The Secrets Inside Self-Reported Seizure Diaries

Using a seizure diary helps individuals with epilepsy and their neurologists track and record seizure-related details. Self-reported seizure diaries represent the primary means by which people with epilepsy evaluate their degree of illness. Diaries offer a unique perspective on diverse features of seizures and epilepsy which will be elaborated below. They also assist in evaluating the effectiveness of antiseizure medications (ASMs). In randomized clinical trials (RCTs), seizure diaries have been helpful in assessing response to both the medication under investigation and placebo. This article summarizes several unexpected and revealing findings that have been uncovered in seizure diaries (Table).

DIARY PATTERNS

Electronic seizure diaries capture the time of occurrence and features of seizures. There are often common findings across diaries. Patient diaries are the clinical standard for the basis of ASM selection and clinical management. However, self-reported diaries have considerable limitations. Seizures are believed to be underreported in ~50% of diaries,¹ although estimates range from 0% to 100% depending on the individual, the recording modality, and the study.

Several popular e-diaries exhibit differing user interface features.² Common data elements define the minimum information required to ensure interoperability.³ Statistical models of individual diaries have revealed more variability than expected from simple random models.^4-7 Across a large cohort (n=10,186), there was an uneven distribution of seizure frequencies⁸ (Figure 1). This skewed distribution may explain the higher-than-expected seizure rates reported in RCTs.⁶ Patterns in self-reported infantile spasms have been reproduced in video-EEG data,⁹ reinforcing that subjective self-reported diary reports align with objective data. Similarly, reported timing patterns in self-reported seizure clusters^8-10 have been validated against EEG data.^9,11

Figure 1. Simulated seizure rate properties. Population histogram of average monthly seizure rates from 50,000 virtual participants (A). The median seizure rate of the entire population and that of a subpopulation with >4 seizures/month are highlighted. A previous study8 found the population median to be 2.7 seizures/month for adults and 3.5 seizures/month for children.
The L relationship is also shown (B). Each dot represents 1 of the 50,000 simulated participants. The position of each dot represents the logarithm of the mean and the logarithm of the standard deviation of the monthly seizure rate of that individual. A regression line is plotted through the entire population of points, with a y-intercept of 0. The slope is shown. A previous study¹⁷ found this slope to be 7.3 to 8.3.
Reproduced with permission from Goldenholz DM, Westover MB. Flexible realistic simulation of seizure occurrence recapitulating statistical properties of seizure diaries. Epilepsia. 2023;64(2):396-405.⁶

Other findings have been reported related to cyclical variations. There was no clear evidence of shifts in seizure rates due to daylight savings time.¹² However, day/night cycles,¹³ as well as weekly cycles,⁸ were present in e-diaries. A study using implanted responsive neurostimulation electrodes found multiple coexisting seizure risk cycles, with circadian and multiday cycles in people with drug-resistant focal epilepsy (n=37).¹⁴ Self-reported diaries have revealed analogous risk cycles in many individuals (n=1118: circadian, 82%; weekly, 21%; longer than 3 weeks, 22%).¹⁵ A similar pattern has been validated using long-term intracranial EEG.¹⁵

E-diaries have also revealed a predictable summary statistics feature.^16,17 The average and standard deviation of seizure counts have shown tight coupling, called the “L relationship” (Figure 1), which has been observed in multiple data sets (eg, intracranial recordings, longitudinal study data, wearable device data).^7,17

Building on these observations, an open-source realistic e-diary simulator has been developed,⁶ which accounts for the uneven distribution of seizure rates across individuals;⁸ statistical features,⁴ including the L relationship;¹⁷ multiple coexisting seizure risk cycles;^14,15 seizure clusters;^10,18 and maximum allowable seizure rates.^8,19,20 By including all these elements, the simulator can generate an unlimited set of realistic patient diaries (Figures 1 and 2).

CLINICAL CARE

Several aspects of clinical care with ASMs have been affected by e-diary research. A method was described for estimating the severity of a seizure disorder using an e-diary calculator, epiSAT (Epilepsy Seizure Assessment Tool),^21,22 which was found to have greater accuracy than expert clinicians (87.5% vs 74.7%). This tool may be valuable for making ASM adjustments.

