Blood biomarkers of seizures in epilepsy: machine learning reveals altered plasma proteome Authors: Sarah Akel, Saman Hosseini Ashtiani, Mayuresh Anant Sarangdhar, Markus Axelsson, Johan Zelano Published: 2025 ABSTRACT AbstractBackgroundFluid biomarkers are emerging for diagnosis and monitoring in neurology. Epilepsy remains an exception, despite seizures being known to result in inflammation and altered levels of several brain proteins. We aimed to identify candidate biomarkers through a comprehensive proteomic analysis of proteins associated with brain disease or inflammation.MethodsCross-sectional analysis of plasma from adult participants in the PREDICT biobank study in Västra Götaland, Sweden. Participants aged 18–50 with epilepsy and seizures (n=88) were compared with those with epilepsy who had been seizure-free (n=88) for >1 year using four OlinkExplore384 panels, analysing 1447 proteins.ResultsTwo machine-learning models, one linear and one non-linear, identified protein expression differences through feature selection, resulting in 51 unique proteins between the models. Twenty-three proteins were considered differentially expressed after false discovery rate adjustment (p≤0.05), including neurofilament light and several cytokines. Protein–protein interaction (PPI) analysis identified clusters among 51 unique proteins, with the largest clusters primarily associated with inflammatory processes. In addition, machine learning–independent gene set enrichment analysis identified 34 gene sets, mainly related to immune and inflammatory processes, that were also enriched in participants with seizures.ConclusionsPersons with epilepsy and seizures had different plasma protein profiles compared with those seizure-free, primarily suggesting altered inflammatory/immune processes. This fits well with growing evidence of inflammation as a key process in epilepsy. In addition to new therapeutic targets, immune processes need further exploration as candidate biomarkers for monitoring of treatment response in epilepsy.Keywords: EPILEPSY WHAT IS ALREADY KNOWN ON THIS TOPIC Evidence suggests an association between inflammation and epilepsy, emphasising the importance of investigating inflammatory pathways and proteins. Alterations in blood–brain injury markers, such as neurofilament-light and S100 calcium-binding protein B, have also been reported after seizures in epilepsy. This has generated interest in understanding the relationship and extent of brain injury and neurodegeneration associated with epilepsy. Although proteomic studies on epilepsy have been conducted, we still lack established fluid biomarkers for the condition and a thorough understanding of blood proteome changes resulting from seizures. Evidence suggests an association between inflammation and epilepsy, emphasising the importance of investigating inflammatory pathways and proteins. Alterations in blood–brain injury markers, such as neurofilament-light and S100 calcium-binding protein B, have also been reported after seizures in epilepsy. This has generated interest in understanding the relationship and extent of brain injury and neurodegeneration associated with epilepsy. Although proteomic studies on epilepsy have been conducted, we still lack established fluid biomarkers for the condition and a thorough understanding of blood proteome changes resulting from seizures. WHAT THIS STUDY ADDS Proteomic analysis identified 51 proteins associated with seizure status, of which 23 were differentially expressed (p≤0.05) between groups. Clusters identified in protein–protein interaction analysis were mainly associated with inflammation. Additionally, out of 34 gene sets identified through gene set enrichment analysis, several were also associated with immune and inflammatory processes. Proteomic analysis identified 51 proteins associated with seizure status, of which 23 were differentially expressed (p≤0.05) between groups. Clusters identified in protein–protein interaction analysis were mainly associated with inflammation. Additionally, out of 34 gene sets identified through gene set enrichment analysis, several were also associated with immune and inflammatory processes. HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY This study shows that individuals with epilepsy and seizures have different protein profiles in the plasma compared with those who are seizure-free. The study provides further insight into blood proteome changes that occur as a result of ongoing seizures in individuals with epilepsy. Further investigations into the role of the identified proteins and inflammatory pathways as biomarkers of epilepsy are needed. This study shows that individuals with epilepsy and seizures have different protein profiles in the plasma compared with those who are seizure-free. The study provides further insight into blood proteome changes that occur as a result of ongoing seizures in individuals with epilepsy. Further investigations into the role of the identified proteins and inflammatory