A 27B model reads whole EEG recordings locally: ERIA tunes only its prompt policy, lifting protected-test report scores from 0.5052 to 0.6024
Synopsis
Clinical EEG interpretation must cover technical quality, background and sleep states, interictal and ictal abnormalities across a whole recording, then synthesize a report, so event detection alone is not enough. ERIA has a locally deployed 27B language-model controller call deterministic signal analyses, learned clinical models and local knowledge retrieval, while controller weights, tools and safety constraints stay fixed and only the natural-language analysis policy is rewritten by GEPA using feedback from expert reports. On 200 paired clinical EEGs, the physician-aligned report score on the protected test set rose from 0.5052 to 0.6024, and the policy still performed on CerebraGloss-Bench and BrainBench Foundational Analysis.
Interpretation
ERIA turns clinical EEG interpretation into a whole-recording reporting task led by a single language-model controller, which does not read waveform samples directly but selects tools, inspects their structured outputs and keeps gathering evidence until a structured report can be written. Earlier EEG-to-language systems mapped EEG representations into report text with an analysis path fixed after training, while earlier agentic systems relied on larger controllers, hosted frontier models or multi-agent coordination. Among 200 routine clinical EEGs with paired reports, 50 cases were used for optimization, 50 for validation and 100 as a protected test set; the optimized policy raised the aggregate clinical score from 0.5099 to 0.6277 on validation and from 0.5052 to 0.6024 on the protected test set.
The gain comes from the analysis policy rather than model scale: controller weights, specialist tools and safety constraints stay fixed, and only the mutable natural-language analysis policy is optimized. The raw policy is the initial mutable analysis policy before GEPA optimization and serves as the comparison baseline. GEPA used expert-written EEG reports as the feedback signal and ran for about 4 to 5 days with 750 metric calls on a single NVIDIA DGX Spark; after optimization, tool calls per case rose from 33.72 to 46.78 while controller iterations fell from 10.58 to 9.34.
The optimized policy changes how evidence is gathered: it covers behavioral state across early, middle and late portions of the recording more systematically, localizes poor-quality channels or intervals more actively, and corroborates suspicious abnormalities. Under the raw policy only 40% of validation cases sampled sleep/wake evidence from all three portions, compared with 98% after optimization. On the validation set, sleep_wake_staging calls rose from 3.42 to 7.12 and assess_signal_quality calls from 2.18 to 5.62, while state-related retrieval queries rose from 46.9% to 63.2% of retrieval calls.
A policy optimized only for whole-recording reporting still performs on independent short-segment and general EEG benchmarks. On CerebraGloss-Bench, multiple-choice accuracy held at 71.11%, conversational QA rose from 3.60 to 3.87, and description ROUGE-1 fell modestly from 30.70 to 30.17. The BrainBench Foundational Analysis difficulty-weighted score rose from 85.86 to 86.89, against the highest score of 84.97 reported under CodeAct in the original BrainBench evaluation.
Perspective
The evaluation covers 200 retrospective paired EEG-and-report studies from a single institution, 100 of them in a protected test set; external evaluation is limited to three language tasks of CerebraGloss-Bench and the BrainBench Foundational Analysis subset, which contains 40 tasks and 950 instances.
The physician-aligned score measures agreement with a reference report rather than independent waveform-level truth, reference reports may omit findings, and two clinically acceptable interpretations can differ in wording or detail. The GEPA-selected policy also introduced one instruction unsupported by the current toolset, encouraging inspection of sleep spindles, K-complexes and vertex waves although no dedicated detector exists, and that policy was kept unchanged for the reported evaluations.