EEG-GRAPH: A Factor-Graph-Based Model for

Capturing Spatial, Temporal, and Observational

Relationships in Electroencephalograms

Yogatheesan Varatharajah ∗ Min Jin Chong∗ Krishnakant Saboo∗ Brent Berry†

Benjamin Brinkmann† Gregory Worrell† Ravishankar Iyer∗

Abstract
This paper presents a probabilistic-graphical model that can be used to infer char-
acteristics of instantaneous brain activity by jointly analyzing spatial and tempo-
ral dependencies observed in electroencephalograms (EEG). Specifically, we de-
scribe a factor-graph-based model with customized factor-functions defined based
on domain knowledge, to infer pathologic brain activity with the goal of identify-
ing seizure-generating brain regions in epilepsy patients. We utilize an inference
technique based on the graph-cut algorithm to exactly solve graph inference in
polynomial time. We validate the model by using clinically collected intracra-
nial EEG data from 29 epilepsy patients to show that the model correctly iden-
tifies seizure-generating brain regions. Our results indicate that our model out-
performs two conventional approaches used for seizure-onset localization (5–7%
better AUC: 0.72, 0.67, 0.65) and that the proposed inference technique provides
3–10% gain in AUC (0.72, 0.62, 0.69) compared to sampling-based alternatives.

1 Introduction
Studying the neurophysiological processes within the brain is an important step toward understand-
ing the human brain. Techniques such as electroencephalography are exceptional tools for studying
the neurophysiological processes, because of their high temporal and spatial resolution. An elec-
troencephalogram (EEG) typically contains several types of rhythms and discrete neurophysiolog-
ical events that describe instantaneous brain activity. On the other hand, the neural activity taking
place in a brain region is very likely dependent on activities that took place in the same region at pre-
vious time instances. Furthermore, some EEG channels show inter-channel correlation due to their
spatial arrangement [1]. Those three characteristics are related, respectively, to the observational,
temporal, and spatial dependencies observed in time-series EEG signals.
The majority of the literature focuses on identifying and developing detectors for features relating to
the different rhythms and discrete neurophysiological events in the EEG signal [2]. Some effort has
been made to understand the inter-channel correlations [3] and temporal dependencies [4] observed
in EEG. Despite these separate efforts, very little effort has been made to combine those depen-
dencies into a single model. Since those dependencies possess complementary information, using
only one of them generally results in poor understanding of the underlying neurophysiological phe-
nomena. Hence, a unified framework that jointly captures all three dependencies in EEG, addresses
an important research problem in electrophysiology. In this paper, we describe a graphical-model-
based approach to capture all three dependencies, and we analyze its efficacy by applying it to a
critical problem in clinical neurology.
∗
Department of Electrical and Computer Engineering, University of Illinois at Urbana-Champaign, Urbana,
Illinois 61801. Email: {varatha2, mchong6, ksaboo2, rkiyer}@illinois.edu
†
Department of Neurology, Mayo Clinic, Rochester, Minnesota 55904. Email: {Berry.Brent,
Brinkmann.Benjamin, Worrell.Gregory}@mayo.edu
31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.
Graphical models in general are useful for representing dependencies between random variables.
Factor graphs are a specific type of graphical models that have random variables and factor functions
as the vertices in the graph [5]. A factor function is used to describe the relationship between two or
more random variables in the graph. Factor graphs are particularly useful when custom definitions
of the dependencies, such as in our case, need to be encoded in the graph. Hence, we have chosen to
adopt a factor graph model to represent the three kinds of dependencies described previously. These
dependencies are represented via three different factor functions, namely observational, spatial,
and temporal factor functions. We assess the applicability of this model in localization of seizure
onset zones (SOZ), which is a critical step in treating patients with epilepsy [6]. In particular, our
model is utilized to isolate those neural events in EEG that are associated with the SOZ, and are
eventually used to deduce the location of the SOZ. However, in a general setting, with appropriate
definitions of factor functions, one can utilize our model to describe other neural events of interest
(e.g., events related to behavioral states or memory processing). Major contributions of our work
are the following.

1. A framework based on factor graphs that jointly represents instantaneous observation-

based, temporal, and spatial dependencies in EEG. This is the first attempt to combine
these three aspects into a single model in the context of EEG analysis.
2. A lightweight and exact graph inference technique based on customized definitions of fac-
tor functions. Exact graph inference is typically intractable in most graphical model repre-
sentations because of exponentially growing state spaces.
3. A markedly improved technique for localizing SOZ based on the factor-graph-based model
developed in this paper. Existing approaches utilize only the observations made in the EEG
to determine the SOZ and do not utilize spatial and temporal dependencies.