pathways as biomarkers of epilepsy are needed. Background Epilepsy affects over 50 million people worldwide, out of which one-third continue to have seizures despite antiseizure medications.1 Drug-resistant epilepsy is not only linked to seizure-related injury and death but also cognitive symptoms such as memory impairment.2 Blood protein levels are increasingly used in the study of brain disorders like Alzheimer’s disease and multiple sclerosis, particularly proteins linked to neurodegeneration and inflammation.3,5 Epileptic seizures leave short-term biochemical traces, such as lactic acidosis, but persistent biomarkers are not yet available.6 We have recently described that neurofilament-light (NEFL) levels are higher if seizures persist in epilepsy.7 Emerging literature suggests that in addition to neuronal injury markers like NEFL, cytokines increase following seizures.7,9 Based on this finding, along with evidence suggesting that drug-resistant epilepsy features chronic inflammation in affected brain tissue, we hypothesised that there are chronic alterations in the plasma proteome of patients with epilepsy who have seizures compared with those who are seizure-free.10 11 Inflammatory mediators can, in turn, make neurons hyperexcitable, potentially creating a bidirectional interaction between the blood and the brain in seizure-promoting states.12 13 This highlights the need for a more detailed understanding of seizure-induced changes in the plasma proteome. In addition to enhancing pathophysiological understanding, these insights could potentially offer scalable methods for monitoring disease activity and brain health in epilepsy through blood tests. Proteomic studies in epilepsy are emerging, but further research is still needed to address seizure-induced changes to the proteome.14,16 We analysed 1447 plasma proteins in 176 individuals with epilepsy using four Olink Explore 384 panels. In this exploratory study, we aim to compare plasma protein profiles between persons with and without seizures. Using machine learning, we identified 51 proteins related to seizures. Among these, 23 proteins were significantly differentially expressed between the seizure and seizure-free groups, including the already known seizure-related NEFL. Network and machine learning–independent enrichment analyses suggested that inflammatory processes were primarily associated with recent seizures. Methods Study design and participant selection The Prospective Regional Epilepsy Database and Biobank for Individualised Clinical Treatment (PREDICT) (clinicaltrials.org, NCT04559919) aims to identify biomarkers of clinical utility in epilepsy and has recruited participants at five neurology outpatient clinics in Västra Götaland region of Sweden since December 2020. The cohort has a comparable age and sex distribution to all adult patients seen at neurology departments for epilepsy in the region and Sweden as a whole.16 The study is approved by the Swedish Ethical Review Authority (approval number 2020–00853), and all participants provided informed consent before their inclusion. An overview of the study design is illustrated in figure 1. Overview of the study design. (A) Plasma samples from 176 participants selected from the PREDICT biobank were analysed using four Olink Explore 384 panels. Experimental analysis was performed using the proximity extension assay and next-generation sequencing. (B) A total of 1447 proteins underwent computational analysis where the values of these proteins were assessed and compared between participants with and without seizures. Two machine-learning models were used: random forest and sPLS-DA. Further analysis involved a moderated t-test to identify statistically significant differentially expressed proteins selected from the machine-learning models along with protein–protein interaction and enrichment analyses. FDR, false discovery rate; sPLS-DA, sparse partial least squares discriminant analysis We included 176 participants from PREDICT, aged 18–50 years, with an epilepsy diagnosis as per the current definition from the International League Against Epilepsy.17 The upper age limit was set to reduce variability from age-related brain changes. Participants were selected based on their seizure status. Participants who reported seizures ≤1 year at the last clinic visit/or blood sampling date (n=88) were categorised as having ‘seizures’, whereas those who reported no seizures for >1 year (n=88) were categorised as ‘seizure-free’. All participants had their blood sampling within 10 days or less of their latest clinic visit. None of the participants had an acquired cause of their epilepsy, such as stroke or traumatic brain injury. Blood collection and storage On recruitment to PREDICT, blood samples were collected into EDTA tubes and centrifuged for 10 min at 2000×g at room temperature. Plasma supernatant was collected and aliquoted, with all samples stored at −80°C in the regional healthcare-integrated Biobank Väst (registration number 