Association Between Alzheimer Pathology, Sleep, and Cognition in Patients With Late-Onset Unexplained Epilepsy

Abstract

Background and Objectives

Late-onset unexplained epilepsy (LOUE) has been linked to accelerated cognitive decline. However, the role of neurodegenerative proteins and whether they exert their effects through sleep-related mechanisms remain unexplored in this population. The goal of this study was to investigate the association of plasma p-tau217, a measure of Alzheimer disease (AD) pathology, with cognition in LOUE and its association with sleep.

Methods

Participants with LOUE, with new-onset unprovoked seizures, age at onset 55 years or older, and absence of cortical lesions on MRI, were prospectively recruited. They underwent cognitive testing which included the extended Preclinical Alzheimer Cognitive Composite (PACC5). A 24-hour EEG was obtained, and sleep manually scored. Measures of sleep macroarchitecture and N2/N3 microarchitecture were extracted including slow oscillation (SO), fast spindle (FS), slow spindle (SS), and spindle-SO coupling measures.

Results

Eighty-five participants enrolled (mean age 71.3 ± 7.1 years; 49% female). The average PACC5 Z score (±SD) was −0.63 ± 1.0. Higher plasma levels of log-transformed p-tau217 (per 1 unit increase) were associated with poorer cognition, PACC5 (β = −0.66; 95% CI −1.19 to −0.13; p = 0.0017), after adjusting for age, sex, and education. The severity of epilepsy modulated this relationship: individuals with both elevated p-tau217 and medication-refractory epilepsy exhibited worse cognitive performance (refractory epilepsy * p-tau217 interaction term β = −1.49; 95% CI −2.94 to −0.05; p = 0.04). Higher plasma p-tau217 levels were associated with reduced spindle-SO coupling. Mediation analysis indicated that approximately 24% of the association between p-tau217 and cognition was accounted for by spindle-SO coupling (95% CI 3%–85%, p = 0.032).

Discussion

In individuals with LOUE, elevated plasma p-tau217 was associated with both cognitive impairment and disrupted sleep microarchitecture, with the most pronounced cognitive deficits observed in individuals with medication refractory epilepsy. These findings support the utility of plasma p-tau217 as a biomarker in this population and underscore the role of sleep processes in the relationship between neurodegenerative pathology and cognition in older adults with epilepsy.

Introduction

The average incidence of unprovoked seizures and epilepsy with onset in older adults is high with a rate of up to 2 per 1,000 persons.¹ Individuals with late-onset unexplained epilepsy (LOUE) have been found to have worse memory and executive function compared with controls,² and a course of accelerated cognitive decline.³ The exact mechanisms contributing to this cognitive dysfunction and accelerated course in individuals with LOUE remain poorly understood, although one hypothesis is that Alzheimer disease (AD) pathology is present in a subset of these individuals.⁴

Sleep plays a key role in memory consolidation and has been linked to cognitive performance in older adults.^5,6 Sleep spindles, one of the characteristic neurophysiologic features of nonrapid eye movement (REM) sleep, are generated by the thalamus and are divided into fast (around 13–16 Hz) and slow (around 11–13 Hz) based on their frequency. Different spindle frequency bands exhibit distinct regional predominance and engage separate cortical networks.⁷ Other pertinent oscillations during non-REM sleep include slow oscillations (SO) <1 Hz, which are generated by the cortex.⁵ Importantly, spindle-SO coupling has been found to facilitate synaptic plasticity, and mistimed coupling is associated with worse memory consolidation in healthy older adults.⁸

In older adults, sleep and neurodegeneration are also closely linked. The neurodegenerative proteins, amyloid beta and phosphorylated tau (p-tau), are the hallmark pathologic features of AD.⁹ Studies show that these proteins are also correlated with sleep physiology and that alterations in sleep microarchitecture may represent biomarkers of early neuronal dysfunction.¹⁰ For example, high levels of CSF tau have been linked with lower fast spindle density in N2 sleep,¹⁰ and individuals with elevated frontal amyloid beta levels based on PET have mistimed spindle-SO coupling and accelerated cognitive decline.¹¹

Previously, most studies examining the association between tau, amyloid beta, and sleep have relied on expensive PET ligands or invasive CSF studies. More recently, plasma p-tau, specifically p-tau217, has emerged as a robust indicator of AD pathology in the brain and is currently considered a biomarker of response to AD-related treatments in clinical trials and as an acceptable alternative to CSF and PET.^12,13

In this study of individuals with LOUE, we sought to determine (1) whether plasma p-tau217 correlated with cognitive performance, (2) whether it was associated with epilepsy-related variables, and (3) whether it correlated with sleep macroarchitecture and microarchitecture. We hypothesized that in individuals with LOUE, higher levels of plasma p-tau217 would be associated with worse cognitive performance, more severe epilepsy, and worse measures of spindle-SO coupling.

