Written by Silvia Pavelli
A large language model trained on routine electronic health records separated epilepsy from functional seizures with high accuracy. The more instructive result is what happened when clinicians used it: alone, the model matched senior neurologists; combined with human judgement, it beat them.
Roughly one in five patients referred to a tertiary centre for drug-resistant epilepsy turns out not to have epilepsy at all. Psychogenic non-epileptic seizures, also called functional or dissociative seizures, look like epileptic events but arise from different mechanisms and require different care. The gold standard for telling them apart is prolonged inpatient video-electroencephalography, which costs more than $15,000 per admission in the United States and carries waits of eight to twelve weeks.
A study published in npj Digital Medicine asks whether that bottleneck can be relieved with information clinicians already have. The team built EpiScreen, a system that classifies epilepsy against functional seizures using only the unstructured narrative in electronic health records. Open-source models including Qwen-2.5-32B, DeepSeek-R1-32B, Llama-3.3-70B and Med42-70B were fine-tuned with quantised low-rank adaptation across 13,633 records from two institutions.
The headline numbers are strong. EpiScreen reached an area under the curve of up to 0.992 in within-institution testing and 0.980 across institutions, outperforming conventional baselines. It also produced explanations, using sentence-level attribution to flag which parts of the note drove each decision.
Then comes the part that matters. Asked to classify 300 cases unaided, four medical teams scored between 76.3% and 81.7%. Standalone EpiScreen scored 81.0%. Those differences were not statistically significant. The model, in other words, did not out-diagnose the neurologists.
When the same clinicians were given the model’s prediction alongside its sentence-level attributions, resident teams rose to between 87.7% and 89.0%, and the attending pair to 89.3%. The largest gain was about 12.7 percentage points. Error patterns shifted in complementary directions: one resident group cut missed epilepsy cases from 36 to 14, while the attending pair cut false positives from 37 to 19.
That is the finding worth taking seriously, and it is the standard Brainlayer holds NeuroAI to. Neural and physiological measurements can be useful research inputs, but they are not a shortcut around behavioural evidence, artifact control, or proof on real systems. A standalone accuracy figure is not evidence that a deployment performs better; only a measured comparison with the human in the loop is. EpiScreen is unusual in reporting both.
The caveats are substantial and the authors state them. The work is retrospective. Not every label was confirmed by video-EEG. Documentation bias is possible, because the notes were written by the clinicians whose conclusions were being predicted. Other seizure mimics such as syncope and sleep disorders were not evaluated. The proposed benefits — shorter delays, lower costs, less exposure to unnecessary antiseizure medication — are hypotheses, not measured outcomes.
For anyone building in this space, the methodological lesson generalises beyond epilepsy. A benchmark number says whether a model can perform a task. It says nothing about whether a clinician performs better because the model exists. That question requires the human in the loop to be measured, on held-out cases, with and without assistance — the same evaluation logic Brainlayer applies to human-state models in physical AI, where improvement is only credible when it holds across unseen operators, tasks and conditions.
The next test is prospective and real-time, on minimally curated notes. Until that exists, the honest summary is that this system did not replace neurological judgement. It relocated it — and improved the outcome by doing so.