Our study establishes the feasibility of the factor-graph-based model and demonstrates its application
in SOZ localization on a real EEG dataset collected from epilepsy patients who underwent epilepsy
surgery. Our results indicate that utilizing the spatial and temporal dependencies in addition to
observations made in the EEG provides a 5–7% improvement in the AUC (0.72, 0.67, 0.65) and
outperforms alternative approaches utilized for SOZ localization. Furthermore, our experiments
demonstrate that the lightweight graph inference technique provides a considerable improvement
(3–10%) in SOZ localization compared to sampling-based alternatives (AUC: 0.72, 0.62, 0.69).

2 Related work
Identifying features (or biomarkers) that describe underlying neurophysiological phenomena has
been a major focus of research in the EEG literature [2]. Spectral features [7], interictal spikes [8],
high-frequency oscillations [2], and phase-amplitude coupling [4] are some of the widely used fea-
tures. Although feature identification is an important step in any electrophysiologic study, features
alone often cannot completely describe the underlying physiological phenomena. Researchers have
also looked at spatial connectivity between EEG channels as means of describing neurophysiological
activities [3]. In recent times, because of the availability of long-term EEG recordings, understand-
ing of the temporal dependencies within various brain activities has also advanced significantly [4].
A recent attempt at combining spatial and temporal constraints has shown promise despite lacking
comprehensive validation [9]. Regardless, a throughly validated and general model that captures all
the factors, and is applicable to a variety of problems has not, to our knowledge, been proposed in
the EEG literature. Since the three factors are complementary to each other, a model that jointly
represents them addresses an important research gap in the field of electrophysiology.
Graphical models have been widely used in medical informatics [10], intrusion detection [11], so-
cial network modeling [12], and many other areas. Although factor graphs are applicable in all these
settings, their applications in practice are still very much dependent on problem-specific custom
definitions of factor functions. Nevertheless, with some level of customization, our work provides a
general framework to describe the different dependencies observed in EEG signals. A similar frame-
work for emotion prediction is described in Moodcast [12], for which the authors used a factor graph
model to describe the influences of historical information, other users, and dynamic status to predict
a user’s emotions in a social network setting. Although our factor functions are derived in a simi-
lar fashion, we show that graph inference can be performed exactly using the proposed lightweight
algorithm, and that it outperforms the sampling-based inference method utilized in Moodcast. Our

2
algorithm for inference was inspired by [13], in which the authors used an energy-minimization-
based approach for performing exact graph inference in a Markov random field-based model.

3 Model description
Here we provide a mathematical description of the model and the inference procedure. In a nutshell,
we are interested in inferring the presence of a neurophysiological phenomenon of interest by ob-
serving rhythms and discrete events (referred to as observations) present in the EEG, and by utilizing
their spatial and temporal patterns as represented by a probabilistic graph. Since the generality of
our model relies on the ability to customize the definitions of specific dependencies described by the
model, we have adopted a factor-graph-based setting to represent our model.
Definitions: Suppose that EEG data of a subject are recorded through M channels. Initially, the
data is discretized by dividing the recording duration into N epochs. We represent the interactions
between the channels at an epoch n as a dynamic graph Gn = (V, En ), where V is the set of
|V | = M channels and En ⊂ V × V is the set of undirected links between channels. The state of a
channel k in the nth epoch is denoted by Yn (k), which might represent a phenomenon of interest. For
example, in the case of SOZ localization, the state might be a binary value representing whether the
k th channel in the nth epoch exhibits a SOZ-likely phenomenon. We also use Yn to denote the states
of all the channels at epoch n, and use Y to denote the set of all possible values that Yn (k) can take.
We refer to the EEG rhythm or discrete event present in the EEG as observations and use Xn (k)
to denote the observation present in the nth epoch of the k th channel. Depending on the number of
rhythms and/or events, Xn (k) could be a scalar or vector random variable. The observations made
in all the channels at epoch n are denoted by Xn .
Inference: Given a dynamic network Gn , and the observations Xn , our goal is to infer the states
of the channels at epoch n, i.e., Yn . In our approach, we derive the inference model using a factor
graph with factor functions defined as shown in Table 1. The factor functions are defined using
exponential relationships so that they attain their maximum values when the exponents are zero, and
exponentially decay otherwise. All factor functions range in [0, 1].
Table 1: Factor functions used in our EEG model and their descriptions, definitions, and notations.

Function Description Defnition Notations

−(Yn (k)−φ(Xn (k)))2
Observational: Measures the direct contri- e φ : X → Y is a mapping
f (Yn (k), φ(Xn (k))) bution of the observations from the observations to the
made in a channel to the phenomenon of interest. In
phenomenon of interest. general, it is not an accurate
map, because it is based on
observations alone.
2
− 1
2 (Yn (k)−Yn (l))
d
Spatial: Measures the correlation be- e kl dkl denotes the physical dis-
g(Yn (k), Yn (l)) tween the states of two chan- tance between electrodes (or
nels at the same epoch. channels) k and l.
2
Temporal: Measures the correlation be- e−(Yn (k)−Ωn−1 (k)) Ωn−1 (k) is a function of all
h(Yn (k), Ωn−1 (k)) tween a channel’s current previous states of channel k.
Pn−1
state and its previous states. E.g., Ωn−1 (k) = i=1Yi (k)
n−1

