Predicting seizure recurrence after an initial seizure-like episode from routine clinical notes using large language models: a retrospective cohort study

BACKGROUND: The evaluation and management of first-time seizure-like events in children can be difficult because these episodes are not always directly observed and might be epileptic seizures or other conditions (seizure mimics). We aimed to evaluate whether machine learning models using real-world data could predict seizure recurrence after an initial seizure-like event. METHODS: This retrospective cohort study compared models trained and evaluated on two separate datasets between Jan 1, 2010, and Jan 1, 2020: electronic medical records (EMRs) at Boston Children’s Hospital and de-identified, patient-level, administrative claims data from the IBM MarketScan research database. The study population comprised patients with an initial diagnosis of either epilepsy or convulsions before the age of 21 years, based on International Classification of Diseases, Clinical Modification (ICD-CM) codes. We compared machine learning-based predictive modelling using structured data (logistic regression and XGBoost) with emerging techniques in natural language processing by use of large language models. FINDINGS: The primary cohort comprised 14 021 patients at Boston Children’s Hospital matching inclusion criteria with an initial seizure-like event and the comparison cohort comprised 15 062 patients within the IBM MarketScan research database. Seizure recurrence based on a composite expert-derived definition occurred in 57% of patients at Boston Children’s Hospital and 63% of patients within IBM MarketScan. Large language models with additional domain-specific and location-specific pre-training on patients excluded from the study (F1-score 0·826 [95% CI 0·817–0·835], AUC 0·897 [95% CI 0·875–0·913]) performed best. All large language models, including the base model without additional pre-training (F1-score 0·739 [95% CI 0·738–0·741], AUROC 0·846 [95% CI 0·826–0·861]) outperformed models trained with structured data. With structured data only, XGBoost outperformed logistic regression and XGBoost models trained with the Boston Children’s Hospital EMR (logistic regression: F1-score 0·650 [95% CI 0·643–0·657], AUC 0·694 [95% CI 0·685–0·705], XGBoost: F1-score 0·679 [0·676–0·683], AUC 0·725 [0·717–0·734]) performed similarly to models trained on the IBM MarketScan database (logistic regression: F1-score 0·596 [0·590–0·601], AUC 0·670 [0·664–0·675], XGBoost: F1-score 0·678 [0·668–0·687], AUC 0·710 [0·703–0·714]). INTERPRETATION: Physician’s clinical notes about an initial seizure-like event include substantial signals for prediction of seizure recurrence, and additional domain-specific and location-specific pre-training can significantly improve the performance of clinical large language models, even for specialised cohorts. FUNDING: UCB, National Institute of Neurological Disorders and Stroke (US National Institutes of Health).