our findings showed that a subnetwork maintaining connectivity between dorsomedial prefrontal cortex and parietal regions was the essential component indicating the preictal transition. In temporal lobe epilepsy, nodal strength decreased in the left medial orbital superior frontal gyrus but increased in the right inferior parietal gyrus, with centrality of the latter associated with seizure duration26. The recruitment ability was significantly reduced compared to the origin of left hippocampal sclerosis27. In cingulate gyrus epilepsy, there was increased connectivity, particularly between the left angular gyrus and posterior cingulate cortex28. The symmetrical channel pairs spanning both hemispheres, derived from the preictal network, speculated to belong to the default mode network, aligning with its recognized role in epilepsy dynamics29.Fig. 6Comparison between the default mode network and distribution of minimal reduced channel. (a) The default mode network, which is known to be the largest in the resting state, includes the prefrontal cortex, posterior cingulate cortex, and parietal lobe. (b) The distribution of the numerical region for channels with the reduction was demonstrated. The reduced channels included the prefrontal cortex (Fp1, Fp2, F3, and F4) and parietal regions (P3, P4, P7, and P8). Comparison between the default mode network and distribution of minimal reduced channel. (a) The default mode network, which is known to be the largest in the resting state, includes the prefrontal cortex, posterior cingulate cortex, and parietal lobe. (b) The distribution of the numerical region for channels with the reduction was demonstrated. The reduced channels included the prefrontal cortex (Fp1, Fp2, F3, and F4) and parietal regions (P3, P4, P7, and P8). Our study examined whether a network-based EEG channel selection method yields similar subnetworks across populations with shared epilepsy etiology. Although certain frequency combinations exhibited some variation, nodes encompassing the dorsomedial prefrontal cortex and parietal regions emerged across groups to represent the preictal state. Frontal lobe epilepsy patients showed reduced activity in the ventrolateral and dorsomedial prefrontal, angular gyrus, and anterior temporal regions, which disrupted semantic and emotional processing30. In temporal lobe epilepsy, the described network alterations are likely due to hippocampal connections with posterior cingulate and other default mode network nodes31. Seizures in the parietal lobe were most epileptogenic in the superior parietal lobule for most patients, with the remaining cases involving inferior parietal and opercular regions, indicating variable manifestations despite consistent localization32. Altered network between dorsomedial prefrontal and parietal regions are observed across focal epilepsies. Therefore, our subnetwork, encompassing frontal and parietal nodes, may signify the preictal transition toward ictal states regardless of population, capturing the resting network reconfiguration marking epileptic dynamics. Our study proposes a methodology to select a minimum set of EEG channels that captures epilepsy preictal network traits. While using fewer channels decreases spatial resolution, we found the sub-network preserved global graph characteristics. Selecting four symmetrical channel pairs per hemisphere enables convenient brain signal measurement while maintaining network properties. Despite differing subject populations, this channel set retained global dynamics, coinciding with default mode network regions implicated in seizure onset and propagation. Our reduction methodology extends beyond previous subject-specific approaches to enable application across cohorts. This practical method utilizes symmetry and intuitive channel pairs to facilitate clinical translation, with the potential to extend beyond focal onset localization. Furthermore, the sub-network derivation approach may offer a valuable tool for studying network dysfunction in neurological disorders, including epilepsy. While seizure onset zones vary across individuals, the objective of this study was not to focus on the localization of these zones but to demonstrate that network-level changes associated with epilepsy can be preserved with a reduced number of EEG channels. This approach highlights the potential for a generalized method applicable to studying epilepsy-related networks beyond focal onset localization. Limitation and suggestion Our study has a few limitations as follows. First, in this study, our approach derived functional connectivity networks from cortical EEG channels rather than brain regions at the source level. This means the selected channels might not fully represent all regions of the DMN when reconstructed at the source level. Future research should investigate these results by reconstructing EEG signals at the source level. A second limitation is that Our study exclusively analyzed EEG data from the preictal phase, where alterations in network topology occur prior to seizure onset. Further research is warranted to investigate the applicability of the proposed channel reduction methodology across different phases of the epileptic cycle, including interictal, ictal, and post-ictal stages, to establish its comprehensive generalizability. Lastly, this study’s focus on preserving network characteristics using a bilaterally symmetric channel selection limits its direct applicability for tasks like localizing the seizure onset zone, especially in unilateral focal epilepsy. Validating a channel selection strategy for such specific clinical applications remains an important area for future work. Future studies could explore whether the reduced montage retains sufficient spatial resolution for precise SOZ localization. Moreover, future studies should rigorously evaluate the impact of reduced channel configurations on the accuracy of detecting focal seizures and other region-specific neural activities. Conclusion This study demonstrates the feasibility of reducing EEG channels to four pairs while maintaining network characteristics essential for studying epilepsy using graph theory analysis. Notably, the selected channels significantly covered the Default Mode Network (DMN), known for its association with seizure activity. In addition, the results indicated that the methodology could be applied to different populations of epilepsy patients (i.e., Korea University Hospital database and Siena database). Therefore, the proposed method successfully applied to data from diverse epileptic patients, suggesting its potential generalizability. Future research exploring network dynamics could benefit from utilizing the identified channel locations and number, potentially replicating these findings. These results pave the way for the proposed channel reduction method to enhance understanding of various brain network-related diseases, while improving convenience and efficiency in clinical applications. Supplementary Information Supplementary Information. Supplementary Information.