Association of Epileptic and Nonepileptic Seizures and Changes in Circulating Plasma Proteins Linked to Neuroinflammation

Abstract

Objective

To develop a diagnostic test that stratifies epileptic seizures (ES) from psychogenic nonepileptic seizures (PNES) by developing a multimodal algorithm that integrates plasma concentrations of selected immune response–associated proteins and patient clinical risk factors for seizure.

Methods

Daily blood samples were collected from patients evaluated in the epilepsy monitoring unit within 24 hours after EEG confirmed ES or PNES and plasma was isolated. Levels of 51 candidate plasma proteins were quantified using an automated, multiplexed, sandwich ELISA and then integrated and analyzed using our diagnostic algorithm.

Results

A 51-protein multiplexed ELISA panel was used to determine the plasma concentrations of patients with ES, patients with PNES, and healthy controls. A combination of protein concentrations, tumor necrosis factor–related apoptosis-inducing ligand (TRAIL), intercellular adhesion molecule 1 (ICAM-1), monocyte chemoattractant protein-2 (MCP-2), and tumor necrosis factor–receptor 1 (TNF-R1) indicated a probability that a patient recently experienced a seizure, with TRAIL and ICAM-1 levels higher in PNES than ES and MCP-2 and TNF-R1 levels higher in ES than PNES. The diagnostic algorithm yielded an area under the receiver operating characteristic curve (AUC) of 0.94 ± 0.07, sensitivity of 82.6% (95% confidence interval [CI] 62.9–93.0), and specificity of 91.6% (95% CI 74.2–97.7). Expanding the diagnostic algorithm to include previously identified PNES risk factors enhanced diagnostic performance, with AUC of 0.97 ± 0.05, sensitivity of 91.3% (95% CI 73.2–97.6), and specificity of 95.8% (95% CI 79.8–99.3).

Conclusions

These 4 plasma proteins could provide a rapid, cost-effective, and accurate blood-based diagnostic test to confirm recent ES or PNES.

Classification of Evidence

This study provides Class III evidence that variable levels of 4 plasma proteins, when analyzed by a diagnostic algorithm, can distinguish PNES from ES with sensitivity of 82.6% and specificity of 91.6%.

Accurate diagnosis of epilepsy is challenging because clinicians rarely observe the actual clinical seizure outside of the hospital. Furthermore, psychogenic nonepileptic seizures (PNES) can mimic epileptic seizures (ES), leading to erroneous diagnosis and inappropriate treatments. A critical gap in the diagnostic assessment of seizures is a blood test that can distinguish ES from PNES. Elevated serum prolactin levels have been proposed as a biomarker to distinguish ES from PNES,^1–3 however, a 15-study meta-analysis assessing serum prolactin and other markers, for example, ACTH, cortisol, neuron-specific enolase, creatinine kinase, ghrelin, and brain-derived neurotrophic factor, demonstrated that although these serum protein levels may change following ES, the lack of diagnostic sensitivity, specificity, and accuracy precluded use in clinical practice.⁴ In contrast, alternate approaches have been explored that use patient characteristics and clinical histories to create an algorithm to distinguish PNES from ES.^5–7

Growing evidence demonstrates that neuroinflammation is highly associated with ES.⁸ ES can activate both systemic and brain proinflammatory pathways, including enhanced interleukin (IL)–1β production, activation of Toll-like receptor 4, mammalian target of rapamycin, and mitogen-activated protein kinase cascades, attraction of activated lymphocytes, activation of microglia and macrophages, and alteration of astrocyte physiology.^9–11 A recent meta-analysis of 149 research articles supports the use of blood-based inflammatory marker measurements for ES diagnosis.¹² Furthermore, blocking inflammation reduces or prevents seizures in preclinical models, and new treatments for epilepsy are being developed to target inflammatory pathways.^13–17 We hypothesized that altered levels of select plasma markers of neuroinflammation would help distinguish ES from PNES.

Methods

Standard Protocol Approvals, Registrations, and Patient Consents

The University of Pennsylvania (Philadelphia) institutional review board approved this study. Informed consent was obtained for all participants prior to study initiation. The study was registered with clinicaltrials.gov (identifier NCT02269397).

Study Participants

A total of 137 patients were evaluated in the epilepsy monitoring unit (EMU) for clinically indicated differential diagnosis or presurgical evaluation and 29 healthy volunteers were prospectively enrolled in this all-comers study between 2014 and 2015. The healthy control (HC) samples were collected from friends and family members of admitted EMU patients.