Figure 2. Randomized clinical trial (RCT) simulations. The results of simulating 10,000 RCTs comprising 200 participants each (1:1 for 0 efficacy placebo vs 30% efficacy drug), including the simulated 50% responder rates (RR50) and median percentage change (MPC) values, are shown. These are compared with corresponding values from 23 historical RCTs, also summarized here. The results show that simulated RCTs without a simulated placebo effect can reproduce historical RCT outcomes. This suggests that placebo effects may be less important for understanding the placebo response in epilepsy RCTs.
Reproduced with permission from Goldenholz DM, Westover MB. Flexible realistic simulation of seizure occurrence recapitulating statistical properties of seizure diaries. Epilepsia. 2023;64(2):396-405.⁶

Seizure underreporting of varying degrees has been documented in patient-reported diaries, ranging from 0% to 100%.¹ To better understand the effects of underreporting,¹ ASM adjustments in clinic were simulated with 100,000 virtual participants over 10 years. Analysis showed that most individuals receive reasonable clinical care, provided they report ≥10% of their seizures.²³

Because seizures appear to cluster in different ways for different individuals per e-diaries, a personalized seizure cluster tool—ClusterCalc (EpilepsyAI, San Francisco, CA)—was developed.¹⁰ Using this approach, an estimated 38% to 60% of clusters defined by traditional criteria were consistent with chance rather than true clustering.¹⁰ The tool also identified longer-duration clusters that fall outside classical definitions.¹⁸ Individualized cluster identification may help guide treatment strategies for individuals with recurrent clusters.

A 12-month history of frequent generalized tonic-clonic seizures is a known marker of sudden unexpected death in epilepsy (SUDEP) risk.²⁴ Using >11 years of diaries from 12,402 participants, 12-month histories of recurrent seizures fluctuated over time; therefore, the SUDEP risk fluctuated as well.²⁵ Among participants initially classified as high risk, 59% had decreased SUDEP risk over time, whereas 23% of those initially classified as low risk became high risk over time, suggesting that “high-risk” and “low-risk” SUDEP designations are time-dependent.

A realistic simulation showed that natural fluctuations in seizure frequency may lead clinicians to erroneously discontinue an effective ASM.²⁶ A potentially effective ASM given to 100,000 virtual participants caused apparent worsening of the seizure rate in 12%. Of those, 76% were on an effective ASM at the time of the observed “failure.” The same situation likely arises in clinical practice, resulting in valuable, effective drugs being discontinued erroneously due to perceived inefficacy.

PLACEBO RESPONSE

Placebo-controlled RCTs are designed to separate true treatment effects from changes unrelated to the treatment under study. In epilepsy, a substantial part of the “placebo response” appears to come from natural ups and downs in seizure frequency and regression to the mean rather than a psychological response to receiving an intervention. When simulations built trial-like timelines from large sets of individual e-diaries outside of actual trials, many individuals appeared to “respond” even though no treatment was given.²⁷ The same apparent responses showed up when e-diaries were analyzed backwards in time.²⁸ This forward/backward symmetry supports a statistical explanation for improvement rather than a causal placebo mechanism. These findings were replicated across self-reported diaries, RCT diary data, and intracranial EEG-derived seizure records.²⁸ Simulations^5,6 reproduced placebo responses seen in historical epilepsy RCTs without adding any extra placebo effect (Figure 2). Simulation also showed that trial design can reduce regression to the mean effects,²⁹ potentially shortening studies, lowering costs, and meaningfully reducing required sample sizes.

Collectively, these findings suggest a strong hypothesis: that placebo response in epilepsy may not result from a placebo effect mechanism at all.^30-33 Under this lack-of-placebo-effect assumption, simulations showed that key parameters of RCTs can substantially change placebo response, required sample size, and overall cost.^34,35

RCT ANALYSIS

RCT performance has also been evaluated using e-diaries. Selecting different RCT outcome measures can result in better analytics.^16,29,36-38 Some measures (eg, median percent change) resulted in lower needed sample sizes, much lower costs, and lower SUDEP risk, whereas others (eg, 50% responder rate) led to increased sample sizes, costs, and risk.³⁷ This highlights that regulator-endorsed measures could have dramatic financial implications for ASM research.

Several RCT design improvement approaches have been explored, including strategies to reduce placebo response,²⁷ optimize RCT parameters,³⁵ use intracranial EEG,^29,34 decrease regression to the mean,²⁹ and measure effects with wearables.³⁹ Even moderate-sensitivity wearables (≥80%) are predicted to be beneficial for RCTs due to the longer timescales involved.³⁹

Machine learning has also enabled the generation of exploratory hypotheses. One study found that seizure rate fluctuations can be harnessed to predict RCT outcomes before the study begins.⁴⁰ In theory, if all participant data were available on the first day of the RCT, it would be possible to predict a successful or unsuccessful outcome without running the study at all.

