Epilepsy patients hospitalized in the Affiliated Hospital of Xuzhou Medical University from January 2014 to October 2019 underwent extensive clinical examination, including full medical record evaluation, laboratory examination, electroencephalography (EEG) and magnetic resonance imaging (MRI). According to the criteria proposed by the International League against Epilepsy (25), we first screened 205 patients with focal to bilateral tonic-clonic seizures of unknown etiology from January 2014 to December 2017 as a modeling cohort. Thereafter, we screened 49 patients with focal to bilateral tonic-clonic seizures of unknown etiology from January 2018 to October 2019 as a validation cohort. The enrolled were newly diagnosed or monotherapy epilepsy patients. We collected a total of 254 fresh-frozen serum samples and 20 cerebrospinal fluid (CSF) samples. Written informed consent to participate in the study was obtained from all patients or their legal guardians before sample collection. Major exclusion criteria were patients with epilepsy related to structural abnormalities (hippocampal sclerosis, focal cortical dysplasia, brain malformations, cerebrovascular arteriovenous malformations, cerebral infarction, cerebral hemorrhage, craniocerebral trauma, and brain tumors, etc.), metabolic encephalopathy, inherited disease, severe cognitive impairment, the acute phase of central nervous system infection, autoimmune encephalitis (NMDA, AMPA1, AMPA2, LGI1, CASPR2, lgLON5, GABAB, GAD65, GlyR1, mGluR1, and mGluR5), demyelinating diseases, systemic lupus erythematosus, paraneoplastic syndrome, and psychogenic nonepileptic seizures. The patients were followed up for ≥ 24 months. The Morisky Medication Adherence Questionnaire (MMAS-8) was applied to evaluate the compliance of patients (26). MMAS-8 scores of < 6, 6-7, and 8 indicated poor, moderate, and high adherence, respectively. Drug-responsive epilepsy (n=166) was defined as a seizure-free period of at least 12 months or three times the longest seizure interval, whereas DRE (n=76) was defined as seizures at any frequency in the last 12 months after the administration of ≥ 2 drugs at optimal doses (27). Clinical data and serum samples as well as cerebrospinal fluid samples were collected. This study was approved by the Ethics Committee of Affiliated Hospital of Xuzhou Medical University (XYFY2016-KL017-02 and informed consent was obtained from each subject).

Serum GluR3B antibody from epilepsy patients was examined in a blinded fashion by ELISA as described previously (23). Immunoplates (Nunc, Roskilde, Denmark) were coated with 100 µl per well of amino acids 372-395 of human GluR3B peptide with the sequence of NEYERFVPFSDQQISNDSSSSENR (20 mg/l, synthesized by Shanghai Qiangyao Biotechnology Company, Shanghai, China). Other plates were coated with 100 µl per well of phosphate-buffered saline (PBS) with 1% bovine serum albumin (BSA). Plates were incubated overnight at 4°C and then washed with PBS containing 1% BSA. To block nonspecific binding, 100 µl PBS containing 1% BSA was added to each well of the enzyme plates and incubated for 2 h at 37°C. After the microtiter wells were washed twice, serum samples from epilepsy patients were added to microtiter wells (100 µl/well) and incubated overnight at 4°C. After washing, 100 µl of horseradish peroxidase (HRP)-conjugated goat anti-human IgG (1:1000, Jackson Immunoresearch, West Grove, PA, USA) was added to each well of the GluR3B-coated and BSA-coated plates, and the plates were incubated for 2h at 37°C. Then, 100 µl peroxidase substrate (Kierkegaard and Perry Laboratories, Gaithersburg, MD) was added to each microtiter well. Optical density (OD) at 405 nm was measured using a microplate reader (Bio-Rad, Hercules, CA, USA). Repeated OD measurements were recorded 3 times. GluR3B antibody levels in each serum or CSF were determined using the following equation: [average OD of duplicate wells in the GluR3B-coated plate] - [average OD of duplicate wells in the BSA-coated plate].

The statistical analysis was performed by SPSS software (version 18.0, SPSS Inc., Chicago, IL, USA) and R (version 3.6.0). The nonparametric Mann-Whitney U test was applied to compare the average GluR3B antibody serum level between DRE and drug-responsive epilepsy. The clinical value of the GluR3B antibody level in discriminating DRE from drug-responsive epilepsy was determined by logistic regression analysis. Specifically, we constructed the full model (including the serum GluR3B antibody level as well as age, sex, epilepsy duration, age at first seizure, frequency of seizures before treatment, and initial therapy response) in the training data (n = 193) and validated the performance of the fitted model using the test data (n = 49) by calculating the area under the curve (AUC) of the ROC curve. The GluR3B antibody level with the maximum AUC was selected as the cutoff value. Afterwards, we transformed the GluR3B antibody level from a continuous variable into a binary variable in terms of this clinical value. Logistic regression models were further applied to evaluate the independent importance of the binary GluR3B antibody level while adjusting for age, sex, epilepsy duration, age at first seizure, frequency of seizures before treatment, and initial therapy response. Finally, decision curve analysis (DCA) was utilized to evaluate the clinical net benefit in predicting DRE according to Vickers et al. (28). The significance level was set to be 0.05 throughout our study.