Eligibility requirements for EMU patients were aged 12 and over, able to provide written informed consent or accompanied by a legally authorized representative (LAR) who could provide informed consent on their behalf, and agreement from the investigator or the patient's neurologist that it was medically appropriate to invite the patient to participate in experimental blood draws. For the HC samples, eligibility requirements were aged 18 and over, able to provide informed consent on their own behalf without an LAR, a self-reported history of no lifetime seizures or suspected seizures (a single febrile seizure before age 2 was allowed) and no treatment with an antiepileptic drug (AED) within the 2 months prior to blood draw.

Blood Sampling and Analysis

Each patient contributed 4 mL of blood every morning during the duration of the EMU admission and an additional tube of blood after experiencing a clinical event that was determined to be ES or PNES; patients who refused an individual blood draw could remain in the study (a “missed” blood draw did not constitute a protocol violation). HC were asked to contribute a maximum of two 4-mL tubes of blood when they signed consent. Blood was collected into lavender-topped vacutainer blood collection tubes containing K₂EDTA as an anticoagulant (BD Biosciences). The blood collection tubes were placed on ice at 4°C for 10–15 minutes before centrifuging at 1,000 RCF for 10 minutes at 4°C. Plasma supernatant was aliquoted into sterile 2-mL microtubes (Sarstedt, type I) and frozen at −70°C to −80°C. All sample transport was executed on dry ice.

An extensive literature review was used to identify 51 inflammation response related proteins that were previously linked to seizures, proinflammatory pathways, cell adhesion, and the systemic immune response, including C-reactive protein (CRP), calbindin, cytokeratin-8, eotaxin, eotaxin-2, eotaxin-3, granulocyte-macrophage colony-stimulating factor, intercellular adhesion molecule 1 (ICAM-1), interferon (IFN)–γ, IL-1β, IL-1α, IL-2, IL-4, IL-5, IL-6, IL-7, IL-8, IL-10, IL-12/IL-23 p40, IL-12 p70, IL-13, IL-15, IL-16, IL-17A, IFN-γ-inducible protein 10, macrophage colony-stimulating factor (M-CSF), monocyte chemoattractant protein (MCP)–1, MCP-2, MCP-4, macrophage-derived chemokine, macrophage migration inhibitory factor, macrophage inflammatory protein (MIP)–1β, MIP-1α, MIP-5, matrix metalloproteinase (MMP)–1, MMP-3, MMP-9, Nectin-4, Osteoactivin, osteonectin, P-cadherin, serum amyloid protein A, stem cell factor (SCF), thymus and activation regulated chemokine, tumor necrosis factor (TNF)–α, TNF-β, TNF–receptor 1 (R1), TNF–receptor 2 (R2), TNF–related apoptosis-inducing ligand (TRAIL), vascular cell adhesion molecule 1, and vascular endothelial growth factor A. Levels of these 51 plasma proteins were assayed by sandwich ELISA with electrochemiluminescent detection using 6 multiplex plates from Meso Scale Discovery (MSD; Gaithersburg, MD) including V-PLEX Proinflammatory Panel 1 (catalog K15049D), V-PLEX Cytokine Panel 1 (catalog K15050D), V-PLEX Chemokine Panel 1 (catalog K15047G), and V-PLEX Vascular Injury Panel 2 (catalog K15198D) and 2 custom panels comprising a combination of U-PLEX and R-PLEX assays configured into 2 plates, containing MMP-3, calbindin, eotaxin-2, MIP-5, MMP-1, Osteoactivin, P-cadherin, TNF-R1 and TNF-R2 or MCP-2, M-CSF, Nectin-4, osteonectin, SCF, and TRAIL.

Panels were processed using a common assay protocol (MSD). Subsequent to an initial dilution step, plasma samples were added to assay plates precoated with capture antibodies for the target proteins. Sealed plates were shaken at room temperature for 2 hours. Plates were washed with MSD wash buffer (catalog R61AA), detection antibody solution was added, and plates were shaken at room temperature for 2 hours. Finally, MSD electrochemiluminescent read buffer (catalog R92TC) was added and incubated for 10 minutes. Plates were imaged using the MSD Sector S600 to determine electrochemiluminescent signals, then converted to concentrations using an intra-assay standard curve. All assays were run in duplicate with internal standard to control for interassay variability with an interassay coefficient of variation below 20%. As defined below, only qualifying samples (i.e., collected within 24 hours of a qualifying event) were processed using the above method.