With these definitions, the state of a channel is spatially related to the states of every other channel,
temporally related to a function of all its previous states, and, at the same time, explained by the
current observation of the channel. These dependencies and the factor functions that represent them
are illustrated in Fig. 1a and 1b respectively. (Note that Fig. 1b illustrates only the factor functions
related to Channel 1 and that similar factor functions exist for other channels although they are
not shown in the figure.) Provided with that information, for a particular state vector Y , we can
write P (Y |Gn ) as in Eq. 1, where Z is a normalizing factor. In general, it is infeasible to find the
normalizing constant Z, because it would require exploration of the space |Y|M .
 
M
1 Y Y
P (Y |Gn ) = g (Y (k), Y (i)) × f (Y (k), φ(Xn (k))) × h (Y (k), Ωn−1 (k)) (1)
Z
k=1 i6=k

3
 Current observation Previous states of
(events, rhythms) the same region

Current state of a
brain region

States of nearby
regions

(a) Factors that explain the state of a brain region. (b) Dependencies as factor functions.
Figure 1: The dependencies observed in brain activity and a representative factor graph model.
Therefore, we define the following predictive function (Eq. 2) for inferring Yn with the highest
likelihood per Eq. 1.
 
M
Y Y
Yn = arg max  g (Y (k), Y (i)) × f (Y (k), φ(Xn (k))) × h (Y (k), Ωn−1 (k)) (2)
Y ∈Y M k=1 i6=k

Still, finding a Y that maximizes this objective function involves a discrete optimization over the
space |Y|M . A brute-force approach to finding an exact solution is infeasible when M is large.
Several methods, such as junction trees [14], belief propagation [15], and sampling-based methods
such as Markov Chain Monte Carlo (MCMC) [16, 17], have been proposed to find approximate
solutions. However, we show that this can be calculated exactly when the aforementioned definitions
of the factor functions are utilized. We can rewrite Eq. 2 using the definitions in Table 1 as follows.
 
M Y − 12 (Y (k)−Y (l))2
Y 2 2
Yn = arg max  e dkl × e−(Y (k)−φ(Xn (k))) × e−(Y (k)−Ωn−1 (k))  (3)
Y ∈Y M k=1 l6=k

Now, representing the product terms as summations inside the exponent and using the facts that the
exponential function is monotonically increasing and that maximizing a function is equivalent to
minimizing the negative of that function, we can rewrite Eq. 3 as:
 
Yn = arg min (4)
PM 1
(Y (k)−Y (l))2 +(Y (k)−φ(Xn (k)))2 +(Y (k)−Ωn−1 (k))2
P
k=1 l6=k d2
kl
Y ∈Y M

Although the individual components in this objective function are solvable optimization problems,
the combination of them makes it difficult to solve. However, the objective function resembles
that of a standard graph energy minimization problem and hence can be solved using graph-cut
algorithms [18]. In this paper, we describe a solution for minimizing this objective function when
|Y| = 2, i.e., the brain states are binary. Although that is a limitation, the majority of the brain state
classification problems can be reduced to binary state cases when the time window of classification
is appropriately chosen. Regardless, potential solutions for |Y| > 2 are discussed in Section 6.
Graph inference using min-cut for the binary state case: We constructed the graph shown in
Fig. 2a with two special nodes in addition to the EEG channels as vertices. The additional nodes
function as source (marked by 1) and sink (marked by 0) nodes in the conventional min-cut/max-flow
problem. Weights in this graph are assigned as follows:
• Every channel is connected with every other channel, and the link between channels k and
l is assigned a weight of d12 (Y (k) − Y (l))2 based on the distance between them.
kl
• Every channel is connected with the source node, and the link between channel k and the
2 2
source is assigned a weight of (1 − Ωn−1 (k)) + (1 − φ (Xn (k))) .
• Every channel is also connected with the sink node, and the link between channel k and the
2
sink is assigned a weight of Ω2n−1 (k) + (φ (Xn (k))) .
Proposition 1. An optimal min-cut partitioning of the graph shown in Fig. 2a minimizes the objec-
tive function given in Eq. 4.

4
 (a) New graphical structure (b) Min-cut partitioning
Figure 2: Graph inference using the min-cut algorithm.
Proof: Suppose that we perform an arbitrary cut on the graph shown in Fig. 2a, resulting in two sets
of vertices S and T . The energy of the graph after the cut is performed is:
XM h i X X 1 
2 2 2
Ecut = (Y (k) − Ωn−1 (k)) + (Y (k) − φ (Xn (k))) + (Y (k) − Y (l))
d2kl
k=1 k∈T l∈S

It can be seen that, for the same partition of vertices, the objective function given in Eq. 4 attains the
same quantity as Ecut . Therefore, since the optimal min-cut partition minimizes the energy Ecut , it
minimizes the objective function given in Eq. 4.
Now suppose that we are given two sets of nodes {S ∗ , T ∗ } as the optimal partitioning of the graph.
Without loss of generality, let us assume that S ∗ contains the source and T ∗ contains the sink. Then,
the other vertices in S ∗ and T ∗ , are assigned 1 and 0 as their respective states to obtain the optimal
Y that minimizes the objective function given in Eq. 4.

