Network topology in brain tumor patients with and without structural epilepsy: a prospective MEG study

Abstract

Background:

It was proposed that network topology is altered in brain tumor patients. However, there is no consensus on the pattern of these changes and evidence on potential drivers is lacking.

Objectives:

We aimed to characterize neurooncological patients’ network topology by analyzing glial brain tumors (GBTs) and brain metastases (BMs) with respect to the presence of structural epilepsy.

Methods:

Network topology derived from resting state magnetoencephalography was compared between (1) patients and controls, (2) GBTs and BMs, and (3) patients with (PSEs) and without structural epilepsy (PNSEs). Eligible patients were investigated from February 2019 to March 2021. We calculated whole brain (WB) connectivity in six frequency bands, network topological parameters (node degree, average shortest path length, local clustering coefficient) and performed a stratification, where differences in power were identified. For data analysis, we used Fieldtrip, Brain Connectivity MATLAB toolboxes, and in-house built scripts.

Results:

We included 41 patients (21 men), with a mean age of 60.1 years (range 23–82), of those were: GBTs (n = 23), BMs (n = 14), and other histologies (n = 4). Statistical analysis revealed a significantly decreased WB node degree in patients versus controls in every frequency range at the corrected level (p_1–30Hz = 0.002, p_γ = 0.002, p_β = 0.002, p_α = 0.002, p_θ = 0.024, and p_δ = 0.002). At the descriptive level, we found a significant augmentation for WB local clustering coefficient (p_1–30Hz = 0.031, p_δ = 0.013) in patients compared to controls, which did not persist the false discovery rate correction. No differences regarding networks of GBTs compared to BMs were identified. However, we found a significant increase in WB local clustering coefficient (p_θ = 0.048) and decrease in WB node degree (p_α = 0.039) in PSEs versus PNSEs at the uncorrected level.

Conclusion:

Our data suggest that network topology is altered in brain tumor patients. Histology per se might not, however, tumor-related epilepsy seems to influence the brain’s functional network. Longitudinal studies and analysis of possible confounders are required to substantiate these findings.

Keywords

brain metastasis brain tumor epilepsy magnetoencephalography network topology resting-state

Introduction

Several symptoms observed in patients with brain tumors, such as attention deficits or reduced psychomotor speed, affecting quality of life, cannot be explained solely by the location of the lesion but might be explained by changes in the underlying neuronal network.¹

Large-scale brain networks are often characterized by means of functional connectivity (FC) and network topology.^2–6 FC describes how information is transmitted locally as well as between different areas and is defined by statistical interdependencies of signals resulting from brain activity.^7,8 These signals can be measured by functional neuroimaging techniques such as functional magnetic resonance imaging (MRI), or neurophysiological procedures, such as electroencephalography (EEG) or magnetoencephalography (MEG).

Network topology describes how the elements defining the network are arranged and how they interact. It is based on a mathematical graph theory that applies to various networks. This model has been increasingly applied in the fields of cognitive and clinical neuroscience, showing that functional and neuronal networks have so-called small-world properties.^3,9,10 Small-world networks are characterized by short path lengths and high clustering coefficient.¹¹ Douw et al.¹⁰ calculated a small-world coefficient (ratio between local clustering coefficient and average shortest path length) from resting-state MEG and performed neuropsychological testing in healthy controls and found that higher small-worldness significantly correlated with better cognitive performance. The average shortest path length is a node property that indicates that nodes of a network relate to each other via short paths. The local clustering coefficient gives the fraction of a node’s neighbors that are neighbors of each other. Furthermore, a high clustering describes that the probability of several nodes to form connections is high.^9,12,13 The node degree is also a node property, another network topological parameter, which is an absolute value that specifies the number of connections of one single node.¹⁴

Among adult intraparenchymal brain tumors, brain metastases (BMs) are the most common, followed by gliomas.¹⁵ Even though prior studies already investigated the functional network of patients with glial brain tumors (GBTs), non-glial lesions and meningiomas, BMs – to the best of our knowledge – have not yet been investigated regarding their impact on the brain’s network.^16–25

GBTs can be roughly subdivided into low-grade gliomas (LGGs) and high-grade gliomas (HGGs), that both derive from neuroepithelial tissue, but differ in terms of malignancy/clinical outcome, growth patterns, molecular signature, and protein expression.²⁶ Some of these aspects have been previously investigated to identify reasons for network changes in brain tumor patients.^18,22,27 Inter alia it was proposed that histology and tumor grade may explain and directly impact the extent of network changes.^18,22 Van Dellen et al.²² found that LGGs had a different influence on the network characteristics than HGGs, namely decreased synchronizability and reduced global integration. Furthermore, some commonalities have also been described in patients with non-glial lesions, LGGs and HGGs. All these patients showed altered FC and impaired network topology, especially in the theta band, compared to healthy controls.^22,27 These changes were also associated with impaired cognition and were not found to be explicable by the tumors’ anatomical location.²²

In contrast to GBTs, BMs are secondary neoplasms that derive from completely different cell lines. Tumors from the lung, prostate, breast, melanoma, and renal cell carcinoma most commonly spread to the brain.¹⁵ The growth pattern of BMs is distinct from GBTs, which can be visualized radiologically and intraoperatively. High-grade, as well as LGGs mostly infiltrate the brain parenchyma, whereas BMs usually only displace the surrounding tissue. Due to these considerable differences between GBTs and BMs, it seems obvious that they would also reflect on the brain’s functional network. That appears even more so regarding the previously published data that described differently impacted networks of high-grade and LGGs. Therefore, an analysis of the networks of patients with BMs could lead to a better understanding of how networks become disrupted by brain tumors.

Another explanatory mechanism for network changes in brain tumor patients might be structural (symptomatic) epilepsy. Tumors causing structural epilepsy might have an even greater negative impact on the functional network compared to those that do not. It has been suggested that LGGs have a greater impact on network topology than HGGs.^17,20,22,27 This corresponds to the fact that LGGs also cause structural epilepsy more often than HGGs. Douw et al.²⁰ further investigated tumor-related epilepsy and found it to be interdependent with network characteristics and even to be related to protein expression. Previous MEG studies on epilepsy found increased FC for focal and generalized epilepsy, as well as evidence for an increase in seizure frequency by augmented dissociation of the functional network.^28,29 So far, however, there has not yet been any investigation published that compared the network topology of patients with versus without tumor related epilepsy.

