114 IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS, VOL.

11, 2025

A Novel Spectral Graph Distance Measure and Its

Applications in Biomedical Signal Processing
Jyoti Maheshwari , Shiv Dutt Joshi , and Tapan K Gandhi

Abstract—In many real time scenarios, it becomes necessary to requires node correspondence [4]. Whereas spectral distances
compare complex networks. Graph distances measure the degree of are defined using eigenvalues and eigenvectors of the edge
similarity between any two networks. We conjecture that construc- weight matrices [2], [3]. It captures the global changes in the
tion of the edge weight matrix of any network should consider the
underlying physics of the problem. In addition to that, the selection networks.
of eigenvectors chosen for analysis also play an important role in Specifically, we explore spectral distances between any two
comparing any two networks. In this paper, we explore the most complex networks. There are various types of spectral distances
naturally formed networks in various real time applications, such based on the properties of eigenvalues and eigenvectors. One of
as tracking the transitions of brain states and real time seizure such graph distance is lp distance on eigenvalues. As we know
onset detection, and propose a graph distance measure based on
the spectral characteristics of the edge weight matrix. We have that ordering of the nodes plays a crucial role in finding the dis-
conducted several experiments to show the efficacy of the proposed tance between two complex networks, permutation invariant dis-
graph distance measure. It has an application in comparing any two tance is needed. Interestingly, eigenvalues are found to be permu-
brain networks while studying brain dynamics. Our results clearly tation invariant, and hence, eigenvalues based distances are very
show that the proposed graph distance outperforms the existing useful [2], [4], [5]. Other well known spectral distance is span-
graph distances.
ning tree dissimilarity which measures the degree of similarity
Index Terms—Graph distances, sub-band CRV, brain networks, by comparing their number of spanning trees between any two
spectral distance, eigenvectors. networks [6]. The major lacuna with the existing spectral graph
distances is that they do not consider the physics of the underly-
I. INTRODUCTION ing problem. We conjecture that there are many non unique ways
OMPLEX networks appear in diverse real world appli- in which complex networks can be formed and construction of
C cations ranging from the study of biological networks to
analysing computer networks [1]. Study of complex networks
real world networks should depend on the physics of the underly-
ing problem. In addition, we hypothesize that the information of
requires some kind of a similarity measure for comparing two only a few eigenvalues and eigenvectors would suffice to find the
networks which has been an active area of research. In order distance between any two networks. Selection of the eigenvalues
to measure the degree of similarity between any two networks, and eigenvectors should depend on what physical entity they
many graph distances are proposed in the literature such as: represent.
graph edit distances, matrix distances, spectral distances and In this endeavour, we have explored the most naturally formed
feature based distances etc [2], [3], [4]. In particular, graph edit dynamic networks in various real world applications. We found
distances measure the similarity between networks by finding that the used connectivity measure and the selection of eigen-
the number of structural changes required to make the networks vectors play an important role in comparing any two real time
isomorphic. One of the famous graph edit distances is Hamming dynamic networks. In this direction, we propose a novel graph
distance [2]. It is defined as the number of inserted and deleted distance measure which can be used to find the degree of simi-
edges required to transform one graph into another. It captures larity between any two graphs having node correspondence. We
the local changes in the network. On the other hand, matrix find the similarity between networks by calculating the distance
distance is defined as the lp norm of the difference between between their sub-band characteristic response vector (sub-band
edge weight matrices of the two brain networks. Matrix distance CRV) formed by summing a few relevant eigenvectors weighted
with their corresponding eigenvalues. The selection of eigenvec-
Received 11 January 2024; revised 15 September 2024 and 9 December 2024;
tors to form the sub-band CRV depends on what physical entity
accepted 10 January 2025. Date of publication 29 January 2025; date of current the eigenvectors and eigenvalues represent. To show the effec-
version 18 February 2025. The associate editor coordinating the review of this tiveness of the proposed algorithm, we have explored graph dis-
article and approving it for publication was Dr. Wei Hu. (Corresponding author:
Jyoti Maheshwari.)
tances to address a well known problem in the field of brain com-
Jyoti Maheshwari is with the Institute of Technology, Nirma University, puter interface which is tracking the transitions of brain states.
Ahmedabad 382481, India (e-mail: [email protected]). Also, we have explored the most naturally formed dynamic brain
Shiv Dutt Joshi and Tapan K Gandhi are with the Department of Electrical
Engineering, Indian Institute of Technology Delhi, Delhi 110016, India (e-mail:
networks in detecting real time seizure onset utilising graph
[email protected]; [email protected]). distances which has been an active area of research in the study of
Digital Object Identifier 10.1109/TSIPN.2025.3536085 epilepsy.

