Seizures are a significant risk factor for neonates with critical illnesses, and early detection and treatment are associated with improved outcomes. However, the gold standard for seizure detection, the continuous electroencephalogram (CEEG), is resource-intensive and not universally available. In The Lancet Digital Health, Jillian L McKee and colleagues report on a model to predict seizures in neonates, which can facilitate better resource allocation and early CEEG discontinuation where seizures are unlikely to occur.

Previous studies have shown that clinical features and initial laboratory values cannot predict which infants will develop seizures in the first days after birth. Models have attempted to predict neonatal seizures by combining clinical and early EEG data but with limited sensitivity and specificity. Additionally, these models rely primarily on retrospective, manual review of charts and EEGs to extract the key data to input into the model. This is time-consuming and might not be possible in clinical practice.

McKee and colleagues harnessed the power of their electronic medical record (EMR)-embedded EEG documentation system, which has reported more than 42,000 EEGs since 2018, including those from 1,117 neonates included in their study, 150 of whom had hypoxic ischaemic encephalopathy. The authors show that compliance with this documentation system has steadily increased since its introduction to more than 98%. This large trove of standardized EEG reports provides the basis for seizure-prediction model development.

The authors began with traditional logistic regression models, which showed seizure prediction accuracy of 84% (95% CI 78–89) in the overall cohort. However, machine learning algorithms outperformed logistic regression, achieving seizure prediction accuracy of up to 90% (95% CI 83–94), with recall (sensitivity) of up to 97% (91–100) in a random forest model. The model also performed well in the subset of neonates with hypoxic ischaemic encephalopathy, with an accuracy of 97% (88–99) and recall of 100% (100–100).

This model applies machine learning algorithms to widely available clinical data – in this case, CEEG reports. The authors show that machine learning algorithms might be well-suited to process the density of large EMR datasets. Recent studies have highlighted the potential for machine learning algorithms to detect seizures in real-time and predict seizures in infants with hypoxic ischaemic encephalopathy. McKee and colleagues expand this technology further, applying it not just to infants with hypoxic ischaemic encephalopathy or to direct analysis of digital EEG recordings. Instead, they apply it to broader groups of neonates at risk of seizures and use only summary CEEG reports from the EMR.

This work supports the promise of machine learning for neonatal seizure prediction, although much work must be done before clinical applications are possible. The authors’ institutional EMR-embedded EEG reporting system is valuable for their model but also has limitations. Because it is an institution-specific system, there are no comparable datasets from other centers to validate this model. Similarly, this limits clinicians’ ability to implement their model at other institutions without similar EMR-embedded EEG documentation systems. It is not known whether neonatal CEEG reports across other institutions have the same standardization and predictive ability as those of this single institution. In a gesture toward generalizability, the authors provide an online calculator for seizure prediction based on their model. This calculator allows clinicians to input a patient’s CEEG characteristics to view the classifier decision representing the estimated seizure risk. However, it is not shown whether this is valid for CEEG reports outside of their standardized institutional template.