Clinical Data Collection and Analysis

To define the EMU event diagnosis as ES or PNES, continuous video-EEG recordings were reviewed as part of clinical assessment. Clinical events were reviewed in concert with video-EEG to determine if the events were ES, defined by ictal changes on the EEG, or PNES, in which a clinical event was observed but the EEG remained normal. All patients with ES had EEG-confirmed seizures with distinct regional onset patterns, frontal or temporal lobe, and all were treated with at least 1 antiepileptic drug. All patients with ES exhibited interictal spikes or sharp waves on the EEG. Average seizure duration was 2.6 minutes and consisted of focal impaired awareness or generalized tonic-clonic seizure types.

At the completion of the hospitalization, EMU reports were deidentified and data extracted into a custom database that included age, sex, race, age at event onset, years with events, etiology (although very often unknown), seizure frequency, a complete list of EMU event diagnoses, a final epilepsy diagnosis, comorbidities, and current medications. A review of the patient records retrospectively identified which blood draws occurred within 24 hours of an EEG-diagnosed ES or nonepileptic seizure (including PNES) event.

Statistical Analysis

The statistical analysis was designed to explore differences in patient characteristics and plasma protein levels of 51 proteins and model the likelihood of seizures between ES and PNES within 24 hours of a clinical event. Of the 51 proteins screened, only 32 could be reliably quantified within validated ranges, as defined by the standard curve, and were carried forward for additional analysis. Proteins were excluded if more than 5 samples had concentrations less than the lower limit of quantitation. All analysis was performed using custom processing scripts created in-house using the following modules: Python 2.7.11, Pandas (version 0.18.0-7), Numpy (version 1.11.3-3), Statsmodels (version 0.6.1-16), and Scikit Learn (version 0.19.1-2). Prism 7.00 for Windows (GraphPad Software, La Jolla, CA) was used to visualize the results and create figures. Comprehensive statistical analyses and methodologies are presented in the supplemental data^18–21 (additional Methods, Data available from Dryad, doi.org/10.5061/dryad.xgxd254dz).

PNES Risk Factors

Clinical histories may aid in the identification of patients with PNES.^5–7 Comorbidities and patient-reported historical factors may help differentiate ES and PNES. Combining literature survey results with the information available from the EMU reports in our study, we assessed the following PNES risk factors: major depressive disorder, posttraumatic stress disorder, cluster B personality disorder, polyallergies, conversion disorder, sex, fibromyalgia, migraine, pain, and asthma. Risk factors for each patient were summed to generate an aggregate PNES risk score. Using the sum of risk factors reduced the number of variables that were included into the logistic regression to limit the possible overfitting of the data. The aggregate risk factor score was incorporated into a diagnostic algorithm with the protein combination identified above. The algorithm was refined and evaluated using the same metrics as above. For comparison, an algorithm using only the aggregate PNES risk factor was refined using a logistic regression.

Classification of Evidence

Our primary research question was whether ES can cause detectable changes in plasma proteins linked to inflammation. This study provides Class III evidence that when analyzed using a specific diagnostic algorithm, levels of 4 plasma proteins can distinguish ES from PNES.

Data Availability

The datasets generated or analyzed during the current study are available from the corresponding author on reasonable request.

Results

Patient Cohorts

All patient samples were retrospectively analyzed for inclusion into patient cohorts that experienced an event within 24 hours prior to a blood draw (figure 1). A total of 137 patients were enrolled into the study and upon analysis of EMU reports 48 were excluded because no definitive diagnosis was reported. The remaining patients (n = 89) had a final diagnosis of either epilepsy (n = 55) or PNES (n = 34). Patients exhibiting both ES and PNES were categorized in the epilepsy group. An additional group of 42 patients were excluded because they did not have a qualifying blood draw within 24 hours after an EEG diagnosed event. The final cohorts (table 1) consisted of 23 patients with epilepsy (65.2% female with mean age of 44.23 years) and 24 patients with PNES events (79.2% female with a mean age of 44.16 years); both cohorts represent multiple races and ethnicities (e.g., White, Black).

Figure 1. Patient Inclusion Flowchart.

