In a retrospective single-center case–control study, data of 112 patients were included. We selected 39 patients (18 women) with a late-onset first epileptic seizure in its clinical appearances defined by the ILAE in 2017 (18) at the age of 60 years or older (LOFES group). The participants had been treated as inpatients at the Department of Neurology and Clinical Neurophysiology at the Helios University Hospital Wuppertal, Germany, between 2015 and 2020 and had neither evidence of potential epileptogenic cortical lesions nor other plausible explanations for the seizures. Patients with a history of epilepsy or psychogenic seizures, neurodegenerative diseases, amyloid angiopathy, hypoxic brain damage, severe electrolyte disorders or other reasons for acutely symptomatic seizures such as inflammatory diseases of the brain, severe hypoglycemia, substance withdrawal or poisoning were excluded by reviewing the diagnosis lists, anamnesis, laboratory findings and the MRI in the specialist discharge report. Only first-time, unprovoked seizures were considered. Among the patients, 20 (51%) experienced a focal impaired awareness seizure, while 3 (8%) were diagnosed with a focal aware seizure. Additionally, 16 patients (41%) presented with a seizure of generalized onset.

In order to compare WML load between different disease entities, two control groups were chosen. The first group comprised 38 patients (19 women) with a transient ischemic attack (TIA) with a clinically apparent cerebrovascular incident without any cortical lesions in FLAIR and T1-weighted MRI (TIA group). The TIA group included 23 (60%) patients with a cerebrovascular event in the middle cerebral artery (MCA) territory and 15 (40%) in the vertebrobasilar (VB) territory. The second group consisted of 35 patients (21 women) without any clinically apparent epileptic or cerebrovascular incident or cortical lesions (patient control (PC) group). These patients received MRI scans, due to vertigo or headache as part of routine clinical diagnostics. Basic demographic data is shown in Table 1.

WML were identified and segmented in structural FLAIR images by use of the automated lesion prediction algorithm (19) as implemented in the LST toolbox (20) based on the Statistical Parametric Mapping (SPM) software (21) and under careful visual control with manual lesion adjustment. This approach ensured that the WML were captured as comprehensively as possible. Investigators performing lesion-marking were not blinded to group membership. WML were further dichotomized into juxtacortical and (remaining) ‘distacortical’ as well as into periventricular and deep lobar WML using the distance map tool of the FMRIB Software Library (FSL) (22). To additionally examine possible alterations in cortical gray matter, structural T1-weighted MRI sequences were processed using the automated SPM based Computational Anatomy Toolbox (CAT12) (23) software to compute mean cortical thickness and to perform a cortical voxel-based morphometry (VBM) analysis.

In addition to MRI, the following patient data were collected: age at time of MRI acquisition (≥ 60 years in all cases), gender, presence of diabetes mellitus, hypertension, hypercholesterolemia/dyslipidemia, atrial fibrillation and coronary artery disease. Obesity and smoking were not listed, as it was not included in diagnostic lists as default. The study was authorized by the local ethics committee.

Structural MRI scans were acquired in 1.5 T scanners (Avanto FIT/ Aera, Siemens Erlangen, Germany) equipped with a 20-channel phased-array head coil and were extracted in Digital Imaging and Communications in Medicine (DICOM) format.

In accordance with the setting of a retrospective study, MRI protocol selection based on the referral indication and thus varied among the groups: It included a 3D ultrafast gradient echo T1 sequence (MPRAGE with TR/TE = 2.200/2.97 msec; voxel size 1.0 × 1.0 × 1.0 mm; 26 LOFES / 4 TIA / 13 PC), a 2D T1-weighted spin echo sequence (SE with TR/TE = 599.0/15.0 msec; voxel size 0.6 × 0.6 × 5.0 mm; 3 LOFES / 1 TIA / 5 PC), a 2D T1-weighted gradient echo incoherent gradient spoiled sequence (FLASH with TR/TE = 354.0/24.76 msec; voxel size 0.4 × 0.4 × 5.0 mm; 10 LOFES / 33 TIA / 17 PC) and a (3D) T2-weighted fluid attenuated turbo spin echo sequence (FLAIR with TR/TE = 9.000/97 msec; voxel size 0.9 × 0.9 × 4.0 mm; 14 LOFES / 34 TIA / 22 PC or 5.000/404 msec; voxel size 0.8 × 0.8 × 1.0 mm; 25 LOFES / 4 TIA / 13 PC).

