Multimodal LLMs Tested on Smartphone Videos to Separate Epileptic from Functional Seizures

A study published in Nature has measured how accurately multimodal large language models can differentiate epileptic seizures from functional (psychogenic nonepileptic) seizures using only smartphone-recorded video of the events.

The work targets one of the most consequential bottlenecks in epileptology: the wait for scalp video-EEG monitoring. Patients with functional seizures typically spend years carrying an epilepsy diagnosis, receiving antiseizure medications that cannot help them, before inpatient telemetry clarifies the picture. Scalp video-EEG remains the diagnostic standard precisely because the two conditions can look nearly identical on clinical inspection — and because clinicians are rarely present when a spell occurs.

Smartphone video is already abundant in practice. Families record events on phones and bring the clips to appointments, and the recordings are increasingly shared with epilepsy centers. The study asks whether a multimodal model — an AI system that processes visual and textual input together — can read those clips and reliably sort epileptic from functional events.

The published findings report diagnostic accuracy metrics for the model on the video classification task, comparing its performance against the reference standard of expert evaluation. The study appeared in Nature, and its title frames the output directly as a diagnostic-accuracy assessment rather than a technology demonstration.

Why the epileptic-versus-functional question matters

Epileptic seizures and functional seizures overlap heavily in semiology. Generalized shaking, unresponsiveness, and prolonged motor episodes occur in both. Distinguishing them usually requires capture of a typical event during long-term scalp video-EEG, an inpatient, resource-intensive procedure with waiting lists at most centers.

An automated video-based triage tool could shorten that pathway. If a model trained on smartphone footage can flag which patients are more likely to have functional seizures, clinicians could route them earlier to appropriate diagnostic monitoring and, where confirmed, to seizure-psychology programs instead of escalating drug regimens.

What the study evaluated

The researchers applied multimodal large language models to seizure videos recorded on smartphones — the same class of consumer recordings neurologists already review in clinic. The outcome measure was diagnostic accuracy in the binary task: epileptic versus functional seizure.

The framing matters for how the results should be read. This is classification research on recorded video, not a regulatory clearance and not a bedside device. Any clinical deployment would require prospective validation on independent cohorts, across varied cameras, angles, lighting, and patient populations, before regulators would consider it a diagnostic aid.

Where this fits in the reading room

For clinical neurophysiologists, the study adds to a growing literature on computational seizure analysis, much of it built on EEG and ECoG signals. This line of work is different: it places the analysis upstream of the EEG lab, in the video that families capture spontaneously.

It also raises practical questions reading rooms will face if such tools mature:

• Who curates and labels the video cohorts used for training, and against which gold standard — scalp video-EEG, intracranial confirmation, or both?
• How does the model perform on events with atypical semiology, partial recordings, or poor framing?
• Does the output function as a triage suggestion for clinicians, or as a patient-facing tool?

The study's publication in Nature signals peer-reviewed validation of the approach at the level of the reported metrics. What it does not yet establish is real-world performance in an unselected monitoring-unit population.

What comes next

The logical next steps the field will watch for are prospective trials of video-based AI triage against expert scalp video-EEG diagnosis, and formal regulatory evaluation of any resulting device. If those hold up, smartphone video — already sitting in every patient's pocket — could become the first diagnostic filter in the seizure pathway, well before electrodes are applied.