890) until further analysis. Clinical data collection Clinical data were collected into a pseudonymised clinical report form from medical records and recruitment appointments by a neurologist. The variables used in this study included age, sex, radiology findings (clinical reports from MRI (n=130) and/or CT (n=19)), seizure frequency and type of epilepsy (focal, generalised or unknown onset). Olink proteomics Markers were quantified in the plasma at SciLifeLab Uppsala (Uppsala University, Sweden) using four Olink Explore384 (Uppsala, Sweden) panels: Neurology, Neurology II, Inflammation and Inflammation II. To keep the sample size within the capacity of two plates, a total of 176 participants were selected for this study. Olink uses a dual-recognition immunoassay based on their proximity extension assay (PEA) method.18 Two oligonucleotide-coupled antibodies (PEA probes) bind closely to their protein target, enabling the oligonucleotides to hybridise. This creates a distinct DNA template quantified with next-generation sequencing. The output is expressed as normalised protein expression (NPX) values on a log2 scale indicating relative quantification. Data preprocessing Data analyses were conducted in the R programming environment (version 4.4.1). We merged the four Olink panels, and in the case of overlapping assays (n=6), the assay with the highest NPX value was kept in the resulting dataset. Nineteen proteins failed to meet Olink’s quality control criteria and were excluded. The final dataset comprised 1447 proteins. Machine learning Machine-learning algorithms were chosen due to their ability to manage large sets of features, selecting the most informative ones. Machine-learning approaches can model more complex and non-linear interactions, uncovering associations that traditional statistical methods may overlook. The primary objective in this study was to use machine-learning models for protein discovery instead of clinical prediction. To capture both linear and non-linear associations between protein abundances and seizure status, two machine-learning approaches were applied: sparse partial least squares discriminant analysis (sPLS-DA) using the R package mixOmics (version 6.28.0)19 and the Random Forest (RF) R package (version 4.7–1.2).20 The incorporation and adjustment for age and sex were tested in both models in terms of the area under the receiver operating characteristic curve (AUROC) values. Given the limited number of samples, the separation of the data into training, validation and test subsets and even the nested cross-validation would lead to worse classification power in terms of AUROC values. Therefore, 10-fold cross-validation was used for model evaluations. The statistical significance of the AUROC was tested using the non-parametric DeLong approach.21 One-sided and two-sided hypothesis tests were implemented against the null hypothesis of AUROC=0.5, that is, no discrimination. 95% CIs for AUROCs were also obtained using the DeLong method. Sparse partial least squares discriminant analysis sPLS-DA was employed to achieve both dimension reduction and feature selection, identifying proteins that best differentiated participants with and without seizures. This method constructs components that maximise covariance between the response variable (seizure status) and predictors (protein levels). The incorporation and adjustment of age and sex into this model would worsen the classification performance evaluated by AUROC; therefore, age and sex were not used in the final model. The optimal number of variables to keep was determined by grid search and repeated cross-validation. Through the mentioned model tuning, the balanced error rate criterion was used to choose the number of variables to select in the model (online supplemental figure S1A). Random forest Complementary to the sPLS-DA model, RF was also employed. This ensemble method leverages multiple decision trees to enhance predictive accuracy and mitigate overfitting. Each tree was developed using a bootstrap sample, with a subset of features randomly chosen at each node to determine splits. The model was designed to classify participants based on seizure status. The adjustment of age and sex into this model would worsen the classification performance evaluated by AUROC. However, the incorporation of age and sex improved the AUROC. Therefore, age and sex were incorporated into the final model. Using 5000 trees, we set the number of proteins considered at each split to 14, determined through a preliminary grid search using cross-validation for improving model performance (online supplemental figure S1B). We selected an equal number of proteins used by the RF model as the sPLS-DA model for classification to keep the balance in capturing linearity and non-linearity. Model validation Model performance metrics, including sensitivity, specificity and AUROC values, were calculated to assess the classification performance for each model (online supplemental figure S1C). Feature