For image preprocessing, we used the Anatomical Processing Script (fsl_anat, BETA version) provided by FSL with FLAIR images as input files. We mainly kept the default settings of preprocessing which included a reorientation of the images to the standard (MNI) orientation, a bias-field correction, a registration to standard space (linear and non-linear), a brain-extraction, a tissue-type segmentation and last a subcortical structure segmentation. The automated cropping was disabled since the FLAIR images did not cover cervical tissue.

Also, the VBM analysis required preprocessing of the T1-weighted sequences. Because voxel resolution has to be better than 5 mm in any dimension, T1-weighted images with a lower resolution obtained from T1-weighted SE or FLASH sequences had to be upsampled. This was done using a B-spline interpolation in SPM. In the following VBM analysis the default options of the standard preprocessing pipeline of CAT12 [version 1715, (23)] were applied. Described very briefly, first an affine regularization was performed based on the SPM12 tissue probability maps. Within the extended preprocessing options, the strength of corrections affecting the affine preprocessing, for example, the local adaptive segmentation or the internal resampling were adjusted by maintaining the default values. A partial segmentation into gray matter, white matter and cerebrospinal fluid was conducted in a modulated normalized space to compensate for volume changes caused by spatial normalization. As recommended, bias, noise and globally intensity corrected T1-weighted images were written in normalized space as well as partial volume effect label image volumes to ensure a quality control. One patient in the LOFES group had to be excluded due to an insufficient imaging resolution.

Statistical analyses were performed in (R Core Team, 2020). We first report basic demographic data (see Table 1). We then compare the gender and age adjusted risks for diabetes mellitus, hypertension, hypercholesterolemia−/dyslipidemia, atrial fibrillation, coronary artery disease and radiologically reported brain atrophy between the LOFES, TIA and PC groups. To do so, we fitted a generalized linear model with a binomial link function for each variable. The predictors in the model were group, gender and age. The effect of group was then assessed by running an analysis of deviance on the fitted generalized linear model. Next, we compared the three patient groups on the following measures: mean cortical thickness, tWML/TIV ratio, jWML/TIV, pWML/TIV, dWML/TIV, dpWML/TIV, jWML/tWML, pWML/tWML, dpWML/tWML and dWML/tWML ratios as well as the jWML/dWML ratio. The group means were first adjusted for the effects of gender and age using linear regression models. All dependent variables were first transformed to satisfy the normality assumption of the fitted linear regression models. We used power and logarithmic transformations. The assumptions for a linear regression were checked with the help of the gvlma package (25). The only exception was the dWML/tWML variable for which a beta regression was fitted. As in all other models, we used group, gender and standardized age as predictors. The beta regression was fit using the betareg package (24) Using the fitted models, the estimated marginal means were compared using the emmeans package (26), adjusting for the effects of gender and age and applying the Bonferroni correction for multiple comparisons. Plots were created using the ggplot2 (27) and patchwork (28) packages for R.

To investigate complex interrelationships between different variables and the risk of a late-onset first epileptic seizure, we used structural equation models, specifically path analyses models and fitted them using the lavaan R package (29). We then looked for the model which best balanced parsimony, clinical significance, statistical power and interpretability (see Supplementary materials for model specifications). Statistical model comparisons were conducted using the ‘compareFit’ function from the ‘semTools’ (30) package (for results see Supplementary Table 2). Model fit was assessed using the χ² model fit statistic, the Root Mean Square Error of Approximation (RMSEA), the comparative fit index (CFI) and the Tucker Lewis index (TLI). A non-significant χ², RMSEA <0.08, CFI > 0.90 and TLI > 0.90 were considered to denote a good fit.