All-comers enrollment was 137 and through a retrospective analysis resulted in cohort sizes of 23 patients with epileptic seizures (ES) and 24 patients with psychogenic nonepileptic seizures (PNES). Patients were excluded if they did not experience a diagnostic event in the epilepsy monitoring unit (EMU) and if they did not contribute blood within 24 hours after a qualifying event (e.g., ES or PNES event).

Table 1.

Demographic Information on Patient and Control Cohorts

For patients donating more than 1 qualifying blood sample, each was processed independently and the resulting concentration values were averaged. Note that the time elapsed between event and blood draw averaged 9.0 and 11.6 hours with an SD of 5.7 and 7.4 for patients with epilepsy and patients with PNES, respectively. All samples were collected within 24 hours of an event. To assess longitudinal variation of protein concentration, the Pearson correlation coefficient was calculated between elapsed time and protein concentration. No correlations greater than 0.38 were observed.

A total of 85 samples were analyzed, comprising 31 samples from patients with ES, 25 samples from patients with PNES events, and 29 samples from HCs (table 1). There were no significant differences observed for age, sex, or race between the cohorts (p > 0.05). The number of AEDs that the ES and PNES cohorts were prescribed upon enrollment in the trial differed; 70% of the ES cohort were prescribed 2 or more AEDs and 79% of the PNES cohort were prescribed 1 or fewer. This difference likely arises because 83% of patients with ES in the EMU had drug-resistant epilepsy and had already progressed through AED monotherapy. All AEDs were sorted by drug class and counted to explore potential differences between the cohorts (table 2). In addition, the prevalence of nonsteroidal anti-inflammatory drugs (NSAIDs) was considered: 6 patients (26%) in the ES cohort and 7 patients (29%) in the PNES cohort were currently taking either over-the-counter or prescription NSAIDs. No patient was taking other immunosuppressant drugs. The ES cohort patients averaged 1.7 seizures per patient, with 8 of the 24 patients having a combination of focal impaired awareness and focal to bilateral tonic-clonic seizures. The counts for each seizure type were as follows: focal impaired awareness 19, focal to bilateral tonic-clonic 10, focal aware 5, generalized motor seizure 3, and focal with unknown awareness 1. All patients enrolled in the ES cohort and 79% of patients in the PNES had a motor manifestation component to their characteristic seizure. Patients with events such as syncope or tremor were not included in this study.

Table 2.

Prevalence of Antiepileptic Drug Prescription in Psychogenic Nonepileptic Seizures (PNES) and Epileptic Seizures (ES) Cohorts

Changes in Plasma Protein Concentrations After ES

To assess the changes in protein concentration as a result of ES compared to PNES, we analyzed the concentration of 51 proteins in plasma samples collected from patients experiencing an event in the prior 24 hours. Of the 51 proteins tested, 32 were analyzed and 19 proteins had limited dynamic range, where many samples fell below the lower limit of quantitation and were therefore excluded from analysis (figure 2). Of the 32 quantified proteins analyzed, 5 proteins had effect size magnitudes >0.6 with associated 95% confidence intervals (CIs) that did not cross zero. These proteins include MCP-2, TRAIL, ICAM-1, MMP-1, and TNF-R1. The effect size was also calculated between the 2 cohorts and the controls. TRAIL, ICAM-1, and MCP-2 all had effect size magnitudes greater than 0.5 and less than 0.35 for ES–control and PNES–control, respectively, indicating that the controls and PNES trend together. Note that TRAIL and ICAM-1 had effect sizes of −1.0 and −0.8, indicating decreased concentrations in the ES cohort compared to PNES. TNF-R1 had an effect <0.1 for the ES–control comparison and 0.4 for the PNES–control comparison. MMP-1 effect sizes were less than 0.1 and greater than 0.7 for ES–control and PNES–control comparisons, respectively, indicating that control sample concentrations trend with the epilepsy cohort.

Figure 2. Comparison of Protein Concentrations for the Epileptic Seizures (ES) and Psychogenic Nonepileptic Seizures (PNES) Cohorts.