Network changes, that is, could be the cause or consequence of diseases. It might be lesions, like brain tumors, or pathological processes, like in Alzheimer’s/Parkinson’s Disease or Epilepsy, that provoke alterations in functional networks.^28,30,31 Conversely, the distortion of the network could also generate the symptoms, or rather the diseases themselves.

Compared to the different impact of LGGs and HGGs on FC, we hypothesized that BMs would interact even more differently with their environment, respectively within the brain’s functional network, in comparison to the group of GBTs.

And since previous studies found the FC in epilepsy to be clearly impaired,^{14,28,29,32,33} we now aimed to further investigate the effect of tumor-related epilepsy on network topology.

Study objective

The main objective of this study was to investigate whether network topology is altered in patients with brain tumors and whether histology or the presence of structural epilepsy have roles in these alterations.

We also aimed to confirm results from previous studies by applying current and most common network parameters and intended to add further information to the understanding of (pathological) network topology by investigating BMs.

We hypothesized that there is a relevant difference in network topology of patients with brain tumors compared to healthy controls, that BMs have a different influence on the brain’s network compared to GBTs and that structural epilepsy relevantly influenced network topology compared to patients without symptomatic seizures.

Methods

Study population

We consecutively included all adult patients who were referred to the Department of Neurosurgery of the Christian-Doppler-University Hospital in Salzburg for neurosurgical resection or biopsy of a single, supratentorial tumor with MRI-radiological suspicion of either being a GBT or BMs from 4 February 2019 to 21 March 2021.

We excluded patients who had undergone prior neurosurgery or radiotherapy of the brain, patients with unrelated psychiatric, neurological, endocrine comorbidities, or implanted (non-titanium) metallic devices.

We matched every patient with a healthy control in gender and age (±5 years). The control group consisted of staff of the University Hospital Salzburg, as well as of students from the Paris Lodron University, Salzburg, Austria.

Both groups underwent a single 20 min resting-state MEG recording. For coregistration with MEG data, a template MRI was used for controls, whereas patients’ T1-pre- and post-gadolinium or T2-sequences were employed for the tumor group.

Clinical data were retrieved from patients’ electronic medical charts. In particular, the presence of structural epilepsy was recorded. In most patients with structural epilepsy (PSEs), tumors had been initially detected due to the occurrence of seizures. Whenever this was unclear, an EEG was performed, and a neurological consultation was conducted.

The patient cohort was divided into two subgroups: GBTs and BMs, and PSEs and patients without structural epilepsy (PNSEs).

MEG recordings

We recorded the brains’ magnetic signal at 1000 Hz (hardware filters: 0.1–330 Hz) in a standard passive magnetically shielded room (AK3b, Vacuumschmelze, Germany) using a whole head MEG (Neuromag Triux, MEGIN Oy, Finland). Signals are captured by 102 magnetometers and 204 orthogonally placed planar gradiometers at 102 different positions. All subjects wore nonmetallic clothes and had five head position indicator coils attached on the glabella, forehead, as well as preauricularly bilaterally. They all underwent head shape acquisition prior to the recording, and the head shape points were later used for coregistering the anatomical MRI scan. During the measurement, subjects were seated and instructed to keep their eyes open. We periodically checked if subjects stayed awake via a camera in the magnetically shielded room. Every measurement consisted of one block of 20 min.

Data analysis

To preprocess the data, we employed a signal space separation (SSS) algorithm that is implemented in the Maxfilter program (version 2.2.15), provided by the MEG manufacturer (MEGIN Oy, Espoo, Finland).³⁴ Here, external noise from the MEG signal is removed (mainly 16.6 Hz train mains supply, and 50 Hz plus harmonics) and data is realigned to a common standard head position (-trans default Maxfilter parameter). This is performed across different blocks based on the measured head position at the beginning of each block using the five coils that have been placed before the measurement.

We used Brainstorm [Version 181008 (08-Oct-2018)] to further preprocess the data.³⁵ At first, we applied a high-pass filter at 0.1 Hz (sixth order zero-phase Butterworth filter) to the continuous data. Subsequently, for signal space projection (SSP), continuous data were chunked in 10 s blocks.³⁶ The resulting data were manually scrutinized to identify remaining eye blinks, eye movements, heartbeat, and the 16 and 2/3 Hz train power supply artifacts and to create appropriate projectors. Finally, the artifactual components were projected out, and the continuous data was segmented in 2 s chunks. Then, data were again visually scrutinized to find bad epochs that were not corrected by the SSS and SSP combination, and further removed. An average of 15.4 (range 0–152.6) trials, corresponding to approximately 0.5 min were therefore removed, leaving roughly 19 min of artifact-free data per subject (range 15–20 min) that was used for further analysis.

The subsequent steps were then performed using the fieldtrip toolbox (git version 20180124).³⁷ At first, the data were downsampled to 300 Hz to speed up the following computations of the source space connectivity and graph theoretical values.

