Missing Not at Random: Why EEG Data Are So Precious

Charlie has an inpatient EEG next week; “three to five days,” starting Wednesday. I put out the "Bat Signal" to the other epilepsy families on Facebook.

Epilepsy families:

Charlie is doing an inpatient EEG next week. He is like a toddler in a grown man's body. For the 20 minute EEG they had to put him in the wrap thing the whole time because the minute they let one arm out those leads were history.

What is your best advice for big kids who cannot be easily distracted? I tried to explain the situation to the EEG people. Bless them, they are sending in Child Life but so far the only options they recommend are toys and "distractions" but that's not how Charlie rolls. He doesn't follow a TV show. Like...based on that conversation they don't seem to get it which is weird since I KNOW he is not the only kid like this. He will rip those leads off of his head the very second he has an option and I can't physically hold his hands for three days. I messaged the team to tell them to eat their Wheaties and don't overschedule that day...I am planning on sitting right in the bed with him the whole time and am wearing my stretchy pants because it will be a marathon.

I guess we will try to do our best but if anyone wants to weigh in for a blog post on best practices from experienced parents it could help others, let me know!

What the experienced parents said

I am de-identifying everyone here, but I want you to hear the collective voice, because it is remarkably consistent:

My son is exactly like yours. We had to leave a sleep study in the middle of the night. The leads never stood a chance and he wasn't sleeping anyway. He's bigger and stronger now and I can't just hold him down.

We had a home study scheduled. The tech spent three hours trying to get him hooked up and he took off every single lead. We ended up inpatient. And sedated.

I can't get her to sit still either. She'll rip stuff off in the blink of an eye. We've had to sedate her to get through any lengthy test like this.

I'm dreading the next time they suggest an EEG. It's just not possible anymore. Even if you manage to get the leads on, he'll just pull them off.

He dislikes people being in his face or trying to touch his head, unless you're massaging his scalp, which he's fine with.

And more than one person said the same thing about Child Life, with real affection: they mean so well, and I stress out anyway, because he just throws the toy across the room.

Several people had no advice at all. They just said: I've been through it, it isn't easy, I cried, you are not alone. Which, honestly, helped as much as anything.

But there were also a lot of specific, hard won, genuinely clever suggestions, and I do not want a single one of them to get lost. So here is the full list:

For engagement and distraction

• FaceTime with people he loves and recognizes. This came up more than once, and from families whose kids do not track a TV show.
• Game shows on a loop. Family Feud and Wheel of Fortune were both named. Fast, loud, repetitive, low plot.
• Music. A basic MP3 player, a Yoto or similar, and a tip I had not thought of: you can convert audio from YouTube for the non-traditional things your kid loves that are not on any streaming service. A keyboard or other instruments if the unit allows it.
• Plastic eggs with treats inside, hidden around the room for an egg hunt.
• Favorite items placed within arm's reach of the bed so he has to find them.
• Cards, blocks, drawing, if those are things your kid engages with, even in short bursts.
• Food. Do not underestimate food.
• Visits from family and friends on a rotation, if the unit permits it.

For the leads themselves

• A felt backpack, sewn by a friend, to hold the lead pack. This is one of those solutions that only ever comes from a parent.
• Elbow immobilizers, the elbow no-nos. Two different families named these independently. One had used them to keep a chest tube in place. They prevent the arm from bending up to the head without full restraint.
• Ask whether he can get up and walk around the room while connected. Some setups allow more mobility than you would assume.
• Ask about cap EEGs, if precision requirements allow it.

For staffing and timing

• Bring backup. One mom brings her own mother to every overnight EEG so there are two adults in the room. She described it as the single thing that makes it survivable.
• Ask the hospital for a sitter in the room during all awake hours, for extra hands. One family has this arranged now as standard, and it changed everything for them. This is a real, requestable resource in many hospitals and almost nobody tells you it exists.
• Time the placement strategically. One family got through hookup after an ER trip when their daughter was already loaded on Ativan. Not a plan you can schedule, but if you are already in a situation, use it.
• Sedation, when nothing else works. Several families have gone this route for lengthy studies. It is a legitimate choice and I will not hear otherwise.

And one from the outside-the-box file

• Think about your kid's idiosyncratic triggers. After several days inpatient with no seizures captured, one mom mentioned to staff, half joking, that sugar was her daughter's trigger and a sucrose bag might do the job. They tried it. It worked.

I am not just Charlie's mom. I am the Data Mom.

I spend my professional life working with observational health data. Registries, electronic health record (EHR) data, common data models, real world evidence. And when I read that Facebook thread, I did not just see a hard week coming. I saw a structural gap in the epilepsy evidence base, described in the plainest possible language by the people living inside it.