(A) Cohen d effect size was used to explore differences in mean concentrations between the ES and PNES cohorts. The bar plot shows the effect size and 95% confidence interval of each protein sorted in descending order. (B) Select proteins demonstrate the magnitude of changes observed between cohorts. This figure also shows how noncanonical responses are observed for inflammation-related proteins, such as tumor necrosis factor–related apoptosis-inducing ligand (TRAIL) and intercellular adhesion molecule 1 (ICAM-1), which can be used to identify patients experiencing a seizure. CRP = C-reactive protein; IL = interleukin; IP-10 = IFN-γ-inducible protein 10; MCP = monocyte chemoattractant protein; MIP = macrophage inflammatory protein; MMP = matrix metalloproteinase; SAA = serum amyloid protein A; SCF = stem cell factor; TARC = thymus and activation‐regulated chemokine; TNF-R1 = tumor necrosis factor–receptor 1; TNF-R2 = tumor necrosis factor–receptor 2; VCAM = vascular cell adhesion molecule; VEGF = vascular endothelial growth factor.

The above effect sizes demonstrate a difference between ES and PNES cohorts, and to further understand each protein's contribution to predicting a change between ES and PNES, a random forest (RF) was generated and percent contributions of each protein were calculated. The RF was generated using 1,000 trials of random sampling with replacement and the relative importance of each protein was calculated by the mean decrease in impurity. Two generated data sets, constructed from random noise, were included to aid in differentiating protein with potential classification ability from spurious results. The noise values had an average relative importance of 2%. To provide a margin from the noise floor, values with a relative importance greater than 3% were considered and include 11 proteins: MCP-2 (8.5%), TRAIL (8.1%), ICAM-1 (5.4%), TNF-R1 (5.0%), Osteoactivin (4.7%), M-CSF (4.1%), MMP-1 (3.9%), P-cadherin (3.6%), TNF-α (3.5%), CRP (3.1%), and Eotaxin-2 (3.0%), where the values in parentheses are the relative importance.

Diagnostic Algorithm to Retrospectively Diagnose ES

The overall objective of our analysis was to create a diagnostic algorithm that would provide a concise retrospective diagnosis of ES. To create a diagnostic algorithm, multiple protein concentrations were combined using logistic regression. Potential proteins to include into the diagnostic algorithm were selected from the above effect size and RF analyses. Using the strict criteria that the effect size CI cannot cross zero and a relative importance greater than 3%, 5 proteins were selected: MCP-2, TRAIL, ICAM-1, MMP-1, and TNF-R1. To determine which proteins to include into the final algorithm, all combinations of the above 5 proteins were generated and refined (31 combinations). Algorithm results were evaluated using Akaike information criterion (AIC; see Supplemental Methods, doi.org/10.5061/dryad.xgxd254dz) to select the algorithm with the optimal information content relative to the number of proteins included, with the objective of selecting an algorithm with the fewest components while retaining the maximum amount of information. An AIC comparison threshold of 2 was used as defined by Burnham and Anderson.²¹ The combination of all 5 proteins had the minimum AIC value, but the combination of MCP-2, TRAIL, ICAM-1, and TNF-R1 had an AIC difference of <1 and therefore the 5-protein algorithm was rejected. This combination compared to the combination of MCP-2, TRAIL, ICAM-1, and MMP-1 had very similar AIC (difference of 0.4) but because TNF-R1 had a higher relative importance in the RF analysis the prior combination was selected. Algorithms with fewer proteins were also considered but had AIC differences >4 and were rejected. The results of the selected combination of MCP-2, TRAIL, ICAM-1, and TNF-R1 are shown in figure 3. The coefficients for the algorithm were −0.0153, −0.0108, 0.1456, and 2.3963 for the concentrations of TRAIL, ICAM-1, MCP-2, and TNF-R1, respectively, and an intercept of 0.6194 was used. Each regression coefficient had a p value < 0.05. The algorithm performance was robust (figure 3). The median score for the ES cohort was 90.7 (95% CI 66.7–97.5) and the median score for the PNES cohort was 6.8 (95% CI 2.7–23.5). Q1 through Q3 quartiles of the 2 score distributions did not overlap, indicating stratification of the cohorts. A threshold of 64 was selected to simultaneously maximize the sensitivity and specificity (figure 3B). At this threshold, the overall accuracy was 87.2%, sensitivity was 82.6% (95% CI 62.9–93.0), and specificity was 91.7% (95% CI 74.2–97.7). The receiver operating characteristic (ROC) curve was calculated and resulted in an area under the ROC curve (AUC) of 0.94 ± 0.07 (figure 3C). The leave-one-out cross-validation had an average accuracy of 80.9%. In the case that the control cohort was grouped with the PNES cohort, the algorithm had an accuracy of 80.3%.