Then the coordinate frame of the MEG data was coregistered to a template MRI scan in the case of healthy controls, and to T1-contrast enhanced or T2 anatomical images (depending on best visualization of the tumor) in patients, using the right and left preauricular points, nasion, and the head shape points measured during preparation. A single-shell source model was generated based on the segmented anatomical scan and per subject source model with resolution 10 mm was created.³⁸ Next, the data were Fourier transformed for each of the frequency bands of interest (‘broadband’ = 1–30 Hz, gamma = 25–40 Hz, beta = 15–25, alpha = 8–15 Hz and theta = 4–7, delta = 1–4), was computed and used for a partial canonical coherence beamformer, with a noise covariance regularization factor of 5%.^37,39

From the resulting source activity, for each position of the source model, the component of the dipole moment projected onto the dominant direction was taken and subsequently used to compute the imaginary part of the coherence as an all-to-all whole brain (WB) voxel-wise connectivity metric. The 95th percentile of coherence values was used as a threshold for the construction of the adjacency matrix for computing node degree, local clustering coefficient, and average shortest path length. All source, connectivity, and network analyses were carried on using the fieldtrip toolbox (git version 20180124).³⁷

To minimize the spurious connectivity effects driven purely by the power difference, we decided to stratify the data to be sent to source space for connectivity and network analysis. For selecting the strata, two big pools containing the power spectral densities values were computed overall 2 s blocks on the magnetometers, for the patients and for the healthy controls, respectively. Then, the fieldtrip function ft_stratify (with the parameters method = ‘histogram’, equalbinavg = ‘yes’, numbin = 20) was employed to select the trials and equalize the two power spectral distributions, per each frequency band of interest.

Surgery and histopathology

All decisions on neurosurgical treatment (consisting of either microsurgical tumor resection or biopsy) were agreed upon in the interdisciplinary neuro-oncological tumor board. Patients gave their written, informed consent prior to surgery.

Histopathological examinations (frozen section, formalin-fixed and paraffin-embedded tissue, immunohistochemical, and molecular genetic analysis) were performed according to the 2016 World Health Organization Classification of Tumours of the central nervous system, the current edition available during the inclusion period/histological sampling.²⁶

Statistical analysis

For statistical analysis, we used Jamovi, Version 1 and R studio, Version 2022.02.1.^40,41 As a dependent variable, the maximum value of each network measure, per subject/patient and per frequency band of interest, was taken, that is node degree, local clustering coefficient, and average shortest path. We tested for differences in age in the patient population by means of Student’s t-test and Welch’s test. We designed a contingency table for one subgroup (PSE versus PNSE in relation to the respective histological classification of these patients) – for which it was applicable – and calculated Fisher’s exact test. For the comparison between the groups (1) patients versus controls, (2) GBTs versus BMs, and (3) PSEs versus PNSEs Shapiro–Wilk test for normality distribution was calculated. When the normality assumption was met, we performed independent Student’s t-tests, when it was violated, Mann–Whitney U-tests were employed. Consequently, effect sizes were calculated by Cohen’s d or Rank biserial correlation (RBC), depending on the test employed. As customary, for p-values lower than 0.05 the null hypothesis was rejected, and the test was considered significant. We additionally provide the Bayes Factors (BFs) to inform on the relative levels of evidence for the null/alternative hypotheses. At last, we corrected for multiple comparisons by calculating the false discovery rate (FDR).

Methods prior to statistical analysis are roughly summarized by a graphical depiction in Figure 1.

Figure 1.

Summary of data collection and analysis. Patients with glial brain tumors and BMs were consecutively included and matched with healthy controls. Presence of structural epilepsy was recorded. MEG examination was performed several days preoperatively, whole brain connectivity was analyzed, and resultant analysis of network topological parameters (node degree, average shortest path length, local clustering coefficient) was performed with Fieldtrip, Brain Connectivity MATLAB toolboxes,42 and in-house built scripts.

Results

Clinical data

A total of 42 patients were initially enrolled in the study. One patient diagnosed with acute disseminated encephalomyelitis by stereotactic biopsy had to be excluded from further analysis. Thus, the final study population consisted of 41 patients (21 men), as well as 41 healthy controls, who were matched in gender and age. Tumor histology included GBTs (n = 22), dysembryoplastic neuroepithelial tumor (DNET) (n = 1), BMs (n = 14), meningiomas (n = 2), and primary central nervous system lymphomas (PCNSLs) (n = 2). The patient with DNET was investigated together with patients who had GBTs, so that the final group of GBTs included 23 subjects.

None of the patients had undergone prior tumor treatment, besides corticosteroids or antiepileptogenic drugs at the time of MEG measurement.

Mean age in patients was 60.5 years (±15.2 years). T-test and Welch’s test revealed that there was no significant difference in age between patients with GBTs and BMs (p = 0.260 and p = 0.195, respectively).

A total of 27 patients were suffering from structural epilepsy: 15 with GBTs, 10 with BMs, and 2 with primary CNS lymphomas. Fisher’s exact test revealed no significant difference in the analysis of the contingency table.

Patient characteristics are shown in Tables 1 to 4.

Table 1.

Description of the study population on subject level.

Patient ID	Age (years)	Sex (m/f)	Histology	Presence of struct. epilepsy (y/n)	Control ID	Age (years)	Sex (m/f)
1	23	f	DNET	y	03	26	f
2	71	m	BM (Melanoma)	y	24	75	m
3	25	m	GBM IV IDH neg	y	33	29	m
4	70	m	BM (Lung)	n	34	67	m
5	67	f	GBM IV IDH neg	y	25	70	f
6	52	f	BM (Lung)	y	15	57	f
7	39	m	LGG NOS	n	14	40	m
8	67	m	GBM IV IDH neg	y	37	63	m
9	64	f	ODG II IDH pos	y	18	66	f
10	82	f	GBM IV IDH neg	n	21	77	f
11	51	m	AC III IDH neg	y	13	50	m
12	73	m	GBM IV IDH neg	y	29	74	m
13	64	m	GBM IV IDH neg	y	39	60	m
14	61	m	GBM IV IDH neg	n	32	58	m
15	73	m	PCNSL	n	26	78	m
16	58	m	BM (Neuroendoc.)	y	05	55	m
17	46	f	GBM IV IDH neg	y	09	39	f
18	56	m	GBM IV IDH neg	y	16	57	m
19	73	f	GBM IV IDH neg	n	10	68	f
20	51	f	BM (Breast)	y	11	56	f
21	76	f	BM (Renal)	n	20	71	f
22	70	f	GBM IV IDH neg	y	19	65	f
23	58	m	GBM IV IDH neg	n	02	55	m
24	54	m	ODG II IDH pos	y	38	51	m
25	78	m	GBM IV IDH neg	y	22	82	m
26	49	f	Meningioma I	y	08	52	f
27	78	f	GBM IV IDH neg	n	23	73	f
28	58	f	BM (Rectum)	y	04	58	f
29	61	f	BM (Adeno.phar.)	n	12	58	f
30	27	m	ODG II IDH pos	n	06	30	m
31	43	f	Meningioma I	y	31	45	f
32	72	m	GBM IV IDH neg	n	01	71	m
33	63	m	BM (Lung)	n	27	58	m
34	65	m	GBM IV IDH neg	y	07	66	m
35	55	f	BM (Breast)	y	30	60	f
36	38	f	AC III IDH neg	y	28	33	f
37	78	f	PCNSL	n	36	77	f
38	59	m	BM (Renal)	y	35	63	m
39	76	f	BM (Lung)	y	40	73	f
40	77	m	BM (Lung)	y	41	72	m
41	69	f	BM (Lung)	y	42	65	f

Latin numbers I–IV indicate WHO grade of the respective tumor.

