Introduction

Epilepsy is a serious neurological disorder characterized by unprovoked seizures, which affects up to 1% of individuals worldwide (WHO, 2019), with children and the elderly being particularly affected. Although epilepsy can be caused by acquired conditions such as stroke, tumor or head injury, most cases (ca. 70–80%) are due to genetic influences1, including rare and common genetic variants. Diagnosing epilepsy is often challenging2,3,4 and multiple individuals are initially misdiagnosed5. An epilepsy diagnosis is potentially lifesaving with a 3x elevated mortality risk in epilepsy (WHO, 2019). Epilepsy-related deaths can be prevented by antiseizure medication (ASM) which however often have adverse effects6. Thus, correct epilepsy diagnosis is crucial, but the most widely-used diagnostic tool in epilepsy, the electroencephalogram (EEG), has quite variable sensitivity and specificity in different clinical settings ranging about 17–58% and 70–98%, respectively7,8 and moderate inter-rater agreement9, illustrating a need for additional biomarkers3. Due to the great importance and challenge specific ‘first seizure clinics’ are solely dedicated to investigating an epilepsy diagnosis after a newly onset seizure2. Until 2014, epilepsy was defined as having two unprovoked seizures >24 h apart by the International League Against Epilepsy10. This definition was then extended to the following additional scenarios of having 2) a diagnosis of an epilepsy syndrome or 3) one unprovoked seizure and a probability of further seizures with a recurrence risk of at least 60% over the next 10 years10.

The common epilepsies can be broadly categorized into genetic generalized (GGE) and non-acquired focal epilepsies (NAFE), where the latter originate from a particular brain area10. First-degree relatives of patients with GGE had an 8.3-fold increased risk of developing GGE while first-degree relatives of patients with NAFE had a 2.5-fold increased risk of developing NAFE, compared to the general population, respectively11. In agreement, the SNP-heritability (i.e., the variance of GGE attributed to common genetic variants) is approximately 30–40%12,13,14 which is relatively high compared to other common diseases15. The same measure is more moderate for NAFE with SNP-heritability of about 9-16%12,13. Previous genome-wide association studies have shown that common variants contribute more substantially to the more common forms of epilepsy12. There is only a modest burden of ultra-rare genetic variants in GGE and NAFE; rare variants likely contribute only a small fraction towards their heritability16 and there are few Mendelian disease genes exclusively associated with them17.

Recent research has shown that common genetic variants with small effects on specific diseases can be combined into “polygenic” risk scores (PRSs) with high disease-specific PRSs conferring comparable risk as rare monogenic variants18. Thus, interest in PRSs is growing as a potential clinically important diagnostic tool19,20,21,22,23. It was recently shown that individuals with epilepsy had a significantly higher epilepsy PRS compared to unaffected controls24,25. However, investigating how epilepsy PRSs may predict epilepsy risk in specific clinical scenarios has so far been lacking. Here, we thus investigate how epilepsy PRSs can stratify epilepsy risk across lifetime and after unspecified seizure events.

Results

Electronic health records accurately represent epilepsy diagnoses

We investigated epilepsy PRSs in detailed longitudinal electronic health records (EHR) from the FinnGen project26,27 using FinnGen data freeze R12 (n = 520,105, 282,064 females) and the Estonian biobank28 as a validation cohort (for further study sample characteristics, see Table 1). We further explored the BioMe cohort29 with regards to Non-European ancestries. Our phenotype data was derived from ICD codes and ASM purchases and reimbursements of official state registries spanning up to 50 years. We defined non-acquired focal epilepsy (=NAFE) by having ≥2 NAFE-specific ICD codes and genetic generalized epilepsy (=GGE) by having ≥2 GGE-specific ICD codes, respectively. Additionally, we require ≥2 ASM purchases for a NAFE or GGE category (Supplementary Fig. 1). Further details on epilepsy case definitions can be found in Supplementary Tables 1 and 2 and Methods. Sample numbers are given in Table 1. We also investigated 226 individuals with ≥2 traditional idiopathic generalized epilepsy diagnoses (=IGE)30 i.e. Childhood/Juvenile Absence Epilepsy (ICD 40.33/35), Juvenile Myoclonic Epilepsy (ICD 40.36) and Generalized Tonic–Clonic Seizures Alone (ICD 40.35). Individuals’ age at first epilepsy diagnosis was in line with the known age of onset of respective IGE syndromes supporting our EHR-derived diagnoses (Supplementary Fig. 2).