To increase statistical power and to focus on the prediction of first-time seizures, we merged the TIA and patient controls into one group and constructed the variable Epileptic seizure, with 1 denoting LOFES and 0 denoting no seizure. As lavaan requires ordered factors as dependent variables, Epileptic seizure was defined as an ordinal variable. The model was fit using the unweighted least squares estimator and the nonlinear minimization subject to box constraints optimization routine. All other model fitting options were left at lavaan default settings.

Extended demographic and medical data can be assessed in the Supplementary Table 1. Besides gender and age, cardiovascular risk factors such as presence of Diabetes mellitus, hypertension, hypercholesterolemia−/dyslipidemia, atrial fibrillation and coronary artery disease is listed there. The prevalence of these medical data in the individual groups are shown in Table 2. Obesity and smoking were not systematically registered and are therefore missing. For LOFES group, also information on pathological EEG findings is provided. The analyses of deviance showed no group differences for diabetes mellitus ( χ 2 = 1.08 , d f = 2 , p = 0.58 ), hypertension ( χ 2 = 5.74 , d f = 2 , p = 0.06 ), hypercholesterolemia−/dyslipidemia ( χ 2 = 2.90 , d f = 2 , p = 0.23 ), atrial fibrillation ( χ 2 = 0.40 , d f = 2 , p = 0.82 ), coronary artery disease ( χ 2 = 1.01 , d f = 2 , p = 0.60 ) and brain atrophy ( χ 2 = 3.20 , d f = 2 , p = 0.20 ).

We first compared the groups on the tWML/TIV ratio and global cortical thickness. The LOFES patients had a higher tWML/TIV ratio than the PC (M_est = 0.0081, SE = 0.0014 vs. M_est = 0.0029, SE = 0.0005, t₍₁₀₆₎ = 3.42, p = 0.003) but not the TIA group (M_est = 0.0057, SE = 0.0010). The PC and TIA groups also differed significantly (t₍₁₀₆₎ = −2.47, p = 0.046). Regarding the global cortical thickness, we found that the marginal means of the LOFES (M_est = 2.55, SE = 0.02) and TIA (M_est = 2.61, SE = 0.02) patients as well as the LOFES and PC (M_est = 2.52, SE = 0.03) did not differ. The PC group, however, had a lower estimated global cortical thickness than the TIA group (t₍₁₀₆₎ = −0.09, p = 0.032).

Next, we investigated if the LOFES patients differed from the control groups with respect to the ratios of specific WML volumes to TIV and the total WML volume. The comparisons are visualized in Figure 2, while the following text reports the marginal means, t-statistics and p values for the comparisons. The comparison of estimated marginal means showed that LOFES patients had a higher jWML/TIV ratio than PC (M_est = 0.0009, SE = 0.00022 vs. M_est = 0.0001, SE = 0.00004, t₍₁₀₆₎ = 3.54, p = 0.002). The LOFES and TIA patients did not differ significantly (M_est = 0.0009, SE = 0.00022 vs. M_est = 0.0003, SE = 0.00010). Likewise, the PC and TIA groups did not differ significantly. The LOFES group also had a higher jWML/tWML ratio than the PC (M_est = 0.010, SE = 0.012 vs. M_est = 0.053, SE = 0.009, t₍₁₀₇₎ = 3.18, p = 0.006) and the TIA group (M_est = 0.052, SE = 0.009, t₍₁₀₇₎ = 3.31, p = 0.004). The PC and TIA groups did not differ significantly.