Methods

Study Population

This was an observational study of older adults with new onset seizures prospectively recruited between 2019 and 2024 from epilepsy and neurology clinics at Brigham and Women’s and affiliated hospitals including the main campus and community hospitals (Faulkner Hospital, South Shore Hospital). Clinics were systematically screened, and eligibility was determined based on review of clinical documentation in the electronic medical record, including review of EEG recordings and neuroimaging. All eligible participants were contacted for enrollment.

Inclusion criteria included at least 1 unexplained seizure with age at onset older than 55, and a prior MRI of the brain with contrast with no identifiable cortical lesion (i.e., cortical stroke, primary or metastatic neoplastic lesion, encephalomalacia, or vascular malformations). Cortical microhemorrhages and hippocampal sclerosis were not exclusionary. Participants with a remote history of seizures, with a nonlesional MRI, and resolution of epilepsy (10-year seizure free and 5 years of antiseizure medications)¹⁴ were eligible for enrollment if seizures recurred after the age of 55. There were no predetermined cognitive cutoffs for study entry.

Exclusion criteria included non-English speaking, provoked seizures, prior diagnosis of a neurodegenerative condition, cortical stroke, CNS tumors, a diagnosis of dementia, suspected autoimmune encephalitis, comorbid psychotic symptoms, seizures due to traumatic brain injury, and ongoing alcohol or polysubstance abuse.

Neuropsychological Assessment

All participants underwent a baseline cognitive battery, which included the Mini Mental State Examination (MMSE), Logical Memory–delayed recall (LogMemDR) task from the Weschler memory scale,¹⁵ Free and Cued Selective Reminding Task,¹⁶ category fluency (animals, fruits, and vegetables), and Digit Symbol Substitution Test (DSST).¹⁷ Performance on each test was transformed into a Z score, with the Harvard Aging Brain cohort,¹⁸ a community-based cohort of cognitively unimpaired older adults with a similar age range and no neurologic disease, used as a reference.

Scores were then extracted for the extended Preclinical Alzheimer cognitive composite (PACC5), a measure optimized to detect amyloid-beta–related cognitive decline. 19 The PACC5 consists of the average Z score of the 5 cognitive tests mentioned above. Participants also underwent the Clinical Dementia Rating scale,²⁰ which is a measure of global functioning, and the geriatric depression scale (GDS).²¹ The GDS consists of 30 items with yes/no responses; scores range from 0 to 30, with higher scores indicating greater depressive symptoms. All participants also completed the Epworth sleepiness scale (ESS) to assess for excessive daytime fatigue.²² The ESS consists of 8 questions, each scored from 0 (no chance of dozing) to 3 (high chance of dozing), resulting in a total score ranging from 0 to 24, with higher scores indicating more daytime fatigue.

Clinical and Demographic Variables

Clinical and demographic variables extracted from the electronic medical record included self-reported race and ethnicity, known diagnosis of obstructive sleep apnea (OSA), number of antiseizure medications, epilepsy duration, lateralization, lifetime-generalized tonic clonic seizures and focal seizures with impaired consciousness, and body mass index (BMI). Epilepsy localization was based on seizure semiology, and review of prior and current EEG findings. LOUE participants were considered medication refractory at enrollment if they fulfilled the International League Against Epilepsy (ILAE) criteria for medication refractoriness.²³

EEG Recordings and Sleep and Epilepsy Variables

All participants underwent prolonged EEG monitoring. Scalp EEG electrodes were placed using the International 10–20 system with additional anterior temporal electrodes (T1, T2). Twenty-four-hour ambulatory EEG recordings were acquired with XLTEK TREX hardware (Natus Medical Inc, Pleasanton, CA), sampling at 200 Hz or with Arc Apollo EEG hardware (Cadwell, Kennewick, WA), sampling at 250 Hz. If participants had an epilepsy monitoring unit admission within 6 months of enrollment, then that data were used instead of the ambulatory EEG, and the first 24 hours were used for analysis. EEGs were visually reviewed separately by 2 board-certified epileptologists (H.N.L., R.A.S.) who marked whether the study was normal and whether it had any interictal epileptiform abnormalities (IEAs), including spike waves, sharp waves, and temporal lateralized rhythmic delta activity. A consensus by the 2 reviewers was reached in case of disagreement on initial review.

Extraction of Sleep Variables From EEG

Sleep macroarchitecture and microarchitecture features were extracted using Luna.²⁴ Sleep staging of the whole recording was performed by an expert sleep technician, in non-overlapping 30-second epochs according to standards by the American Academy of Sleep Medicine, with 5 stages: wake, non-REM (NREM) stage 1 (N1), NREM stage 2 (N2), NREM stage 3 (N3), and REM.²⁵

Artifact Rejection

After restricting each recording to overnight N2/N3 epochs and re-referencing all EEG electrodes to the linked mastoids, EKG artifact was suppressed using a previously described algorithm,²⁶ and signals were then bandpass filtered between 0.3 and 50 Hz. We then performed a series of automated steps to retain artifact-free 30-second epochs. We first masked channel/epoch pairs for which that channel was more than 5 SD units from the mean for all channels for that epoch, for any of the 3 Hjorth parameters (activity, mobility, and complexity). Second, among all remaining unmasked epochs and within each channel, we iteratively flagged epochs that were 4 SD units from the mean for that channel for any Hjorth parameter, followed by a second iteration based on a 3 SD threshold. We then removed epochs with at least 1 masked channel for all channels to create a uniform data set (all channels with the same number of passing epochs).