Table 1 Descriptive Statistics of main study and replication cohort
Full size table

Epilepsy PRS is most elevated in GGE, specifically IGE, with the same effect sizes as in clinically curated epilepsy cohorts

We then calculated epilepsy PRSs to determine individuals’ genetic burden for epilepsy. Here, we used the International League Against Epilepsy (ILAE) genomewide association study’s (GWAS) 2023 summary statistics13 as discovery data, i.e. to determine which genetic variants increase or decrease epilepsy risk. We then summed 1000 s of genetic risk/protective variants for epilepsy with individually small effects into a single epilepsy PRS per individual. Here, we constructed separate focal epilepsy PRS (PRSNAFE) and generalized epilepsy PRS (PRSGGE). We found a significant elevation of PRSGGE in 924 individuals with GGE (Fig. 1A) which was particularly pronounced in IGE (see also next paragraph and Fig. 1B). We also found a significant elevation of PRSNAFE in 5509 individuals with NAFE (Fig. 1C), but no significant elevation of PRSGGE in individuals with unspecified seizures (Fig. 1D). Similarly, we found a high correlation (Pearson’s correlation coefficient = 0.91, p-value = 4 × 10−8) between PRSGGE decile and GGE prevalence in our data (Fig. 2). Overall, we can thus confirm previous studies that used PRS as a marker for genetic liability of common epilepsy types. Importantly, we find very similar respective effect sizes of PRSGGE and PRSNAFE on GGE and NAFE as reported in previous cohorts from Epi25 or the Cleveland Clinic24 when using the same GWAS12 in an earlier version of the manuscript31 or the updated GWAS13 (see Table 2). We thus consider it likely that the epilepsy phenotypes in our biobank data are comparable to the phenotypes curated according to clinical criteria in these cohorts.

Fig. 1: Epilepsy PRS of epilepsy cases (red) compared to population controls (gray) (n = 273,974) (density curves).
figure 1

A PRSGGE of GGE (n = 924) compared to controls. B PRSGGE of IGE (n = 226) compared to controls. C PRSNAFE of NAFE (n = 5509) compared to controls. D PRSGGE of unspecified seizure without epilepsy (n = 2485) compared to controls. On top of each panel: odds ratios (95%- confidence intervals in brackets) and p-values of standard deviation increase of epilepsy PRS on case versus control status. Mean of the PRS distributions is shown as vertical dotted lines. Method: logistic regression. OR; odds ratio, p; p-value.

Full size image
Fig. 2: Correlation of PRSGGE decile and GGE prevalence.
figure 2

The bins of adjacent PRS deciles are overlapping, as labeled. Data points are also provided as a Source Data file.

Full size image
Table 2 Enrichment of epilepsy cases in individuals with increasing epilepsy PRS
Full size table

High PRS is associated with epilepsy across lifetime and after unspecified seizure events

We next investigated the effect of epilepsy PRS on epilepsy rates across lifetime, separately for PRSGGE and PRSNAFE. We stratified our cohort into bins of epilepsy PRS standard deviations (SD) (Fig. 3) and compared the cumulative epilepsy incidence in each SD bin to the rest of the cohort (for increased power). Individuals with a PRSGGE > 2 SD (ca. 2% of the cohort) had a more than 4-fold lifetime risk of developing GGE than the rest of the cohort (Hazard ratio [HR]: 4.2, Confidence Interval [CI]: 3.2-5.4, p-value: 4 × 10−27, method: cox proportional hazard model32 [coxph], Fig. 3, panel A). The epilepsy risk decreased proportionally with the decreasing PRSGGE SD bin. Overall, the HR increased by 1.73 per increased SD of PRSGGE (Table 3, 95%-CI 1.62–1.86, p-value = 8 × 10−55). When restricting to IGE the HR per PRSGGE SD was 2.4 (95%-CI 2.1–2.7, p-value 2 × 10−34). Individuals with a PRSGGE > 1 SD had a HR of 12.1 for IGE compared to those with PRSGGE < −1 SD (95%-CI 6-25, p-value 3 × 10−11, IGE rate in PRSGGE < −1 SD: 8/75,114, IGE rate in PRSGGE > 1 SD: 88/75,505). PRS discriminated GGE cases versus controls with a concordance index (C-index) of 0.64 (95%-CI 0.61–0.68) adjusting for the same covariates (birth year, sex, batch and PCs). Overall, we thus showed PRSGGE as a significant biomarker for lifetime epilepsy risk.

