Seizure prediction using real world data – a learning health system realized

Neonatal seizures. Neonatal seizures can lead to serious consequences for newborns, including long-term morbidity and mortality. In high-resource neonatal intensive care units, screening for seizures with CEEG has become commonplace and is considered standard of care. Accurate seizure prediction can help optimize the allocation of CEEG resources and improve care for critically ill neonates. In our recent study, we aimed to develop seizure prediction models using data extracted from standardized EEG reports. Here is a brief overview of our findings using real-world data to predict seizures in neonates.

Figure 1. Model performance for all neonates, published in (McKee et al., Lancet Digital Health, 2023). (A) Performance values are displayed for the logistic regression (LR1–LR4), decision tree (DT), and random forest (RF1–RF15) models tested on the entire neonatal cohort. Accuracy, precision, recall, F1, AUC, AUCPR, and Cohen’s κ scores are provided for each model. Lighter colors represent better performance. (B) The precision of each model plotted over recall for all 20 models, coded by type, AUC, and accuracy. (C) The relative importance of each feature in the model and the AUCPR are shown for model RF15. AUC=area under the curve. AUCPR=area under the precision–recall curve.

Learning health systems. The Learning Health System (LHS) approach is a comprehensive framework that integrates various sources of data, information, and knowledge to continuously improve the quality and safety of healthcare. These data can be used to develop and refine clinical algorithms and decision support tools that can assist clinicians in the diagnosis and management of neonatal seizures. By leveraging the LHS approach, our study collected and analyzed data from the electronic medical record (EMR) to build models to predict neonatal seizures that can learn over time with data generated through routine clinical care.

Standardized reporting. We implemented a novel CEEG reporting system in the EMR that incorporated standardized terminology from the American Clinical Neurophysiology Society (ACNS). This allowed for consistent documentation of key features that could be easily extracted from the EMR. Our study evaluated 1117 neonates, including 150 with hypoxic-ischemic encephalopathy (HIE), a condition with a high incidence of seizures. The use of a standardized reporting system led to more than 95% reporting of key EEG features, and the completion rate of the templated reports improved overtime.

Computational models. While no simple combination of features could adequately predict seizure risk, we turned to computational models to complement clinical identification of high-risk neonates. We developed logistic regression, decision tree, and random forest neonatal seizure prediction models using EEG features reported on day one to predict seizures on future days. The random forest models incorporating background features performed with classification accuracies of up to 90% for all neonates and 97% for neonates with HIE. The recall (sensitivity) of these models was up to 97% for all neonates and 100% for neonates with HIE (Figure 1). Our model is hosted on a public website and we hope others will try it out and provide feedback.

Clinical implications. Our findings demonstrate that standardized EEG reports can be used to predict neonatal seizures with high accuracy. This information can help guide decisions about the necessity of continuing CEEG monitoring beyond the first day, improving the allocation of limited resources. Additionally, our study highlights the benefits of standardized clinical data collection, which can contribute to personalized CEEG utilization and drive learning health system approaches.

A data-driven future. The ability to predict seizures in neonates, especially those with HIE, using standardized EEG reports can revolutionize the way we allocate CEEG resources and improve care for critically ill neonates. Our findings pave the way for automated predictions and dashboard development for use at scale in real-time clinical care. By embracing standardized clinical data collection and computational models, we can work to optimize resource allocation and provide the best possible care for our youngest patients.

Jillian McKee

Epilepsy genetics fellow at the Children’s Hospital of Philadelphia