Macroarchitecture

We then extracted the total sleep time (TST in minutes), minutes in each sleep stage, wake after sleep onset (WASO in minutes), sleep fragmentation index (SFI, defined as sleep-to-wake transition count/TST), Bout_MN (mean duration in minutes of continuous stretches of a stage, to measure consolidated periods of sleep), and conditional probabilities of transitioning from one state (wake, any non-REM state, or REM) to another.

Spindles/Slow Oscillations

Spindles were detected using a wavelet method as previously used²⁶ with specific center frequencies of 11 Hz for slow spindles (SS) and 15 Hz for fast spindles (FS) targeting approximately ± 2 Hz. Features extracted included density (count per minute), mean duration, and frequency.

SOs were identified by detecting zero-crossings in the 0.3–4 Hz bandpass-filtered EEG signals based on specific temporal criteria (zero-crossing leading to a negative peak was between 0.3 and 1.5 seconds long; a zero-crossing leading to a positive peak was not longer than 1 second) and adaptive/relative amplitude thresholds (twice the size of the mean peak-to-peak and negative peak amplitudes). For each channel, SO density, peak-to-peak amplitude, and duration were computed.

To assess SO/spindle coupling, the SO phase at the spindle peak (point of maximal wavelet coefficient) was estimated using a filter-Hilbert method: the intertrial phase clustering (magnitude) metric was used to quantify the consistency of coupling. SO/spindle overlap was also measured as the proportion of spindles overlapping with a detected SO. To account for differences in spindle and SO density, we used 1,000 randomized surrogate time series to allow for coupling magnitude and overlap metrics as Z-scores, relative to the empirical null distribution. We also calculated the density of coupled spindles to SO (per minute). The A1+A2 electrodes were used as the reference electrode for all calculations noted above.

Asymmetry Analysis

To determine whether there were sleep microarchitecture differences based on the epileptogenic hemisphere, we selected participants with a known unilateral epileptogenic focus and labeled their hemispheres as ipsilateral or contralateral to the epileptogenic focus. We then created 6 electrode sets: left frontal (Fp1+F3), left centro-temporal (F7+C3+T3+T5), left posterior (P3+O1), right frontal (Fp2+F4), right centro-temporal (F8+C4+T4+T6), and right posterior (P4+O2). We restricted the analysis to spindle, SO, and spindle-SO metrics in the ipsilateral vs contralateral homologous electrode sets.

Blood-Based Measurements

Plasma samples were drawn on the day of cognitive testing in K2EDTA tubes, centrifuged within 30 minutes of collection, and then stored in a −80°C freezer. Plasma p-tau217 was quantified using the pTau217 S-PLEX assay (Cat: K151APFS, MSD, Rockville, MD)²⁷ using a sandwich immunoassay format with monoclonal antibodies and electrochemiluminescence (ECL) detection. The lower limit of detection (LLOD) was defined as the concentration yielding a signal 2.5 standard deviation (SD) above the mean of the blank.

APOEε4 carrier status was determined by the presence of at least 1 e4 allele.

Statistics

1. Plasma p-tau217 levels were naturally log-transformed to approximate a more normal distribution. To examine the associations between log-transformed plasma p-tau217 levels, the PACC5, and other epilepsy-related variables and APOEε4 carrier status, we conducted bivariate analyses. To reduce the influence of extreme cognitive test scores, PACC5 values were capped at a Z score of −3 (n = 4). Of the 85 participants, 2 did not complete cognitive testing and were excluded from analyses involving the PACC5. Among the remaining 83 participants, missing data were minimal, with 6 individuals missing 1 cognitive test score each. Given the limited extent, these values were imputed using the mice package in R.

Associations between continuous variables were assessed using Spearman rank correlation. Differences in continuous variables across categorical groups were assessed using 2-sample t tests or analysis of variance, as appropriate. Finally, we performed multivariable linear regression analyses with PACC5 as the dependent variable, log-transformed p-tau217 as the predictor of interest, adjusting for age, sex, and years of education. Based on previous findings linking medication refractory epilepsy to cognitive impairment,²⁸ we additionally tested the interaction between log-transformed p-tau217 and refractory epilepsy status. Statistical significance was defined as p < 0.05, corrected for multiple comparisons using the Benjamini-Hochberg false discovery rate (FDR) procedure.