Fig. 3: Epilepsy PRS as a marker for epilepsy risk across lifetime and after unspecified seizure events.
figure 3

A GGE (n = 924) across lifetime, (B) GGE (n = 266) after an unspecified seizure, (C) NAFE (n = 5509) across lifetime, (D) NAFE (n = 1290) after an unspecified seizure. In (A, B), we investigate PRSGGE, in (C, D), PRSNAFE. In each panel, on the left are density curves that display how samples are partitioned into six bins of PRS standard deviations. Survival curves in the middle give the cumulative epilepsy incidence (y-axis) across time (x-axis [years]) stratified for epilepsy PRS bins. In (B), the two lowest PRS bins are fused in the survival curve as epilepsy case counts were too low. The rightmost forest plots show epilepsy risk of each epilepsy PRS bin compared with the rest of the cohort. Here, the point estimates represent hazard ratios (method: cox proportional hazard’s model), error bars show the 95%-confidence intervals (CIs). Summary statistics are provided as a Source Data file.

Full size image
Table 3 Effects of PRS on epilepsy risk in biobanks FinnGen and Estonian biobank
Full size table

However, the absolute risk of developing epilepsy is small across lifetime (<1%, see Fig. 3A/C), even for individuals with high epilepsy PRS. Lifetime risk prediction is thus less clinically meaningful. When considering the subset of individuals that were diagnosed with an unspecified seizure corresponding to ICD code R56.8/7803A at an age <40 years their absolute risk for GGE increased compared to baseline (Fig. 3B). Within 10 years after the unspecified seizure, the GGE rate reached 42% in > 2 SD PRSGGE compared to 4% in <−2 SD PRSGGE (or 27% in > 1 SD PRSGGE versus 8% in < −1 SD PRSGGE). PRSGGE affected relative epilepsy risk similarly after an unspecified seizure (HR per PRSGGE SD: 1.5, 95%-CI: 1.3–1.8, p-value = 1 × 10−9, C-index 0.60, 95%-CI 0.53-0.67) as across lifetime. Similarly, PRSNAFE had a significant but more modest effect on NAFE cumulative lifetime incidence (HR per PRSNAFE SD: 1.13, 95%-CI: 1.09-1.17, p-value = 3 × 10−10) and after unspecified seizure (HR per PRSNAFE SD: 1.075, 95%-CI: 1.014–1.14, p-value = 0.02), in line with a lower heritability of focal epilepsy (Fig. 3C, D). In addition, we tested the effect of a PRSall-epilepsy computed from a GWAS of all epilepsy phenotypes, including unclassified epilepsy, on lifetime epilepsy. Unfortunately, we found only limited association with lifetime risk of GGE, NAFE or any epilepsy (Supplementary Table 3).

We replicated analyses of PRSGGE effects across lifetime in the Estonian biobank28 (Estonia, European ancestry, Supplementary Fig. 3), obtaining similar estimates (see Table 3) and thus validating our results. We further explored the effects of PRSGGE in individuals with diverse ancestries in the BioMe biobank (Supplementary Fig. 4, Supplementary Note), a biobank that links genetic and EHR data for more than 30,000 individuals from diverse ancestral and cultural backgrounds recruited primarily in the Mount Sinai Health System in New York City. While the effect of PRSGGE in BioMe followed similar trends, our analyses were underpowered and thus did not reach significance. Further analyses are needed to investigate the portability of epilepsy PRS effects to other ancestry groups.