importance was determined for the RF model using the mean decrease in impurity and for sPLS-DA using the loading weights of selected features. Statistical analysis Using the moderated t-test from the limma R package (version 3.58.1),22 23 we identified proteins selected from the union of sPLS-DA and RF that were differentially expressed between participant groups after FDR adjustment (adjusted p≤0.05), controlling for age and sex. We also re-ran the same machine-learning analysis and moderated t-test by categorising participants into focal and generalised epilepsy subgroups to determine whether the proteomic profiles differed between these two types of onset. To evaluate whether plasma protein levels vary with the number of days since the last seizure, independent of the seizure group classification, we fitted a pooled linear model to the NPX values for all proteins. The main predictor was log2 (1+days since last seizure). Log2 transformation was applied to stabilise skewness and enable interpretation as the change in NPX per doubling of seizure-free duration. Age and sex were included as covariates. The model was fitted using the limma framework, which estimates coefficients via least squares and applies empirical Bayes shrinkage to improve variance stability across proteins. Multiple testing correction was performed using the Benjamini–Hochberg FDR method. Analysis using the untransformed ‘days since last seizure’ variable was also performed, and results are provided in the supplementary material (online supplemental table S1). Protein–protein interaction network A protein–protein interaction (PPI) network was generated using the STRING database (accessed 6 December 2024, v12.0).24 The input gene set included the 51 proteins identified via the union of feature selection from sPLS-DA and RF models. The default interaction confidence score threshold (0.4) was applied. We visualised the first shell of interactors in the network, consisting of up to 20 interactor proteins per query protein, to increase the network’s size and improve clustering and enrichment analyses. Markov Cluster Algorithm (MCL) was applied to identify protein clusters (inflation parameter=4). Functional enrichment analyses were performed on the clusters (FDR-adjusted p≤0.05). Sub-group analysis: focal and generalised epilepsy A PPI network with clustering (MCL, inflation parameter=1.5) was also constructed for proteins identified as classifiers of seizure status in both focal and generalised epilepsy. Functional enrichment analyses were performed on the clusters (FDR-adjusted p≤0.05). No additional first shell interactors were included. Gene set enrichment analysis Gene set enrichment analysis (GSEA) was conducted using GSEA v4.3.3 (Broad Institute, Cambridge, MA, USA).25 The NPX values of 1447 proteins were provided as input in a gct file. Patient phenotype information, categorised as ‘with seizure’ and ‘seizure-free,’ was supplied in a cls file. P-values for the enrichment scores were calculated through an empirical, phenotype-based permutation test. The number of permutations was set to 1000, and the permutation seed was set to 149. Weighted enrichment statistics were used to calculate the enrichment scores. The C5 ontology gene set database was used as input, which consists of 16 107 gene sets, and all other parameters were kept at their default settings. Gene sets with nominal p<0.05 and FDR q-value <25% were considered statistically significant enriched sets. Results Cohort description The median age of the 176 participants was 31 (range, 18–50) years, and most participants in the seizure group had focal epilepsy (table 1). The majority of participants had normal radiology (n=104) or a non-acquired abnormality unrelated to the epilepsy (n=24) on imaging. There were 17 participants with an epileptogenic lesion on imaging. Thirty-one individuals had no clinical radiology reports, most of whom had generalised epilepsy (n=18). Table 1Participant demographics and characteristics (n=176)Seizure-free group (n=88)Seizure group (n=88)Median age (range)30 (18–49)32 (19–50)Sex (%) Male29 (33.0)44 (50.0) Female59 (67.0)44 (50.0)Epilepsy type (%) Focal31 (35.2)53 (60.2) Generalised31 (35.2)18 (20.5) Unknown26 (29.5)17 (19.3)Median days since last seizure (range)1793 (367–9472)n=85 with available data22 (0–360)n=81 with available dataMedian seizure frequency in the 2 months prior to blood sampling/last clinic visit (range)*NA2 (1 – 98)n=65 with available dataAnti-seizure medication (%) Monotherapy70 (79.5)39 (44.3) Polytherapy18 (20.5)47 (53.4) No current ASM/unknownNA2 (2.3) Lamotrigine46 (52.3)43 (48.9) Levetiracetam27 (30.7)38 (43.2) Carbamazepine11 (12.5)12 (13.6) Valproic acid5 (5.7)12 (13.6) Lacosamide3 (3.4)17 (19.3) Oxcarbazepine5 (5.7)1 (1.1) Topiramate6 (6.8)5 (5.7) Other†5 (5.7)31 (35.2)*Seizure frequency in individuals who experienced seizures in the 2 months leading up to the sampling/last clinic visit (n=71).