Publications

Periodic discharges (PDs) and rhythmic delta activity (RDA) are common EEG patterns in critically ill patients that require characterization of lateralization, spatial localization, and frequency. Manual characterization is subjective and time-consuming. We present an automated system for all four major subtypes: lateralized and generalized PDs (LPD, GPD) and lateralized and generalized RDA (LRDA, GRDA).&#xD;&#xD;Approach. The PD-Profiler combines a per-channel convolutional neural network, hemisphere-specific learned evidence traces, dynamic programming with an approximately periodic prior, and discharge-locked topographic localization. The RDA-Profiler uses iterative narrowband Hilbert refinement for frequency and lateralization, with phase-locking-value analysis for spatial extent. Both pipelines were trained and evaluated on 12,425 EEG segments from 11,729 patients.&#xD;&#xD;Main results. Four electroencephalographers independently scored 200 stratified segments per subtype. The algorithm exceeded expert-expert (EE) agreement on LPD frequency (expert-algorithm (EA) ICC 0.931 versus EE 0.897; p<0.001) and was statistically indistinguishable from EE on GPD frequency (EA 0.980 versus EE 0.973; p=0.061), GRDA frequency (p=0.056), LRDA frequency (p=0.435), LPD laterality (EA kappa 0.945 versus EE 0.974; p=0.079), and LRDA laterality (EA kappa 0.961 versus EE 0.993; p=0.269). No attribute fell significantly below the EE ceiling. PD discharge timing achieved an F1 of 0.817 at single-sample precision (mean absolute timing error 3.1 ms). PD spatial localization reached 97.3% of expert inter-rater Jaccard agreement, and RDA spatial extent matched EE agreement.&#xD;&#xD;Significance. GROND reaches expert-level reliability across lateralization, spatial localization, discharge timing, and frequency for PDs and RDA. In discordant cases, experts judged its frequency estimates more accurate than the original labels in 94% of cases, supporting reproducible characterization at scale.

BACKGROUND: Clinicians increasingly use AI systems to search the medical literature, but current benchmarks do not jointly test whether a response identifies the originating paper and its supporting passage. LitBench evaluates four dimensions: stated versus interpretation-requiring facts, the number and composition of competing papers, one-versus two-paper evidence, and refusal when evidence is absent.

METHODS: LitBench comprises 1,980 open-access papers (1,000 epilepsy papers holding the answers, 980 decoys from unrelated fields) and 9,472 human-reviewed facts, giving 2,188 single-paper questions and 170 requiring a fact from each of two papers. Difficulty rose by burying the answer among up to 2,000 distractors, then again over live PubMed Central. The same questions were then asked with the answering paper removed, so that refusal was the correct response. Four systems were tested: Gemma-4B, Gemma-12B, Sonnet-5, and DeepSeek-V4-Flash (refusal conditions only). Three model judges scored each answer by majority; two-paper questions counted only when both facts were found.

RESULTS: Across fixed-corpus single-paper conditions, accuracy ranged from 55.9% to 71.9% for Gemma-4B, 61.2% to 75.5% for Gemma-12B, and 90.0% to 91.3% for Sonnet-5. Similar epilepsy papers were harder than mixed candidate sets for both Gemma configurations. Two-paper accuracy was 5.7%, 17.1%, and 37.8%, respectively. With evidence absent, Gemma-12B refusal fell from 88.3% to 39.9% as candidate sets grew; Gemma-4B almost never refused. Sonnet-5 refused on 94.9% of sampled single-paper and all sampled two-paper cells. DeepSeek-V4-Flash also refused frequently, but on 80.3% of matched answer-present controls.

CONCLUSIONS: No system was reliable across every axis tested: accuracy fell as competing papers were added, interpretation-requiring facts were harder than stated ones, two-paper synthesis was rarely achieved, and only some systems recognized absent evidence. What counts as "successful" literature search carries considerable nuance; LitBench can evaluate proposed tools at a fine-grained level.

UNLABELLED: Automated seizure detection devices are increasingly plausible tools for epilepsy trials, but no device is perfect. We used CHOCOLATES, a realistic seizure diary simulator, to examine how device sensitivity and false alarm rate (FAR) affect regression-to-the-mean (RTM) and placebo median percentage change (MPC) in a simulated randomized trial design. For each device condition, 100,000 potential participants were generated; eligibility was assessed during a 2-month baseline, followed by a 3-month test period. With FAR fixed at 0, reducing sensitivity from 100% to 10% increased the fraction of eligible participants exhibiting RTM from 38.2% to 64.8% and increased placebo MPC from 14.7% to 48.1%. With sensitivity fixed at 100% and expected FAR correction, increasing FAR from 0 to 1 alarm/day increased RTM from 38.2% to 53.2% and placebo MPC from 14.7% to 31.3%. Imperfect seizure detection can therefore change the apparent placebo response expected from RTM.

SHORT SUMMARY FOR TABLE OF CONTENTS: In simulated epilepsy trials, imperfect seizure detection altered regression to the mean and placebo median percentage change. Trial planning should model detector sensitivity and false alarm rate before device-derived seizure counts are used as endpoints.