AC, Astrocytoma; BM, brain metastasis; DNET, dysembryoplastic neuroepithelial tumor; f, female; GBM, Glioblastoma multiforme; IDH, Isocitrate dehydrogenase; m, male; n, no; NOS, not other specified; ODG, Oligodendroglioma; PCNSL, primary CNS lymphoma; y, yes.

Table 2.

Characteristics of the patient- and control group.

Characteristic	GBT	BM	Meningioma	PCNSL	Controls
No. patients	23	14	2	2	41
Mean age (years)	58.2	64.1	46	75.8	60.8
Gender
No. male	14	6	0	1	21
No. female	9	8	2	1	20

BM, brain metastasis; GBT, glial brain tumor; PCNSL, primary CNS lymphoma.

Table 3.

Characteristics of the patient group.

Lesion type		No. patients	Mean age (years)	No. male	No. female
Glial brain tumors	HGG	18	60	11	7
	LGG	4	46	3	1
	DNET	1	21	0	1
Brain metastases	Lung	6	68	3	3
	Other	3	59	1	2
	Breast	2	53	0	2
	Renal	2	68	1	1
	Melanoma	1	70	1	0

DNET, dysembryoplastic neuroepithelial tumor (this patient was analyzed together with the GBTs); HGG, high-grade glioma; LGG, low-grade glioma.

Table 4.

Characteristics of the patients with and without structural epilepsy-split by histology.

Histology	No. patients with structural epilepsy	No. patients without structural epilepsy	Total
Struct. epilepsy	No. patients with structural epilepsy	No. patients without structural epilepsy	Total
No. GBTs	15	8	23
No. BMs	10	4	14
No. Meningiomas	2	0	2
No. PCNSL	0	2	2
Total	27	14	41

Fisher’s exact test revealed no significant difference for the contingency table (p = 0.287).

BMs, brain metastases; GBTs, glial brain tumors; PCNSL, primary CNS lymphoma.

Network characteristics

We computed the resting state power spectral density analysis for patients and healthy controls that revealed significant differences (p < 0.001, RBC = 0.65, BF₁₀ = 907). Therefore, to minimize the spurious connectivity effects driven purely by the power difference, we decided to stratify the data to be sent to source space for connectivity and network analysis. In the literature, there seems to be no standard way to summarize the WB network parameters, being mean, median, or maximum. Here, for simplicity, we report only the maximum values of the WB network parameters, but the usage of the mean and median delivers similar results.⁴³

Results for all network topological parameters analyzed are shown in Table 5 (patients and controls) and seven (patients with and patients without structural epilepsy).

Table 5.

Results from network topological analysis summarized in groups (patients and controls) for each parameter’s maximum value and frequency band.

Frequency band	Network characteristic	Max
Frequency band	Network characteristic	Patients	Controls
Broadband	ND	227^a	287^a
	CC	0.236^a	0.217^a
	L	2.70	2.69
Gamma	ND	302^a	391^a
	CC	0.270	0.265
	L	2.78	2.75
Beta	ND	288^a	397^a
	CC	0.281	0.269
	L	2.83	2.82
Alpha	ND	300^a	398^a
	CC	0.326	0.314
	L	2.96	2.96
Theta	ND	348^a	399^a
	CC	0.372	0.334
	L	3.09	3.07
Delta	ND	367^a	454^a
	CC	0.389^a	0.338^a
	L	3.14	3.20

Maximum values are given.

Values that were found to be significant in later analysis.

CC, local clustering coefficient; L, average shortest path length; ND, node degree.

We found a highly significant lower node degree at the uncorrected level in patients compared to healthy controls (p_1–30Hz < 0.001, d = −0.86, BF₁₀ = 110.80; p_γ = 0.002, d = −0.72, BF₁₀ = 19.35; p_β < 0.001, d = −1.13, BF₁₀ = 6458.72; p_α < 0.001, d = −0.99, BF₁₀ = 767.93; p_θ = 0.024, RBC = 0.29, BF₁₀ = 2.59; p_δ < 0.001, RBC = 0.42, BF₁₀ = 27.26; depicted in Table 6). After the correction for multiple comparison, we still found a highly significant result in all six frequency bands as shown in Figure 1. Further analysis at the uncorrected level demonstrated significantly higher local clustering coefficient (p_1–30Hz = 0.031, RBC = 0.49, BF₁₀ = 1.83; p_δ = 0.013, d = 0.32, BF₁₀ = 9.07) in patients compared to healthy controls. However, the latter results of higher local clustering coefficient in patients showed a tendency but did not stand the correction (p_1–30Hz = 0.093, p_δ = 0.078).

Table 6.

Results from Student’s t-test or Mann–Whitney U-test comparison (depending on normality distribution) at the uncorrected level between patients and healthy controls.