2. To study the association between cognition, plasma p-tau217, and sleep, we used a generalized permutation-based approach to investigate whether the variables of interest (log p-tau217 and PACC5) were individually associated with sleep macroarchitecture features and sleep microarchitecture metrics at each of the 19 electrodes. We controlled for age, sex, and years of education. Missing sleep metrics were imputed using a kNN-based approach with k = 3. Least square linear models were fit with statistical significance assessed and adjusted for multiple testing using the false discovery rate correction with a cutoff <0.05.

We performed a sensitivity analysis including the additional covariates: number of antiseizure medications, epilepsy duration in years, medically refractory epilepsy, GDS score, time between cognitive testing and EEG, ESS, and BMI. The ESS and BMI were added as surrogates for obstructive sleep apnea.

To visualize the spatial distribution of statistical effects across EEG channels for the sleep microarchitecture variables, we generated a topographic map of t-values using the MNE-Python library.

We then performed an exploratory mediation analysis using the R mediation package to examine whether spindle-slow oscillation coupling statistically accounts for the association between plasma p-tau217 and cognition (PACC5), adjusting for age, sex, and years of education. Estimates were obtained using 500 bootstrap samples. The coupling metric was selected based on its observed associations with both PACC5 and p-tau217.

3. To evaluate hemispheric asymmetries in spindle, slow oscillation, and spindle-slow oscillation coupling metrics relative to the epileptogenic hemisphere, we restricted our analysis to individuals with a confirmed unilateral epileptogenic focus. For each participant, we averaged the relevant metrics across predefined electrode sets (frontal, centro-temporal, and posterior) for the hemisphere ipsilateral and contralateral to the epileptogenic focus. We then computed the difference between homologous electrode sets (ipsilateral minus contralateral) and performed paired t tests to assess statistical significance. A threshold of p < 0.05, adjusted for multiple comparisons using the FDR correction, was applied. For visualization, we derived an asymmetry index defined as (ipsilateral − contralateral)/(ipsilateral + contralateral).²⁹

Standard Protocol Approvals, Registrations, and Patient Consents

The study was approved by the Mass General Brigham Institutional Review Board (#2019P00143), and written informed consent was obtained from all participants. We adhered to the STROBE checklist for study reporting.

Data Availability

Data that support the findings of this study are available from the corresponding author on reasonable request, subject to internal review to protect patient confidentiality, and after completion of a data-sharing agreement in accordance with Mass General Brigham institutional guidelines.

Results

Cohort Characteristics

We enrolled 85 adults with new-onset seizures after age 55 (eFigure 1). The cohort was predominantly non-Hispanic White, 88% fulfilled ILAE¹⁴ criteria for epilepsy, while the rest had 1 isolated seizure. Two participants were APOEε4 homozygous carriers. One participant had left mesial temporal sclerosis. Fifteen participants had known OSA at time of study enrollment, 13 treated with CPAP, 1 with a mouth guard, and 1 untreated.

On average, the cohort had negative Z transformed scores on the PACC5 suggesting lower cognitive performance compared with controls, as summarized in Table 1.

Table 1.

Demographic and Clinical Characteristics of the Cohort

	LOUE (n = 85)
Age at cognitive testing (years ±SD)	71.3 ± 7.1
Sex
Female	42 (49%)
Male	43 (51%)
Education (years ±SD)	15.7 ± 2.9
Ethnicity
Hispanic	3 (3%)
Non-Hispanic	82 (97%)
Race
White	84 (99%)
Black	0
Asian	1 (1%)
BMI (kg/m2) ± SD	26.6 ± 4.6
GDS ±SD	5.7 ± 4.7
ESS ±SD	5.4 ± 3.8
APOEε4 carrier	% (n = 76): 27%
CDR = 0	59 (72%)
CDR = 0.5	23 (28%)
PACC5 Z score (±SD)	−0.63 ± 1.0
Log transformed plasma p-tau217 pg/mL (±SD)	2.17 ± 0.49 (n = 80)

Abbreviations: BMI = body mass index; CDR = clinical dementia rating scale; ESS = Epworth sleepiness scale; GDS = geriatric depression scale; LOUE = late-onset unexplained epilepsy; PACC5 = extended Preclinical Alzheimer Cognitive Composite.

Plasma P-Tau217, Cognition, and Epilepsy Severity

Higher log transformed plasma p-tau (per 1-unit increase, an approximately 2.7-fold increase on the original scale) was associated with worse PACC5 scores (Figure 1A). Using multivariable analysis, findings remained significant when adjusting for age, sex, education (β = −0.66, 95% CI [−1.19 to −0.13], p = 0.0017). Plasma p-tau217 levels were higher in APOEε4 carriers compared with noncarriers (mean log p-tau217 2.40 vs 2.07; p = 0.01). PACC5 was lower in APOEε4 carriers (mean −0.96 vs −0.46), but this difference did not reach statistical significance (p = 0.10).

Figure 1. Association Between Plasma P-Tau217 and Cognition, Overall and by Medication Refractory Status.

