Seizure Diaries Don't Work Very Well. Here's What the Evidence Says.

Seizure diaries are the standard outcome measure in epilepsy research. They're also, by the evidence, fairly unreliable — and the problem has been documented for almost twenty years.

The foundational study is Hoppe, Poepel & Elger (2007) in Archives of Neurology. They recorded 91 inpatients on video-EEG and compared what the monitoring unit captured against what the patients wrote in their diaries. Roughly half of seizures went undocumented by patients, including the majority of nocturnal events and most focal impaired-awareness seizures. A more recent review by Elger and Hoppe (2024) reports that this hasn't changed much — undercounts between 40% and 60% are still typical, and for some seizure types (nocturnal, focal aware) it's substantially worse.

If you've worked on an epilepsy trial, none of this is new. But it's worth sitting with what it actually means for the studies we design around seizure frequency.

It isn't a patient problem

One reaction to this literature is to blame the patients — they forget, they don't care enough, they fill in the diary at the last minute. That framing misses the mechanism.

A person having a focal seizure often doesn't know it's happening. Post-ictally, they can't recall it. Nocturnal seizures happen during sleep. Even for generalized tonic-clonic events with full recovery, autobiographical memory for medical events decays fast and reconstructs itself at diary-completion time rather than event time. Patients aren't unreliable reporters because they're bad patients. They're unreliable reporters because you're asking them to report events their brain either never encoded or actively suppressed.

The fix is not "remind patients harder." It's to stop relying exclusively on post-hoc self-report for an endpoint the brain is structurally bad at recalling.

What the undercount does to your endpoint

Here's the part that matters for protocol design.

If your primary endpoint is ≥50% reduction in seizure frequency — which it probably is — and baseline diary accuracy is around 50%, then your baseline count is already a coin flip away from reality. So is your follow-up count. You're computing a ratio of two noisy estimates and treating a threshold cross as a treatment effect.

This doesn't mean every historical epilepsy trial result is wrong. Large effects show up even through noisy instruments. It does mean your study is much less sensitive to smaller, real effects than your power calculation suggests — and that responder-rate analyses, which are the most common secondary endpoints, are particularly vulnerable. A 5-point difference in responder rate between arms is easy to explain away when the measurement instrument itself is wrong half the time.

There's also a subtler problem, documented in Fisher et al. (2016): patients who enter a trial are often hyperaware at baseline and report more faithfully than they did before. As the study wears on, that vigilance fades and documentation drops off. That looks, in the data, like a treatment effect.

What actually changes with real-time capture

The case for logging seizures as they happen — rather than reconstructing them on Sunday night — isn't a technology argument. It's about what kind of data you get.

When a patient or caregiver logs an event in the moment, you get a timestamp, not a day-of-week. That timestamp lets you look at things a weekly count can't: medication-seizure interval, sleep duration on the preceding night, clustering patterns, circadian distribution. None of these are possible from a weekly number, and all of them are interesting for secondary analyses and label expansion.

For nocturnal events specifically, smartphone-based accelerometer data can flag probable events overnight that the patient and caregiver can then review and confirm in the morning. This doesn't match the gold standard of continuous EEG, but it's substantially better than a morning-after memory check, and it addresses the part of the undercount that's caused by the patient being asleep.

A 2024 study in Scientific Reports showed that patients using an app diary reported seizures with much higher precision and temporal resolution than paper diaries — not perfect agreement with objective recording, but a real improvement and a richer dataset for downstream analysis.

Burden is the wrong objection

The common pushback on event-driven capture is that it's more burdensome than a weekly diary.

In our experience building Forma, and consistently with the EMA literature going back to Shiffman, Stone & Hufford (2008), the opposite is true. A 60-second post-event prompt feels lighter than a 10-minute weekly reconstruction. The cognitive work is distributed to the moment when it's cheapest — right after the event, when memory is fresh and patients are often motivated to record what just happened. Adherence holds up because you're asking for small amounts of effort at the right moments, not a single large obligation at an arbitrary one.

The design principle worth stating explicitly: the post-event log should capture the three or four fields that matter most (type, duration, rescue medication, post-ictal state) in a handful of taps. Save the longer characterization for a follow-up an hour later, when the patient has recovered and can answer more carefully. If your in-moment form takes more than 90 seconds, you have a design problem, not a burden problem.

If you're designing a protocol

A few things we've found useful when working with teams on epilepsy studies:

Keep validated instruments, add real-time capture. QOLIE-31 and NDDI-E have established psychometric properties and regulatory acceptance. Don't replace them. Run them on their intended schedule and use real-time logging as the primary efficacy capture layer.

Design for the caregiver. For pediatric epilepsy, post-ictal patients, and anyone with significant focal-onset events, the caregiver is often the better reporter. Build this into the consent and data structure from day one rather than bolting it on later.

Pre-specify how timestamps map to your endpoint. Real-time logs produce a richer data structure than a weekly count. Write the analytical crosswalk at the protocol stage so your statistical reviewers aren't surprised.

Don't over-promise on the undercount. Real-time capture improves on paper diaries, but it doesn't reach video-EEG fidelity. State this plainly in your protocol. Reviewers respect clarity about instrument limits more than they respect a claim of completeness that the literature contradicts.

The accuracy problem with seizure diaries has been documented since Hoppe 2007 and it hasn't gone away. The tools to materially reduce it are available. The friction is in adopting them, not in building them.