Frequency band	Network characteristic	p-Value	Effect size	BF₁₀
Broadband^a	ND^a	<0.001	d = −085	110.80
	CC^a	0.031	d = 0.49	1.83
	L	0.623	d = 0.11	0.26
Gamma^a	ND^a	0.002	d = −0.72	19.35
	CC	0.875	RBC = 0.02	0.23
	L	0.523	d = 0.14	0.28
Beta^a	ND^a	<0.001	d = −1.12	6458.72
	CC	0.479	d = 0.15	0.29
	L	0.711	d = 0.08	0.24
Alpha^a	ND^a	<0.001	d = −0.99	767.93
	CC	0.705	RBC = 0.05	0.30
	L	0.796	RBC = 0.03	0.23
Theta^a	ND^a	0.024	RBC = 0.29	2.59
	CC	0.08	d = 0.39	0.89
	L	0.717	d = 0.08	0.24
Delta^a	ND^a	<0.001	RBC = 0.42	27.26
	CC^a	0.013	RBC = 0.32	9.07
	L	0.218	d = −0.27	0.45

Values that were found to be significantly different between groups.

BF10, Bayes factor₁₀; CC, local clustering coefficient; d, Cohen’s d; L, average shortest path length; ND, node degree; RBC, Rank biserial correlation

For the comparison of the two tumor groups (GBTs versus BMs, we excluded the four patients with primary CNS lymphomas and meningiomas, respectively, leaving 23 patients in the cohort of GBTs and 14 patients in the cohort of BMs (Table 2)). There was no discrepancy in power between the two groups (p = 0.676, RBC = 0.09, BF₁₀ = 0.41), and in further network analysis, we did not find a significant difference in any of the parameters and frequency bands.

We also compared patients with against PNSEs. For this analysis, we investigated all 41 patients – 27 patients with versus 14 PNSEs (these patient and network characteristics are depicted in Tables 4 and 7). Here no power spectral difference between the groups was found (p = 0.091, RBC = 0.33, BF₁₀ = 0.66), so that no further stratification was needed. There was significantly higher maximum local clustering (p_θ = 0.048, RBC = 0.38, BF₁₀ = 1.12), but significantly lower maximum degrees (p_α = 0.039, RBC = 0.40, BF₁₀ = 1.09) at the uncorrected level in PSE (Table 8). However, after calculating the FDR, the two results did not stand the correction (local clustering coefficient: p_θ = 0.288; node degree: p_α = 0.095).

Table 7.

Results from network topological analysis summarized in groups (patients with and without structural epilepsy) for each parameter’s maximum value and frequency band.

Frequency band	Network characteristic	Max
Frequency band	Network characteristic	PSE	PNSE
Broadband	ND	210	250
	CC	0.214	0.215
	L	2.70	2.68
Gamma	ND	299	351
	CC	0.260	0.273
	L	2.87	2.80
Beta	ND	291	343
	CC	0.282	0.289
	L	2.90	2.91
Alpha^a	ND^a	306^a	354^a
	CC	0.325	0.340
	L	2.99	2.97
Theta^a	ND	370	373
	CC^a	0.400^a	0.361^a
	L	3.19	3.15
Delta	ND	342	390
	CC	0.357	0.363
	L	3.16	3.11

Maximum values are given.

Values that were found to be significant in later analysis.

CC, local clustering coefficient; L, average shortest path length; ND, node degree; PNSEs, patients without structural epilepsy; PSEs, patients with structural epilepsy.

Table 8.

Results from Student’s t-test or Mann–Whitney U-test comparison (depending on normality distribution) at the uncorrected level between patients with and without structural epilepsy.

Frequency band	Network characteristic	p-Value	Effect size	BF₁₀
Broad band	ND	0.055	RBC = 0.37	2.22
	CC	0.935	d = −0.02	0.32
	L	0.488	RBC = 0.14	0.37
Gamma	ND	0.079	d = 0.60	1.12
	CC	0.581	d = −0.18	0.36
	L	0.148	d = 0.49	0.74
Beta	ND	0.078	d = −0.60	1.12
	CC	0.881	RBC = 0.03	0.33
	L	0.923	d = −0.03	0.32
Alpha^a	ND^a	0.039^a	RBC = 0.40	1.09
	CC	0.470	d = −0.24	0.39
	L	0.712	d = 0.12	0.34
Theta^a	ND	0.956	RBC = 0.01	0.32
	CC^a	0.048^a	RBC = 0.38	1.12
	L	0.494	d = 0.23	0.38
Delta	ND	0.067	RBC = 0.35	0.66
	CC	0.786	d = −0.09	0.33
	L	0.694	RBC = 0.08	0.41

Values that were found to be significantly different between groups.

BF, Bayes Factor₁₀; CC, local clustering coefficient; L, average shortest path length; ND, node degree; RBC, Rank biserial correlation (effect size).

Results are listed in Tables 6 (comparison of patients and healthy controls) and 8 (comparison of patients with versus patients without structural epilepsy) as well as in Figures 2 and 3.

Figure 2.

T-test or Mann–Whitney U-test comparison (depending on normality distribution) between patients (red) and healthy controls (blue) revealed significantly lower node degree in patients at the corrected level (p_1–30Hz = 0.0015, p_γ = 0.0024, p_β = 0.0015, p_α = 0.0015, p_θ = 0.024, p_δ = 0.0015).

Figure 3.

Representative coherence plot thresholded at 75% of the maximum respective coherence of a patient and its matched control: on the left is patient 2, who was diagnosed of a metastasis in the right parietal lobe, on the right is its matched control depicted.

Discussion

It was previously described that different kinds of neurological disorders, such as brain tumors lead to deviations in the brain’s functional network.^{16,18–20,22,24,31,44–48}

In this prospective study, we investigated 41 patients with brain tumors and 41 matched healthy controls. We were able to identify significant differences in the network topology of patients compared to healthy controls, as well as between patients with and without structural epilepsy. Unexpectedly, we found no differences in the functional networks of patients with GBTs compared to BMs.

Comparison of patients and healthy controls

In the comparison of patients and healthy controls, we found a significantly lower node degree in patients in all frequencies analyzed. This result is also consistent with the definition of node degree since the impact of the tumor on the functional network will impair the formation of connections and therefore lead to lower degree networks.¹⁴

We found an increased local clustering coefficient in brain tumor patients at the descriptive level, but this result did not stand up to correction for multiple testing.

