Kevin Xie

Self-repairs – in-utterance revisions in which a speaker abandons and reformulates their speech – are a promising interpretable marker for speech-based cognitive screening. Detecting them automatically is difficult because a self-repair is defined by its relationship to surrounding speech rather than by fixed lexical cues. On the DementiaBank ADReSS corpus, we compared the capability of generative LLMs under a five-condition prompt ablation against a fine-tuned DistilBERT token classifier at detecting self-repairs. GPT-5 performed best (test F1 = 0.73) and was largely insensitive to prompt design, whereas the LlaMA (open-weight alternative) was both weaker and far more prompt-sensitive (test F1 = 0.47). DistilBERT, nearly 100 times smaller, matched the open-weight LLM at a fraction of the computational cost. These results suggest that a locally deployable encoder, given sufficient in-domain annotation, is a more plausible route to clinical self-repair detection than scaling model size or prompt complexity.

Background and ObjectivesInpatient video-EEG epilepsy monitoring aims to provide a safe environment for antiseizure medication (ASM) tapering to precipitate seizures, which is often required for di...

One-third of the world’s 70 million people with epilepsy have seizures that are not controlled by medication; and implantable devices are an exciting option for treatment. These devices improve seizure control and can detect impending attacks, missed medication, and impaired cognition. Unfortunately, they have no way to share this information with their hosts in real-time – a limitation common to most medical devices. This is a missed opportunity for implants and wearables to learn from patients, focus on what matters most to them, and teach them how their behavior affects their health. Here, we present a device platform that converses with patients and learns to co-manage epilepsy. The inpatient pro-totype links scalp and intracranial EEG (electroencephalograms) to secure large language models that communicate freely and bidirectionally with their hosts through a smartphone app. An AI agent ingests biomarkers of sleep, medication level, cognition, and seizure risk extracted from brain activity. It con-verses with patients to inform them of clinical events and physiological trends, records their symptoms, responses, and behaviors, and automatically retrains itself to improve performance. Both patients and the AI agent can initiate conversations to teach each other and personalize interactions. We demon-strate this platform in 13 patients undergoing inpatient video-EEG monitoring for epilepsy and validate its performance. Algorithms for detecting seizures optimized their precision over several days without expert intervention – in contrast to the months of iterative, in-person physician programming currently required. Patients responded positively to messages regarding sleep, cognition, and seizure risk while rating the system as highly usable. The platform includes several safeguards, including a system for further algorithm fine-tuning using efficient expert review, and features that ensure data security and regulate communication content. Further work will link other biosensors to measure behavior, improve performance, and optimize therapeutic stimulation. We propose this system as a scalable platform for medical devices that can rapidly adapt to patient and provider needs; one that is broadly adaptable to improving care for many medical conditions.

Objective. New implantable and wearable devices hold great promise to help patients manage their seizure disorders. One proposed application is measuring the rate of interictal epileptiform discharges as a biomarker of medication levels and seizure risk. This study aims to determine whether interictal epileptiform spike rates (spikes) are independently associated with anti-seizure medication (ASM) levels and evaluate whether spike rates are a reliable biomarker for ASM levels. Approach. We conducted a retrospective analysis of 69 patients with drug resistant epilepsy undergoing intracranial EEG monitoring during ASM taper in the epilepsy monitoring unit. An automated spike detection algorithm, validated seizure annotations, and a model of ASM load were used to assess the relationship between spike rates, ASM level, state of consciousness, and seizure timing using linear mixed effects models. Event related analysis and seasonal autoregressive models were applied to evaluate both short-term and longer-term effects of ASMs on spike rates, respectively. Main Results. Spike rates were found to increase following seizures, during sleep, and in patients experiencing early seizures, with no significant association between ASM load and spike rates after controlling for these confounders. Initial models without controls showed a positive association between ASM load and spike rates. However, this relationship disappeared when the effects of sleep and seizures were accounted for. Medication-specific analysis revealed that only levetiracetam showed a significant impact on spike rates during taper. Significance. This study demonstrates that interictal spike rates are more strongly influenced by seizure activity and sleep than by ASM levels, suggesting limited utility of spike rates as a biomarker for ASM load. However, spikes may still serve as important markers for seizure control and disease severity.

