A probabilistic approach for pediatric epilepsy diagnosis using brain functional connectivity networks

Epilepsy is a neurological disorder characterized by recurrent seizures with unspecified causes. The Center for Disease Control and Prevention (CDC) estimates that more than 2.3 million adults and half a million children in the United States are affected by Epilepsy [1]. This number is projected to dramatically increase every year by about 0.15 million newly diagnosed epilepsy cases [2]. Although the impact of seizures varies from person to person, physical and mental functions of the affected person could be severely altered. A systematic approach for epilepsy diagnosis could improve the planning for a treatment process and thus relieve the burden already imposed on the overall healthcare system. Scalp Electroencephalography (EEG) recording at resting state condition has been widely perceived as an effective preliminary tool for non-invasive study of the brain in individuals with epilepsy. Analysis of Scalp EEG during resting state condition, without performing a cognitive task and with the absence of external stimuli, has gained significant prominence for assessing brain function and related disorders. Applications include tasks that require assessing responses of the brain under the influence of different drug therapies [3], and tasks that rely on determining the 3D source localization of epileptic seizures which exploits techniques in the time/frequency domains for analysis of individual EEG electrode recordings [4, 5]. Assessment of brain functional connectivity network in patients suffering with various neurological disorders through modalities such as EEG recording, Magnetoencephalography (MEG) and functional Magnetic Resonance Imaging (fMRI) has elicited new findings in ways of underlying distinctions that delineate epileptic from control populations [6–14]. The high temporal resolution of EEG renders it as an indispensable tool in the primary diagnosis of epilepsy and in visualization of characteristic temporal events like interictal spikes which are closely associated with epileptic foci [7, 15]. Additionally, EEG has been utilized to distinguish focal and generalized neurophysiologic correlates of epilepsy [16]. Diagnosis of epilepsy by the means of scalp EEG visual inspection often involves long term recording and investigation by the EEG expert to search for abnormal activities.

However, visual inspection and interpretation of continuous temporal EEG recordings is tedious, time consuming and prone to human error. Furthermore, epilepsy diagnosis based on visual inspection of EEG has been shown to be very subjective to the expert opinion [17]. This has led to the general cohort of adopting various computer aided techniques with the help of machine learning for medical applications [18–20]. Artificial Neural Network (ANN) based classifiers have received the most attention towards epilepsy diagnosis using scalp EEG recordings [21–23] where the accuracy rate of 0.92 [21] and 0.8 [22, 24] are reported which involved the existence of training set and a priori knowledge. The general routine of ANN based techniques is to process each isolated EEG signal with the aim of extracting a set of discriminating features as input to train an ANN in the design of an optimal classifier and predictor of the diagnosis. Considering the fact that Epilepsy is a complex disease with multifactorial causes, makes the diagnostic process much more complicated than simply relying on solely model driven knowledge. Furthermore, the human brain includes a complex web of neuronal interconnectivity and discrete anatomical regions that function together to generate brain activity [11, 25]. This underlying functional brain infrastructure suggests that a methodology for enhanced epilepsy diagnosis needs to consider the whole brain network as described by its patterns of functional connectivity networks (FCNs). Thus, FCNs seek to define the patterns of cross-correlation between discrete functionally characterized brain regions to give statistical importance to anatomical connectivity (upon the existence of physical connection) and subsequently determining inter-regional neurophysiological inferences. FCNs could be grouped into two broad categories: Directed and Undirected. Undirected FCNs finds the dynamic associations among functional brain regions without considering the hypothetical causalities among them. Whereas Directed FCNs, sometimes referred to as effective connectivity [26], assesses the influence of one cerebral region upon another and therefore gives direction to the calculated associations. Current trends in adopting FCNs for understanding the complex brain are placed toward developing data driven methodologies for constructing FCNs which benefits from a robust parcellation of functional data of the brain and an objective formulation of the hypothesized association among functional parcels [6, 26]. The crucial role of time delay in the dynamics of large scale networks [27] such as brain networks is well motivated, due to the large scale property of brain connectivity networks including discrete sub-networks [28], but yet not fully understood and incorporated in constructing the brain networks and decision making processes [29]. Time delay is coupled with the fact that on large scale networks such as brain networks, the recorded signal at each electrode is in fact showing the summation of the variance of the brain area closer to the electrode and a lagged version of variances from other brain areas.

