Identification of nonconvulsive status epilepticus in the ictal-interictal continuum using artificial intelligence: a prospective observational cohort study

Article information

encephalitis. 2026;6(2):48-59

Publication date (electronic) : 2026 March 16

doi : https://doi.org/10.47936/encephalitis.2025.00157

Sungyeong Ryu, Dong Ah Lee, Kang Min Park

Department of Neurology, Inje University Haeundae Paik Hospital, Inje University College of Medicine, Busan, Korea

Correspondence: Kang Min Park Department of Neurology, Inje University Haeundae Paik Hospital, Inje University College of Medicine, 875 Haeundae-ro, Haeundae-gu, Busan 612896, Korea E-mail: smilepkm@hanmail.net, H00246@paik.ac.kr

Received 2025 December 8; Revised 2026 January 6; Accepted 2026 January 13.

Abstract

Introduction

The term “ictal-interictal continuum (IIC)” was first introduced by Pohlmann et al. in 1996 [1]. It describes a spectrum of electroencephalography (EEG) patterns between interictal and ictal states, ranging from relatively benign, transient epileptiform activity to more malignant, evolving electrophysiological changes with harmful clinical consequences [2]. With the growing use of continuous EEG monitoring in intensive care units, IIC patterns are being recognized more frequently in critically ill patients [3]. Despite over 25 years of research, however, the pathophysiology, epileptogenic potential, and clinical significance of IIC EEG patterns in this population remain poorly understood [2,4].

Patients on the IIC can be classified as those with nonconvulsive status epilepticus (NCSE) and those with coma-IIC. NCSE is characterized by altered consciousness without major convulsive movements caused by an epileptic condition, whereas coma-IIC is a state of coma caused by the underlying pathology [5,6]. Accordingly, anti-seizure medications (ASMs) are generally required for NCSE but are often unnecessary for coma-IIC [3,7]. However, distinguishing these entities remains challenging, creating a clinical dilemma regarding whether and how aggressively to treat patients on the IIC. EEG patterns with fast rhythms (>3 Hz) or rapidly evolving activity are widely considered ictal and warrant ASM treatment, as they are linked to increased metabolic demand, tissue hypoxia, and impaired cerebral autoregulation [8]. Additional multimodal approaches can be helpful for distinguishing between patients with NCSE and those with coma-IIC, such as structural brain magnetic resonance imaging (MRI), fluorodeoxyglucose-positron emission tomography (PET), perfusion MRI, single-photon emission computed tomography, or magnetoencephalography [9-11].

Increasing evidence from electrophysiology, histopathology, and neuroimaging supports the concept of epilepsy as a disease of the brain network, with alterations of connectivity [12,13]. Therefore, many efforts have been made to analyze the structural and functional connectivity of the brain at the individual and group levels. Among various connectivity analysis techniques, graph theoretical analysis of structural and functional data offers the opportunity to understand the details of complex networks and derive models for these networks using a mathematical framework to model pairwise communications [14]. This approach has provided compelling evidence that epilepsy is characterized by marked alterations in the topology of large-scale networks, which is associated with clinically relevant parameters, including disease duration and medical or surgical outcomes, suggesting the possible utility of topological markers in the diagnosis and management of epilepsy [12,14-16].

Machine learning is a type of artificial intelligence (AI) in which algorithms learn patterns directly from data rather than being explicitly programmed [17,18]. Deep learning, a subfield of machine learning, further reduces the need for handcrafted feature specification by humans [19,20]. With the increasing utilization of AI in the field of medical science, a few neurology researchers have used AI for seizure detection and diagnosis of epilepsy as well as the identification of sleep disorders [21-23]. However, to date, no studies have used machine learning or deep learning to distinguish between patients with NCSE and those with coma-IIC.

In this study, we analyzed functional connectivity based on EEG data using graph theory in patients on the IIC (both patients with NCSE and those with coma-IIC) and healthy controls. The aim of this study was to investigate differences in functional connectivity between patients on the IIC and healthy controls, and between patients with NCSE and those with coma-IIC. In addition, we evaluated the ability of machine learning based on functional connectivity measures and a deep learning approach to classify patients with NCSE and those with coma-IIC. We hypothesized that functional connectivity based on EEG would differentiate between patients with NCSE and those with coma-IIC, and that machine learning and deep learning approaches would be useful for distinguishing between the two groups.