Anti-seizure medications (ASMs) are often prescribed using a trial-and-error approach with a similar sequence for many patients. Comparative effectiveness data beyond the first ASM prescription are limited. Artificial intelligence can automatically extract information from electronic health records (EHRs) and augment clinical trials. We conducted a retrospective cohort comparative effectiveness study of currently available ASMs using emulated clinical trials through a causal inference- and informatics-based framework. We extracted data from EHRs using natural language processing algorithms to collect epilepsy covariates and seizure outcomes. We compared ASMs using weighted Kaplan–Meier analyses and log-rank tests with Bonferroni corrections. We compared ASMs across seizure freedom over 2 years of follow-up, and drug retention rate over 5 years of follow-up. Emulated trials included 1596 patients (8379 patient-years) for seizure freedom as outcome and 2945 patients (14 238 patient-years) for retention as outcome. For all epilepsy types, among first-line ASMs, levetiracetam and lamotrigine were superior to oxcarbazepine for seizure freedom (P < 0.001), and levetiracetam was superior to lamotrigine and oxcarbazepine for retention (P < 0.001); among second-line ASMs, lacosamide had the best retention (P < 0.001) but lower seizure freedom rate than topiramate (P = 0.032); among third-line ASMs, cenobamate had longer retention than brivaracetam, but lower seizure freedom rate than clobazam and brivaracetam (P < 0.001). For focal epilepsies, first-line ASMs achieved similar rates of seizure freedom, but levetiracetam remained superior to lamotrigine and oxcarbazepine for retention (P < 0.001); among second-line ASMs, topiramate and lacosamide were both superior to zonisamide for seizure freedom (P < 0.001), while lacosamide remained superior to topiramate and zonisamide for retention (P < 0.002); among third-line ASMs, cenobamate, brivaracetam and clobazam achieved similar retention and seizure freedom. For generalized and unclassified epilepsies, first-line ASMs achieved similar seizure freedom, but levetiracetam remained superior to lamotrigine and oxcarbazepine for retention (P < 0.02); topiramate had the best seizure freedom among second-line ASMs (P < 0.001), while lacosamide again had better retention than zonisamide (P < 0.006); there were insufficient patients included in third-line ASM trials for this subgroup. In conclusion, emulated clinical trials provide a useful framework for studying the comparative effectiveness of anti-seizure medications using real-world observational data. Our findings of differences in seizure freedom and retention between ASMs help to generate hypotheses that should be tested in future prospective, randomized comparative effectiveness trials.

Background: Pleomorphic xanthoastrocytomas (PXAs) are rare brain tumors that are often associated with seizures. There are limited data characterizing epilepsy phenotypes in relation to PXA tumor biology and survival outcomes. Methods: This is a retrospective observational study of 35 patients with PXA who received treatment at the University of Pennsylvania or Dana-Farber Cancer Institute. Demographic and clinical features were assessed in PXA patients with or without seizures and with respect to seizure freedom following tumor resection. Results: During their clinical course, 27 (77%) developed tumor-related epilepsy (TRE), with 25 (71%) initially presenting with a seizure. Compared to those without TRE, patients with TRE were more likely to have a BRAF-mutated PXA and less likely to have frontal lobe tumor localization. Patients with TRE who became seizure-free after the initial resection up to the time of progressive disease were found to have a lower age of seizure onset, smaller tumor diameter, and more likely to have BRAF-mutated tumors compared to those who were not seizure-free. However, following the last tumor resection and accounting for tumor recurrences, there were no significant differences in clinical features between those who were seizure-free and those who were not. Overall survival was 88% after 5 years and 59% after 10 years, with similar survival rates between patients with and without TRE. Conclusion: These findings indicate that BRAF-mutated and BRAF-wildtype PXAs have distinct epilepsy phenotypes. Further investigation of the interplay between tumor biology and seizures may help guide counseling and targeted therapeutic strategies for PXA-related epilepsy.