In the current study, there were 193 patients in the modeling cohort and 49 patients in the validation cohort with focal to bilateral tonic-clonic seizures. In the modeling cohort, 12 patients were excluded from the statistical analysis due to poor adherence to the prescribed medication or loss to follow-up. The clinical features of the patient cohort are summarized in Table 1 .

Our final study modeling cohort (n=193) included 60 patients (31.1%) with drug-resistant seizures and 133 patients (68.9%) with drug-responsive seizures. In the modeling cohort, we found that 70 (36.3%) patients had elevated levels of GluR3B antibody. Among them, the proportion of GluR3B antibody positive for newly diagnosed and monotherapy epilepsy were 26.9% (11/41) and 38.8% (59/152), respectively. The mean age at first seizure was 22.6 and 29.2 years, whereas the mean epilepsy duration was 10.1 years in patients with DRE and 5.5 years in patients with drug-responsive epilepsy. High frequency of seizures before therapy was present in 18 drug-resistant patients (30.0%) compared to 8 drug-responsive patients (6.0%). Poor response to initial therapy was present in 42 drug-resistant patients (70.0%) compared to 18 drug-responsive patients (13.5%). In this regard, the average numbers of drugs administered to drug-resistant patients and drug-responsive patients were 2.2 and 1.2, respectively. Levetiracetam (41.2%), oxcarbazepine (30.8%), carbamazepine (25.4%), valproic acid (22.5%) and lamotrigine (21.6%) were the most frequently used.

Our final study validation cohort (n=49) included 16 patients (32.7%) with drug-resistant seizures and 33 patients (67.3%) with drug-responsive seizures. The proportion of patients with GluR3B antibody positive was 38.8% (19/49). There were 7 (38.9%) newly diagnosed patients and 12 (38.7%) monotherapy epilepsy with GluR3B antibody positive. The mean age at the first seizure was 22.7 and 23.5 years, whereas the mean epilepsy duration was 9.3 and 4.8 years in patients with DRE and drug-responsive epilepsy, respectively. High frequency of seizures before therapy was present in 8 drug-resistant patients (50.0%) compared to 5 patients with a favorable drug response (15.2%). There were 11 drug-resistant patients (68.8%) with poor response to initial therapy compared to 3 drug-responsive patients (9.1%). Regarding medication, the average numbers of drugs administered to drug-resistant patients and drug-responsive patients were 2.3 and 1.1, respectively. Levetiracetam (44.6%), oxcarbazepine (29.5%), carbamazepine (25.0%), valproic acid (23.6%) and lamotrigine (18.4%) were the most frequently used.

For the patients in the modeling cohort, serum GluR3B antibody levels were 0.40 ± 0.22 OD (95% confidence interval [CI] 0.35-0.46, n=60) in drug-resistant patients and 0.30 ± 0.18 OD (95% CI 0.27-0.33, n=133) in drug-responsive patients. Drug-resistant patients had much higher levels of GluR3B antibody than drug-responsive patients (P=0.0049, Figure 1A ). Twenty patients had CSF samples. CSF GluR3B antibody levels were 0.23 ± 0.11 OD (95% CI 0.11-0.34, n=6) in drug-resistant patients and 0.12 ± 0.06 OD (95% CI 0.09-0.16, n=14) in drug-responsive patients. Similarly, CSF GluR3B antibody levels significantly increased in drug-resistant patients compared with drug-responsive patients (P<0.05, Figure 1B ). In addition, we compared GluR3B antibody levels in CSF and serum samples from 20 epilepsy patients, and the results indicated that CSF GluR3B antibody levels (0.16 ± 0.10 OD) were lower than serum GluR3B antibody levels (0.27 ± 0.13 OD, P=0.0024, Figure 1C ). In the modeling cohort, the proportion of GluR3B antibody-positive epilepsy patients with duration ≤2 years was 29.2% (21/72); the proportion of GluR3B antibody-positive epilepsy patients with duration >2 years was 40.5% (49/121). The level of GluR3B antibody in patients with duration ≤ 2 years and more than 2 years were 0.32 ± 0.21 vs 0.34 ± 0.20 (P=0.52). In order to understand the correlation between GluR3B antibody level and DRE in patients with the same duration of epilepsy and frequency of seizures prior to enrolment, we selected 56 patients with the same duration of epilepsy and frequency of seizures prior to enrolment, matched age and sex, including 28 patients with drug-resistant epilepsy and 28 patients with drug-response epilepsy. Serum GluR3B antibody levels were 0.42 ± 0.22 OD (95% CI 0.33-0.50, n=28) in drug-resistant patients and 0.21 ± 0.09 OD (95% CI 0.18-0.25, n=28) in drug-responsive patients. Drug-resistant patients had much higher levels of GluR3B antibody than drug-responsive patients (P=0.0003). Finally, the analysis of GluR3B antibody levels with epilepsy duration, ages at first seizure, frequency of seizures before treatment, as well as with initial therapy response didn’t highlight any significant association (data not shown).