Epilepsy PRS has sex-specific effects on epilepsy subtypes

In other diseases than epilepsy, studies previously reported sex-specific PRS effects and larger effects of PRS on disease in earlier age groups33. Thus, we sought to investigate the effect of age at onset and sex on PRS effects on epilepsy. We found a significant interaction of sex and PRSGGE on GGE case (n = 924) status (cox model p-value 0.002, regression p-value 0.02). So we next investigated the effect of PRS on lifetime epilepsy separately for men and women. PRSGGE had a larger influence on lifetime GGE in females (HRfemale per PRS SD: 1.9, 95%-CI 1.7–2.0, p-value = 1 × 10−47, nGGE = 543) than in males (HRmale per PRS SD: 1.5, 95%-CI 1.3–1.7, p-value = 2 × 10−11, nGGE = 381, Supplementary Fig. 5). We further found a higher prevalence of epilepsy in females, specifically with onset in the teenage—young adult range (Supplementary Fig. 6). Exploring sex-specific effects on specific epilepsy types we found no significant effect of sex (p = 0.4) nor PRS*sex interaction (p = 0.7) in IGE (n = 226) but found a significant effect of sex (p-value 7 × 10−4) and PRS*sex interaction (1 × 10−3) in non-IGE GGE (n = 657). Similarly, the effect of PRSGGE on non-IGE GGE was substantially higher in females (HRfemale 1.78, 95%-CI 1.60–1.99, p-value 5×10−26, HRmale 1.32, 95%-CI 1.16–1.50, p-value 2 × 10−5) while it was quite comparable for IGE (HRfemale 2.35, 95%-CI 1.16–1.60, p-value 1 × 10−25; HRmale 2.57, 95% CI 1.92–3.44, p-value 2 × 10−10). We also found a significant interaction of sex and PRSNAFE on NAFE (p-value 0.008) with slightly higher PRSNAFE effects on NAFE in females (HRmale: 1.10, 95%-CI 1.06–1.15, p-value = 2 × 10−6; n = 2706, HRfemale: 1.18, 95%-CI 1.13–1.22, p-value = 2 × 10−17, n = 2806).

Epilepsy PRS has a larger effect when epilepsy onset is earlier

We further explored whether epilepsy PRS effects were potentially different for different ages of epilepsy onset. We thus divided our cohort into quintiles of age at first epilepsy diagnosis and found significant effects of PRSGGE on GGE and of PRSNAFE on NAFE case status in all age at onset bins except GGE onset > 60 years and NAFE onset > 80 years (method logistic regression, see Supplementary Table 4). We found the largest effects of PRSGGE when individuals had earlier ages at first diagnosis, e.g. for GGE effects were largest at onset 0-20 (OR 1.9, 95%-CI 1.7–2.1, p-value 9 × 10−38) and of PRSNAFE on NAFE at onset 20-40 years (OR 1.21, 95%-CI 1.14–1.28, p-value 4 × 10−10). This is in line with other illnesses33. We next investigated, if the large genetic influences on IGE described in the paragraph above could be explained by a higher proportion of individuals with younger age at epilepsy onset in the IGE group. So within the GGE group, we compared the effect of PRSGGE on IGE versus non-IGE and still found a higher effect of PRSGGE on IGE even when accounting for age at first epilepsy diagnosis (OR 1.58, 95%-CI 1.29–1.95, p-value 2 × 10−5).

