Faculty Bibliography

Large Language Models (LLMs) offer a framework for understanding language processing in the human brain. Unlike traditional models, LLMs represent words and context through layered numerical embeddings. Here, we demonstrate that LLMs' layer hierarchy aligns with the temporal dynamics of language comprehension in the brain. Using electrocorticography (ECoG) data from participants listening to a 30-minute narrative, we show that deeper LLM layers correspond to later brain activity, particularly in Broca's area and other language-related regions. We extract contextual embeddings from GPT-2 XL and Llama-2 and use linear models to predict neural responses across time. Our results reveal a strong correlation between model depth and the brain's temporal receptive window during comprehension. We also compare LLM-based predictions with symbolic approaches, highlighting the advantages of deep learning models in capturing brain dynamics. We release our aligned neural and linguistic dataset as a public benchmark to test competing theories of language processing.

Recent research demonstrates that large language models can predict neural activity recorded via electrocorticography during natural language processing. To predict word-by-word neural activity, most prior work evaluates encoding models within individual electrodes and participants, limiting generalizability. Here we analyze electrocorticography data from eight participants listening to the same 30-min podcast. Using a shared response model, we estimate a common information space across participants. This shared space substantially enhances large language model-based encoding performance and enables denoising of individual brain responses by projecting back into participant-specific electrode spaces-yielding a 37% average improvement in encoding accuracy (from r = 0.188 to r = 0.257). The greatest gains occur in brain areas specialized for language comprehension, particularly the superior temporal gyrus and inferior frontal gyrus. Our findings highlight that estimating a shared space allows us to construct encoding models that better generalize across individuals.

BACKGROUND:Sudden unexpected death in epilepsy (SUDEP) is the leading cause of epilepsy-related mortality. Generalised-particularly nocturnal-convulsive seizures, longstanding epilepsy, and solitary living have been identified retrospectively as risk factors. No definitive electroclinical biomarkers have been prospectively ascertained. This study aimed to identify SUDEP risk markers using multimodality data with long-term follow-up. METHODS:This prospective, multicentre, observational cohort study, conducted at nine centres (eight in the USA and one in the UK), recruited children and adults with epilepsy who were undergoing prolonged video-electroencephalographic (EEG) monitoring. Inclusion criteria were diagnosis of epilepsy by an epilepsy specialist, with or without drug resistance; age older than 2 months; admission to the epilepsy monitoring unit of a participating centre, with video-EEG monitoring; and completion of at least one 6-month follow-up. Demographic, electroclinical, and cardiorespiratory data were collected at baseline. Participants were followed up long term through routine clinic visits, review of electronic health records, and telephone interviews to collect information about seizure frequency, medication status, and mortality. The primary endpoint was time to SUDEP. Cox proportional hazards models were used to assess significant risk factors. FINDINGS/RESULTS:Between Sept 17, 2011, and Dec, 30, 2021, 2632 children and adults with epilepsy were enrolled in this study; 164 were lost to follow-up. 38 (1·54%) of 2468 participants died from SUDEP (12 definite, 18 probable, and eight possible SUDEP cases) and two had near-SUDEP events. Incident SUDEP mortality rate was 4·76 (95% CI 3·37-6·53) cases per 1000 person-years, from a cohort of 7982 person-years. Living alone (hazard ratio 7·62, 95% CI 3·94-14·71), three or more generalised convulsive seizures in the previous year (3·1, 1·64-5·87]), longer ictal central apnoea (1·11, 1·05-1·18), and longer postictal central apnoea (1·32, 1·14-1·54]) were significant predictors of increased SUDEP risk. In a subanalysis excluding possible and near-SUDEP cases, longer ictal central apnoea was not significant. INTERPRETATION/CONCLUSIONS:This study shows an association between premortem peri-ictal apnoea and increased SUDEP risk. Cardiorespiratory monitoring during seizures might benefit assessments of epilepsy mortality risk. Together with solitary living and convulsive seizure frequency, peri-ictal apnoea (>14 s for postictal central apnoea and >17 s for ictal central apnoea) could inform the development of a validatable SUDEP risk index. FUNDING/BACKGROUND:US National Institutes of Health.