Methods

Machine learning approach

We assessed whether machine learning could distinguish patients with NCSE from those with coma-IIC. The classifier was trained using graph-theoretical network measures (assortativity, average degree/strength, characteristic path length, clustering coefficient, diameter, eccentricity, global and local efficiency, modularity, radius, small-worldness, and transitivity) across the delta, theta, alpha, beta, and gamma bands. All features were standardized using z-scores, and a supervised support vector machine (SVM) model was applied [32]. Data were split into training (80%) and testing (20%) sets, and all analyses were conducted in MATLAB R2020b. Classification performance was evaluated using accuracy and receiver operating characteristic curves. The overall workflow of functional connectivity analysis is shown in Figure 1.

Figure 1.

Process of functional connectivity analysis

SVM, support vector machine.

Results

Clinical characteristics

We enrolled 72 patients on the IIC with a mean age of 67.5±13.8 years. Thirty-four patients (47.2%) were male and 38 patients (52.7%) were female. Median GCS was 6.5 (range, 2–15), and median mRS was 5 (range, 1–6). The most common etiology of coma was anoxic-hypoxic encephalopathy (25 patients), followed by epileptic encephalopathy (15 patients), metabolic encephalopathy (14 patients), and others (18 patients). Of the 72 patients on the IIC, 53 patients were assigned to the NCSE group, and 19 patients to the coma-IIC group.

With regard to EEG location, generalized IIC (59 patients) was the most common type, followed by lateralized (11 patients) and bilateral independent (two patients) types. In the EEG patterns, periodic discharges were the most common type of IIC, followed by rhythmic delta activities (18 patients) and sharp waves (10 patients). The median frequency of IICs was 1.5 Hz (range, 1–4.5 Hz). Forty patients had modifiers in addition to IIC EEG patterns.

Table 1 shows the differences in clinical characteristics between patients with NCSE and those with coma-IIC. Compared with the NCSE group, the proportions of anoxic-hypoxic encephalopathy and male patients were significantly higher in the coma-IIC group, whereas GCS scores were higher and EEG modifiers more frequent in the NCSE group.

Table 1

Differences in clinical characteristics between patients with NCSE and those on the coma-IIC

Differences in functional connectivity between patients on the ictal-interictal continuum and healthy controls

Table 2 shows the differences in functional connectivity between patients on the IIC and healthy controls. In patients with IIC EEG patterns, the average strength and global efficiency of the alpha band were decreased, whereas its modularity, radius, characteristic path length, and eccentricity were increased compared to healthy controls. The assortative coefficient of the beta band was decreased in patients on the IIC compared to healthy controls. The characteristic path length, diameter, eccentricity, and radius of the theta band were decreased in patients on the IIC, whereas the average strength, mean clustering coefficient, global efficiency, local efficiency, modularity, small-worldness index, and transitivity of this band were increased. The characteristic path length, diameter, eccentricity, and radius of the delta band were decreased in patients on the IIC, whereas the assortative coefficient, average strength, mean clustering coefficient, global efficiency, local efficiency, modularity, and transitivity were increased. The modularity of the gamma band was decreased in patients with IIC EEG patterns, whereas the average degree, average strength, and global efficiency were increased compared to healthy controls.

Table 2

Differences in functional connectivity between patients on the IIC and healthy controls

Differences in functional connectivity between patients with nonconvulsive status epilepticus and those in coma on the ictal-interictal continuum

Table 3 shows the differences in functional connectivity between patients with NCSE and those with coma-IIC. With regard to the alpha band, the average strength, mean clustering coefficient, global efficiency, local efficiency, and transitivity were lower in patients with NCSE than those in patients with coma-IIC. The average strength, global efficiency, and local efficiency of the beta band were decreased in patients with NCSE compared to those in patients with coma-IIC, whereas the average strength, mean clustering coefficient, global efficiency, local efficiency, small-worldness index, and transitivity of the theta band were decreased but its radius increased in patients with NCSE compared to those in patients with coma-IIC. Similarly, the mean clustering coefficient, local efficiency, and transitivity of the delta band were lower in patients with NCSE than those in patients with coma-IIC, whereas the average strength, mean clustering coefficient, global efficiency, local efficiency, and transitivity of the gamma band were decreased, but the characteristic path length, diameter, eccentricity, and radius of this band increased in patients with NCSE compared to patients with coma-IIC.