PRSGGE is specifically associated with GGE while PRSNAFE is more heterogeneous

We aimed to investigate the phenotypes associated with a genetic epilepsy liability that are not epilepsy, in a hypothesis-free approach to elucidate whether genetic factors influence GGE/NAFE in a disease-specific manner. We thus performed a phenome-wide association study (PheWAS) testing the effect of PRSGGE and PRSNAFE on 2139 distinct disease phenotypes (method: logistic regression, FinnGen data freeze: R6, GWAS: ILAE 201812, Fig. 4). GGE (labeled as ‘Generalized Epilepsy’) is the only phenotype that is significantly affected by PRSGGE after Bonferroni correction. We thus argue that PRSGGE is very specifically associated with GGE increasing its potential diagnostic utility. While PRSNAFE is expectedly associated with NAFE, multiple other phenotype associations are unexpected. The most significant ones are related to back pain, but also include hypertension, cardiovascular disease and depression medications, with lower significance. We tested the genetic correlation of NAFE and the 19 traits that were significant in our PheWAS (method: LD score regression34,35, Supplementary Fig. 7). After multiple testing correction none remained significant. However, phenotypes ‘other anxiety disorders’ (rg = 0.54, p-value = 0.02), ‘all anxiety disorders’ (rg = 0.44, p-value = 0.02) and ‘depression medications’ (rg = 0.33, p-value = 0.04) were genetically correlated with nominal significance.

Fig. 4: Phenome-wide association study testing the effect of epilepsy PRS on 2139 distinct disease phenotypes in FinnGen.
figure 4

In (a) we display a PheWAS of PRSGGE. GGE, here labeled as ‘Generalized Epilepsy’, is the only phenotype that is significant after Bonferroni correction. In (b) we display a PheWAS of PRSNAFE. In both panels, the phenotypes are grouped and colored by clinical field (x-axis). The y-axis shows the -log10 p-value of the association (method: logistic regression). The dashed orange horizontal line at p-value 2.3 × 10−5 is the significance threshold after multiple testing correction (Bonferroni). Summary statistics are provided as a Source Data file.

Full size image

Discussion

The diagnosis of epilepsy is an important yet challenging clinical task; thus the need for novel biomarkers remains high. Recent studies demonstrated a genetic burden in the form of an elevated PRSGGE in epilepsy cases versus controls24,25 which we replicate in our data. The effect of PRSGGE has however not been studied outside the case control setting. Here, we investigate the effect of PRSGGE longitudinally; on lifetime epilepsy, on epilepsy after an unspecified seizure event and on 1000 s of disease endpoints in other clinical areas.

In this study, we could demonstrate that common genetic variants, in the form of PRSGGE have a significant quantitative effect on GGE lifetime cumulative incidence that we could reproduce in another biobank with hazard ratios of 3-4 for the upper tails of the PRSGGE distribution in line with previous studies24 and after unspecified seizure events. Predictions are modest with C-indices of ca. 0.6, but comparable to the performance of models using clinical variables (C-indices in similar ranges of ca. 0.6 reported in the MESS trial4 or in EEG studies36,37). Thus, we expect PRSs to have potential utility as a supportive but not standalone tool. In our data, the effect of PRSNAFE on NAFE across lifetime was also significant but more modest than for PRSGGE and GGE, in line with other studies.

We were surprised to find that the effect of PRSGGE on GGE was substantially larger in females than males which was not previously reported. Previous studies reported a higher incidence of GGE in women38,39 which we also observed in our cohort. These could be caused by a different epilepsy susceptibility in males and females mediated by biological or environmental sex-specific factors. This is likely not caused by different pathomechanisms as a recent study found a high correlation of genetic effects on epilepsy in males and females13. However, we find sex-specific PRS effects predominantly for non-IGE suggesting sex-specific genetic factors may differentially influence risk for specific epilepsy subtypes. Thus, further research is needed to elucidate how genetic factors may differently influence epilepsy between sexes.

The effect of PRSGGE on GGE was quite specific with no significant effects on other diseases. However, we did not test the effects on non-disease phenotypes. Previously, high PRSGGE and high PRSNAFE were both associated with low educational attainment and neuroticism-related personality traits40 which could result from epilepsy or side effects of ASMs or may also be pleiotropic effects. Apart from NAFE, PRSNAFE had effects on other diseases including back pain, which was not previously reported; and anxiety/depression-related traits. Here, nominal significant genetic correlations of NAFE with anxiety disorders and depression medications are in line with previous reports in the UK biobank that individuals with high PRSNAFE but without a NAFE diagnosis had more likely experienced anxiety or depression40 pointing to a potential pleiotropic effect. Co-morbidities of chronic pain and depression have been previously reported41.