†Phenytoin, perampanel, clobazam, phenobarbital, gabapentin, cenobamate, eslicarbazepine, zonisamide, clonazepam, brivaracetam, pregabalin and ethosuximide.ASM, anti-seizure medication; NA, not applicable. Participant demographics and characteristics (n=176) Seizure frequency in individuals who experienced seizures in the 2 months leading up to the sampling/last clinic visit (n=71). Phenytoin, perampanel, clobazam, phenobarbital, gabapentin, cenobamate, eslicarbazepine, zonisamide, clonazepam, brivaracetam, pregabalin and ethosuximide. ASM, anti-seizure medication; NA, not applicable. Machine learning identifies proteins associated with seizures Out of the 1447 proteins, sPLS-DA and RF each identified 30 proteins as important classifiers. The union of these two sets resulted in 51 unique proteins, with nine common to both algorithms (online supplemental table S2). Among these, 23 proteins were differentially expressed after a moderated t-test with FDR adjustment (table 2, figure 2). Samples were projected onto the two-dimensional space by their first two components calculated using sPLS-DA (online supplemental figure S1D). The AUROC of the sPLS-DA model was 0.65 (95% CI 0.58 to 0.75, DeLong). The one-sided DeLong test showed that the AUROC was significantly greater than chance (p<0.01), being in compliance with our a priori hypothesis that the model would outperform random classification. The two-sided test for deviation from chance also achieved statistical significance (p<0.01), demonstrating that the observed discrimination has not occurred under the null hypothesis of no association. The random forest model achieved an AUROC of 0.61 (95% CI 0.54 to 0.7, DeLong). The one-sided and two-sided DeLong tests for the RF model also showed statistical significance (p<0.01). Table 2Description of differentially expressed proteins identified via moderated t-test with FDR adjustmentUniprot accessionGene symbolProtein nameLog fold changeP value*Q9UQV4LAMP3Lysosome-associated membrane glycoprotein 30.6360.001Q96A32MYLPFMyosin regulatory light chain 2, skeletal muscle isoform0.6340.001P35613BSGBasigin0.3350.002Q07065CKAP4Cytoskeleton-associated protein 40.4870.002O95971CD160CD160 antigen0.5770.003P22102GARTTrifunctional purine biosynthetic protein adenosine-30.3730.003Q13241KLRD1Natural killer cell antigen CD940.5910.003P00813ADAAdenosine deaminase0.4290.004P55291CDH15Cadherin-150.1970.005Q10588BST1ADP-ribosyl cyclase/cyclic ADP-ribose hydrolase 20.2020.006Q13478IL18R1Interleukin-18 receptor 10.5340.006Q03403TFF2Trefoil factor 20.3460.006Q04609FOLH1Glutamate carboxypeptidase 20.6870.007Q12866MERTKTyrosine-protein kinase Mer0.4870.007P07196NEFLNeurofilament light polypeptide0.2990.007P09466PAEPProgestagen-associated endometrial protein (glycodelin)−0.6940.007Q92834RPGRX-linked retinitis pigmentosa GTPase regulator0.6760.007O60279SUSD5Sushi domain-containing protein 50.1540.009Q969V3NCLN†Nicalin0.5920.01O43493TGOLN2Trans-Golgi network integral membrane protein 20.110.011O95715CXCL14†C-X-C motif chemokine 140.2490.013Q08174PCDH1Protocadherin-10.4020.015P26951IL3RAInterleukin-3 receptor subunit alpha0.1430.022*These p-values remained significant after FDR adjustment (p≤0.05).†Proteins with>30% of samples below the limit of detection (see online supplemental table S11).FDR, false discovery rate. Description of differentially expressed proteins identified via moderated t-test with FDR adjustment These p-values remained significant after FDR adjustment (p≤0.05). Proteins with>30% of samples below the limit of detection (see online supplemental table S11). FDR, false discovery rate. Tukey box and whisker plot illustrating the 23 significant proteins between the seizure and seizure-free groups. The median is represented by the centre line of the box and whisker plot. The lower and upper hinges represent the 25th and 75th percentiles, respectively, and the whiskers extend 1.5 times the IQR from both hinges. Analysis was done using a moderated t-test with FDR adjustment (p≤0.05) and performed on the 51 unique proteins identified through machine-learning analysis. Arrows depict the direction of values that appear outside of the visualised plot range. ADA, adenosine deaminase; BSG, basigin; BST1, ADP-ribosyl cyclase/cyclic ADP-ribose hydrolase 2; CD160, CD160 antigen; CDH15, cadherin-15; CKAP4, cytoskeleton-associated protein 4; CXCL14, C-X-C motif chemokine 14; FDR, false discovery rate; FOLH1, glutamate carboxypeptidase 2; GART, phosphoribosylglycinamide formyltransferase, phosphoribosylglycinamide synthetase, phosphoribosylaminoimidazole synthetase (trifunctional purine biosynthetic protein adenosine-3); IL18R1, interleukin-18 receptor 1; IL3RA, interleukin-3 receptor subunit alpha; KLRD1, killer cell lectin-like receptor subfamily D, member 1 (natural killer cells antigen CD94); LAMP3, lysosome-associated membrane glycoprotein 3; MERTK, tyrosine-protein kinase Mer; MYLPF, myosin regulatory light chain 2, skeletal muscle isoform; NCLN, nicalin; NEFL, neurofilament light polypeptide; NPX, normalised protein