Differences in functional network between focal onset nonconvulsive status epilepticus and toxic metabolic encephalopathy: application to machine learning models for differential diagnosis

We investigated the differences in EEG network properties during periods of periodic discharges between focal-onset NCSE and TME. The major findings were as follows: (1) the property of EEG functional connectivity was more integrated and efficient in TME, relative to NCSE; (2) machine learning algorithms, including the EEG global graph measures as input features, could classify the NCSE and TME with high accuracy, and gradient boosting was the most accurate classification model with an AUC > 0.9; and (3) in addition to laboratory findings suggesting metabolic derangement, EEG graph measures reflecting integrated and efficient network properties (i.e., high degree and global efficiency in the alpha band, high local efficiency, and low modularity in delta band) were found to be contributing features in the gradient boosting model to predict TME.

Recent advances in graph theoretical network analysis enable the assessment of the topological architecture of complex human brain networks (Farahani et al. 2019; Sporns 2013b). Therefore, graph theoretical analysis has been widely applied to quantify characteristics of functional network in terms of efficiency and integration in various neurological disorders, as well as conditions affecting brain function (Bullmore and Sporns 2009; Choi and Kim 2021; Kim et al. 2021; Park et al. 2020; van Diessen et al. 2013). Our findings of relatively high degree, strength, and global efficiency in TME suggest that network properties of TME might be more integrated and efficient, relative to those of focal-onset NCSE (Bullmore and Sporns 2009). Modularity is regarded as a global graph measure of network segregation (Sporns 2013a), which is calculated by partitioning a network into groups of modules with high connectivity within modules relative to the connectivity between regions in distinct modules (Girvan and Newman 2002; Newman 2006). In addition, high assortativity could be interpreted as high-degree nodes have a high tendency to connect with each other (Bullmore and Sporns 2009). Taken together, the higher modularity and assortativity found in focal-onset NCSE compared with that in TME suggest that network properties of focal-onset NCSE may tend to be segregated into functionally stable regions and irritative zones having strong local connection around the ictogenic focus (Bialonski and Lehnertz 2013), whereas those of TMEs may be globally integrated (Sporns 2013a).

The mechanism underlying the more integrated and efficient network properties in patients with TME, relative to those with NCSE, is not fully understood. It has been widely accepted that GPDs with triphasic morphology, the EEG hallmark of TME, may originate from dysfunction of the thalamocortical circuits with recruitment of metabolically impaired cortical neurons (Karnaze and Bickford 1984). Based on the hypothesis regarding the generation of periodic discharges in the TME, it is plausible that the simultaneous involvement of pathophysiological mechanisms in widespread brain areas might explain the altered EEG network with highly synchronized and efficient properties in the TME. There are dynamic changes in network topology in patients with seizures (van Diessen et al. 2013). The ictal period could be characterized by a more synchronized and integrated network configuration of the epileptic brain (Ponten et al. 2009; Schindler et al. 2008; van Diessen et al. 2013). The number of connections gradually declines to preictal levels after the midictal phase, resulting in a less synchronizable and disintegrated network topology after ictal termination (Ponten et al. 2009; Schindler et al. 2008; van Diessen et al. 2013). Therefore, our findings of a more segregated network property may be in accordance with the network configuration of postictal or interictal periods of focal-onset NCSE. Given that focal-onset NCSE and TME may have different pathophysiologies in terms of network, graph theoretical measures could be considered as important biomarkers to differentiate focal-onset NCSE and TME.

We found that machine learning algorithms based on EEG graph measures could classify focal-onset NCSE and TME with relatively high accuracy. These findings suggest that quantitative analysis indicators that utilize network properties can complement the limitations of the visual interpretation of EEG for differential diagnosis between the two conditions. Indeed, differences in graph measures between focal-onset NCSE and TME were observed predominantly in the delta band, suggesting the presence of invisible pathophysiological differences within periodic discharges with delta frequency between the conditions. Furthermore, the accuracy of differential diagnosis was beyond 90% using only the short EEG data for 20 s, thus implying that the graph theoretical analysis could be an optimal framework for quantitatively differentiating network properties between focal-onset NCSE and TME, and that graph measure-based machine learning algorithms may be clinically useful for distinguishing the conditions.

Among the variables, four EEG graph measures (i.e., degree in the alpha band, global efficiency in the alpha band, local efficiency in the delta band, and modularity in the delta band) played important roles in the differential diagnosis, which were included in the top 10 most contributing features. Considering that laboratory findings may differ according to the causative condition of TME, EEG graph measures could contribute to the machine learning model as pivotal features to consistently differentiate between focal-onset NCSE and TME. To verify the accuracy of differential diagnosis of the machine learning algorithm using EEG graph measures and to determine the contributing power of EEG graph measures, a subgroup analysis of the TME for each causative condition is required with a larger study population.

The present study had several limitations that should be considered when interpreting the results of this study. First, the sample size was relatively small. Second, we could not externally validate the accuracy of machine learning models. Therefore, the generalizability of the accuracy of the differential diagnosis between focal-onset NCSE and TME might be limited, although internal validation has been statistically performed. Third, as an inherent limitation of the retrospective study design, we could not identify the difference in network properties according to the change in severity in patients with TME. Further longitudinal studies are required to develop a machine learning algorithm for the differential diagnosis between focal-onset NCSE and TME by confirming the changes in network properties associated with the severity of TME. Finally, there are various etiologies of the development of NCSE (Lee et al. 2021); therefore, the property of the NCSE network might overlap with that of TME when the occurrence of NCSE is due to metabolic derangement. Additional analyses are required to determine whether there are distinct network characteristics that can be used for the differential diagnosis of each cause of NCSE.

One strength of our study is that EEG global graph measures were not dependent on the location of the ictogenic focus. Therefore, the proposed gradient boosting model based on EEG global graph measures in our study could be applied in general to distinguish focal-onset NCSE from TME, regardless of the location of ictogenic focus. The application of an interpretable and explainable model is another strength of our study. The model enables physicians to recognize what the important clinical and EEG variables are, which can promote the clinical usefulness of applying the model as a decision support system in the differential diagnosis between focal-onset NCSE and TME. It may also provide insights for further studies to understand the mechanisms of generating periodic discharges in the TME and focal-onset NCSE from the perspective of functional networks.

留言 (0)

沒有登入
gif