We see the highest potential clinical utility of epilepsy PRS in patient groups with a high absolute risk of having epilepsy such as after an unspecified seizure event. Current clinical guidelines require at least one unprovoked seizure and at least a 60% chance of a second seizure to diagnose epilepsy10. In a clinical setting, the diagnosis is often not as quantifiable as the definition suggests and is heavily dependent on clinical expertise. We find, as an example, that individuals with a PRSGGE > 2 SD have a > 3x increased risk of being diagnosed with GGE than the rest of the population. This includes individuals with unspecified seizure events who are at elevated risk for a later epilepsy diagnosis. After the exclusion of reversible causes for their unspecified seizure, a high PRSGGE could support stratifying groups at risk for a second seizure in conjunction with an EEG while other biomarkers are currently sparse3. Other recent studies suggest that PRSGGE have additional value to the information of family history42,43. Practically, genetic testing is regularly done in pediatric epilepsy and generation of PRS could thus potentially be integrated in an existing workflow. Here, integrating PRS with rare variants could also improve disease prognosis as genetic background has been shown to influence how severely carriers of genetic variants with large disease effects44 such as Dravet syndrome45 are affected. Another advantage is a high cost-effectiveness as PRS can be generated from genotype data that can also be repurposed from other disease areas23.

Our study has several limitations. We have conducted most of our analyses in cohorts with European ancestry. As has been previously described for other diseases, the predictive ability of polygenic risk scores is heavily dependent on genetic ancestry14. While the effect of PRSGGE on epilepsy showed similar trends in the primarily non-European BioMe cohort sample sizes remained prohibitive. Further studies in diverse populations are thus needed. Another limitation is that our phenotype data is derived from EHRs. We can thus not verify how many epilepsy cases have been confirmed by epileptologists. However, we obtain similar PRS effect sizes as in clinical cohorts24, which thus validates our case definitions by combining EHR diagnoses with ASM purchase and reimbursement data. The central registry of Finnish EHR data have the unique advantage that reimbursements for ASMs are always based on a certificate made by a neurologist. In addition, while we excluded individuals that were also part of the discovery GWAS we did not have the option to directly compare individual-level data between the discovery GWAS and our validation cohorts. We could thus not control for any potential relatedness between the cohorts with the potential to inflate our results46.

Our data thus proposes an interesting potential for epilepsy PRS, specifically for PRSGGE, as a biomarker for epilepsy risk where it could—combined with clinical markers such as the EEG—improve epilepsy risk prediction. Our data outlines how this could be specifically useful in situations of elevated epilepsy risk such as an unspecified seizure event. Ultimately, this needs to be investigated in a clinical setting.

Methods

This study complies with all relevant ethical regulations; the Ethics Committee of the Hospital District of Helsinki and Uusimaa approved the study protocol for FinnGen (Nr HUS/990/2017), the Estonian Committee on Bioethics and Human Research for Estonian biobank (protocol 1.1-12/624) and the Icahn School of Medicine at Mount Sinai Institutional Review Board (IRB; approval STUDY-19-00951) for BioMe.

Data and definition of epilepsy cases and controls

Here, we define epilepsy case and control status from detailed longitudinal EHR of the FinnGen project27 using data freeze R12 as a main cohort (n = 520,105) and Estonian biobank28 as an additional validation cohort (n = 210,382). We use phenotype data derived from official state registries. These include 9,313 individuals with epilepsy ICD codes, 2,485,702 ASM purchases and 12,695 ASM reimbursements of ATC codes N03A*. We list an overview of case definitions and numbers in Supplementary Table 1. 94.7% of individuals with ≥ 2 generalized seizure ICD codes and 93.7% of individuals with ≥ 2 focal seizure ICD codes purchased ≥ 2 ASMs, while only 16.4% of individuals without epilepsy diagnoses purchased ≥ 2 ASMs (see Supplementary Fig. 1). This cross-validates our EHR data.

Reimbursement rights for epilepsy are derived from the Social Insurance Institution of Finland (KELA), Finland’s national authority. All persons with newly diagnosed epilepsy are eligible for ASM reimbursement, which is also routinely applied for, necessitating a detailed statement by a neurologist and investigations at a specialist clinic. The statement is checked and approved by specialist physicians at the reimbursement institution KELA before the right is granted. Epilepsy diagnoses in Finland are made according to national guidelines, which are updated according to ILAE epilepsy definitions.