Figure 3. Epileptic Seizures (ES)/Psychogenic Nonepileptic Seizures (PNES) Diagnostic Algorithm Performance Refined Using Selected Protein Concentrations.

Performance of a diagnostic algorithm that incorporates tumor necrosis factor–related apoptosis-inducing ligand, intercellular adhesion molecule 1, monocyte chemoattractant protein 2, and tumor necrosis factor–receptor 1 protein concentrations. The algorithm was refined using a logistic regression. (A) The score distribution for ES and PNES cohorts shown as a bar and whisker plot using the Tukey method. (B) The performance matrix and (C) the receiver operating characteristic (ROC) curve demonstrate a good tradeoff between sensitivity and specificity with an area under the ROC of 0.94. These results demonstrate the algorithms’ ability to stratify patients with ES from those with PNES. NPV = negative predictive value; PPV = positive predictive value.

To understand the effect of confounding measures on algorithm performance, an algorithm was refined incorporating each of the following measures: sex, age, race, seizure type, seizure motor component, NSAID use, and AED prescription count. The AIC difference between the algorithm with and without the confounding measure and coefficient p values were calculated for each algorithm. All coefficient p values indicate that there is not a significant relationship for any of the confounding measures to stratify ES from PNES. Consistently, none of the algorithms with confounding measures had an improved AIC (e.g., all differences greater than 2) and were therefore excluded from the final algorithm. Seizure frequency was investigated by comparison with protein concentrations using the Pearson correlation coefficient for patients with ES and patients with PNES separately and no correlations were observed (r < 0.4). The mean event frequency for the ES cohort was 7.4 with a SD of 9.7 and a mean of 8.9 with a SD of 15.6 for the PNES cohort.

Because some AEDs may influence the level of inflammation-associated cytokines, a detailed analysis of confounding effects of AEDs on the algorithm performance was performed. All AEDs that patients were taking at time of EMU admission were recorded and classified according to chemical composition. The prevalence of each AED class for the ES and PNES cohorts is shown in table 2. Each of the drug types was then combined with the protein concentration data and a logistic regression algorithm was refined (supplemental table, doi.org/10.5061/dryad.xgxd254dz). None of the algorithms had coefficients with significant p value for the AED families. The difference in AIC between the algorithm with only protein concentrations was compared for each of the algorithms with AED families and no reductions greater than 2 were observed. Interestingly, the carboxamides family (carbamazepine, oxcarbazepine, and eslicarbazepine) had an AIC difference of 1.7. Although these results are likely a result of 30% of patients with ES taking one of these drugs compared to only 4% of patients with PNES, Cohen d effect size was calculated to compare patients with ES on drug vs not on drug. The effect sizes were −0.4, 0.1, 0.3, and 0.0 for TRAIL, ICAM-1, MCP-2, and TNF-R1, respectively. Multiple classification techniques are available to stratify cohorts; to compare the efficiency of these techniques, the logistic regression (LR) model was compared to K nearest neighbor (KNN), naive Bayes (NB), decision tree (DT), and RF models. A 10-fold cross-validation and leave-one-out cross-validation was used to compare the models. Average AUCs of the 10-fold cross validation are as follows: LR 0.81, KNN 0.64, NB 0.93, DT 0.63, and RF 0.79. Mean accuracies are LR 81%, KNN 66%, NB 81%, DT 68%, and RF 81%.

Influence of PNES Risk Factors on Diagnostic Algorithm Performance

We next assessed whether identified clinical risk factors for PNES^5–7 were synergistic with protein biomarkers in stratifying our ES and PNES cohorts. Based on previously published data identifying clinical features that correlate with a PNES diagnosis, a list of 11 measures was used to calculate a total risk score for each patient (table 3). The risk score was refined against the patient diagnosis (ES or PNES) using a logistic regression both alone and in combination with the protein biomarkers to assess the diagnostic efficiency (figure 4).

Table 3.

Psychogenic Nonepileptic Seizures (PNES) Risk Factor Prevalence, %

Figure 4. Epileptic Seizures/Psychogenic Nonepileptic Seizures (PNES) Diagnostic Algorithm That Integrates Both Clinical Risk Factors and Protein Concentrations.

