Seizure count forecasting to aid diagnostic testing in epilepsy

Area of Science:

• Neurology
• Data Science
• Biomedical Engineering

Background:

• Epilepsy monitoring unit (EMU) admissions are crucial for evaluating drug-resistant epilepsy.
• Current seizure forecasting methods predict seizure occurrence but not frequency, potentially leading to non-diagnostic EMU admissions.
• Predicting seizure counts can enhance diagnostic yield and mitigate patient morbidity.

Purpose of the Study:

• To evaluate a state-space method for predicting future electrographic seizure counts.
• To assess the performance of a Bayesian negative-binomial dynamic linear model (DLM) for forecasting daily seizure numbers.

Main Methods:

• Developed a Bayesian negative-binomial dynamic linear model (DLM).
• Utilized seizure count data from 19 patients with responsive neurostimulation (RNS) devices.
• Employed holdout validation for forecast horizons of 1-7 days.

Main Results:

• The negative-binomial DLM improved seizure count prediction over chance by 73.1% for one-day-ahead forecasts.
• Accurate predictions were maintained for up to 7-day forecast horizons.
• Previous day's seizure count and laterality of electrographic detections were key predictors.

Conclusions:

• Dynamic linear models can predict electrographic seizure counts days in advance with above-chance accuracy.
• This forecasting approach shows promise for optimizing EMU admissions and improving patient care.