2373-776X © 2025 IEEE. All rights reserved, including rights for text and data mining, and training of artificial intelligence and similar technologies.
Personal use is permitted, but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: VIT University. Downloaded on May 17,2025 at 14:36:36 UTC from IEEE Xplore. Restrictions apply.
MAHESHWARI et al.: NOVEL SPECTRAL GRAPH DISTANCE MEASURE AND ITS APPLICATIONS IN BIOMEDICAL SIGNAL PROCESSING 115

II. EXISTING GRAPH DISTANCES The spanning tree dissimilarity between two graphs G and G
is defined as
A. Graph Edit Distances
It measures the similarity between networks by finding the dST (G, G ) = |log(ΓG ) − log(ΓG )| (7)
number of structural changes required to make the networks
isomorphic. One of the famous graph edit distances is Hamming III. PROPOSED GRAPH DISTANCE
distance [2]. It is defined as the number of inserted and deleted A. Mathematical Preliminaries
edges required to transform one graph into another. It captures
A recent study [9] proposed a novel feature termed as char-
the local changes in the network. Let A0 and A0 be two adja-
acteristic response vector (CRV) by extending the idea of the
cency matrices associated with the graphs G and G respectively
impulse response of linear time-invariant systems to the edge
on M nodes. The i, j-th element of adjacency matrix is given
weight matrices. In this paper, CRV utilises the information
as [7]
 of eigenvectors along with the eigenvalues of the correlation
1, if node i and j are connected coefficient matrix to capture the dynamics of brain’s electrical
a0i,j = (1)
0, otherwise activity. CRV is defined as the sum of all eigenvectors weighted
with their respective eigenvalues which is given by
Then the normalised Hamming distance is given as
r = ΣM
m=1 λm um (8)
 ||A0 − A0 ||1
dH (G, G ) = (2) th
M (M − 1) Here, λm is the m eigenvalue of the correlation coefficient
matrix and um is the corresponding eigenvector. Motivated
B. Matrix Distance by the concept of CRV, we propose a novel graph distance to
It is defined as the lp norm of the difference between edge measure the degree of similarity between any two networks
weight matrices of the two brain networks. Matrix distance re- having node correspondence. The proposed graph distance is
quires node correspondence [4]. Let the matrices A and A be the a spectral distance between any two networks which utilises the
adjacency matrices for weighted graphs G and G respectively. eigenvalues and eigenvectors of the corresponding edge weight
The idea behind the matrix distances is that the matrix A and A matrices.
carries important structural information about the graphs. The
i, j-th element of adjacency matrix is given as B. Formation of Sub-Band CRV
 Let A and A be two edge weight or connectivity matrices as-
wi,j , if node i and j are connected
ai,j = (3) sociated with the graphs G and G respectively on M nodes. Let
0, otherwise M eigenvalues associated with the edge weight matrices of net-
Then the matrix distance between two networks can be given works G and G are {λ1 , λ2 , λ3 . . .λM } and {λ1 , λ2 , λ3 . . .λM }
as respectively. Let the corresponding eigenvectors be given as
{u1 , u2 , u3 . . .uM } and {u1 , u2 , u3 . . .uM } respectively.
dM (G, G ) = ||A − A ||p (4) Since eigenvalues and eigenvectors are rich source of in-
Here ||.||p represents pth norm of the matrix. formation, we form the sub-band CRVs r Q and r Q for both
the networks by taking the sum of a few eigenvectors, having
C. Spectral Distances relevant information, weighted with their respective eigenvalues.
1) lp Distances on Eigenvalues: Let M eigenvalues associ- r Q (s) = Σl∈Q λl (s)ul (s) (9)
ated with the functional connectivity matrices of networks G
Here, Q represents the set of indices of eigenvectors contain-
and G be denoted as λm and λm (1 ≤ m ≤ M ) respectively.
ing the relevant information. In general, one can choose those
The lp distances on eigenvalues are defined as
eigenvectors that have the distinguishing capability as per the
M
 specific requirement. Particularly, our proposed graph distance
dlp (G,G ) = |λm − λm |p (5) is highly dependent on two factors:
m=1 1) Chosen edge weight matrix
2) Spanning Tree Dissimilarity: It finds the similarity be- 2) Selection of eigenvectors to form sub-band CRV
tween two networks by comparing their number of spanning 1) Chosen Edge Weight Matrix: We hypothesize that the
trees [6] which can be found using Kirchhoff’s matrix tree physical quantity that the edge weight matrix represents, plays a
theorem. It states that number of spanning trees of a connected very crucial role. We should choose edge weights between nodes
graph is given as normalised product of non-zero eigenvalues in such a way that they capture the essence of the problem. Vari-
of the Laplacian matrix. Let λm be the eigenvalues of the ous edge weight matrices have different physical interpretations.
Laplacian matrix such that λM = 0 ≤ λM −1 ≤ . . . ≤ λ2 ≤ λ1 . For instance, Pearson correlation coefficient matrix represents
The number of spanning trees for a graph G is given as the strength of linear relationship between any two nodes, graph
M −1 Laplacian matrix measures how much the graph differs at one
1  node from its nearby nodes, heat kernel is the fundamental solu-
ΓG = λm (6)
M m=1 tion of the network diffusion equation which captures temporal