Table 3

Differences in functional connectivity between patients with NCSE and those on the coma-IIC

Discussion

In this study, we investigated the functional connectivity of patients with IIC EEG patterns and healthy controls using graph theory. Overall, patients with IIC EEG patterns showed increased global functional connectivity compared with controls, particularly in the theta, delta, and gamma bands. In addition, we observed significant differences in functional connectivity between patients with NCSE and those with coma-IIC. Patients with NCSE had decreased global functional connectivity compared to patients with coma-IIC. Furthermore, the machine learning classifier based on functional network measures had an accuracy of 92.8% in differentiating patients with NCSE from those with coma-IIC, and the accuracy of the deep learning of CNN models to classify them was 73.9%.

To date, structural and functional connectivity in patients on the IIC has not been investigated, although connectivity can be assessed using modalities such as diffusion tensor imaging for structural networks and EEG, magnetoencephalography, functional MRI, or PET for functional networks. Among these methods, EEG offers a high temporal resolution, making it particularly useful for determining how neural signals change with time [14]. It can show the organization of intrinsic brain activities and functional brain organization [33]. In patients with IIC EEG patterns, global efficiency was increased and characteristic path length decreased compared to the EEG patterns of healthy controls, indicating enhanced network integration. Integration reflects the efficiency of global information transfer across the network, whereas segregation reflects local specialization; accordingly, increased local efficiency and mean clustering coefficient suggest strengthened network segregation in patients on the IIC [14,34]. In addition, the small-worldness index was higher in patients on the IIC than in healthy controls, suggesting a more optimal balance between network integration and segregation and more efficient information transfer at relatively low wiring cost [34,35]. All of these findings suggest increased functional connectivity in patients with IIC EEG patterns compared to healthy controls. These findings are in agreement with a previous study that identified the characteristic abnormalities of connectivity during status epilepticus (SE) in a pilocarpine-induced rodent model [36]. The study showed that the coherence connectivity among the frontal cortex, hippocampus, and thalamus significantly increased during SE episodes in almost all EEG frequency bands, demonstrating aberrant network communication and enhanced connectivity in SE [36].

We also investigated differences in functional connectivity between patients with NCSE and those with coma-IIC. The average strength, global efficiency, and local efficiency of the alpha, beta, theta, and gamma bands were lower in patients with NCSE than in patients with coma-IIC. In addition, the mean clustering coefficient of the alpha, theta, delta, and gamma bands was decreased in patients with NCSE compared to patients with coma-IIC. All of these findings suggest decreased functional connectivity in patients with NCSE than in patients with coma-IIC. Although seizures typically involve hypersynchronization, our study showed decreased functional connectivity in patients with NCSE compared to those with coma-IIC. This apparent paradox can be explained by the distinct underlying pathophysiology and structural etiology of the two groups. In our cohort, the coma-IIC group had a significantly higher prevalence of anoxic-hypoxic encephalopathy (63.1%) compared to the NCSE group. Severe diffuse brain injury, such as anoxic encephalopathy, is known to cause selective loss of inhibitory interneurons and widespread cortical disinhibition [37,38]. Furthermore, recent computational models suggest that impaired astrocytic gap-junction coupling may contribute to pathologically enhanced synchronization or bistable network dynamics, potentially reflecting disrupted homeostatic regulation [39].

Conversely, the NCSE group comprised a higher proportion of patients with focal structural lesions, such as acute stroke or tumors. These focal lesions are known to disrupt white matter tracts, causing local network disconnection. Consequently, even in the presence of ictal activity, the global functional connectivity in NCSE patients may be attenuated by these structural disconnections, rendering it lower than the pathologically elevated global synchronization seen in the diffusely damaged brains of the coma-IIC group. A recent study found significantly lower global network efficiency in focal-onset NCSE than diffuse encephalopathy, highlighting the divergent topological effects of focal versus diffuse pathology [40]. Furthermore, a metabolic study demonstrated that NCSE is characterized by focal hypermetabolism, reinforcing the notion that NCSE is a spatially restricted process distinct from the globally hypersynchronous state observed in the diffusely damaged brains of coma-IIC patients [41].

Clinically, patients with coma-IIC often show a poorer response to ASM than those with NCSE. We hypothesize that the pathologically elevated functional connectivity observed in the coma-IIC group might reflect a state of ‘network failure’ or extensive damage that is less likely to respond to standard synaptic modulation by ASM. However, as we did not directly test ASM efficacy in the current study, this potential link between hyperconnectivity and pharmaco-resistance remains speculative and warrants further investigation.