(A, B) Performance characteristics of a diagnostic algorithm that was refined using both the plasma concentration of tumor necrosis factor–related apoptosis-inducing ligand, intercellular adhesion molecule 1, monocyte chemoattractant protein 2, and tumor necrosis factor–receptor 1 and the sum of PNES risk factors. Risk factors include major depressive disorder, posttraumatic stress disorder, cluster B personality disorder, multiple allergies, conversion disorder, sex, fibromyalgia, migraine, pain, and asthma. (C, D) Performance of a diagnostic algorithm that was refined using only the PNES risk factors listed above. AUC = area under the receiver operating characteristic curve; NPV = negative predictive value; PPV = positive predictive value.

The average risk score for the ES cohort was 1.16, while that for the PNES for was 2.88. Given the substantial difference in the mean number of risk factors observed between cohorts (effect size 1.48) and to reduce the degrees of freedom when an algorithm was refined, the sum of all risk factors present in each patient was used in algorithm development. The sum of the risk factors for each patient was refined against their diagnosis using logistic regression to generate an algorithm to stratify ES and PNES. The final algorithm had a coefficient for the risk factors of −1.4177 and an intercept of 2.98 (p < 0.005; figure 3). The overall accuracy was 82.1% when a threshold of 82 was selected. Accordingly, the sensitivity was 74.2% (95% CI 56.8–86.3), the specificity was 92.0% (95% CI 75.0–97.7), and the AUC was 0.877 ± 0.086.

A second logistic regression algorithm was generated that incorporated both PNES risk factors and the protein biomarkers identified above to predict patient diagnosis. The algorithm performance (figure 3) had an overall accuracy of 94% when a threshold of 41 was selected. The algorithm produced a sensitivity of 91.3% (95% CI 73.2–97.6), specificity of 95.8% (95% CI 79.8–99.3), and AUC of 0.97 ± 0.046.

Discussion

We demonstrate that plasma levels of 4 proteins (TRAIL, ICAM-1, MCP-2, and TNF-R1) linked to neuroinflammation can distinguish patients with EEG-confirmed ES from PNES when assayed within 24 hours after the clinical event. A logistic regression algorithm that included known clinical risk factors for PNES augmented the accuracy, sensitivity, and specificity of the algorithm. These results suggest that clinical assay of TRAIL, ICAM-1, MCP-2, and TNF-R1 might provide diagnostic information to aid in distinguishing ES from PNES in select individuals. Our cohort was enrolled from the EMU because this provided EEG (gold standard) diagnosis for clinical events. All EMU reports were reviewed by 3 epileptologists to aid in control for inter-rater reliability. Thus, we could definitively identify ES vs PNES. The study design resulted in a representative population of patients typically seen in the EMU. Both focal and generalized seizures were represented, and seizure frequency ranged from fewer than 1 per year to multiple seizures per day.

The premise of our targeted proteomic screen was the established link between neuroinflammation and seizures. We postulated that ES, but not PNES, would cause an inflammatory response that was detectable in plasma. We identified 5 proteins that distinguished the 2 cohorts and have effect sizes greater than 0.6. The proteins were further evaluated using an RF to confirm that they had a relative importance that differentiated them from the noise floor. Combinations of 1 or more of these 5 proteins were considered to refine a diagnostic algorithm. The algorithm with the optimal information content and performance included TRAIL, ICAM-1, MCP-2, and TNF-R1. Confounding variables, such as age, race, seizure type, and number of AEDs prescribed, were considered and none provided strong support for inclusion into the algorithm, or influenced the overall diagnostic accuracy of the algorithm. In addition, various model-generating techniques were compared to the logistic regression and although NB and RF had a similar performance the discrete probability score that a patient experienced a seizure that is generated by the logistic regression is preferable for this application.