Authorized licensed use limited to: VIT University. Downloaded on May 17,2025 at 14:36:36 UTC from IEEE Xplore. Restrictions apply.
116 IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS, VOL. 11, 2025

Fig. 1. Three network topologies [8].

Fig. 2. Various graph distances between the reference network having rewiring probability = 0.2 and other small world networks with varying rewiring probability
pws .

as well as spatial information and phase locking value matrix Pearson correlation coefficient matrix as they represent
measures the variability in the relative phases of two signals, the directions of maximum variance.
present at any two nodes, over time etc. 2) If our objective is to capture the static information of the
2) Selection of Eigenvectors to Form Sub-Band CRV: Selec- network, we should choose the eigenvectors correspond-
tion of the eigenvectors, whether corresponding to dominant ing to smaller eigenvalues as they represent the direction
eigenvalues or smaller eigenvalues, depends on what physical of minimum variance.
entity the eigenvectors represent. For example, 3) If we are looking for finer details of the network, we should
1) If we want to capture the highly varying information of go for eigenvectors corresponding to higher eigenvalues of
the graph, we should select the dominant eigenvectors of graph Laplacian matrix, since the eigenvectors associated

Authorized licensed use limited to: VIT University. Downloaded on May 17,2025 at 14:36:36 UTC from IEEE Xplore. Restrictions apply.
MAHESHWARI et al.: NOVEL SPECTRAL GRAPH DISTANCE MEASURE AND ITS APPLICATIONS IN BIOMEDICAL SIGNAL PROCESSING 117

Fig. 3. The proposed graph distance (for different selections of eigenvectors to form sub-band CRV starting from the dominant one) between the reference
network having rewiring probability = 0.2 and other small world networks with varying rewiring probability pws .

as they represent low spatial frequencies and correspond-

ing eigenvectors are smooth and vary slowly across the
graph [11], [12].

C. Euclidean Distance Between Sub-Band CRVs

We propose that degree of similarity between the two net-
works G and G can be given by finding Euclidean distance
between corresponding sub-band CRVs r Q and r Q .

dQ = ΣM k  k 2
k=1 (rQ − rQ ) (10)

 k
where rQ k and rQ are the k th elements of r Q and r Q respec-
tively.
In short, eigenvectors represent the intrinsic directions of con-
nectivity within a network, depending on the chosen functional
connectivity measure such as Pearson correlation coefficient,
graph Laplacian, heat kernel, coherence, phase locking value
etc. Sub-band CRV provides the aggregated information across
specific eigenvectors, each reflecting unique dynamic properties
(e.g., high-or low frequency variations, steady-state or transient
behaviours, high or low variance directions etc). By selecting
eigenvectors relevant to the problem and weighting them appro-
Fig. 4. Various distance with increasing absolute difference between pws and
p0ws .
priately, sub-band CRV encapsulates the network’s most critical
characteristics for a given problem. Euclidean distance between
Sub-band CRVs of two networks measures the dissimilarity
with larger eigenvalues representing high spatial frequen- in their dynamic representations. If any two networks exhibit
cies oscillate more rapidly [10], [11] across the graph. similar dynamics, their corresponding sub-band CRVs should be
4) And if we are interested in overall structure or coarse in- closer which will result into smaller Euclidian distance between
formation of the network, we should choose eigenvectors them. In other words, we have made an attempt to represent any
corresponding to smaller eigenvalues of Laplacian matrix network in the eigenspace, where sub-band CRV acts as a feature