I do not want to overstate it. Plenty of people with epilepsy, including plenty with significant disabilities, get through EEGs just fine. Some kids nap right through placement. Some adults have done it so many times it is routine. This is not a universal problem and I am not claiming it is.

But it is a very common one. And it clusters. The people who cannot tolerate the study are disproportionately the people with the most complex, most severe, most treatment resistant epilepsy. The exact group we most need to understand.

Here is the thing about EEG data. It is one of the richest, most objective signals we have. It is how seizure types get characterized. It is how surgical candidacy gets decided. It is increasingly the substrate for machine learning work, including seizure detection, seizure forecasting, EEG based biomarkers, and automated phenotyping. Enormous methodological energy is pointed at EEG right now, and rightly so, because it is dense, longitudinal, and physiologically real in a way a diagnosis code will never be.

But many of the people in those very data sets could barely tolerate the recording. That is not a small caveat. That is a selection mechanism operating on exactly the axis we most need to study. How many never made it into the data set at all?

Where the missingness comes from

Some never get the study at all. A neurologist who knows the family looks at the situation, weighs the odds, and does not order it. That decision leaves almost no trace in the data. There is no code for we didn't try because we knew how it would go.

Some start it and abort it. Three hours of attempted placement, a middle of the night departure, an incomplete recording. In most systems that becomes a cancelled encounter, a short unusable file, or nothing.

Some get it under sedation. Which is often the right clinical call and I will never second guess a family who made it. But sedated EEG is not necessarily the same measurement. If you are training a model or defining a phenotype and you do not know sedation status, you have quietly mixed two different things together.

Some get through it on the strength of a parent sitting in the bed for seventy two hours. That is an enormous input to data quality, and it is completely invisible in the resulting file.

So the EEG data we have are not a random sample of epilepsy. They are a sample weighted toward people who can sit still, and toward families with the reserves to make sitting still happen. That is a bias with a social gradient sitting on top of a clinical one.

This is exactly why we built REST-LGS

I have lived it from the research side too. I am one of the authors on the Refractory Epilepsy Screening Tool for Lennox-Gastaut Syndrome, the REST-LGS. If you have not run into it, the short version is that LGS is an epileptic encephalopathy that starts in childhood and is classically defined by a triad: multiple seizure types with childhood onset, intellectual disability, and a characteristic EEG pattern, specifically slow spike and wave discharges under 2.5 Hz.

Notice the third item. The diagnosis leans on an EEG finding.

Now think about who has LGS. Adults with drug resistant epilepsy and intellectual and developmental disabilities. Many of them living in group homes or residential settings. Many of them, to put it plainly, exactly like Charlie. The syndrome also evolves, so the classic pediatric EEG signature can fade as people age, which means you often cannot simply repeat the study in adulthood and get the answer. And the childhood records frequently did not survive the transition from pediatric to adult care.

So we built a screening tool that does not depend on any single finding. Eight criteria, four major and four minor, developed through a modified Delphi process with epileptologists, epilepsy nurses, residents, and scientists. It was designed to be usable by non specialists, including primary care providers, and to be populated from whatever records actually exist.

Then we tested it in the real world, in a retrospective chart review of 100 adults with drug resistant epilepsy and intellectual and developmental disabilities. And the missingness was stunning.

Slow spike and wave history on EEG: unavailable for 46 percent.

Age at seizure onset: unavailable for 42 percent.

Both of those are major criteria. Other EEG abnormalities, a minor criterion, was unavailable for 49 percent.

Nearly half the sample had no usable answer for the single most syndrome specific feature in the definition.

We had to add a weighted scoring system, giving major criteria three points and minor criteria one, precisely so the tool would still function when the EEG data were gone.

And it found people. Of the 74 adults in that review with no prior LGS diagnosis, 42, which is 57 percent, screened as possible or likely LGS. Those are people who had been living for years, sometimes decades, with drug resistant epilepsy and no syndrome level diagnosis. A syndrome diagnosis is not academic. It changes which medications are indicated and covered by insurance, it can reduce polypharmacy, it connects families to a community, and in a population with elevated risk of sudden unexpected death in epilepsy it changes how seriously the seizure burden gets treated.

So when I say EEG data are precious, I am not being sentimental about it. I am telling you that we had to design a diagnostic instrument around its absence.

Making EEGs easier to obtain and share

If you are in a position to improve any of this, here is what I would like you to do.