4 Application of the model in seizure onset localization

Background: Epilepsy is a neurological disorder characterized by spontaneously occurring
seizures. It affects roughly 1% of the world’s population, and many do not respond to drug treatment
[19]. Epilepsy surgery, which involves resection of a portion of the patient’s brain, can reduce and
often eliminate seizures [20]. The success of resective surgery depends on accurate localization of
the seizure-onset zone [21]. The conventional practice is to identify the EEG channels that show the
earliest seizure discharge via visual inspection of the EEG recorded during seizures, and to remove
some tissue around these channels during the resective surgery. This method, despite being the cur-
rent clinical standard, is very costly, time-consuming, and burdensome to the patients, as it requires
a lengthy ICU stay so that an adequate number of seizures can be captured. One approach, which
has recently become a widely researched topic, utilizes between-seizure (interictal) intracranial EEG
(iEEG) recording to localize the seizure onset zones [22, 6]. This type of localization is preferable
to the conventional method, as it does not require a lengthy ICU stay.
Interictal SOZ identification methodology: Like that of the conventional approach, the goal here
is to identify a few channels that are likely to be in the SOZ. Channels situated directly on or close to
a SOZ exhibit different forms of transient electrophysiologic events (or abnormal events) between
seizures [23]. The frequency of such abnormal neural events plays a major role in determining
the SOZ. However, capturing these abnormal neural events that occur in distinct locations of the
brain alone is often not sufficient to establish an area in the brain as the SOZ. The reason is that
insignificant artifacts present in the EEG may show characteristics of those abnormal events that are
associated with SOZ (referred to as SOZ-likely events). In order to set apart the SOZ-likely events,
their spatial and temporal patterns could be utilized. It is known that SOZ-likely events occur in a
repetitive and spatially correlated fashion (i.e., neighboring channels exhibit such events at the same
time) [6]. Hence, the factor-graph-based model described in Section 3 can be applied to capture and
utilize the spatial and temporal correlations in isolating the SOZ-likely events.

5
Identifying abnormal neural events: Spectral characteristics of iEEG measured in the form of
power-in-bands (PIB) features have been widely utilized to identify abnormal neural events [24,
6, 7]. In this paper, PIB features are extracted as spectral power in the frequency bands Delta
(0–3 Hz), Low-Theta (3–6 Hz), High-Theta (6–9 Hz), Alpha (9–14 Hz), Beta (14–25 Hz), Low-
Gamma (30–55 Hz), High-Gamma (65–115 Hz), and Ripple (125–150 Hz) and utilized to make
observations from channels. As described in Section 3, a φ function is used to relate the observations
to abnormal events. In Section 6, we evaluate different techniques for obtaining a mapping from
extracted PIB features to the presence of an abnormal neural event. However, a mapping obtained
using observations alone is not sufficient to deduce SOZ because in addition to SOZ-likely events,
signal artifacts will also be captured by this mapping. This phenomenon is illustrated in Fig.3, in
which PIB features show similar characteristics for the events related to both SOZ and non-SOZ.
Therefore, we utilize the factor graph model presented in this paper to further filter the detected
abnormal events based on their spatial and temporal patterns and isolate the SOZ-likely events.

Non-SOZ
Channels

SOZ

SOZ

SOZ

1 2 3 4 5
Time (sec)
Non-SOZ Signal

0.5 1
SOZ Signal

0.5
0
0

-0.5 -0.5
0 0.5 1 1.5 0 0.5 1 1.5
Time(sec) Time(sec)
Normalized PIB

Normalized PIB

8 8
7 7
6 6
5 5
4 4
3 3
2 2
1 1
0 0.5 1 1.5 0 0.5 1 1.5
Time(sec) Time(sec)