A simulation showed that RCT outcomes, including treatment response and treatment-related symptoms, could be obtained using large-language models to review hundreds of clinical encounters.⁴¹

Evidence of underreporting bias in RCT publications has been identified regarding the fraction of individuals with worsening seizure frequency.⁴² Among 16 ASM studies published between 2000 and 2024, only 23% of peer-reviewed publications included graphs showing worsening, compared with 63% of Food and Drug Administration labels.

FORECASTING SEIZURES

One important question is whether e-diaries can be harnessed to predict future seizure timing. Long-term seizure freedom after epilepsy surgery has been modeled using presurgical measures, including MRI, routine EEG, ictal video-EEG, and intracranial EEG. In longitudinal data from 118 participants, a combined multivariable model did not achieve sufficient accuracy to be clinically useful.⁴³ Postoperative EEG was also evaluated as a predictor of longer-term seizure freedom using diaries from 83 participants, but no robust predictive signal was identified.⁴⁴ Collectively, these results suggest that meaningful surgical seizure-free outcome prediction likely requires very large data sets.⁴⁵

Short-term seizure risk has been examined using a review of the available evidence. Studies based on seizure diaries have proposed potential triggers (eg, medication nonadherence, exercise, hormonal cycling), but substantial gaps in study design, measurement, and replication limit confidence in the magnitude (or even presence) of these effects.⁴⁶ A separate systematic review⁴⁷ assessed objective evidence for the widely taught association between sleep deprivation and increased short-term seizure risk. Only 5 studies were identified that directly evaluated causality. Of those, only 1 RCT incorporated objective EEG measures, and did not detect an effect of sleep deprivation. Overall, if there is a causal link between sleep deprivation and seizure risk, it has yet to be proven.

Based on promising studies of cyclical risk,^8,13-15 e-diary–only short-term seizure forecasting was evaluated as a potential clinical tool. An initial 24-hour forecasting model using deep learning on retrospective self-reported diaries suggested early promise, achieving what appeared to be good performance.⁴⁸ However, subsequent simulations⁶ demonstrated that clustering and seizure cycles can strongly influence predictive performance. This suggests that whereas these 2 features may contribute to prediction, their prediction accuracy alone appears insufficient for clinical use.

To address these issues, stringent benchmarks for seizure forecasting evaluation have been developed using simulations, diary data, and wearable-derived data.⁴⁹ The “napkin method” (so named because it is simple enough to be calculated on a piece of paper [ie, napkin] to provide a quick, intuitive risk estimate) is a simple calculation that can be performed at the bedside to identify seizure risk. Calculating the number of seizure days in the past 90 days and dividing by 90 results in a forecasted 24-hour risk score. For example, an individual with a seizure frequency of 1 every 10 days would have 9 seizures in 90 days, and their 24-hour risk would be computed as 9/90 (ie, 10%). As previously stated, a deep-learning forecasting tool looked promising retrospectively.⁴⁸ However, in a follow-up prospective forecasting study (using the appropriate number of individuals⁵⁰), the tool failed to outperform the napkin method.⁵¹ This suggests inadequate added value from the deep-learning model.

To test whether this limitation was model-specific or more general, 5 additional forecasting methods were evaluated using 5501 retrospective and 36 prospective e-diaries; none consistently outperformed the napkin benchmark.⁵² Notably, several methods appeared favorable when compared with a chance-level comparator, underscoring that weak comparators can artificially inflate perceived performance. A review of other forecasting studies⁵³ found that almost all relied on chance comparators rather than strong, clinically meaningful benchmarks.

CONCLUSIONS

Self-reported seizure diaries contain substantial information beyond simple event counts. There are practical lessons that can be learned from analyses of self-reported seizure diaries. Tools have been developed to support clinician interpretation of diaries, simulation, trial design, and forecasting evaluations. These have implications for outpatient care, clinical trials, and seizure risk forecasting. Key examples include approaches to safer ASM management in the setting of underreporting; the central role of natural fluctuations in seizure frequency; and end point choices that enable faster, lower-cost, and potentially safer RCTs. In addition, strong benchmarks are needed when evaluating seizure forecasting models. As diary data sets expand and methods mature, additional clinically relevant insights are likely to emerge from patient-reported seizure diaries.