Make acquisition easier. Some of this is technology. Dry electrode systems, cap based setups, ambulatory and home recording, subscalp and implanted devices, wearables that capture something useful even when scalp EEG is impossible. Some of it is much less glamorous: adequate staffing, a room that is not overstimulating, techs who have done this before with a big kid who cannot be reasoned with, scheduling that does not stack the placement against three other procedures. And some of it is designing acquisition protocols with families instead of handing them a plan. The felt backpack, the egg hunt, the game show marathon, the FaceTime rotation. Those came from people who have run more trials of "how do we keep the leads on" than any research team ever will.

Then make sharing easier. This is the piece that gets skipped, and it is the one that makes me the craziest. We fought for that EEG. Three days, one parent's whole week, a hospital bed, a child pushed to the edge of what he can handle. And then the study lives as a PDF report in one health system's chart, in a proprietary raw format nobody outside that system can open, with no clear, easy path for the family to obtain a copy of the actual data. When Charlie transitions to adult care, most of that could be gone. That is the 46 percent.

I have written up the practical mechanics of this in detail, including how families can request their own records and what the different sharing pathways look like: Research Onboarding for Rare Epilepsy Communities: Registry Demo and Toolkit.

If you take nothing else away

EEG data generally do not live where the rest of your records live. Reports are sometimes available through MyChart. The actual study almost never is in your portal. In most systems you have to go through a separate office, often the same one that handles imaging. In my experience those offices are pretty accommodating, and I have had good interactions with them. But it still takes phone calls, forms, follow up, and time.

That bottleneck is a solvable informatics problem, and nobody owns it.

Who is going to carry this

Here is my honest prediction. As with everything else in rare disease, advocacy will end up bearing the burden of education and facilitation. Again. Because we have a deeply fragmented system and there is no other actor positioned to sit with a family and walk them through it.

I would like that to be less true. So here are the things I am working on in my own space, and the things I would like to see other people fund, review favorably, and build.

• Treat EEG retrieval as an informatics priority in its own right. When you are reviewing grants or defining funding opportunities, this is a great area to focus on. It is unglamorous, it is high yield, and it is currently nobody's job. There are imaging experts, especially in the Observational Health Data Sciences and Informatics community, known as OHDSI, who would be glad to support this work and who I am sure would have excellent ideas.
• Lean on the informatics community. They have a genuinely good handle on interoperability and data sharing. This is their home turf and they are underused here.
• Support self download and self sharing. Encourage research activities that let families obtain and share their own data, rather than defaulting to expensive vendor solutions that take years to implement. As much as I love AI and all the automation it enables, this is one example where the automated path has repeatedly turned out to be the slower, more expensive, more time consuming path. Sometimes the old fashioned way is the better option. I write about this a lot in Ten Habits of Great Data Analysts: the least complicated approach is very often the best one. Until we get our act together as a research community and can facilitate free or inexpensive EEG transfer for research studies, the direct approach is what actually works.
• Build federated methods for studying EEG. The OHDSI community is the right one to solve this, and I say that as someone working the same problem from the International League Against Epilepsy side. The federated pattern is already built there and already proven at scale. Analysis travels to the data, patient level records stay at the site, and only aggregate results leave. That pattern is exactly what waveform research needs, and the community that runs it has the interoperability expertise, the vocabulary infrastructure, and the network of sites already standing.
  