BACKGROUND:Sudden unexpected death in epilepsy (SUDEP) is the most common category of epilepsy-related mortality. Centrally mediated respiratory dysfunction has been observed to lead to death in the majority of cases of SUDEP. SUDEP also mainly occurs during nighttime sleep. This study seeks to identify sleep EEG and sleep-related respiratory biomarkers of SUDEP risk. METHODS:In this case-control study, we compared demographic, clinical, EEG, and respiratory data from people with epilepsy who later died of SUDEP (the SUDEP group) with data from age and sex-matched living people with epilepsy, classified as high risk of SUDEP (with ≥1 generalised tonic-clonic seizure [GTCS] per year), low risk of SUDEP (no history of GTCS), and non-epilepsy controls. These data were prospectively collected as part of a multicentre National Institutes of Health study. We analysed sleep macroarchitecture and microarchitecture features and measured sleep homoeostasis by calculating overnight change in slow wave activity (SWA; 0·5-4·0 Hz) in non-rapid eye movement (NREM) sleep during seizure-free nights using linear regression models. We also analysed sleep respiratory metrics, including inter-breath interval variability. We used receiver operating characteristic analysis to assess the individual discriminative performance of demographic, clinical, sleep EEG, and sleep-related respiratory features to predict the risk of SUDEP. FINDINGS/RESULTS:Between Sept 1, 2011, and Oct 15, 2022, 41 participants who later died of SUDEP and 123 matched controls (41 people living with epilepsy at hight risk of SUDEP, 41 people living with epilepsy at low-risk of SUDEP, and 41 non-epilepsy controls) were enrolled. The SUDEP group showed an abnormal lack of overnight decline and an increase in the slope of SWA power compared with the other groups (SUDEP group mean 0·005 standardised error of the mean [SEM] 0·003; high-SUDEP risk group -0·005, 0·002; low-SUDEP risk group -0·003, 0·002; non-epilepsy controls -0·007, 0·003; p=0·017). The overnight increase in the SWA slope was more pronounced in males compared with females (males mean 0·012, SEM 0·001; females 0·001, 0·002; p=0·005). The variability of the inter-breath interval was significantly higher in the SUDEP (coefficient of variation mean 0·15, SD 0·09; SD mean 0·54 s SD 0·35 s) and high-SUDEP risk groups (0·11, 0·03; 0·46 s, 0·19 s) compared with low-SUDEP risk group (0·08, 0·03; 0·30 s, 0·14 s) and non-epilepsy controls (0·08, 0·02; 0·31 s, 0·11 s; p<0·0001). The coefficient of variation of inter-breath interval had the greatest predictive power of SUDEP risk (between-group point estimate difference 0·30, AUC 0·80; 95% CI 0·70-0·90; p<0·0001). INTERPRETATION/CONCLUSIONS:This study identifies impaired sleep homoeostasis in the form of altered SWA progression during NREM sleep overnight in people with epilepsy who later died of SUDEP, and an increase in respiratory variability during NREM sleep in people with epilepsy who later died of SUDEP and in people with epilepsy at high risk of SUDEP. Multiday polysomnography studies are needed to validate sleep homoeostasis and respiratory variability during sleep as potential biomarkers of SUDEP risk. Further studies are required to explore possible sleep interventions that could mitigate SUDEP risk. FUNDING/BACKGROUND:National Institutes of Health-National Institute of Neurological Disorders and Stroke.

Syntax, the abstract structure of language, is a hallmark of human cognition. Despite its importance, its neural underpinnings remain obscured by inherent limitations of non-invasive brain measures and a near total focus on comprehension paradigms. Here, we address these limitations with high-resolution neurosurgical recordings (electrocorticography) and a controlled sentence production experiment. We uncover three syntactic networks that are broadly distributed across traditional language regions, but with focal concentrations in middle and inferior frontal gyri. In contrast to previous findings from comprehension studies, these networks process syntax mostly to the exclusion of words and meaning, supporting a cognitive architecture with a distinct syntactic system. Most strikingly, our data reveal an unexpected property of syntax: it is encoded independent of neural activity levels. We propose that this "low-activity coding" scheme represents a novel mechanism for encoding information, reserved for higher-order cognition more broadly.