In this study, a new algorithm based on time-varying associations among channels of scalp EEG using a moving window is examined for its ability to identify multichannel EEG segments with epilepsy driven characteristics without a priori knowledge assumed about the diagnosis. Undirected FCNs estimate the association between the channels of EEG using a geometrical distance between a pair of channels. Undirected FCNs can be represented in the form of undirected graphs. Each electrode is considered as a node and the functional associations among them are the edges of the corresponding graph. Different applications of graph theory and small world networks [30], with causality analysis combined with network analysis [31] and time-frequency coupling detection among isolated scalp EEG recordings [32, 33] are a few of the widely used model-based solutions for studying FCNs using scalp EEG recordings. This study also assesses the merit of graph theory based features, extracted from the graphs corresponding to the FCNs, in identifying EEG segments recorded from patients with epilepsy from segments recorded from the pediatric control group. The main contribution of the study is in introducing a novel approach based on purely machine learning techniques to facilitate the screening process of potential epileptic patients. The emphasis of developing the new decision support system for pediatric epilepsy diagnosis was placed on the accurate diagnosis without a priori knowledge. The probabilistic decision on the subject diagnosis makes the system more applicable in clinical environments. An overview of the proposed decision support system for computer aided diagnosis of pediatric epilepsy is as shown in Figure 1.

Figure 3 shows the constructed undirected FCNs averaged for both pediatric epilepsy (PE) and pediatric control (PC) groups with their corresponding bar plots of connectivity distances (degrees) and under different degrees of strengths. The symmetric property of the adjacency matrices is due to the nature of the constructed FCNs which do not take into consideration the possible causality among brain cortical region. The change in the pattern of functional associations is observed by increasing the distances among some brain regions such as cz-o1 and cz-o2 pairs. A two-sided two sample student t-test was conducted to assess whether there is a significant difference between the undirected mean FCNs of the PC group and the PE group. The test results indicated that there is a statistically significant difference (t(340) = -9.89, p < 0.0001) between the mean FCNs of the PC group (μ _PC = 44.56°, σ _PC = 13.75° σ _PC = 13.75°) and the PE group (μ _PC = 57.74°, σ _PE = 10.70°). Furthermore, we performed a two samples student t test to inspect the existence of statistical significant difference among each connection between source electrode and target electrode (index of electrode pair in Figure 3(b) and 3(d)) across the pediatric control and pediatric epilepsy group. Results are visualized in Figure 3(e). Existence of statistically significant (p < 0.00001 Bonferroni adjusted for multiple comparisons) difference in the connection strength (degree) of a specific source-target pair of electrodes is shown as a black box unit, insignificant difference in the connection strength is shown as a white box unit. As the large number of black boxes in Figure 3(e) states, the brain acts as a complex network and epilepsy affects more than single connection. However this needs to be more investigated on a subject basis as the type of epilepsy (focal or generalized) could vary from subject to subject.

Adjacency matrices shown in Figure 3 are mapped into the graph representations shown in Figure 4. Analysis of the single connections of the graphs is beyond the scope of this study; however the reduction in the pattern of inter-connections among isolated brain regions is visually observable by comparing the two graphs; Table 3 provides the statistics of the features extracted for both graphs and tests for the differences among PC and PE groups. We also performed a connection density assessment on left and right hemispheres and the interconnection among the hemispheres across subjects. Density of connections on each hemisphere were calculated by counting the number of edges with the weight less than or equal to 45° (mid-point in connectivity strength). The results summarized in Table 4 show alterations in the wiring pattern of the brain functional network caused by epilepsy. This alteration is in the form of reduced density of connectivity in both left and right hemisphere as well as inter-hemispheric connectivity. Note also how the connectivity maps for both the PE and PC groups as the threshold of the connectivity strength is changed.

This mis-identification could be due to different factors such as the window size, number of segments required to accurately diagnose or the type of epilepsy which needs more investigation.

The assumption of the existence of no priori knowledge on the diagnosis in the clinical environment could be relieved by assuming the existence of symptoms and the minimal knowledge of a training set of multi-channel EEG segments which could be considered as a tuning approach in the decision making process. A training set was composed including twenty randomly chosen multi-channel EEG segments from the total set of EEG segments and Support Vector Machine (SVM) with a linear kernel were trained to classify the segments. The testing set was then given to the system after self tuning and the results showed 100% accuracy in classification accuracy of the multi-channel EEG segments.

A novel decision support system for computer aided diagnosis of pediatric epilepsy using machine learning techniques was presented. The approach taken in the system was based on constructing functional connectivity networks (FCNs) of the brain and analyzing graph theoretical based features to identify the wiring pattern differences among pediatric control (PC) and pediatric epilepsy (PE) groups. The system is designed to provide clinicians with initial screening results about the likelihood of a given subject to be epileptic or not. The time-varying FCNs implementation increases the resolution by segmenting the multichannel EEG. This created the potential of diagnosing epilepsy without need for long EEG recording session and time-consuming visual inspection as conventionally employed. The main contribution of the study is the reliance of the algorithm on machine learning techniques to facilitate the screening process of potential epileptic patients without need of a priori knowledge and without need for a training phase. The suggested window length in constructing FCNs, the number of principal components (dimension of GMM) to be used, and the inspection of possible causal relationships among cortical brain regions are areas that need further investigation on the basis of a larger dataset.