Authorized licensed use limited to: VIT University. Downloaded on May 17,2025 at 14:36:36 UTC from IEEE Xplore. Restrictions apply.
118 IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS, VOL. 11, 2025

Fig. 5. Stimuli for experiment 2.

Fig. 6. Various graph distances between consecutive functional brain networks of eyes opened and eyes closed experiment for subject 7.

vector capturing the core properties of the network. The efficacy regular, small world and random as shown in Fig. 1. A regular
of the proposed graph distance over existing graph distances has network is a non-random network in which each node connects
been shown in the result section. to all of its nearest nodes. But many real time networks are
neither regular nor random, they lie somewhere between these
IV. RESULTS two networks, and are called small world networks. The small
world networks can be generated by introducing randomness in
We have conducted three experiments to demonstrate the
the regular network. It has been reported in the literature that
efficacy of the proposed method. The complete analysis is
small world networks are highly clustered similar to regular
performed in MATLAB environment.
networks and also have short average path length as that of
To show the efficacy of the proposed graph distance measure
random graphs.
over the existing graph distance measures, we have considered
To demonstrate the performance of proposed graph distance,
one simulated and two real time experiments in which our
firstly we analyzed the similarity between small world networks.
objective was to track the brain state transitions. Our results
In this endeavour, we considered Watts–Strogatz (WS) model [8]
show that the existing graph distances, such as hamming, matrix
which can generate small world networks by rewiring edges of
distances and eigen spectrum based graph distances [4] fail in
the regular network, having N = 500 nodes and k = 10 edges
tracking the brain states transitions during different tasks as well
per node, with a probability pws . This model generates a regular
as in detecting seizure onsets. But the proposed graph distance
network at pws = 0 and a random network pws = 1.
performs well in both the cases.
We have generated a reference network G(p0ws ) from the
WS model with a fixed rewiring probability p0ws = 0.2. Then
A. Experiment 1 we have generated a group of networks G(pws )s with a range
Based on the connection probability of the nearest neighbour of values of rewiring probability such that 0 ≤ pws ≤ 1. The
nodes and average path length, there are 3 types of networks [8]: generated network is represented by its binary adjacency matrix

Authorized licensed use limited to: VIT University. Downloaded on May 17,2025 at 14:36:36 UTC from IEEE Xplore. Restrictions apply.
MAHESHWARI et al.: NOVEL SPECTRAL GRAPH DISTANCE MEASURE AND ITS APPLICATIONS IN BIOMEDICAL SIGNAL PROCESSING 119

Fig. 7. The proposed graph distance for different selections of eigenvectors to form sub-band CRV between consecutive functional brain networks of eyes opened
and eyes closed experiment (subject 9).