BACKGROUND:It remains difficult to predict who will succumb to Sudden Unexpected Death in Epilepsy (SUDEP). As the mechanisms for SUDEP remain unknown, there are not adequate strategies to prevent SUDEP. Thus, some providers are reluctant to discuss SUDEP risk with patients. Public health surveillance and prevention efforts are limited. The SUDEP Summit aimed to identify gaps in the field and prioritize recommendations to advance basic science, clinical care, and public health approaches to mitigate SUDEP. METHODS:In 2020, a diverse group of stakeholders formed the four SUDEP Summit workgroups: 1. Clinical Action, 2. Awareness and Behavior Change, 3. Public Health and Epidemiology, and 4. Basic Science. RESULTS:Each workgroup defined priorities for action and necessary resources and partners; outlined challenges and barriers; defined metrics of success; and developed short and long-term goals. Workgroups discussed methods to prioritize SUDEP research and develop educational materials for healthcare professionals to raise awareness about the risks of SUDEP. Since the meeting, progress has been made in alignment with the workgroups' recommendations. These include studies examining the use of wearables, clinical trials reporting SUDEP rates, tools to improve SUDEP education, policies to improve SUDEP reporting, SUDEP risk calculators, new clinically relevant models, and standardization of data collection. SIGNIFICANCE/CONCLUSIONS:Advancements in SUDEP awareness, education, epidemiology, and causal mechanisms require interdisciplinary collaborative approaches between funding agencies, advocacy groups, providers, and researchers; and the development of new partnerships. More work remains to achieve the recommendations from the Summit, which highlight the fundamental importance of coordinating efforts to mitigate and end SUDEP.

BACKGROUND:Sudden Unexpected Death in Epilepsy (SUDEP) is a rare and tragic outcome in epilepsy, identified by those with the condition as their most serious concern. Although several clinical factors are associated with elevated SUDEP risk, mechanisms underlying SUDEP are poorly understood, making individual risk prediction challenging, especially early in the disease course. We hypothesised that common genetic variation contributes to SUDEP risk. METHODS:Genetic data from people who had succumbed to SUDEP was compared to data from people with epilepsy who had not succumbed to SUDEP and from healthy controls. Polygenic risk scores (PRSs) for longevity, intelligence and epilepsy were compared across cohorts. Reactome pathways and gene ontology terms implicated by the contributing single nucleotide polymorphisms (SNPs) were explored. In the subset of SUDEP cases with the necessary data available, a risk score was calculated using an existing risk prediction tool (SUDEP-3); the added value to this prediction of SNP-based genomic information was evaluated. FINDINGS/RESULTS:Only European-ancestry participants were included. 161 SUDEP cases were compared to 768 cases with epilepsy and 1153 healthy controls. PRS for longevity was significantly reduced in SUDEP cases compared to disease (P = 0·0096) and healthy controls (P = 0·0016), as was PRS for intelligence (SUDEP cases compared to disease (P = 0·0073) and healthy controls (P = 0·00024)). The PRS for epilepsy did not differ between SUDEP cases and disease controls (P = 0·76). SNP-determined pathway and gene ontology analysis highlighted those related to inter-neuronal communication as amongst the most enriched in SUDEP. Addition of PRS for longevity and intelligence to SUDEP-3 scores improved risk prediction in a subset of cases (38) and controls (703), raising the area-under-the-curve in a receiver-operator characteristic from 0·699 using SUDEP-3 alone to 0·913 when PRSs were added. INTERPRETATION/CONCLUSIONS:Common genetic variation contributes to SUDEP risk, offering new approaches to improve risk prediction and to understand underlying mechanisms. FUNDING/BACKGROUND:The Amelia Roberts Fund; CURE Epilepsy; Epilepsy Society, UK; Finding A Cure for Epilepsy and Seizures (FACES).

Sharp wave-ripples (SPW-Rs) are critical to hippocampal function, and the same is true of gonadal steroids, but the interactions are unclear. We find that surgical removal of the gonads greatly reduces SPW rates in both sexes. Ripples are greatly reduced also. Testosterone treatment rescues SPW and ripple rates in males, and 17β-estradiol restores SPW rates in females. We also find that male SPW rates are higher than females but have less power. Furthermore, in intact females, SPW rates fluctuate with the stage of the ovarian cycle. These data demonstrate that hippocampal SPWs are significantly affected by gonadal removal, testosterone, and 17β-estradiol. In addition, there are sex differences. The data are consistent with past demonstrations that testosterone and 17β-estradiol play central roles in hippocampus and significantly expand the views of hormone action and SPW-Rs.