Figure 3: EEG events related to both SOZ and non-SOZ are captured by PIB features because they
possess similar spectral characteristics.
Spatial and temporal dependencies in SOZ localization: Although artifacts show spectral charac-
teristics similar to those of SOZ-likely events, unlike the latter, the former do not occur in a spatially
correlated manner. This spatial correlation is measured with respect to the physical distances be-
tween the electrodes placed in the brain. Therefore, the same definition of the spatial factor function
described in Section 3 is applicable. If a channel’s observation is classified as an abnormal neural
event and the spatial factor function attains a large value with an adjacent channel, it would mean
that both channels likely show similar patterns of abnormalities which therefore must be SOZ-likely
events. In addition, the SOZ-likely events show a repetitive pattern, which artifacts usually do not.
In Section 3, we described the temporal correlation as a function of all previous states. As such, the
temporal correlation here is established with the intuition that a channel that previously exhibited a
large number of SOZ-likely events is likely to exhibit more because of the repetitive pattern. Hence,
temporal correlation is measured as the correlation between the state of a channel and the observed
P n−1
i Y (k)
frequency of SOZ-likely events in that channel until the previous epoch, i.e., Ωn−1 (k) = i=1
n−1 .
Therefore, when Ωn−1 (k) is close to 1 and the observation made from channel k is classified as an
abnormal neural event, the event is more likely to be a SOZ-likely event than an artifact.

5 Experiments
Data: The data used in this work are from a study approved by the Mayo Clinic Institutional Review
Board. The dataset consists of iEEG recordings collected from 29 epilepsy patients. The iEEG
sensors were surgically implanted in potentially epileptogenic regions in the brain. Patients were

6
implanted with different numbers of sensors, and they all had different SOZs. Ground truth (the
true SOZ channels) was established from clinical reports and verified independently through visual
inspection of the seizure iEEGs. During data collection, basic preprocessing was performed to
remove line-noise and other forms of signal contamination from the data.

Channel k 2-hour data segment

3-sec window 3-sec window 3-sec window

PIB feature extraction

Feature classification

Factor graph inference

Figure 4: A flow diagram illustrating the SOZ determination process.

Analytic scheme: Two-hour between-seizure segments were chosen for each patient to represent
a monitoring duration that could be achieved during surgery. The two-hour iEEG recordings were
divided into non-overlapping three-second epochs. This epoch length was chosen because it would
likely accommodate at least one abnormal neural event that could be associated with the SOZ [6].
Spectral domain features (PIB) were extracted in the 3-second epochs to capture abnormal neural
events [6]. Based on the features extracted in a 3-second recording of a channel, a binary value
φ (Xn (k)) ∈ {0, 1} was assigned to that channel, indicating whether or not an abnormal event was
present. Section 6 provides a comparison of supervised and unsupervised techniques used to create
this mapping. In the case of supervised techniques, a classification model was trained using the PIB
features extracted from an existing corpus of manually annotated abnormal neural events. In the
case of unsupervised techniques, channels were clustered into two groups based on the PIB features
extracted during an epoch, and the cluster with the larger cluster center (measured as the Euclidean
distance from the origin) was labeled as the abnormal cluster. Consequently, the respective epochs
of those channels in the abnormal cluster were classified as abnormal neural events. The factor graph
model was then used to filter the SOZ-likely events out of all the detected abnormal neural events. A
factor graph is generated using the observational, spatial, and temporal factor functions described
above specifically for this application. The best combination of states that minimizes the objective
function given in Eq. 4, Yn , is found by using the min-cut algorithm. In our approach, we used the
Boykov-Kolmogorov algorithm [25] to obtain the optimal partition of the graph. The states Yn here
are binary values and represent the presence or absence of SOZ-likely events in the channels. This
process is repeated for all the 3-second epochs and the SOZ is deduced at the end using a maximum
likelihood (ML) approach (described in the following). This whole process is illustrated in Fig. 4.
Maximum likelihood SOZ deduction: We model the occurrences of SOZ-likely events in channel
k as independent Bernoulli random variables with probability π(k). Here, π(k) denotes the true bias
of the channel’s being in SOZ. We estimate π(k) using a maximum likelihood (ML) approach and
use π̂(k) to denote the estimate. Each Yn (k) that results from the factor graph inference is treated
as an outcome of a Bernoulli trial and the log-likelihood function after N such trials is defined as:
"N #
Y
Yn (k) 1−Yn (k)
log (L(π(k))) = log π(k) (1 − π(k)) (5)
n=1

An estimate for π(k) that maximizes the above likelihoodP

function (known as MLE, i.e., maximum
N
Yn (k)
likelihood estimate) after N epochs is derived as π̂(k) = n=1
N .
Evaluation: The ML approach generates a likelihood probability for each channel k for being in
the SOZ. We compared these probabilities against the ground truth (binary values with 1 meaning