of size 500*500 which has been chosen as an edge weight shows the proposed graph distance for different selection of
matrix. We found the similarity between G(p0ws ) and G(pws )s eigenvectors starting from the dominant one. Here, we have
using the existing graph distances such as Hamming distance, formed sub-band CRV using every 20 subsequent eigenvectors
matrix norm-2 distance, l1 , l2 distance on eigenvalues, spanning (chosen experimentally) arranged in descending order and found
tree dissimilarity and the proposed graph distance as shown the distance between G(p0ws ) and G(pws )s. The optimum num-
in Fig. 2(a), (b), (c), (d), (e) and (f) respectively. In order ber of eigenvectors to form sub-band CRV is a trade off between
to find the proposed graph distance, first we have calculated computational efficiency and capturing the core properties of a
Laplacian matrix (=degree matrix - binary adjacency matrix) network. If the number of eigenvectors to make a sub-band CRV
because eigenvalues and eigenvectors belonging to a natural is too small, then more number of sub-band CRVs will be formed
operator (Laplace operator in this case) are more meaningful as for analysis which will be computationally expensive. Whereas
established in the literature [13]. Whereas the binary adjacency if we form sub-band CRVs with more number of eigenvectors,
matrix solely indicates whether two nodes are connected or not. we may not capture the critical characteristics of the network
Eigenvectors corresponding to the smaller eigenvalues of as different eigenvectors are weighted with the corresponding
Laplacian matrix represent the structural properties of a physical eigenvalues. For ex. if we are interested in low variance in-
system i.e., how the network can be divided into subgroups. In formation (given by smaller eigenvectors) of the correlation
other words, eigenvectors corresponding to the smaller eigen- matrix and we form sub-band CRV using all eigenvectors, the
values tell us about the regions or zones that are more tightly resultant will point in the direction of dominant eigenvectors
connected internally than with other parts of the network, (direction of maximum variance) as maximum weightage is
whereas the eigenvectors corresponding to higher eigenvalues given to the eigenvectors corresponding to large eigenvalues.
represent the detailed information of the system. In this ex- In this direction, we have formed sub-band CRV with different
periment, we have binary adjacency matrix, which represents number (2, 10, 20, 50 and 100) of eigenvectors and found that
just the connectedness of any two nodes, hence the dominant 20 eigenvectors to form sub-band CRV is the optimum choice
eigenvectors representing detailed information of the underlying which has been used in experiment 1. It can be noted that the
system are not relevant. Accordingly, we chose the eigenvectors eigenvectors corresponding to smaller eigenvalues represent the
corresponding to smaller eigenvalues demonstrating the struc- connectedness of the network and are hence chosen to form
tural properties of the system to form the sub-band CRV. Fig. 3 sub-band CRV in this experiment.

Authorized licensed use limited to: VIT University. Downloaded on May 17,2025 at 14:36:36 UTC from IEEE Xplore. Restrictions apply.
120 IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS, VOL. 11, 2025

Fig. 8. Various graph distances between consecutive functional brain networks of eyes opened and eyes closed experiment for subject 9.

Ideally distance between G(p0ws ) and G(pws )s should be network (between 0–0.5). Fig. 4 demonstrates the average graph
minimum when pws = 0.2 and the distance should increase distances with varying absolute difference between pws and
as pws takes distant values from the fixed rewiring probability p0ws . It can be noted that high rewiring probabilities generating
p0ws = 0.2. It can be noted that Hamming distance and matrix random reference networks are not included in this analysis
distance show poor performances as they do not have minimum because we can’t talk of similarity of any network with the
value at pws = 0.2. Although l1 , l2 distances on eigenvalues random reference network. Ideally when the absolute difference
and spanning tree dissimilarities are having minimum value at between pws and p0ws is increasing, the graph distance between
pws = 0.2, but these distances are not increasing as rewiring corresponding networks, should also increase. Fig. 4 comprises
probability is taking higher values towards 1. For example, the only the best three graph distances from Fig. 2 which are l1 ,
distance between G(p0ws ) and G(pws )s should be larger when l2 distance on eigenvalues and the proposed graph distance. It
pws = 0.8 (because here the difference between pws and p0ws can be noted that the proposed graph distance is increasing with
is 0.6, more distinct networks) compared to pws = 0 (because increasing difference between pws and p0ws whereas other graph
here the difference between pws and p0ws is 0.2, less distinct distances don’t perform satisfactorily.
networks). Therefore, the graph distance at pws = 0.8 should
be higher than the graph distance at pws = 0. It can be clearly
B. Experiment 2
observed that our proposed graph distance is outperforming all
the existing graph distances. The proposed graph distance is For showing the efficacy of proposed graph distance, second
taking minimum value when rewiring probability is close to 0.2 data set has been taken from Texas Data Repository [14], [15].
and it is taking higher values as rewiring probability is taking This data set comprises scalp recordings at 72 places, out of
distant values as shown in Fig. 2(f). which 64 channels have been used in our analysis. The detailed
Further, in order to compare the performance of proposed information of the data set is given in [14], [15], [16]. In this
graph distance with the existing ones, we have observed the data set, we have continuous EEG recordings for 8 mins having
variations of various graph distances with increasing absolute alternate segments of 1 min each for eyes opened and eyes closed
difference between pws and p0ws as shown in Fig. 4. In this as shown in Fig. 5. The sampling rate is 256 Hz. We have divided
endeavour, we have calculated different graph distances for the continuous EEG data in overlapping, moving window of 1 s
various fixed values of rewiring probability of the reference comprising 256 samples and moved the window by 10 samples