expression; PAEP, progestagen-associated endometrial protein (glycodelin); PCDH1, protocadherin-1; RPGR, X-linked retinitis pigmentosa GTPase regulator; SUSD5, sushi domain-containing protein 5; TFF2, trefoil factor 2; TGOLN2, trans-Golgi network integral membrane protein 2. To explore associations between the time since the last seizure and protein levels, we fit a linear model across all participants, using days since last seizure as the primary predictor, adjusting for age and sex. After applying a log2 transformation to the days since last seizure, several proteins showed significant associations (online supplemental table S3). The proteins largely overlapped with those identified in the group comparison analysis (table 2), implying a temporal component to the observed protein changes. As a sensitivity analysis, the seizure group was restricted to participants with seizures within 6 months of their last clinic visit or blood sampling date (n=82), whereas the seizure-free group was defined as being seizure-free for >1 year (n=88). Age and sex distribution were comparable with those in the original seizure group. The 51 unique union proteins previously identified from the machine-learning models were re-evaluated using the moderated t-test. The majority of proteins showed consistent group differences under the 6-month definition of the seizure group, with two that lost significance (interleukin-3 receptor subunit alpha (IL3RA) and trans-Golgi network integral membrane protein 2) and four newly identified as significant after FDR adjustment (Platelet glycoprotein 4 (CD36), protein-arginine deiminase type-2 (PADI2), prorelaxin H2 (RLN2) and zinc finger protein GLI2 (GLI2))(online supplemental table S4). Protein–protein interaction network and enrichment analysis The PPI network of the 51 proteins identified by sPLS-DA and RF comprised 71 nodes and 291 edges (figure 3A). We identified 10 clusters within the network (online supplemental table S5). The three largest clusters were primarily involved in inflammatory processes. Cluster 1, comprising 18 proteins, was enriched for ‘T-cell activation’ and ‘cytokine–cytokine receptor interaction’. Cluster 2 (13 proteins) was enriched for the ‘regulation of natural killer cell-mediated cytotoxicity’, ‘antigen processing and presentation’ as well as ‘MHC protein complex binding’, while cluster 3 (10 proteins) was enriched for ‘regulation of macrophage activation’ and the regulation of cytokine production (including interleukin (IL)-1, IL-6 and tumour necrosis factor). Several small clusters were generated; cluster 4 was enriched for metabolism, whereas cluster 5 was enriched for cell adhesion. The remaining clusters had limited or no significant enrichment. Protein−protein network and gene set enrichment analysis. (A) Protein−protein interaction network with MCL clustering of proteins identified from machine-learning models. Only proteins with connected nodes were visualised, and first shell interactors (up to 20 interactors per query protein) were included. Nodes represent proteins, and edges represent protein−protein interactions. Dashed lines indicate edges between clusters. (B) Representative gene set enrichment analysis (GSEA) plots. GSEA was based on protein profiles of ‘seizure’ (in red) and ‘seizure-free’ (in blue) participants. GOBP, Gene Ontology Biological Process; GOMF, Gene Ontology Molecular Function; NES, normalised enrichment score; NOM p, nominal p-value. Gene set enrichment analysis Complementary to the machine-learning approach, we used GSEA software to identify gene sets associated with seizures. GSEA revealed altered pathways related to seizures, with a total of 34 gene sets showing significant enrichment. These gene sets predominantly fell under immune (particularly T-cell differentiation) and inflammation categories (such as type II interferon production) (figure 3B, online supplemental table 6). Focal and generalised epilepsy We also performed subgroup analyses on focal (n=84) and generalised (n=49) epilepsy. The two machine-learning models identified 92 proteins as classifiers of seizure status in focal epilepsy and 160 in generalised. A moderated t-test identified 26 proteins significantly different between seizure and seizure-free groups with focal epilepsy (adjusted p≤0.05) (online supplemental table S7); however, no proteins remained significant after FDR adjustment for generalised epilepsy (online supplemental table S8). Thirteen proteins overlapped between focal and generalised epilepsy, 23 between focal and all participants (focal, generalised+unknown onset) and 20 between generalised and all. A PPI network with clustering analysis was generated separately for focal and generalised epilepsy. Both networks showed one dominant cluster (focal=42 proteins, generalised=98 proteins) enriched for various inflammatory processes (focal, online supplemental figure S2; online supplemental table S9; generalised, online supplemental figure S3; online supplemental table S10). There