(A) Association between log-transformed plasma p-tau217 and PACC5 (extended Preclinical Alzheimer Cognitive Composite) Z score showing that elevated p-tau217 levels are associated with worse cognition. The Spearman correlation coefficient (ρ) is displayed. (B) Association between log-transformed plasma p-tau217 and PACC-5 Z scores in individuals with and without medication refractory epilepsy. The figure illustrates that being medication refractory modulates the relationship between plasma p-tau217 and cognition, with higher p-tau217 levels associated with worse cognitive outcomes in those with medication refractory epilepsy. PACC5 = extended Preclinical Alzheimer Cognitive Composite.

On univariable analysis, plasma p-tau217 did not correlate with the frequency of IEAs per hour, the total number of lifetime-generalized tonic clonic seizures + focal seizures with impaired consciousness, epilepsy duration, or number of antiseizure medications (Table 2). IEA frequency ranged between 0 and 250, median: 0, interquartile range 0–5. The time between the cognitive testing/plasma p-tau draw and EEG was 0.4 ± 0.5 years. On average, the epilepsy duration at time of plasma p-tau draw was 1.76 ± 1.59 years. In a multivariable linear regression model including log-transformed p-tau217, medication refractory epilepsy, and their interaction, the interaction term was statistically significant (β = −1.49, 95% CI [−2.94, −0.05], p = 0.04), indicating that the negative association between log-p-tau217 and cognition was stronger among individuals who were medically refractory (Figure 1B, eTable 1). Notably, the association between log-p-tau217 and cognition was also present in the nonrefractory group (β = −0.44, 95% CI [−2.94 to −0.05], p = 0.029). In sensitivity analyses excluding participants with known obstructive sleep apnea, the interaction between log p-tau217 and medication refractory status remained statistically significant with effect sizes identical to the primary analysis (β = −1.49, 95% CI [−2.94 to −0.05], p = 0.043), indicating that known obstructive sleep apnea alone is unlikely to account for the observed association.

Table 2.

Correlation Between Plasma P-Tau217 and Epilepsy Variables

	LOUE (n = 85)	Test statistic
Epilepsy duration (years ±SD)	1.8 ± 1.6	ρ = −0.03	pFDR = 0.99
Medication refractory epilepsy	7 (9%)	t = −1.17	pFDR = 0.57
Lifetime GTC + FIC	2.0 ± 2.0	ρ = +0.12	pFDR = 0.57
EEG
IEA per hour	0.66 ± 1.80	ρ = +0.16	pFDR = 0.57
Antiseizure medications		F = 0.005	pFDR = 0.99
None	6 (7%)
Monotherapy	65 (76%)
Polytherapy	14 (16%)
Epilepsy localization a		F = 0.17	pFDR = 0.99
Left temporal	33 (39%)
Right temporal	9 (11%)
Bitemporal	11 (13%)
Others	2 (2%)
Unknown	30 (35%)

Abbreviations: F = F-statistic from analysis of variance (ANOVA) for categorical predictors; FIC = focal seizures with impaired consciousness; GTC = generalized tonic clonic seizure; IEA = interictal epileptiform abnormalities; LOUE = late-onset unexplained epilepsy; pFDR = false discovery rate–corrected p-value; t = t-statistic from 2-sample comparisons; ρ (rho) = population correlation coefficient.

^aBased on semiology, current and prior EEGs.

Sleep Disruptions and Global Cognition

We next assessed the association between sleep macroarchitecture features and cognitive function in individuals with LOUE. Three sleep disruption metrics—SFI, probability of NREM-to-NREM transition (P_NR_NR), and probability of wake-to-NREM transition (P_W_NR)—showed associations with cognition at the uncorrected level (p < 0.01) but did not survive FDR correction (pFDR = 0.0812 for all 3 variables). SFI and P_W_NR were negatively associated with cognition (β = −3.60, 95% CI −6.29 to −0.95; β = −7.13, 95% CI −12.18 to −2.09, respectively), while P_NR_NR was positively associated (β = +7.28, 95% CI +2.17 to +12.39) (eTable 2). No associations were observed between sleep macroarchitecture and p-tau217 levels (eTable 3). Specifically, N3 sleep was not associated with either PACC5 (β = −0.0013, 95% CI −0.01 to +0.008], pFDR = 0.78) or plasma p-tau217 (β = + 0.001, 95% CI −0.004 to +0.006, pFDR = 0.67).

Sleep Microarchitecture, Cognition, and P-Tau217 in Individuals With LOUE

We next examined whether sleep EEG microarchitectural features were associated with plasma p-tau217 or cognitive performance.