7
that the channel is in the SOZ and 0 otherwise) to generate the area under the ROC curve (AUC),
sensitivity, specificity, precision, recall, and F1-score metrics. First, we evaluated a number of
techniques for generating a mapping from the extracted PIB features to the presence of abnormal
events. We evaluated three unsupervised approaches, namely k-means, spectral, and hierarchical
clustering methods and two supervised approaches, namely support vector machine (SVM) and
generalized linear model (GLM), for this task. Second, we evaluated the benefits of utilizing the
min-cut algorithm for inferring instantaneous states. Here we compared our results using the min-
cut algorithm against those of two sampling-based techniques [12]: MCMC with random sampling,
and MCMC with sampling per prior distribution. Belief-propagation-based methods are not suitable
here because our factor graph contains cycles [26]. Third, we compared our results against two
recent solutions for interictal SOZ localization, including a summation approach [6] and a clustering
approach [22]. In the summation approach, summation of the features of a channel normalized by the
maximum feature summation was used as the likelihood of that channel’s being in the SOZ. In the
clustering approach, the features of all the channels during the whole 2-hour period were clustered
into two classes by a k-means algorithm, and the cluster with the larger cluster mean was chosen
as the abnormal cluster. For each channel, the fraction of all its features that were in the abnormal
cluster was used as the likelihood of that channel being in the SOZ. Both of these approaches utilize
only the observations and lack the additional information of the spatial and temporal correlations.

6 Results & discussion

Table 2 lists the results obtained for the experiments explained in Section 5, performed using a
dataset containing non-seizure (interictal) iEEG data from 29 epilepsy patients. First, a comparison
of supervised and unsupervised techniques for the mapping from PIB features to the presence of
abnormal events was performed. The results indicate that using a k-means clustering approach for
mapping PIB features to abnormal events is better than any other supervised or unsupervised ap-
proach, while other approaches also prove useful. Second, a comparison between sampling-based
methods and the min-cut approach was performed for the task of graph inference. Our results in-
dicate that utilizing the min-cut approach to infer instantaneous states is considerably better than a
random-sampling-based MCMC approach (with a 10% higher AUC and 14% higher F1-score) and
marginally better than an MCMC approach with sampling per a prior distribution (with a 3% higher
AUC and a similar F1-score), when used with k-means algorithm for abnormal event classification.
However, unlike this approach, our method does not require a prior distribution to sample from.
Third, we show that our factor-graph-based model for interictal SOZ localization performs signif-
icantly better than either of the traditional approaches (with 5% and 7% higher AUCs) when used
with k-means algorithm for abnormal event classification and min-cut algorithm for graph inference.

Table 2: Goodness-of-fit metrics obtained for unsupervised and supervised methods for PIB-to-
abnormal-event mapping (φ); sampling-based approaches for instantaneous state estimation; and
conventional approaches utilized for interictal SOZ localization. (“FG/kmeans/min-cut" means that
we utilized a factor-graph-based method, with a k-means clustering algorithm for mapping PIB
featuers to abnormal neural events and the min-cut algorithm for performing graph inference.)

Method AUC Sensitivity Specificity Precision Recall F1-score

Evaluation: techniques for PIB to abnormal event mapping (φ)
FG/kmeans/min-cut 0.72±0.03 0.74±0.03 0.61±0.02 0.39±0.05 0.74±0.03 0.46±0.04
FG/spectral/min-cut 0.68±0.03 0.60±0.07 0.48±0.05 0.31±0.05 0.60±0.07 0.36±0.05
FG/hierarch/min-cut 0.69±0.03 0.52±0.06 0.51±0.05 0.29±0.05 0.52±0.06 0.34±0.05
FG/svm/min-cut 0.71±0.03 0.68±0.06 0.54±0.05 0.36±0.05 0.68±0.06 0.43±0.05
FG/glm/min-cut 0.69±0.03 0.62±0.07 0.47±0.05 0.31±0.05 0.62±0.08 0.37±0.05
Evaluation: sampling vs. min-cut
FG/kmeans/Random 0.62±0.03 0.51±0.08 0.40±0.07 0.35±0.06 0.51±0.08 0.32±0.05
FG/kmeans/Prior 0.69±0.03 0.65±0.04 0.66±0.04 0.40±0.04 0.65±0.04 0.46±0.04
Evaluation: comparison against conventional approaches
Summation 0.67±0.04 0.59±0.05 0.67±0.03 0.38±0.05 0.59±0.05 0.43±0.05
Clustering 0.65±0.04 0.49±0.06 0.72±0.04 0.42±0.06 0.49±0.06 0.44±0.05

8
Significance: Overall, the factor-graph-based model with k-means clustering for abnormal event
classification and the min-cut algorithm for instantaneous state inference outperforms all other meth-
ods for the application of interictal SOZ localization. Utilization of spatial and temporal factor
functions improves the localization AUC by 5–7%, relative to pure observation-based approaches
(summation and clustering). On the other hand, the runtime complexity of instantaneous state infer-
ence is greatly reduced by the min-cut approach. The complexity of a brute-force approach grows
exponentially with the number of nodes in the graph, while the min-cut approach has a reasonable
runtime complexity of O(|V ||E|2 ), where |V | is the number of nodes and |E| is the number of
edges in the graph. Although sampling-based methods are able to provide approximate solutions
with moderate complexity, the min-cut method provided superior performance in our experiments.
Future work: Significant domain knowledge is required to come up with manual definitions of
graphical models, and in many situations, almost no domain knowledge is available. Hence, the
manually defined factor-graphical model and associated factor functions are a potential limitation
of our work, as a framework that automatically learns the graphical representation might result in
a more generalizable model. Dynamic Bayesian networks [27] may provide a platform that can be
used to learn dependencies from the data while allowing the types of dependencies we described.
Another potential limitation of our work is the binary-brain-state assumption made while solving the
graph energy minimization task. We surmise that extensions of the min-cut algorithm such as the
one proposed in [28] are applicable for non-binary cases. In addition, we also believe that optimal
weighting of the different factor functions could further improve localization accuracy and provide
insights on the contributions of spatial, temporal, and observational relationships to a specific appli-
cation that involves EEG signal analysis. We plan to investigate those in our future work.