were several shared inflammatory processes between focal and generalised protein networks such as the ‘regulation of cytokine production’ and ‘T-cell activation’. Discussion The aim of this exploratory study was to investigate plasma changes associated with seizures in individuals with epilepsy using four Olink proteomic panels covering 1447 proteins. Fifty-one unique proteins, among both machine-learning workflows, were identified, of which 23 proteins were considered differentially expressed after FDR adjustment. Our findings demonstrate that the plasma proteome differs between individuals who were experiencing seizures and those who were seizure-free, expanding the potential for future biomarkers beyond known short-term blood alterations after seizures, like lactic acidosis and muscle enzyme release. The findings are in agreement with several more recent reports of changes in neurological or inflammatory markers after seizures.26 Our proteomic approach identified several key markers, including the well-reported neurofilament light polypeptide (NEFL), as well as various inflammatory markers such as interleukins (IL3RA and IL18R1) and other established inflammation markers (eg, killer cell lectin-like receptor subfamily D, member 1, and killer cell lectin-like receptor subfamily F member 1). Enrichment and protein–protein interaction analyses of the machine learning–identified proteins pointed to several inflammatory processes. The central three clusters were predominantly associated with inflammatory processes, including T-cell activation, regulation of cytokine production and natural killer cell–mediated toxicity. The protein network also identified IL-6 and interferon-γ as neighbour proteins interacting with our core protein network. Several studies have reported that they increase in epilepsy, with IL-6 particularly having been associated with seizure severity.27 28 As a confirmatory analysis, a machine learning–independent GSEA also demonstrated that immune and inflammation-related pathways were overrepresented in seizure cases compared with seizure-free cases. The analysis pointed to both adaptive and innate immune responses. The interaction between brain seizure activity and inflammation is gaining interest not only in immune-mediated epilepsy or encephalitis but also with the notion of inflammation perpetuating and maintaining an ictogenic state.12 Inflammation is also of interest in drug-resistant epilepsy.10 29 Clinical trials of anti-inflammatory treatments in drug-resistant epilepsy have shown variable results; however, a substantial proportion do respond.30,32 A more detailed understanding of the inflammatory pathways involved through proteomics could perhaps assist in predicting responses and selecting the right immunotherapy for appropriate patients. Our method also identified NEFL as a differentially expressed protein, which has been previously reported by several investigators to increase after seizures.7 8 Traditionally recognised as a marker of axonal injury, NEFL has been well-studied in conditions such as Alzheimer’s disease and multiple sclerosis.33 34 However, higher levels may also indicate other processes, such as plasticity, rather than solely injury.35 36 In this study, we also found three other proteins associated with seizures as in our previous study: cadherin-15, progesterone-associated endometrial protein (glycodelin) and PHOSPHO1.16 We also analysed focal and generalised epilepsy separately, although the power of the analysis naturally decreased with participant numbers. The exact protein lists identified by machine learning differed, but several gene ontology terms related to inflammatory processes were common to both groups. One possible interpretation is that inflammation detected in the proteome indicates reactions to epileptic brain activity rather than underlying epileptogenic processes, which are different in the two types of epilepsy. Larger studies are needed to explore this observation further, but if confirmed, it suggests the potential of some inflammatory mediators as seizure biomarkers in both focal and generalised epilepsy. We used a dual machine-learning approach for complementary strengths of the respective algorithms. To capture complex dependency and interplay patterns between seizure status and protein NPX levels, we selected RF over gradient boosting methods, such as XGBoost, for its robustness to noise and overfitting and superior feature importance interpretability. RF’s bagging-based approach ensures stable and biologically interpretable feature selection in high-dimensional proteomic datasets.37 Furthermore, empirical evidence supports the reliability of RF in biomarker discovery studies.37 38 However, RF has inherent limitations concerning the reproducibility of feature importance scoring. The randomised nature of the algorithm, including the use of bootstrapped samples and random feature selection at each tree split, can result in variability in the feature importance scores