Authorized licensed use limited to: VIT University. Downloaded on May 17,2025 at 14:36:36 UTC from IEEE Xplore. Restrictions apply.
MAHESHWARI et al.: NOVEL SPECTRAL GRAPH DISTANCE MEASURE AND ITS APPLICATIONS IN BIOMEDICAL SIGNAL PROCESSING 121

Fig. 9. Seizure onset detection using various graph distances between consecutive functional brain networks for subject 3.

implying an overlap of 246 samples. The choice of 1 s window last brain state transitions. On the basis of above reasoning, 8
size is taken from the literature [9]. Further, an attempt has subsequent eigenvectors with lesser variance are used to form
been made to maximize the overlap of samples (i.e. 246) so that the sub-band CRV for tracking brain state transitions.
new information is taken into account without completely losing It can be noted that existing graph distances such as Hamming
the previous information in order to capture the transitions in a distance, matrix distance, l1 , l2 distance on eigenvalues and
smooth manner. For each segment of EEG data, we have formed spanning tree dissimilarity are able to capture all 7 transitions for
functional brain network using Pearson correlation coefficient subject 7 as shown in Fig. 6(a), (b), (c), (d) and (e) respectively.
measure in order to track the brain state transitions. But we can clearly see false alarms close to third and sixth
In this direction, we have found the similarity between con- transitions in all the existing graph distances for subject 7. Since
secutive functional brain networks using various graph distances our proposed graph distance is able to track all the transitions
as shown in Figs. 6 and 8 for two randomly chosen subjects without any false alarm for subject 7, its performance surpasses
7 and 9 respectively. Eigenvectors of the correlation matrix the performance of other graph distances.
represent the directions in which EEG signal varies the most. In case of subject 9, the existing graph distances are able to
In particular, dominant eigenvectors represent the directions of capture only initial four transitions as shown in Fig. 8(a), (b), (c),
maximum variance while eigenvectors corresponding to smaller (d) and (e). Whereas the proposed graph distance is explicitly
eigenvalues represent the direction of lesser variance. In order to capturing all the transitions as shown in Fig. 8(f). Hence our
track the brain state transitions, we need invariant information proposed graph distance is performing better than the existing
about brain states which can be taken from the eigenvectors graph distances in tracking the brain state transitions.
corresponding to smaller eigenvalues. Fig. 7 shows the proposed
graph distance for different selection of eigenvectors to form
sub-band CRV starting from the dominant one. In line with C. Experiment 3
the above discussion, it can be observed that sub-band CRV In our third experiment, we have taken publicly available EEG
formed with eigenvectors corresponding to smaller eigenvalues data from CHB-MIT scalp EEG database. it contains continuous
are able to track all the brain state transitions whereas sub-band EEG data from 22 subjects having intractable seizures. The
CRV formed with dominant eigenvectors fails in detecting the data is acquired at the sampling rate of 256 Hz. The complete

Authorized licensed use limited to: VIT University. Downloaded on May 17,2025 at 14:36:36 UTC from IEEE Xplore. Restrictions apply.
122 IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS, VOL. 11, 2025

Fig. 10. Seizure onset detection using various graph distances between consecutive functional brain networks for subject 8.