The local clustering coefficient is a value that is defined by the probability that neighbors of a node are connected to another. It is an essential characteristic of the small-world concept that is associated with efficient physiological networks.^2,9,10 This result appeared unusual at first sight, since an intact network is assumed to have high local clustering, and we would have expected these parameters to be decreased in a pathologically disturbed network, as it is the case in a patient with a brain tumor. Several previously conducted studies, however, have similarly described increased clustering in brain tumor patients.^17,19,22 Bosma et al.¹⁷ found a higher local clustering coefficient within the theta band in patients with LGGs compared to healthy controls. They contextualized this result in the setting of other neurological diseases such as major depression, autism disorders or Alzheimer’s disease, for which previous investigations have found similar network changes. Also, Derks et al. described higher global clustering in glioma patients compared to controls. They further investigated this finding by differentiating various factors (e.g. age, histology, local and global clustering) but despite this, could not find any relation to increased clustering of any of these aspects.¹⁹ As such, it was concluded that the probable cause of the increased clustering lay with diffuse, WB alterations in these patients.

The augmentation might also be an expression of reactive changes of the network to the tumor. This assumption seems at least conclusive from a clinical point of view since lesions also provoke morphological and immunochemical reactions in the surrounding parenchyma that can be displayed radiologically and by the laboratory. On a functional level, tumor growth may lead to a partial or complete disconnection of network segments. In at least some of these, this then abolished input may have been largely inhibitory, resulting in disinhibition of these partitions. The consequently increased local activity may then strengthen connections between the constituting neuronal populations by means of neuroplasticity. The result would then manifest as higher local clustering coefficients. Hypothetically, such alterations could also be correlated to tumor-related epileptic activity.⁴⁹ Overall, however, these interpretations remain largely speculative, as differences did not reach statistical significance.

Comparison of patients with GBTs and BMs

We further performed two analyses solely within the patient cohort. The first one between patients with GBTs versus BMs. For this, we excluded four patients, due to their histological diagnoses – two showed PCNSL, and the other two had meningiomas. The latter two patients were primarily included since both had an oncological history and from a neuroradiological point of view it had not been possible to rule out a BM.

There was one patient, on whom we had the initial suspicion of an LGG, which then turned out to be a DNET. Although DNETs are a distinct tumor entity and do not belong to the group of gliomas, we kept this case included within our cohort of GBTs because of clinically similar behavior.

Since data on the impact of BMs on the brain’s functional network is lacking, we chose to investigate those patients in this context. We expected metastases to show different patterns of network changes, due to the completely different type of tumorigenesis and growth patterns. However, this hypothesis cannot be supported by our data since no difference in the functional networks of GBTs and BMs was found, neither in the variables nor the frequencies analyzed.

Previous studies had found different network patterns in HGGs and LGGs, which had led to the conclusion that histology plays an important role in network changes.^18,22 Derks et al. even found significant differences in the immunohistochemical profile of tumors with lower alpha FC in IDH-wildtype gliomas compared to IDH-mutated gliomas. Nonetheless, they did not find differences in FC in the theta band between the two groups, and there was also no difference in the alpha band in the comparison of classic histological grading between grade II/III and IV.¹⁸

Our results might be limited by the fact that the two tumor groups were not exactly equally distributed. This is due to the prospective setting of this study and the fact that single (supratentorial) BMs are rare.

Still, it is a noticeable finding, and if metastases and gliomas are indeed comparable regarding their impact on network topology, then the main factors may be the mass effect or growth rate, rather than histology. Metastases usually show less invasive growth patterns compared to intrinsic brain tumors but may cause substantially more mass effect due to common extensive peritumoral edema. Therefore, if these tumors do not lead to different changes in the functional network, histology might not be an explanation for the disturbed network.

Comparison of patients with and without structural epilepsy

Another possible mechanism of the distortion of the functional network could be the presence of seizures. We, therefore, also conducted a comparison between patients with and without structural epilepsy. Here, we found significantly higher local clustering and lower node degree in PSEs at the uncorrected level. After calculating the FDR, these results did not retain significance.

However, we still want to discuss the finding of the lower node degree in PSEs, since it stands in contrast to previous studies that presented data from patients with drug-resistant focal epilepsies that both had used the imaginary part of the coherence as coupling metric for their FC analysis.^14,33 In their data, FC and node degree were augmented in patients with epilepsy, Vogel et al. ¹⁴ also conducted a comparison to healthy controls. One explanation for this contradiction might be that tumor-related structural epilepsy has a different impact on the network than non-tumor-related focal epilepsy. Patients with epilepsy mostly suffer from their disease for many years, in contrast to patients with brain tumors who are primarily structural with seizures and shortly after receiving diagnosis and treatment for the lesion. In the situation of long-lasting epilepsy, with many events of seizures, the network might distort and adapt completely differently than in structural epilepsy. Similarly to our results with a higher local clustering coefficient in PSEs another group found an increase in the small world index in their patient cohort of focal epilepsies compared to healthy controls.⁵⁰ They also discussed contradictory results of earlier publications and concluded that the relationship of graph metrics and pathophysiological mechanisms of epilepsy is far from clear.

In this study, we applied graph theoretical measures since they have proven to be a valuable tool to characterize the network topology of the brain.^2,9,51,52 Some of the most current network topology parameters are local clustering coefficient, path length, and node degree.^{12,14,17,22,24} Until now, there is only one review on the topic by Semmel et al.,²¹ who found that the methodology is often not comparable, and the results are sometimes even contradictory between different publications.

Therefore, we also carried out FC and network topology calculations in source space, compared to other previous studies that computed the connectivity measures in sensor space only. It is crucial to measure in source space, especially in patients with morphological brain pathologies.⁵³

Whenever we found significant results in a network topological parameter, the increase/ decrease compared to healthy controls was concordant in the other frequency bands. To us, this supports the reliability of the results.

On the contrary, another study found, for example, an increase of synchronization and local clustering in theta, but a decrease in beta.¹⁷ They argued that there was more small world organization in lower frequency bands than in higher ones.

It seems that this aspect needs further investigation.