between different runs. This variability may impact the consistency of the identified biomarkers, particularly in high-dimensional datasets with correlated features like proteomics data. To mitigate this limitation, we introduced a fixed random seed before each run of the RF model. Additionally, all analyses were performed on the same machine specifications to eliminate variability arising from differences in computational environments. sPLS-DA was employed as a complementary method to RF to leverage its capacity for simultaneous feature selection and dimension reduction. By focusing on linear relationships, sPLS-DA provides an interpretable model that identifies a minimal set of proteins associated with seizure status while mitigating overfitting through sparsity. However, sPLS-DA’s limitations, such as its sensitivity to parameter tuning, assumption of linearity and potential exclusion of moderate-effect features, should be considered. These characteristics underscore the importance of combining sPLS-DA with non-linear approaches like RF to ensure a comprehensive understanding of the proteomic signatures associated with seizures. Other limitations include the heterogeneity of the cohort. Although this was a large proteomic investigation in epilepsy, the number of patients was possibly too small for adequate statistical power to detect all significantly regulated proteins in individuals with seizures. While we examined protein levels separately in focal and generalised epilepsies, our sample size did not allow for robust stratification by more specific seizure semiology (eg, tonic-clonic seizures). These subtypes are likely to influence biomarker profiles, and seizure type-specific analysis will be important to address in future studies. That a protein was not identified should not be interpreted as having no difference in its expression caused by seizures; our aim was to perform a proof-of-concept analysis of the plasma proteome in individuals with seizures. Our inclusion criteria focused on younger age and non-acquired epilepsy to reduce biological variability, but future proteomic studies are needed on specific epilepsies and patient groups. Reassuringly, some of the differentially expressed proteins have been identified in previous analyses.7 16 Among these are findings from a methodologically distinct mass spectrometry experiment, where we also found more inflammation in the plasma of patients with seizures compared with those who were seizure-free.39 Another limitation is the use of Olink panels as a basis for the analysis, which focused on neurology and inflammation-related proteins. These were selected based on the pre-existing knowledge on inflammation in drug-resistant epilepsy, but naturally, additional proteomic studies that use larger panels containing a more biologically diverse set of proteins would be useful. Conclusion We found differences in the plasma proteome between individuals with epilepsy and seizures and those with epilepsy who are seizure-free. In addition to known markers like NEFL, both machine learning and machine learning–independent analyses identified inflammation-associated clusters. Longitudinal sampling would be an interesting next step, studying the dynamics of protein levels with different seizure frequencies in response to ASMs. Overall, our results suggest protein alterations in the plasma of individuals with seizures, and these proteins may serve as potential markers in epilepsy. Supplementary material Acknowledgements This study was conducted using professional biobank services from Biobank West and Biobank Sweden. We would like to thank the Affinity Proteomics unit (SciLifeLab) at Uppsala (Sweden) for the experimental analysis of the Olink panels. Figure 1 contains an icon provided by Servier Medical Art by Servier (https://smart.servier.com/), licensed under Creative Commons Attribution 4.0 International (https://creativecommons.org/licenses/by/4.0/). Footnotes Funding: This study was funded by the Wallenberg Center for Molecular and Translational Medicine, the Jeansson foundation (JS2014-0032), Swedish Society for Medical Research (SS18-0040), Region Västra Götaland (VGFOUREG-968476), the ALF-agreement (ALFGBG-715781), the Wilhelm & Martina Lundgrens Vetenskapsfond, Hjärnfonden, Amlöv foundation and Ann-Louise och Sven-Erik Beiglers foundation. Provenance and peer review: Not commissioned; externally peer reviewed. Patient consent for publication: Not applicable. Ethics approval: This study involves human participants and was approved by Swedish Ethical Review Authority (approval number 2020-00853). Participants gave informed consent to participate in the study before taking part. Data availability free text: The results of the study contain sensitive personal data (according to Swedish research regulations) and cannot be shared by the authors. Data availability statement No data are available. References Associated Data Supplementary Materials Data Availability Statement No data are available.