  The mechanics are familiar to anyone who has done a network study. Data are extracted, transformed, and loaded into a common data model, most often the Observational Medical Outcomes Partnership model, or OMOP, and then one analytic script runs everywhere it is sent. The payoff in pediatrics would be large, because these studies concentrate in a small number of specialty hospitals. Federation takes the burden off families entirely and makes it far less likely that we miss people. What is missing is not the idea and not the community. It is somebody funding the waveform layer.
• Give families the raw study, not just the report. If a family asks for their own EEG data, the answer should be yes, in a usable format, without a six week fight. Information blocking rules are supposed to address this, but practice has not caught up.
• Use open formats. Open standards for raw EEG already exist and have for decades. When a health system buys an EEG platform, the people in the room are weighing throughput, reading workflow, service contracts, and price, and nobody at that table is representing the multi-site study that will want to pool these recordings in ten years or the family that will ask for a copy. Put portability on the scoring rubric and vendors will respond to it.
• Assign a global unique identifier, or GUID, so the EEG can be connected to everything else. An EEG in isolation is interesting. An EEG linked to that same person's genetic results, seizure diary, medication history, and natural history study enrollment is where the actual discovery happens. Right now those live in separate silos with no reliable way to know they belong to the same human being. A GUID solves this. The general pattern is that a set of stable personal details is hashed to produce a subject identifier, so the underlying information is not what travels and the same person resolves to the same identifier across studies and sites. Several schemes are in use across federal data repositories, rare disease registries, and multi-site research networks, and they differ in how generation and matching are handled, so choose one deliberately rather than inherit whichever one shows up first. It pairs beautifully with federated analysis, because you can establish that two sites hold records on the same person without pooling those records. If you are standing up an epilepsy study or a registry, assign GUIDs at enrollment and agree on the scheme with the people you will eventually want to link to. Retrofitting them later is painful and sometimes impossible.
• Capture the context, not just the signal. Sedation status, restraint or immobilization use, whether a caregiver was present continuously, number of placement attempts, whether the study completed or was truncated and why. All of this is knowable at the bedside and almost all of it is discarded. In a common data model these are perfectly representable. They just have to be treated as something to represent.
• Count the failures. Aborted and cancelled studies are data. A site that can tell you it attempted forty long term EEGs in this population and completed twenty two has told you something important about both the population and the site. Most of us cannot answer that question today.
• Carry the record forward. The single highest yield intervention for adult LGS diagnosis might be making sure the pediatric EEG report survives the handoff to adult care.
• Teach EEG in informatics and data science programs. This is big data at its finest. It is exciting to work with, the methods are interesting, and it could lead to biomarkers that transform how we understand and treat the developmental and epileptic encephalopathies. Students should be getting their hands on it.
• Validate new technology in the hard cases. If the tolerant population is also the validation cohort, we will have built better instruments for the people we could already measure.
• Fund the organizations already doing this work. Support programs that promote education and awareness, facilitate research, and enable data acquisition and sharing across all of the epilepsies. The Rare Epilepsy Network, CombinedBrain, and Data for the Common Good are doing exactly this kind of connective work.

A wicked problem we are positioned to solve

In public administration there is a term for this kind of thing: a wicked problem. Horst Rittel and Melvin Webber introduced the concept in 1973 to describe problems that resist clean definition, have no stopping rule, involve many stakeholders with different framings, and cannot be solved by technical expertise alone. Brian Head's open access book Wicked Problems in Public Policy is an excellent, readable overview if you want the current state of that literature.

Fragmented EEG acquisition and sharing is a textbook case. There is no single owner, no single fix, competing incentives at every layer, and a set of stakeholders who mostly do not talk to each other: neurologists, EEG techs, hospital records offices, device vendors, informaticians, funders, advocacy groups, and families.

But here is the encouraging part. Compared to most wicked problems, this one has an unusually well defined set of levers. We know where the data are. We know who holds them. We know what standards exist. We know which community has the interoperability expertise. And we have families who are highly motivated to participate.

One more thing, for the families

Everything above is about the person in the bed. There is a second half that almost nobody prepares you for, which is what happens to the parent who cannot take their hands off their child for three days. The coffee problem. The shower problem. The meal voucher problem. The fact that you will be asked to step into the hallway to talk and there will be no one to take your place.

I put all of that, along with everything the families sent me, into a free tool: the Hospital Packet Builder. It walks you through building a fact sheet about the accommodations you need, a door sign, a packing list for both of you, a set of questions to ask before you arrive, and wording you can use on the phone or in an email, including the specific language for requesting the raw EEG study rather than the report. Fill it in, print it, hand it over. You can save your version and reopen it before the next admission.

The practical notes I would give another caregiver are in there too, under Notes from a caregiver.

Back to next week

None of this makes Wednesday easier. Wednesday I am going to be in a hospital bed with my nearly adult son, in my stretchy pants, holding his hands and running out the clock.

But I keep thinking about how casually that Facebook thread produced advice: a dozen families, one afternoon, near identical experience, and a near universal sense that the system does not quite believe this is a common problem. Clearly, it is common.

We just built the evidence base out of the studies that made the miraculous journey by being actually ordered, then actually completed, and then actually shared, and then actually being analyzable.

References

Wolf SM, Boyce D, Peña P, Piña-Garza JE, Roland JJ, Thomas B, Zogejani D, McGoldrick PE. Real-world use of the updated refractory epilepsy screening tool for Lennox-Gastaut syndrome. Epilepsia Open. 2024;9:1277-1286. Open access.

Rittel HWJ, Webber MM. Dilemmas in a general theory of planning. Policy Sciences. 1973;4(2):155-169.

Head BW. Wicked Problems in Public Policy: Understanding and Responding to Complex Challenges. Palgrave Macmillan; 2022. Open access: library.oapen.org

If you are building or analyzing EEG data sets in epilepsy and want to talk, my inbox is always open. danielle@boycedatascience.com