description of the data set is given in [17], [18], [19]. For each connectivity measure in this experiment. Further, heat kernel
subject, 9 to 42 EEG recordings are there with mostly 23 chan- is the fundamental solution of the network diffusion equation
nels. In an attempt to detect seizure onset using graph distances, which incorporates temporal as well as spatial information.
we have divided the EEG data into small moving, overlapping Hence, it has been used as an adjacency matrix to find Laplacian
segments of 1 s having 256 samples. We have moved the window matrix.
by 10 samples resulting in the overlap of 246 samples. In this In short, we have analysed the time varying graph Lapla-
paper, we have randomly chosen two subjects from the whole cian matrix, formed using heat kernel as the adjacency matrix,
database. For both the subjects (case 3 and 8), we have found for overlapping and moving data segments in order to detect
the similarity between consecutive functional brain networks seizure and non-seizure brain states. As reported in the lit-
using various graph distances. For finding Hamming distance, erature, eigenvalues of the Laplacian matrix carry the notion
matrix distance, l1 distance on eigenvalues and spanning tree of spatial frequencies [12]. Eigenvectors corresponding to the
dissimilarity, we have considered Pearson correlation coefficient larger eigenvalues represent high spatial frequencies and contain
matrix as the functional connectivity measure. Various graph detailed information of brain’s states. Hence, we have formed
distances between consecutive functional matrices for the two sub-band CRV using dominant eigenvectors and analysed it over
chosen subjects are shown in Figs. 9 and 10. We can observe time. In line with the above discussion, the results (as shown in
only a small drop in the graph distances at seizure onset using Figs. 9 and 10) clearly show that our proposed graph distance
existing approaches. It can be noted that the proposed graph is performing better than the existing graph distances in seizure
distance using invariant eigenvectors of the Pearson correlation onset detection.
coefficient is also not performing satisfactorily. This is due to the
fact that only variance information is not sufficient in detecting
seizure and non seizure brain states. D. Discussion
In order to detect seizure onset, we propose that seizure The present paper shows the importance of choosing edge
progression can be modelled as a diffusion process in the brain. weights and sets of eigenvectors in finding similarity between
Since, Laplacian matrix is the network equivalent of the diffu- any two brain networks. While facing any new practical problem,
sion operator, we have chosen the graph Laplacian matrix as a edge weights or functional connectivity measure should be used

Authorized licensed use limited to: VIT University. Downloaded on May 17,2025 at 14:36:36 UTC from IEEE Xplore. Restrictions apply.
MAHESHWARI et al.: NOVEL SPECTRAL GRAPH DISTANCE MEASURE AND ITS APPLICATIONS IN BIOMEDICAL SIGNAL PROCESSING 123