The clinical value of the GluR3B antibody level in predicting DRE was evaluated by ROC curve and logistic regression analyses. As shown in the ROC curve analysis ( Figure 2A ), the increased GluR3B antibody level showed moderate clinical value to discriminate DRE from drug-responsive epilepsy (area under the curve [AUC] 0.626, 95% CI 0.533-0.719; P=0.0051). Logistic regression analysis was used to evaluate the risk of patients with high GluR3B antibody levels developing DRE. First, we constructed the full model (including the serum GluR3B antibody level, age, sex, epilepsy duration, age at first seizure, frequency of seizures before treatment, and initial therapy response) in the training data and evaluated the performance of the full model by calculating the AUC of the ROC curve. Our data showed that the mean GluR3B antibody level with the maximum AUC in the training data was 0.325 ( Figure 2B ). Meanwhile, the GluR3B antibody level of 0.325 corresponded to the larger AUC of the full model in the validation cohort ( Figure 2C ). Therefore, 0.325 was used as an optimal cutoff value for the serum GluR3B antibody level. Afterwards, we transformed the GluR3B antibody level from a continuous variable into a binary variable (high antibody level and low antibody level) according to the cutoff value. As shown in Figure 3A by univariate analysis, patients with high GluR3B antibody levels (odds ratio [OR] 3.524, 95% CI 1.862-6.669; P<0.0001) had a significantly increased risk of developing DRE compared with patients with low GluR3B antibody levels. As expected, the univariate analysis confirmed the increased risk of developing DRE in patients with poor response to initial therapy (OR 14.907, 95% CI 7.093-31.331; P<0.0001), high frequency of seizures before therapy (OR 6.696, 95% CI 2.714-16.524; P<0.0001), ages at first seizure ≤12 year (OR 4.278, 95% CI 1.942-9.421; P<0.0001), and an epilepsy duration ≥5 years (OR 2.867, 95% CI 1.525-5.392; P=0.001) ( Figure 3A ).

To further evaluate the independent clinical significance of the GluR3B antibody level in predicting DRE, multivariate logistic regression analysis was performed ( Figure 3B ). Logistic regression models were adjusted for binary serum GluR3B antibody level along with patients’ initial therapy response, frequency of seizures before treatment, epilepsy duration, age at first seizure, sex and patient age. Indeed, logistic regression models highlighted the independent clinical significance of increased GluR3B antibody levels in predicting DRE (OR 5.390, 95% CI 2.203-13.189; P<0.0001) ( Figure 3B ).

External validation was performed to evaluate the calibration and discrimination capabilities of this model by the Hosmer and Lemeshow test and ROC curve analysis. We successfully constructed a predictive model for DRE: y=1/(1+e^-z). where y represents the probability of predicting the occurrence of DRE, e represents the exponential function and Z=-12.733+1.725*X₁ +2.865*X₂ +1.685*X₃. X₁, X₂ and X₃ represent frequency of seizures before treatment, initial therapy response, and the level of GluR3B antibody, respectively. The Hosmer and Lemeshow test was used to evaluate the calibration capability of the model in 49 patients from January 2018 to October 2019. The results showed that the P value was 0.868 ( Figure 4B ), indicating that the fitting effect was good. The AUC was used to evaluate the discriminative validity of the model, and the results showed that the AUC was 0.925 (95% CI 0.854-0.996, P<0.0001) ( Figures 4A, B ), suggesting that the predictive model of DRE constructed in this study had a good discriminative ability. Therefore, the model constructed for the prediction of DRE had strong external adaptability and could be a good tool to predict the development of DRE in clinical practice.

The clinical significance of increased GluR3B antibody levels in predicting DRE prompted us to further evaluate their clinical benefit through decision curve analysis. As expected, the univariate analysis showed that the evaluation of GluR3B antibody levels had superior clinical net benefit in epilepsy prognosis ( Figure 5A ). Moreover, the level of GluR3B antibody had superior clinical value in the risk stratification of patients with DRE ( Figure 5A ). Finally, the multivariate logistic regression model incorporating increased GluR3B antibody levels also demonstrated superior clinical net benefit compared to the clinical model in predicting DRE and patient treatment management for threshold probabilities of 70%-80% ( Figure 5B ).