We acknowledge that the relationship between levels of these proteins and epilepsy is not clear and the response we are observing may not exemplify a typical pattern of inflammation given the relative decrease in concentration of TRAIL and ICAM-1 when comparing ES to PNES. TRAIL is linked to apoptosis and cell death via binding to a death receptor (DcR2) and activating a caspase-8 dependent pathway.²² Interestingly, TRAIL has also been shown to exert anti-inflammatory effects.²³ ICAM-1 is a cell surface glycoprotein expressed on a wide array of cell types that functions as an intercellular adhesion molecule that aids in eliciting an immune response.²⁴ Upon degradation to a soluble form, ICAM-1 may function in signal transduction that potentiates proinflammatory pathways. Differences in ICAM-1 have previously been observed in CSF of patients with epilepsy,²⁵ as well as tubers from patients with tuberous sclerosis complex.²⁶ Whereas valproic acid treatment is correlated with a decreased expression of ICAM-1,²⁷ we did not observe any differences in the percent of patients prescribed valproic acid between 2 cohorts studied. MCP-2 (CCL-8) plays a role in immunoregulatory and inflammatory processes binding to multiple chemokine receptors.²⁸ MCP-2 has been implicated to play a role in atopic dermatitis, where, like epilepsy, acute, periodic inflammation correlates with clinical exacerbations.²⁹ Also, consistent with our results, it has been demonstrated that the expression of MCP-2 and TRAIL are inversely correlated.³⁰ Finally, TNF-R1 is a membrane-bound cell surface receptor that binds TNFα and leads to NF-κβ–mediated inflammatory response.³¹ The membrane-bound isoform of TNF-R1 can be cleaved to produce a soluble isoform that has anti-inflammatory properties.³² Previous reports demonstrate that TNF-R1 signaling is activated in mouse seizure models.³³

The performance of the protein concentration–based diagnostic algorithm was improved by including clinical risk factors; that is, 87% accuracy when only protein levels were evaluated compared to 94% when both protein levels and clinical risk factors were considered. For comparison, an algorithm evaluating only clinical risk factors was also generated and had an accuracy of 81%. The clinical risk factor–only algorithm accuracy was consistent with previously published results.^5,6 Whereas a risk factor–only algorithm provides some clinical value, when combined with protein information, the algorithm had superior power to correctly classify patients with simultaneous ES and PNES.

Two false-negatives were observed using our diagnostic algorithm, which could have arisen from confounding variables that were not evaluated as part of this study or not sufficiently powered. Also, given the study design, it is possible that we are observing a chronic immune response in patients having a baseline seizure frequency, that is, epilepsy, and future studies to assess baseline protein levels would be informative. It is also possible that some false-negative seizures were detected by the algorithm but not the gold standard ictal EEG. To fully understand the results presented here, a sufficiently powered study with a reserved validation data set will be required to understand all aspects, including the longitudinal, demographic, and pharmacologic aspects of epilepsy. The results will also need to be tested against other paroxysmal events such as syncope, migraine, and hypoglycemia to further understand the clinical utility of these results.

A targeted proteomic screen identified 4 proteins that were modulated as a result of epileptic seizures within 24 hours of a clinical event, thus providing a clear biomarker signature that might be used to aid clinical diagnosis of ES vs PNES. The performance of the biomarker approach is improved when augmented with clinical and demographic information.

Glossary

AED

antiepileptic drug

AIC

Akaike information criterion

AUC

area under the receiver operating characteristic curve

CI

confidence interval

CRP

C-reactive protein

DT

decision tree

EMU

epilepsy monitoring unit

ES

epileptic seizures

HC

healthy control

ICAM-1

intercellular adhesion molecule 1

IFN

interferon

IL

interleukin

KNN

K nearest neighbor

LAR

legally authorized representative

LR

logistic regression

M-CSF

macrophage colony-stimulating factor

MCP

monocyte chemoattractant protein

MIP

macrophage inflammatory protein

MMP

matrix metalloproteinase

MSD

Meso Scale Discovery

NB

naive Bayes

NSAID

nonsteroidal anti-inflammatory drug

PNES

psychogenic nonepileptic seizures

RF

random forest

ROC

receiver operating characteristic

SCF

stem cell factor

TNF-R1

tumor necrosis factor–receptor 1

TNF-R2

tumor necrosis factor–receptor 2

TRAIL

tumor necrosis factor–related apoptosis-inducing ligand

Appendix. Authors

Study Funding

NIH 1R43NS079029-01A1 Evogen, Inc.

Disclosure

J. Gledhill: employee of Cognizance Biomarkers. E. Brand: employee of Cognizance Biomarkers. J. Pollard: consultant to Cognizance Biomarkers. R. St. Clair: employee of Cognizance Biomarkers. T. Wallach: employee of Cognizance Biomarkers. P. Crino: grant support from NINDS R01NS094596, R01NS099452, scientific advisory board of Cognizance Biomarkers. Go to Neurology.org/Nhttps://n.neurology.org/lookup/doi/10.1212/WNL.0000000000011552 for full disclosures.