We thus chose the following criteria to define GGE:

  • at least two ICD codes of G40.3 (“Generalized idiopathic epilepsy […]”) or corresponding ICD9 codes (Supplementary Table 2) and at least two purchases of ASMs (as defined by N03* ATC codes).

We chose the following criteria to define NAFE:

  • at least two ICD codes of G40.0, G40.1, G40.2 (“Localization-related (focal)(partial) […] epilepsy […]”) or corresponding ICD9 diagnoses (Supplementary Table 2) and at least two purchases of ASMs.

  • excluded possible structural etiology of focal seizures such as stroke, brain tumor, CNS infection and CNS injury (for ICD codes see Supplementary Table 2). Here, we only excluded individuals if they had their first seizure event within one year after the brain-related potential epileptogenic event.

For 1008 individuals with both focal and generalized epilepsy codes we applied the following additional criteria for a GGE diagnosis:

  • more generalized than focal epilepsy codes AND

  • most frequent ICD code is a generalized epilepsy code AND

  • no reimbursement category of focal epilepsy.

We used the same criteria vice versa to define NAFE among individuals with focal and generalized epilepsy codes.

We defined idiopathic generalized epilepsy (IGE) according to ILAE30,47 by at least two ICD codes of 40.33 (Childhood Absence Epilepsy), 40.34 (Generalized Tonic–Clonic Seizures Alone, here using the ICD Code of the formerly known term Generalized Tonic–Clonic Seizures on Awakening) 40.35 (Juvenile Absence Epilepsy), 40.36 (Juvenile Myoclonic Epilepsy). See Supplementary Fig. 2 for age at first diagnosis.

We used individuals without epilepsy-related diagnoses as controls. We excluded individuals who purchased ASMs from the control group.

For the analysis of GGE incidence following an unspecified seizure event, we used the same diagnosis of GGE as described above. We defined an unspecified seizure event with an ICD code of R56.8/7803 A (‘unspecified convulsions’). From the group with a single unspecified seizure event we excluded individuals

  • with any other epilepsy-related diagnoses (G40/G41 ICD codes) AND

  • who purchased or reimbursed ASMs within two years before up to 10 years after event AND

  • who were at any time diagnosed with alcohol-related ICD codes OR who had multiple unspecified seizure events (to exclude potential alcohol withdrawal seizures).

When individuals had 2 seizure diagnoses on the same day we counted them as one seizure event as they most likely represent two labels of the same event. When the 2 seizure diagnoses had discordant ICD labels we labeled them according to the most specific ICD code. (As an example, individuals with diagnoses of unspecified seizure and generalized epilepsy on initial presentation would be classified as diagnosed with ‘generalized epilepsy’ on initial presentation.)

We defined epilepsy cases similarly in the validation cohort Estonian biobank, with the only exception that instead of using the reimbursement data to differentiate between NAFE and GGE in individuals who had both focal and generalized epilepsy codes, we used prescription data. Specifically, we excluded individuals as GGE cases if they had any ASM prescriptions that listed NAFE as a reason for the prescription and vice versa. We performed the unspecified seizure event analysis only in FinnGen where we had a sufficient sample size.

Importantly, we find very similar respective effect sizes of PRSGGE and PRSNAFE on GGE and NAFE as reported in previous cohorts from Epi25 or the Cleveland Clinic24 when using the same GWAS12 in a previous version of the manuscript31 or the updated GWAS13 (see Table 2). We acknowledge, that PRS effects in our cohort may not be directly comparable as we are using a different PRS calculation method (using all 835 K weighted SNPs48 instead of classic clumping and using SNPs <p-value threshold 0.524). However, differences in phenotype definitions have been reported to have larger effects than differences in PRS methods49, specifically for epilepsy13. We thus consider it likely that the epilepsy phenotypes in our biobank data are comparable to the phenotypes curated according to clinical criteria in these cohorts.