Associations between spindle, SO, and spindle-SO coupling metrics are shown in Figure 2. About spindle features, the fast spindles had the most significant associations with the PACC5, with higher density, longer duration, and slower frequency in frontotemporal regions associated with better performance. No association was found between p-tau217 and spindle-specific measures. Findings with PACC5 remained significant after the sensitivity analysis adjusting for number of antiseizure medications, epilepsy duration in years, medically refractory status, GDS score, and time between cognitive testing and EEG, BMI, and ESS (eTable 4): density (fast spindles: Fp1 p_FDR = 0.02, Fp2 p_FDR = 0.04, F8 p_FDR = 0.03), duration (slow spindles: Pz p_FDR = 0.049, T5 p_FDR = 0.04, O1 p_FDR = 0.02), and frequency (fast spindles: Fp1 p_FDR = 0.04). Lower slow oscillation density was associated with higher p-tau217 levels and worse cognition, although these associations did not reach statistical significance (Figure 2).

Figure 2. Summary of Spindle-Slow Oscillation Coupling Metrics Derived From Luna Across the 19-Electrode Montage, Controlling for Age, Sex, and Years of Education.

Density refers to the number of spindles coupled with SOs; magnitude reflects the z-transformed strength of phase coupling; overlap represents the z-score of the probability of spindle-SO co-occurrence. Topographic plots display T-statistic heat maps, with red indicating positive associations and blue indicating negative associations. Black dots denote electrodes with statistically significant associations after FDR correction. Higher FS-SO coupling metrics in frontotemporal regions and SS-SO metrics in posterior regions were positively associated with better cognitive performance. By contrast, plasma p-tau217 levels were negatively associated with both FS-SO and SS-SO overlap. FS = fast spindles; PACC5 = extended Preclinical Alzheimer Cognitive Composite; SS = slow spindles.

Finally, when examining spindle-SO coupling measures, higher fast spindle-coupled density (Fp1, Fp2, Fz, F4, F7, F8, T4, T6, Pz), magnitude of coupling (Fp1, F8, T6), and higher slow spindle-overlap (C3, T3, P3, T5, O1, Pz, C4, P4, T6, O2), Z scores were associated with better cognitive performance. Findings remained significant when adjusting for the previously listed covariates (eTable 5). Overall, higher p-tau217 was associated with worse spindle-SO coupling metrics, with significant associations noted for SS-SO overlap (Fp1, C3, T3, T6) and FS-SO overlap (T6) (Figure 2). Sensitivity analysis showed that SS-SO overlap (T6 p_FDR = 0.03) and FS-SO overlap (T6 p_FDR = 0.04) remained significant (eTable 5).

Exploratory mediation analysis suggested that spindle-slow oscillation (SS-SO; T6) coupling partially accounted for the association between plasma p-tau217 levels and cognition (PACC5), adjusting for age, sex, and education. The estimated indirect association was significant (β = −0.16, 95% CI −0.46 to −0.02; p = 0.032), as was the direct association (β = −0.51, 95% CI −1.01 to −0.03; p = 0.024). The total association between p-tau217 and cognition was −0.67 (95% CI –1.21 to −0.23; p < 0.001), with approximately 24% of this association accounted for by spindle-slow oscillation coupling (95% CI 0.03 to 0.85; p = 0.032) (Figure 3).

Figure 3. Exploratory Cross-Sectional Model Showing Associations Among Plasma P-Tau217, Spindle-Slow Oscillation (T6) Coupling, and Cognition (PACC5), Adjusted for Age, Sex, and Education.

Higher log-transformed p-tau217 was associated with reduced SS-SO coupling (β = −0.95, 95% CI −1.49 to −0.40; p = 0.0009), and greater SS-SO coupling was associated with better cognition (β = 0.17, 95% CI 0.002–0.33; p = 0.047). The total association between p-tau217 and cognition was β = −0.67 (95% CI–1.21 to −0.23; p < 0.001). Inclusion of SS-SO coupling attenuated this association (direct association β = −0.51, 95% CI −1.01 to −0.03; p = 0.024), with approximately 24% of the association accounted for by SS-SO coupling (95% CI 0.03–0.85; p = 0.032). PACC5 = extended Preclinical Alzheimer Cognitive Composite; SO = slow oscillation; SS = slow spindles.

Asymmetry Based on the Lateralization of the Epileptogenic Focus

No differences were noted in spindle and spindle-SO coupling metrics between the hemisphere with epileptogenic focus vs the contralateral hemisphere in the 39 participants with a unilateral focus (33 left, 8 right). In the centrotemporal regions, slow oscillation frequency was higher in the ipsilateral hemisphere (pFDR = 0.03), as indicated by a positive asymmetry index (eFigure 2).