Strengths and limitations

This is a prospective, experimental, single-center study that aims to further increase the neurophysiological understanding of how brain tumors disturb the brain’s functional network.

The study population might be small but is comparable to other studies.^{16–19,22–24} Due to the prospective setting and because we included patients in a consecutive fashion, subjects were not exactly equally distributed in the two groups (GBTs versus BMs and patients with versus without structural epilepsy). On the one hand, this hampers the comparableness between groups, but on the other side we can provide higher data quality and therefore higher significance.

We put a focus on rigorous methodology for best possible comparability with previous and future studies. Therefore, we only investigated patients with a newly diagnosed, singular, supratentorial mass lesion, suspected to be a primary brain tumor or BMs. Furthermore, we arranged MEG measurement for all patients a few days, and up to a maximum of 2 weeks prior to surgery. In this way, we aimed to prevent possible influences of previous treatment, the presence of new lesions or additional tumor growth on the results of network analysis. A defined timing of MEG with respect to treatment can help to define standards in network analysis.

To ensure the significance of our results, we corrected for multiple comparisons, as we compared six different frequency bands for every network parameter. After this, the comparison of node degree in patients and healthy controls was still highly significant. Admittedly, the increased local clustering coefficient in patients was not found to be persistently significant, but there was still a noticeable tendency to be seen. The corrected results of the comparison between patients with and without structural epilepsy were not found to be significant. Nevertheless, we mentioned these results because they differ from previous publications, and we expect this distinction to be an interesting aspect for future investigations.

So far, there exist only seven other studies on the topic of FC and network topology measured by MEG in brain tumors.^{16–19,22–24} Although previous studies tried to draw conclusions from the influence of histology on network alterations, BMs have not been investigated, though, from a pathohistological point of view, they differ the most from primary brain tumors. We aimed to close this gap through our analysis; however, our results did not support the role of histology in network disruption.

Conclusion

We provide data of a prospective study on patients with brain tumors with and without structural epilepsy compared to healthy controls. We found significant differences in several network parameters in the broad band, as well as in all five different frequency bands. We did not find differences in the functional network of primary brain tumors compared to BMs, but there were significant differences at the descriptive level between patients suffering from structural epilepsy compared to those that did not.

Originating from the results of this study, the extent of network disruption in patients with brain tumors seems to be influenced more by the presence of structural epilepsy, than by the histology.

We advocate further prospective studies on even larger cohorts of patients. Especially the differentiation of PSEs, patients suffering from epilepsy and the comparison to healthy controls could be essential.

Footnotes

Acknowledgements

To all my colleagues, friends and family, and the students from Paris Lodron University who took their time to participate in this study as healthy controls and to Dr. Ciara O’Sullivan for providing language help.

Declarations

ORCID iD

Barbara Ladisich

References

Bartolomei

Bosma

Klein

, et al. Disturbed functional connectivity in brain tumour patients: evaluation by graph analysis of synchronization matrices. Clin Neurophysiol 2006; 117: 2039–2049.

Bullmore

Sporns

. Complex brain networks: graph theoretical analysis of structural and functional systems. Nat Rev Neurosci 2009; 10: 186–198.

Stam

de Haan

Daffertshofer

, et al. Graph theoretical analysis of magnetoencephalographic functional connectivity in Alzheimer’s disease. Brain 2009; 132: 213–224.

Stam

. Characterization of anatomical and functional connectivity in the brain: a complex networks perspective. Int J Psychophysiol 2010; 77: 186–194.

Stam

van Straaten

ECW

. The organization of physiological brain networks. Clin Neurophysiol 2012; 123: 1067–1087.

Stam

. Emergence of modular structure in a large-scale brain network with interactions between dynamics and connectivity. Front Comput Neurosci 2010; 4: 133.

Guggisberg

Honma

Findlay

, et al. Mapping functional connectivity in patients with brain lesions. Ann Neurol 2008; 63: 193–203.

Stam

Nolte

Daffertshofer

. Phase lag index: assessment of functional connectivity from multi channel EEG and MEG with diminished bias from common sources. Hum Brain Mapp 2007; 28: 1178–1193.

Stam

Reijneveld

. Graph theoretical analysis of complex networks in the brain. Nonlinear Biomed Phys 2007; 1: 3.

10.

Douw

Schoonheim

Landi

, et al. Cognition is related to resting-state small-world network topology: an magnetoencephalographic study. Neuroscience 2011; 175: 169–177.

11.

Barahona

Pecora

. Synchronization in small-world systems. Phys Rev Lett 2002; 89: 054101.

12.

Boccaletti

Bianconi

Criado

, et al. The structure and dynamics of multilayer networks. Phys Rep 2014; 544: 1–122.

13.

Watts

Strogatz

. Collective dynamics of ‘small-world’ networks. Nature 1998; 393: 440–442.

14.

Vogel

Kaltenhäuser

Kim

, et al. MEG node degree differences in patients with focal epilepsy vs. controls—influence of experimental conditions. Brain Sci 2021; 11: 1590.

15.

Nayak

Lee

Wen

. Epidemiology of brain metastases. Curr Oncol Rep 2012; 14: 48–54.

16.

Bartolomei

Bosma

Klein

, et al. How do brain tumors alter functional connectivity? a magnetoencephalography study. Ann Neurol 2006; 59: 128–138.

17.

Bosma

Reijneveld

Klein

, et al. Disturbed functional brain networks and neurocognitive function in low-grade glioma patients: a graph theoretical analysis of resting-state MEG. Nonlinear Biomed Phys 2009; 3: 9.

18.

Derks

Kulik

Wesseling

, et al. Understanding cognitive functioning in glioma patients: the relevance of IDH-mutation status and functional connectivity. Brain Behav 2019; 9: e01204.

19.

Derks

Kulik

Numan

, et al. Understanding global brain network alterations in glioma patients. Brain Connect 2021; 11: 865–874.

20.

Douw

de Groot

van Dellen

, et al. Local MEG networks: the missing link between protein expression and epilepsy in glioma patients? NeuroImage 2013; 75: 195–203.

21.