Calculation of polygenic risk scores

We calculated epilepsy PRS with the method PRS-CS48. Here, we used the summary statistics from the ILAE GWAS 202313 and ILAE GWAS 201812 (only PheWAS analyses and analyses in the BioMe cohort) as discovery data, i.e. to determine which genetic variants increase or decrease epilepsy risk. We constructed separate focal PRS (PRSNAFE) and generalized epilepsy PRS (PRSGGE). The ILAE 2023 GWAS contained the FinRisk cohort (n > 40k controls, part of FinnGen). We therefore excluded 25,405 FinRisk samples from the controls of our study to avoid overlap with the GWAS discovery cohort. The Finnish GenEpa cohort part of both GWAS was not part of FinnGen. We applied the PRS-CS-auto algorithm to infer posterior effect sizes for the variants for PRS calculation. PRS-CS-auto learns the model’s global scaling parameter ϕ from the data. We used data from the 1000 Genomes50 as a reference panel for linkage disequilibrium. We then weighted and summed all available genetic variants that confer either risk for or protection from epilepsy into a single epilepsy PRS per individual using the PLINK–score command51. The PRS-CS pipeline in FinnGen is described in more detail at https://github.com/FINNGEN/CS-PRS-pipeline. We provide the PRS weights file as Supplementary Data 1.

In FinnGen and EstBB, we restricted our analysis to individuals with European ancestry, while we additionally included African and American continental ancestry groups in the BioMe cohort (Supplementary Note). We inferred population labels based on principle component analysis of the genotype data as described previously27,28.

Statistical analyses

We used the R programming language for all statistical analyses. Pipelines for parallel computing were created using Cromwell-29 and 31 and Wdltool-0.14. Statistical analyses and figures were done using different version of R packages ggplot252, data.table, plyr, survminer, survival, tidyr and Rutils.

In all analyses including PRSs namely logistic regression, survival analyses and concordance index calculations, we included the following covariates: the first 10 principal components of genetic markers (10 PCs) as a proxy for population substructure and ancestry, genotyping batch (only in the FinnGen cohort), sex, birth year, age at last follow up. For analyses that included only individuals with seizures, we included age at the first epilepsy diagnosis as a covariate instead of age at the last follow-up. We tested PRSGGE, PRSNAFE and PRSall-epilepsy as indicated in the results of the manuscript.

All statistical tests were conducted as two-sided hypotheses without assuming a specific direction. No statistical method was used to predetermine sample size, instead the maximum number of available samples from the respective studies were used.

In the PheWas, we defined independent diseases when for any disease category not more than 40% of affected individuals are listed in any other disease category.

We performed survival analyses using the Cox Proportional-Hazards model (Cox-PH)32. Follow-up starts at birth and ends at the age of first epilepsy diagnosis (for individuals with epilepsy), age at last record available in the EHR or death, depending on what happened first. We also performed survival analyses in individuals with an unspecified seizure. Here, follow-up started at the age of the unspecified seizure and ended at the age of first epilepsy diagnosis, age at last record available in the EHR, death or after 10 years, depending on what happened first. We tested for sex differences by including an interaction term of PRS x sex in the Cox-PH model. We used the first 10 PCs, genotyping batch, sex, birth year and age at last follow up as covariates in all survival analyses. We did not exclude individuals that were related as we found in sensitivity analyses of a different project that this did not influence the PRS effect on disease53. As an additional check, we repeated our survival analysis after excluding 320,226 related individuals (corresponding to kinship values > 0.04 and > 3rd degree relatedness using the software KING54) from the FinnGen data. The effect sizes of PRS on lifetime GGE risk (method: cox model, for IGE: HR of 2.39 per SD PRS, 95%-CI 2.1–2.7, p-value 2 × 10e−34, for GGE: HR of 1.74 per SD PRS, 95%-CI 1.6–1.9, p-value 1 × 10e−53) remained almost identical. This may be expected since we found few related individuals among GGE (18 out of 924) and IGE (6 out of 226) cases. We excluded sex in sex-specific survival analyses. We included age at first unspecified seizure in survival analyses of individuals with an unspecified seizure.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.