Discussion

In this study of individuals with late-onset unexplained epilepsy, we found that elevated plasma p-tau217 levels were associated with poorer cognitive performance, particularly among those with medication refractory epilepsy. We also showed that this relationship was partially accounted for by disrupted spindle-slow oscillation coupling, late-onset epilepsy is associated with an increased risk of dementia and can represent one of the first manifestations of AD.³⁰ Previous CSF studies in late-onset epilepsy have suggested blood barrier dysfunction and elevated AD biomarkers in this population compared with controls without epilepsy as measured by lower CSF Aβ42,³¹ higher p-Tau181/Aβ₄₂ ratio, and higher CSF albumin/serum ratio.³² Other groups have also shown a reduction in plasma Aβ42/Aβ40 over time, which may reflect faster amyloid accumulation.³³ Our study is one of the first to explore plasma p-tau217, another AD biomarker, in this patient population. There is accumulating evidence that plasma p-tau217 is associated with cognition and cognitive trajectories. Cross-sectional studies of plasma p-tau217 have shown an association with measures of global cognition, such as the MMSE, and higher levels in those with mild cognitive impairment vs cognitively unimpaired individuals.^34,35 Higher baseline plasma p-tau217 levels have also been associated with higher amyloid and tau PET positivity and memory decline in adults with autosomal dominant AD.³⁶ It was also shown to be a reliable predictor of cognitive decline in cognitively unimpaired individuals using the modified PACC and the MMSE.³⁷ In this study, we extend the utility of this biomarker to individuals with LOUE and show that elevated cross-sectional levels correlate with worse cognitive performance. The mean PACC5 Z score for the cohort was −0.63 ± 1.0, which falls within a range often considered consistent with mild cognitive impairment at the group level. However, there was variability across individuals, with some participants demonstrating more pronounced cognitive deficits on neuropsychological testing. It is important that there were no prespecified cognitive cutoffs for study entry because the cohort was designed to capture the full clinical spectrum of LoUE. Although some individuals exhibited scores in the impaired range, none demonstrated the degree of functional decline required to meet diagnostic criteria for dementia at baseline. Studies have demonstrated a clear relationship between cortical excitability and neurodegeneration,³⁸ with elevated levels of AD-related biomarkers in CSF in AD patients with vs without epilepsy.³⁹ Animal models have also shown a feed-forward cycle of amyloid and tau causing hyperexcitability, which in turn increases their levels.⁴⁰ However, we did not find an association of p-tau217 with epilepsy-related variables, such as epilepsy duration, number of antiseizure medications, and frequency of interictal abnormalities. This may in part reflect this cohort was recruited soon after seizure onset (average 1.6 years), was drug responsive (91%), had a low lifetime burden of seizures (2 on average), and had a low burden of interictal abnormalities on surface EEG. Nonetheless, medication refractoriness appeared to modulate the effect of plasma p-tau217, with individuals who were medically refractory and had elevated p-tau217 levels showing the most pronounced cognitive impairment. Future studies are needed to confirm these findings, particularly given the small number of individuals with medication-refractory epilepsy in our cohort, which increases the risk of model overfitting, especially in models including interaction terms and multiple covariates. These findings should therefore be interpreted with caution. If replicated, they may carry important clinical implications. Specifically, it suggests that individuals with both medically refractory epilepsy and elevated p-tau217 levels might benefit from more aggressive treatment strategies.

Sleep is a modifiable risk factor for poor brain health. Disorders of sleep have been linked to increased risk of cerebrovascular disease and dementia through mechanisms involving increased oxidative stress, neuroinflammation, and diminished glymphatic clearance.^41,42 Measures of disrupted sleep and sleep fragmentation are associated with a significant decline in MMSE performance at 1 year⁴³ and a 1.22 hazard ratio (95% CI 1.03–1.44) of developing AD dementia.⁴⁴ Sleep fragmentation was found to be a mediator of the association between amyloid beta deposition and cognition.⁴⁵ Our results align with this because we show that measures of sleep fragmentation specifically the sleep fragmentation index and the probability of remaining in NREM were associated with poorer memory performance in individuals with LOUE, but these associations did not survive correction for multiple comparisons. However, we did not observe an association between plasma p-tau217 levels and sleep macroarchitecture measures.

Sleep also plays a central role in optimizing cognitive performance. Both synaptic strengthening and synaptic downscaling are believed to occur during sleep to facilitate the strengthening of already acquired information and to enable the acquisition of new information the next day.⁵ Several aspects of NREM sleep neurophysiology have been linked with cognitive outcomes, including spindles, slow oscillations, and sleep-SO coupling. How spindles and SO temporally couple is crucial for the cross-talk between cortical and thalamic regions during sleep.^6,8 Processes affecting both the cortical regions generating the slow oscillations, and subcortical structures generating the spindles including the thalamus will impair timely coupling of SO and spindles. Thus, measures of spindle-SO coupling reflect the integrity of broad cortical and subcortical networks.