7 Conclusion
We described a factor-graph-based model to encode observational, temporal, and spatial dependen-
cies observed in EEG-based brain activity analysis. This model utilizes manually defined factor
functions to represent the dependencies, which allowed us to derive a lightweight graph inference
technique. This is a significant advancement in the field of electrophysiology because a general and
comprehensively validated model that encodes different forms of dependencies in EEG does not ex-
ist at present. We validated our model for the application of interictal seizure onset zone (SOZ) and
demonstrated the feasibility in a clinical setting. Our results indicate that our approach outperforms
two widely used conventional approaches for the application of SOZ localization. In addition, the
factor functions and the technology for exactly inferring the states described in this paper can be
extended to other applications of factor graphs in fields such as medical diagnoses, social network
analysis, and preemptive attack detection. Therefore, we assert that further investigation is necessary
to understand the different usecases of this model.
Acknowledgements: This work was partly supported by National Science Foundation grants CNS-
1337732 and CNS-1624790, National Institute of Health grants NINDS-U01-NS073557, NINDS-
R01-NS92882, NHLBI-HL105355, and NINDS-UH2-NS095495-01, Mayo Clinic and Illinois Al-
liance Fellowships for Technology-based Healthcare Research and an IBM faculty award. We thank
Subho Banerjee, Phuong Cao, Jenny Applequist, and the reviewers for their valuable feedback.

References
[1] C. P. Warren, S. Hu, M. Stead, B. H. Brinkmann, M. R. Bower, and G. A. Worrell, “Synchrony in normal
and focal epileptic brain: The seizure onset zone is functionally disconnected,” Journal of Neurophysiol-
ogy, vol. 104, no. 6, pp. 3530–3539, 2010.
[2] G. A. Worrell, A. B. Gardner, S. M. Stead, S. Hu, S. Goerss, G. J. Cascino, F. B. Meyer, R. Marsh,
and B. Litt, “High-frequency oscillations in human temporal lobe: Simultaneous microwire and clinical
macroelectrode recordings,” Brain, vol. 131, no. 4, pp. 928–937, 2008.
[3] M. Rubinov and O. Sporns, “Complex network measures of brain connectivity: Uses and interpretations,”
Neuroimage, vol. 52, no. 3, pp. 1059–1069, 2010.
[4] C. Alvarado-Rojas, M. Valderrama, A. Fouad-Ahmed, H. Feldwisch-Drentrup, M. Ihle, C. Teixeira,
F. Sales, A. Schulze-Bonhage, C. Adam, A. Dourado, S. Charpier, V. Navarro, and M. Le Van Quyen,
“Slow modulations of high-frequency activity (40–140 [emsp14] hz) discriminate preictal changes in hu-
man focal epilepsy,” Scientific Reports, vol. 4, 2014.