in such a way that it captures the underlying dynamics of the in a variety of applications. In this work, we have explored
problem. Every functional connectivity measure captures the the physical interpretation of the eigenvectors of only Pearson
unique properties of the underlying system. As suggested, we correlation coefficient matrix and graph Laplacian matrix. Other
have further discussed some important connectivity measures functional connectivity measures can be explored in future. The
and interpretation of their eigenvectors which can be used when major limitation of our proposed graph distance is that it requires
faced with a new practical problem. node correspondence, but an attempt can be made in future to
The Pearson correlation coefficient matrix measures the linear make this graph distance node permutation invariant.
dependencies between any two EEG channels. The dominant
eigenvectors of the correlation coefficient matrix represent the REFERENCES
directions in which variance across EEG channels is maximised [1] Y. Shimada, Y. Hirata, T. Ikeguchi, and K. Aihara, “Graph distance for
while smaller eigenvectors point in the directions of decreasing complex networks,” Sci. Rep., vol. 6, no. 1, pp. 1–6, 2016.
variance. [2] C. Donnat et al., “Tracking network dynamics: A survey using graph
distances,” Ann. Appl. Statist., vol. 12, no. 2, pp. 971–1012, 2018.
Other important functional connectivity measures are heat [3] R. C. Wilson and P. Zhu, “A study of graph spectra for comparing graphs
kernel matrix and Laplacian matrix. Heat kernel is a dy- and trees,” Pattern Recognit., vol. 41, no. 9, pp. 2833–2841, 2008.
namic connectivity measure which shows how EEG sig- [4] P. Wills and F. G. Meyer, “Metrics for graph comparison: A practitioner’s
guide,” PLoS One, vol. 15, no. 2, 2020, Art. no. e0228728.
nals spread across brain regions over time (dynamics [5] I. Jovanović and Z. Stanić, “Spectral distances of graphs,” Linear Algebra
of information flow) whereas Laplacian matrix is a net- Appl., vol. 436, no. 5, pp. 1425–1435, 2012.
work’s topology measure which represents flow of infor- [6] A. K. Kelmans, “Comparison of graphs by their number of spanning trees,”
Discrete Math., vol. 16, no. 3, pp. 241–261, 1976.
mation across different nodes of the network how sig- [7] F. R. Chung and F. C. Graham, Spectral Graph Theory. Providence, RI,
nals or information spread through the network. Dominant USA: Amer. Math. Soc., 1997, no. 92.
eigenvectors of the heat kernel matrix represent the rapid [8] D. J. Watts and S. H. Strogatz, “Collective dynamics of ‘small-world’
networks,” in Nature, vol. 393, no. 6684, pp. 440–442, 1998.
changes in the brain signals over a period of time (capturing [9] S. Panwar, S. D. Joshi, A. Gupta, and P. Agarwal, “Automated epilepsy
the transient behaviour) whereas smaller eigenvectors repre- diagnosis using eeg with test set evaluation,” IEEE Trans. Neural Syst.
sent the slow diffusion process over time (capturing the steady Rehabil. Eng., vol. 27, no. 6, pp. 1106–1116, Jun. 2019.
[10] O. Sorkine, “Laplacian mesh processing,” Eurographics (STARs), vol. 29,
state behaviour). On the other hand, dominant eigenvectors of pp. 53–70, 2005.
Laplacian matrix capture the abrupt transitions between nodes [11] N. Saito, “Data analysis and representation on a general domain using
representing local high frequency variations within the network eigenfunctions of Laplacian,” Appl. Comput. Harmon. Anal., vol. 25, no. 1,
pp. 68–97, 2008.
whereas smaller eigenvectors capture the baseline or low fre- [12] D. I. Shuman, S. K. Narang, P. Frossard, A. Ortega, and P. Vandergheynst,
quency oscillations which do not vary across the network. “The emerging field of signal processing on graphs: Extending high-
Some other functional connectivity measures are coherence dimensional data analysis to networks and other irregular domains,” IEEE
Signal Process. Mag., vol. 30, no. 3, pp. 83–98, May 2013.
and phase locking value (PLV) etc. In particular, coherence [13] D. Spielman, “Spectral graph theory,” Combinatorial Sci. Comput., vol. 18,
measures the degree of coupling between various EEG chan- p. 18, 2012.
nels in frequency domain. Dominant eigenvectors of the co- [14] L. Trujillo, “Raw bdf data,” 2017. [Online]. Available: https://doi.org/10.
18738/T8/EG0LJI
herence matrix represent the strong synchronization pattern at [15] L. T. Trujillo, C. T. Stanfield, and R. D. Vela, “The effect of electroen-
a specific frequency whereas small eigenvectors represent the cephalogram (eeg) reference choice on information-theoretic measures of
weaker frequency-specific coupling. While PLV indicates phase the complexity and integration of eeg signals,” Front. Neurosci., vol. 11,
Jul. 2017, Art. no. 425.
synchronisation among various EEG channels and its dominant [16] J. Maheshwari, S. D. Joshi, and T. K. Gandhi, “Tracking the transitions
eigenvectors represent the strong and local phase synchronised of brain states: An analytical approach using EEG,” IEEE Trans. Neural
pattern whereas smaller eigenvectors indicate global phase syn- Syst. Rehabil. Eng., vol. 28, no. 8, pp. 1742–1749, Aug. 2020.
[17] A. H. Shoeb, “Application of machine learning to epileptic seizure onset
chronised pattern across EEG signals. detection and treatment,” Ph.D. dissertation, Massachusetts Inst. of Tech-
In the above reference, this paper proposes that when faced nol., Cambridge, MA, USA, 2009.
with any practical problem in real time, choice of functional [18] A. L. Goldberger et al., “Physiobank, physiotoolkit, and physionet: Com-
ponents of a new research resource for complex physiologic signals,” in
connectivity measure to form the brain network should be based Circulation, vol. 101, no. 23, pp. e215–e220, 2000.
on the underlying physics of the problem. In addition to this, the [19] J. Maheshwari, S. D. Joshi, and T. K. Gandhi, “Real-time auto-
chosen eigenvectors too play a crucial role in finding spectral mated epileptic seizure detection by analyzing time-varying high spa-
tial frequency oscillations,” IEEE Trans. Instrum. Meas., vol. 71, 2022,
graph distance between any two networks. Art no. 4002608.

V. CONCLUSION
In this paper, we proposed a novel graph distance measure
which is providing better results than the existing measures

Authorized licensed use limited to: VIT University. Downloaded on May 17,2025 at 14:36:36 UTC from IEEE Xplore. Restrictions apply.