With aging, there is a change in sleep microarchitecture, which has been linked to cognitive impairment and accelerated cognitive decline.^6,11 Specifically, there is a reduction in SO,⁴⁶ a decrease in spindles^6,8 especially in parietal regions, and diminished spindle-SO coupling.^6,8 Large studies in older adults have shown how this correlates with worse performance on cross-sectional measures of executive function, processing speed, and global cognition.⁶ Similar patterns were observed across spindle measures and spindle-slow oscillation coupling metrics, which were associated with global cognition in individuals with LOUE. Notably, fast spindle features including increased density, longer duration, and lower frequency in frontotemporal regions were associated with better cognitive performance. In addition, coupling metrics for both slow and fast spindles, such as increased coupled density, higher proportion of overlap, and greater coupling magnitude, also showed positive associations with cognition. These relationships remained significant after adjusting for potential confounders, including depression, number of medications, BMI, and Epworth Sleepiness Scale scores.

Studies using PET or CSF biomarkers to explore the association between neurodegenerative proteins have shown a link between spindle-SO coupling and amyloid and tau. Amyloid PET studies show that higher frontal amyloid beta is negatively associated with SO, spindle-SO coupling, and with lower memory consolidation and longitudinal cognitive decline.^11,47 CSF tau measures and PET measures of medial temporal tau also show a negative association with spindle-SO coupling.^10,48 In our study, plasma p-tau217, which correlates well with amyloid PET positivity, also showed this negative association with spindle-SO coupling, but was not a strong predictor of direct spindle or SO disruption. In addition, spindle-slow oscillation coupling was associated with both plasma p-tau217 and cognition and partially accounted for their cross-sectional association, raising the possibility that sleep microarchitecture and neurodegenerative pathology represent interconnected processes related to cognitive performance. Although plasma p-tau217 likely reflects chronic accumulation of AD pathology over years to decades, the cross-sectional design of this study precludes conclusions regarding temporal ordering or causality.

The association between sleep and p-tau217 is likely bidirectional due to synaptic dysfunction,^49,50 and early deposition of neurodegenerative proteins in subcortical sleep-wake structures.⁴⁵ In turn, impaired clearance mechanisms in sleep are associated with elevated neurodegenerative proteins.⁵¹ Although sleep represents a potentially modifiable risk factor, our findings should be interpreted as associative. As plasma biomarkers become more clinically available, these biomarkers will help identify individuals at risk of accelerated cognitive decline and alert physicians to address risk factors including sleep. Evaluation should include screening for common sleep disorders such as obstructive sleep apnea and periodic limb movements of sleep, both of which can disrupt sleep architecture and have been associated with adverse cognitive and biomarker profiles. Management strategies include optimization of sleep hygiene, implementation of behavioral interventions, and treatment of comorbid conditions such as depression. Interventional evidence linking improvements in sleep to changes in neurodegenerative biomarkers remains limited. Emerging evidence suggests that orexin receptor antagonists may influence amyloid and tau dynamics and could represent new therapeutic avenues.⁵²

Finally, we did not observe prominent asymmetries in sleep architecture among individuals with a unilateral epileptogenic focus, apart from a higher slow wave frequency in the centrotemporal region ipsilateral to the focus. Studies examining the relationship between focal epilepsy and sleep spindles have largely focused on young adults, where findings suggest lower spindle rates in the epileptogenic hemisphere.⁵³ This discrepancy may reflect differences in the underlying pathologies and typical sleep microarchitecture between age groups, as well as more prominent neurodevelopmental effects in the younger cohort.

Our study has several strengths, including being one of the largest and most well-phenotyped cohorts of individuals with LOUE. It is also among the first to explore the association between plasma p-tau217 levels and both cognition and sleep in individuals with epilepsy.

This study has several limitations: (1) the cohort mostly consisted of non-Hispanic White individuals with a high level of education, which affects the generalizability of the findings. (2) We did not have oximetry data to control for sleep disordered breathing. (3) The absence of an EMG electrode likely caused an underestimation of REM sleep. (4) The plasma measurements were not performed on the same day as the EEG; therefore, we controlled for the interval between the 2 in our models. (5) The cross-sectional design precludes conclusions regarding temporal ordering or causality among plasma p-tau217, sleep microarchitecture, and cognition.

In conclusion, in individuals with late-onset unexplained seizures, plasma p-tau217 levels correlate with worse cognition especially in those with medically refractory epilepsy and weaker sleep spindle-slow oscillation coupling. These findings suggest the utility of plasma p-tau217 as a biomarker in this patient population and support an association between neurodegenerative proteins, sleep integrity, and cognition in older adults with epilepsy.

Glossary

AD

Alzheimer disease

BMI

body mass index

ESS

Epworth sleepiness scale

FS

fast spindle

GDS

geriatric depression scale

IEAs

interictal epileptiform abnormalities

ILAE

International League Against Epilepsy

LOUE

late-onset unexplained epilepsy

MMSE

Mini Mental State Examination

OSA

obstructive sleep apnea

P_NR_NR

probability of NREM-to-NREM transition

P_W_NR

probability of wake-to-NREM transition

PACC5

extended Preclinical Alzheimer Cognitive Composite

p-tau

phosphorylated tau

SFI

sleep fragmentation index

SO

slow oscillation

SS

slow spindle

TST

total sleep time