Further, the LOFES group had a higher dWML/TIV ratio than the PC (M_est = 0.0071, SE = 0.0012 vs. M_est = 0.0027, SE = 0.0005, t₍₁₀₆₎ = 3.34, p = 0.003) but not the TIA group (M_est = 0.0524, SE = 0.0009). There was also a significant difference between the PC and TIA groups (t₍₁₀₆₎ = −2.48, p = 0.043). Likewise, the LOFES patients had a lower dWML/tWML ratio than the PC (M_est = 0.88, SE = 0.12 vs. M_est = 0.93, SE = 0.15, z = −3.34, p = 0.002) and the TIA group (M_est = 0.93, SE = 0.15, z = −3.45, p = 0.002). The PC and TIA groups did not differ significantly.

With respect to the pWML/TIV ratio, the LOFES patients had a higher ratio than the PC (M_est = 0.007, SE = 0.001 vs. M_est = 0.002, SE = 0.001, t₍₁₀₆₎ = 3.67, p = 0.001) but not the TIA group (M_est = 0.008, SE = 0.001). There was also a significant difference between the PC and TIA groups (t₍₁₀₆₎ = −4.32, p < 0.001). We found no differences between the estimated marginal means of groups for the pWML/tWML ratio (M_LOFES = 0.81, SE_LOFES = 0.02; M_PC = 0.85, SE_PC = 0.02; M_TIA = 0.79, SE_TIA = 0.02).

The dpWML/TIV ratio of the LOFES patients was higher than of the PC (M_est = 0.0015, SE = 0.0004 vs. M_est = 0.0003, SE = 0.0001, t₍₁₀₆₎ = 3.04, p = 0.009) but not the TIA group (M_est = 0.0012, SE = 0.0003). There was also a significant difference between the PC and TIA groups (t₍₁₀₆₎ = −2.69, p = 0.025). We found no differences between the estimated marginal means of groups for the dpWML/tWML ratio (M_LOFES = 0.17, SE_LOFES = 0.02; M_PC = 0.11, SE_PC = 0.02; M_TIA = 0.18, SE_TIA = 0.02).

The LOFES patients also had a higher jWML/dWML ratio than the PC (M_est = 0.12, SE = 0.02 vs. M_est = 0.05, SE = 0.01, t₍₁₀₇₎ = 3.17, p = 0.006) and the TIA group (M_est = 0.05, SE = 0.01, t₍₁₀₇₎ = 3.21, p = 0.005). The PC and TIA groups did not differ significantly (see also Figure 3). In summary, the analyses showed that LOFES patients had the highest ratio of juxtacortical to total WML volume and the jWML/dWML ratio.

The model comparison of our path analysis models indicated that model one best fitted the data (χ² = 0.61, df = 2, p = 0.62; RMSEA = 0.03, 90% CI [0, 0.19]; CFI = 0.99; TLI = 0.98, see also Supplementary Table 3). It included epilepsy as the dependent variable, cortical thickness and jWML/dWML ratio were endogenous and age and gender the exogenous variables (see Figure 4). The model showed that, in our sample, the only significant predictor of a first time seizure was the juxtacortically shifted ratio jWML/dWML ratio (β = 0.509, SE = 0.102, z = 4.998, p < 0.001). The model also showed that higher age increased global cortical atrophy (β = −0.348, SE = 0.097, z = −3.577, p < 0.001). Model six also contained the pWML/tWML ratio as a predictor of a first time seizure. This model did not prove to be a good model for our data (χ² = 5.18, df = 3, p = 0.12; RMSEA = 0.16, 90% CI [0.08, 0.25], CFI = 0.34, TLI = 0.50). Further, the path coefficient from the pWML/tWML ratio to first seizure was not significant (β = −0.089, SE = 0.124, z = −0.719, p = 0.472).

In order to assess between-group differences in regional cortical gray matter volume, we conducted a VBM analysis. LOFES patients showed a significantly decreased gray matter volume in the right precentral and postcentral gyrus when compared to TIA patients (see Figure 5). Anatomical regions were derived from the Neuromorphometrics atlas as implemented in CAT12 by default.