Semmel

Quadri

King

. Graph theoretical analysis of brain network characteristics in brain tumor patients: a systematic review. Neuropsychol Rev 2022; 32: 651–675.

22.

van Dellen

Douw

Hillebrand

, et al. MEG network differences between low- and high-grade glioma related to epilepsy and cognition. PLoS ONE 2012; 7: e50122.

23.

van Nieuwenhuizen

Douw

Klein

, et al. Cognitive functioning and functional brain networks in postoperative WHO grade I meningioma patients. J Neurooncol 2018; 140: 605–613.

24.

Wang

Douw

Hernández

, et al. Effect of tumor resection on the characteristics of functional brain networks. Phys Rev E 2010; 82: 021924.

25.

Yang

Kudulaiti

Chen

, et al. Within and beyond the visual cortex: brain tumors induce highly sensitive plasticity of visual processing in whole-brain neural functional networks. Cereb Cortex 2022; 32: 4422–4435.

26.

Louis

Perry

Reifenberger

, et al. The 2016 World Health Organization classification of tumors of the central nervous system: a summary. Acta Neuropathol 2016; 131: 803–820.

27.

Douw

van Dellen

de Groot

, et al. Epilepsy is related to theta band brain connectivity and network topology in brain tumor patients. BMC Neurosci 2010; 11: 103.

28.

Douw

DeSalvo

Tanaka

, et al. Dissociated multimodal hubs and seizures in temporal lobe epilepsy. Ann Clin Transl Neurol 2015; 2: 338–352.

29.

Li Hegner

Marquetand

Elshahabi

, et al. Increased functional MEG connectivity as a hallmark of MRI-negative focal and generalized epilepsy. Brain Topogr 2018; 31: 863–874.

30.

Engels

MMA

van der Flier

Stam

, et al. Alzheimer’s disease: the state of the art in resting-state magnetoencephalography. Clin Neurophysiol 2017; 128: 1426–1437.

31.

Hyder

Jensen

Højlund

, et al. Functional connectivity of spoken language processing in early-stage Parkinson’s disease: an MEG study. NeuroImage Clin 2021; 32: 102718.

32.

Rampp

Rössler

Hamer

, et al. Dysmorphic neurons as cellular source for phase-amplitude coupling in Focal Cortical Dysplasia Type II. Clin Neurophysiol 2021; 132: 782–792.

33.

Rampp

Kaltenhäuser

Müller-Voggel

, et al. MEG node degree for focus localization: comparison with invasive EEG. Biomedicines 2023; 11: 438.

34.

Taulu

Simola

. Spatiotemporal signal space separation method for rejecting nearby interference in MEG measurements. Phys Med Biol 2006; 51: 1759–1768.

35.

Tadel

Baillet

Mosher

, et al. Brainstorm: a user-friendly application for MEG/EEG analysis. Comput Intell Neurosci 2011; 2011: 1–13.

36.

Uusitalo

Ilmoniemi

. Signal-space projection method for separating MEG or EEG into components. Med Biol Eng Comput 1997; 35: 135–140.

37.

Oostenveld

Fries

Maris

, et al. FieldTrip: open source software for advanced analysis of MEG, EEG, and invasive electrophysiological data. Comput Intell Neurosci 2011; 2011: 1–9.

38.

Nolte

. The magnetic lead field theorem in the quasi-static approximation and its use for magnetoencephalography forward calculation in realistic volume conductors. Phys Med Biol 2003; 48: 3637–3652.

39.

Schoffelen

J-M

Gross

. Source connectivity analysis with MEG and EEG. Hum Brain Mapp 2009; 30: 1857–1865.

40.

S¸ahin

Aybek

. Jamovi: an easy to use statistical software for the social scientists. Int J Assess Tools Educ 2019; 6: 670–692.

41.

Wickham

Averick

Bryan

, et al. Welcome to the Tidyverse. J Open Source Software 2019; 4: 1686.

42.

Rubinov

Sporns

. Complex network measures of brain connectivity: Uses and interpretations. Neuro Image 2010; 52: 1059–1069.

43.

Hillebrand

Stam

. Recent developments in MEG network analysis. In: Supek

Aine

(eds) Magnetoencephalography. Cham: Springer International Publishing, pp. 631–645.

44.

Cho

Peck

Gene

, et al. Resting-state functional MRI language network connectivity differences in patients with brain tumors: exploration of the cerebellum and contralesional hemisphere. Brain Imaging Behav 2021; 16: 252–262.

45.

De Baene

Rutten

GJM

Sitskoorn

. The temporal pattern of a lesion modulates the functional network topology of remote brain regions. Neural Plast 2017; 2017: 1–11.

46.

D’Souza

Hirt

Ormond

, et al. Retrospective analysis of hemispheric structural network change as a function of location and size of glioma. Brain Commun 2021; 3: fcaa216.

47.

Huang

Zhang

, et al. Disturbed small-world networks and neurocognitive function in frontal lobe low-grade glioma patients. PLoS ONE 2014; 9: e94095.

48.

Crosson

, et al. White matter network topology relates to cognitive flexibility and cumulative neurological risk in adult survivors of pediatric brain tumors. Neuroimage Clin 2018; 20: 485–497.

49.

Pedersen

Omidvarnia

Walz

, et al. Increased segregation of brain networks in focal epilepsy: an fMRI graph theory finding. Neuroimage Clin 2015; 8: 536–542.

50.

Hatlestad-Hall

Bruña

Syvertsen

, et al. Source-level EEG and graph theory reveal widespread functional network alterations in focal epilepsy. Clin Neurophysiol 2021; 132: 1663–1676.

51.

Heimans

Reijneveld

. Factors affecting the cerebral network in brain tumor patients. J Neurooncol 2012; 108: 231–237.

52.

Reijneveld

Ponten

Berendse

, et al. The application of graph theoretical analysis to complex networks in the brain. Clin Neurophysiol 2007; 118: 2317–2331.

53.

Stufflebeam

. Clinical magnetoencephalography for neurosurgery. Neurosurg Clin N Am 2011; 22: 153–167.