Machine learning, a branch of AI increasingly adopted in clinical research, has also been applied in epilepsy for automated seizure detection, diagnosis, localization, and prediction of treatment outcomes [21,42-44]. In this study, we successfully demonstrated that a machine learning approach based on functional network measures could differentiate between patients with NCSE and those with coma-IIC with an accuracy of 92.8%, which is a high discriminatory ability. This may assist clinical decision-making in patients with IIC EEG patterns. Interestingly, when we trained the machine learning algorithms with functional network measures in all EEG frequency bands, the ability of machine learning to differentiate patients with NCSE and those with coma-IIC was the highest. Because not all features are informative and some may introduce noise, careful feature selection is often essential in machine learning [45]. However, our results indicate that functional network measures across all EEG frequency bands improved the ability of machine learning to differentiate NCSE from coma-IIC, which suggests that brain network alterations are distributed across multiple frequency ranges.

We analyzed functional connectivity according to EEG frequency bands and observed some differences in connectivity according to band type. Patients on the IIC and those with NCSE had increased global functional connectivity in the theta, delta, and gamma bands, and decreased global functional connectivity in the alpha band compared to healthy controls. A previous study that investigated differences in frequency-specific functional connectivity using graph theory reported that the lower the frequency of the oscillation-based network, the higher the mean network efficiency [46]. These results indicate that network measures vary systematically across frequency bands, potentially reflecting frequency-specific connectivity properties. Another plausible explanation is the effect of artifacts in different frequency bands. High-frequency activity such as that seen in the gamma band obtained by scalp recordings can overlap with face and eye muscle activities and is also affected by volume conduction [47]. Additionally, the delta frequency band is known to be more readily affected by the band filter than the other frequency bands [48].

The accuracy of diagnosing NCSE based on EEG using CNN in this study was about 74%, which is relatively modest. There are several possible reasons for this. One study reported that the ability of deep learning is strongly dependent on total sample size [49]. Moreover, in most studies, there was a positive correlation between the overall accuracy of deep learning and sample size. Thus, our results could be attributed to our relatively small sample size. Another possibility is the use of a ‘light’ CNN model. We applied the specific CNN model VGG19 in this study. The VGG19 model is a variant of VGGNet, created in 2014, which consists of 19 layers [50]. Because VGG19 is a simple model, it is easy to use and has the advantage of short analysis times, and is thus widely used in medical fields. However, its accuracy is relatively low. Accuracy may have been better if a different analysis model had been used, such as NASNet-large or DenseNet-201.

This is the first study to examine functional connectivity in patients on the IIC, demonstrating significant differences in EEG patterns between these patients versus healthy controls and between NCSE and coma-IIC patients. Nevertheless, several limitations should be acknowledged. First, we analyzed functional connectivity based on EEG recordings, which were only recorded for approximately 30 minutes. Given the fluctuating nature of IIC, this may not capture temporal dynamics and could have introduced selection bias. Longer-term EEG recordings with dynamic network analysis are needed [46]. In particular, long-term EEG monitoring lasting more than 24 hours may be more appropriate for functional connectivity analysis. Second, we obtained EEG data at one time point using cross-sectional data. Longitudinal follow-up studies may be more appropriate for investigating changes in functional connectivity. Third, baseline differences such as GCS and etiology are largely intrinsic to the clinical definitions of NCSE and coma-IIC. Due to the limited sample size (n = 19 in the coma-IIC group), robust multivariate adjustment was not feasible, limiting our ability to fully disentangle these effects from the IIC patterns themselves. Fourth, the machine learning classification relied on a single train-test split due to the limited sample size. We acknowledge the potential for overfitting in this small dataset. Thus, the reported accuracy should be interpreted as a preliminary proof-of-concept. Future studies with larger cohorts and robust cross-validation are needed to verify generalizability.

We found that patients with IIC EEG patterns exhibited distinct functional network alterations compared to healthy controls. Specifically, while the overall IIC group showed increased global integration, patients with NCSE exhibited significantly decreased global functional connectivity compared to those with coma-IIC. These divergent network patterns likely reflect underlying pathophysiological differences: focal structural disconnection in NCSE versus pathological hypersynchronization driven by diffuse disinhibition in coma-IIC. Furthermore, the machine learning approach based on these functional network measures demonstrated high diagnostic accuracy in distinguishing between these two conditions. These findings suggested that EEG-based functional connectivity analysis could serve as a valuable biomarker for understanding the complex pathophysiology of the IIC and might aid clinicians in the difficult differential diagnosis between NCSE and coma-IIC.

Supplementary Materials

Article information Continued