9
 [5] B. J. Frey, F. R. Kschischang, H.-A. Loeliger, and N. Wiberg, “Factor graphs and algorithms,” in Proceed-
ings of the 35th Annual Allerton Conference on Communication Control and Computing. University of
Illinois, 1997, pp. 666–680.
[6] Y. Varatharajah, B. M. Berry, Z. T. Kalbarczyk, B. H. Brinkmann, G. A. Worrell, and R. K. Iyer, “Inter-
ictal seizure onset zone localization using unsupervised clustering and bayesian filtering,” in 8th Interna-
tional IEEE/EMBS Conference on Neural Engineering (NER). IEEE, 2017, pp. 533–539.
[7] Y. Varatharajah, R. K. Iyer, B. M. Berry, G. A. Worrell, and B. H. Brinkmann, “Seizure forecasting and
the preictal state in canine epilepsy,” International Journal of Neural Systems, vol. 27, p. 1650046, 2017.
[8] R. Katznelson, “EEG recording, electrode placement, and aspects of generator localization,” Electric
Fields of the Brain, pp. 176–213, 1981.
[9] J. D. Martinez-Vargas, G. Strobbe, K. Vonck, P. van Mierlo, and G. Castellanos-Dominguez, “Improved
localization of seizure onset zones using spatiotemporal constraints and time-varying source connectivity,”
Frontiers in Neuroscience, vol. 11, p. 156, 2017.
[10] L. R. Andersen, J. H. Krebs, and J. D. Andersen, “Steno: An expert system for medical diagnosis based
on graphical models and model search,” Journal of Applied Statistics, vol. 18, no. 1, pp. 139–153, 1991.
[11] P. Cao, E. Badger, Z. Kalbarczyk, R. Iyer, and A. Slagell, “Preemptive intrusion detection: Theoretical
framework and real-world measurements,” in Proceedings of the 2015 Symposium and Bootcamp on the
Science of Security. ACM, 2015, pp. 21:1–21:2.
[12] Y. Zhang, J. Tang, J. Sun, Y. Chen, and J. Rao, “Moodcast: Emotion prediction via dynamic continuous
factor graph model,” in 10th International Conference on Data Mining (ICDM), 2010, pp. 1193–1198.
[13] J. Liu, C. Zhang, C. McCarty, P. Peissig, E. Burnside, and D. Page, “High-dimensional structured feature
screening using binary Markov random fields,” in Artificial Intelligence and Statistics, 2012, pp. 712–721.
[14] W. Wiegerinck, “Variational approximations between mean field theory and the junction tree algorithm,”
in Proceedings of the 16th conference on Uncertainty in artificial intelligence, 2000, pp. 626–633.
[15] J. S. Yedidia, W. T. Freeman, Y. Weiss et al., “Generalized belief propagation,” in Advances in Neural
Information Processing Systems, vol. 13, 2000, pp. 689–695.
[16] W. R. Gilks, S. Richardson, and D. Spiegelhalter, Markov chain Monte Carlo in practice. CRC Press,
1995.
[17] S. Chib and E. Greenberg, “Understanding the Metropolis-Hastings algorithm,” The American Statisti-
cian, vol. 49, no. 4, pp. 327–335, 1995.
[18] V. Kolmogorov and R. Zabin, “What energy functions can be minimized via graph cuts?” IEEE Transac-
tions on Pattern Analysis and Machine Intelligence, vol. 26, no. 2, pp. 147–159, 2004.
[19] R. G. Andrzejak, D. Chicharro, C. E. Elger, and F. Mormann, “Seizure prediction: Any better than
chance?” Clinical Neurophysiology, vol. 120, no. 8, pp. 1465–1478, 2009.
[20] R. W. Lee, G. A. Worrell, W. R. Marsh, G. D. Cascino, N. M. Wetjen, F. B. Meyer, E. C. Wirrell, and E. L.
So, “Diagnostic outcome of surgical revision of intracranial electrode placements for seizure localization,”
Journal of Clinical Neurophysiology, vol. 31, no. 3, pp. 199–202, 2014.
[21] N. M. Wetjen, W. R. Marsh, F. B. Meyer, G. D. Cascino, E. So, J. W. Britton, S. M. Stead, and G. A.
Worrell, “Intracranial electroencephalography seizure onset patterns and surgical outcomes in nonlesional
extratemporal epilepsy,” Journal of Neurosurgery, vol. 110, no. 6, pp. 1147–1152, 2009.
[22] S. Liu, Z. Sha, A. Sencer, A. Aydoseli, N. Bebek, A. Abosch, T. Henry, C. Gurses, and N. F. Ince, “Ex-
ploring the time–frequency content of high frequency oscillations for automated identification of seizure
onset zone in epilepsy,” Journal of Neural Engineering, vol. 13, no. 2, p. 026026, 2016.
[23] M. Stead, M. Bower, B. H. Brinkmann, K. Lee, W. R. Marsh, F. B. Meyer, B. Litt, J. Van Gompel,
and G. A. Worrell, “Microseizures and the spatiotemporal scales of human partial epilepsy,” Brain, pp.
2789–2797, 2010.
[24] G. P. Kalamangalam, L. Cara, N. Tandon, and J. D. Slater, “An interictal eeg spectral metric for temporal
lobe epilepsy lateralization,” Epilepsy Research, vol. 108, no. 10, pp. 1748–1757, 2014.
[25] Y. Boykov and V. Kolmogorov, “An experimental comparison of min-cut/max-flow algorithms for energy
minimization in vision,” IEEE transactions on Pattern Analysis and Machine Intelligence, vol. 26, no. 9,
pp. 1124–1137, 2004.
[26] Y. Weiss and W. T. Freeman, “On the optimality of solutions of the max-product belief-propagation algo-
rithm in arbitrary graphs,” IEEE Transactions on Information Theory, vol. 47, pp. 736–744, 2001.
[27] P. Dagum, A. Galper, and E. Horvitz, “Dynamic network models for forecasting,” in Proceedings of the
8th International Conference on Uncertainty in Artificial Intelligence, 1992, pp. 41–48.
[28] A. Delong and Y. Boykov, “Globally optimal segmentation of multi-region objects,” in 2009 12th IEEE
International Conference on Computer Vision. IEEE, 2009, pp. 285–292.

10