Background: Hyperkinetic movement disorders involve excessive, involuntary movements such as ataxia, chorea, dystonia, myoclonus, tics, and tremor. Recent advances in artificial intelligence (AI) allow investigators to integrate multimodal instrumented movement measurements and imaging techniques and to analyze these data together at scale. In this systematic review, we aim to characterize AI’s performance in diagnosing and quantitatively phenotyping these disorders. Methods: We searched PubMed and Embase using a semi-automated article-screening pipeline. Results: Fifty-five studies met the inclusion criteria (n = 11,946 subjects). Thirty-five studies used machine learning, sixteen used deep learning, and four used both. Thirty-eight studies reported disease diagnosis, twenty-three reported quantitative phenotyping, and six reported both. Diagnostic accuracy was reported in 36 of 38 and correlation coefficients in 10 of 23 studies. Kinematics (e.g., accelerometers and inertial measurement units) were the most used dataset. Diagnostic accuracy was reported in 36 studies and ranged from 56 to 100% compared to clinical diagnoses to differentiate them from healthy controls. The correlation coefficient was reported in 10 studies and ranged from 0.54 to 0.99 compared to clinical ratings for quantitative phenotyping. Five studies had an overall judgment of “low risk of bias” and three had external validation. Conclusion: There is a need to adopt AI-based research guidelines to minimize reporting heterogeneity and bolster clinical interpretability.

Purpose: Pre-trained encoder transformer models have extracted information from unstructured clinic note text but require manual annotation for supervised fine-tuning. Large, Generative Pre- trained Transformers (GPTs) may streamline this process. In this study, we explore GPTs in zero- and few-shot learning scenarios to analyze clinical health records. Materials and Methods: We prompt-engineered LLAMA2 13B to optimize performance in extracting seizure freedom from epilepsy clinic notes and compared it against zero-shot and fine-tuned Bio+ClinicalBERT models. Our evaluation encompasses different prompting paradigms, including one-word answers, elaboration-based responses, prompts with date formatting instructions, and prompts with dates in context. Results: We found promising median accuracy rates in seizure freedom classification for zero-shot GPT models: one-word – 62%, elaboration – 50%, prompts with formatted dates – 62%, and prompts with dates in context – 74%. These outperform the zero-shot Clinical BERT model (25%) but fall short of the fully fine-tuned BERT model (84%). Furthermore, in sparse contexts, such as notes from general neurologists, the best performing GPT model (76%) surpasses the fine-tuned BERT model (67%) in extracting seizure freedom. Conclusion: This study demonstrates the potential of GPTs in extracting clinically relevant information from unstructured EMR text, offering insights into population trends in seizure management, drug effects, risk factors, and healthcare disparities. Moreover, GPTs exhibit superiority over task-specific models in contexts with the potential to include less precise descriptions of epilepsy and seizures, highlighting their versatility. Additionally, simple prompt engineering techniques enhance model accuracy, presenting a framework for leveraging EMR data with zero clinical annotation. Competing Interests: The authors declare no financial or non-financial competing interests.

Objective: Large-language models (LLMs) can potentially revolutionize health care delivery and research, but risk propagating existing biases or introducing new ones. In epilepsy, social determinants of health are associated with disparities in care access, but their impact on seizure outcomes among those with access remains unclear. Here we (1) evaluated our validated, epilepsy-specific LLM for intrinsic bias, and (2) used LLM-extracted seizure outcomes to determine if different demographic groups have different seizure outcomes. Materials and Methods: We tested our LLM for differences and equivalences in prediction accuracy and confidence across demographic groups defined by race, ethnicity, sex, income, and health insurance, using manually annotated notes. Next, we used LLM-classified seizure freedom at each office visit to test for demographic outcome disparities, using univariable and multivariable analyses. Results: We analyzed 84 675 clinic visits from 25 612 unique patients seen at our epilepsy center. We found little evidence of bias in the prediction accuracy or confidence of outcome classifications across demographic groups. Multivariable analysis indicated worse seizure outcomes for female patients (OR 1.33, P ≤ .001), those with public insurance (OR 1.53, P ≤ .001), and those from lower-income zip codes (OR ≥1.22, P ≤ .007). Black patients had worse outcomes than White patients in univariable but not multivariable analysis (OR 1.03, P = .66). Conclusion: We found little evidence that our LLM was intrinsically biased against any demographic group. Seizure freedom extracted by LLM revealed disparities in seizure outcomes across several demographic groups. These findings quantify the critical need to reduce disparities in the care of people with epilepsy.