PHILADELPHIA—Without prior training on the ictal signature, a system that analyzes a patient’s brain waves can detect epileptic seizures within four seconds of onset, according to data described at the 69th Annual Meeting of the American Epilepsy Society. The device uses a dynamically adaptive machine-learning algorithm to distinguish ictal events from nonictal events in real time, and neurologists can use it in a clinical setting, according to the investigators.
“We tried to make a system that will give the earliest alert to the nurse so that they can take care of the patient … in the epilepsy monitoring unit when the patients are undergoing surgical evaluation,” said Daniel Ehrens, a doctoral student at the Johns Hopkins Hospital in Baltimore.
Daniel Ehrens
The performance of machine-learning algorithms designed to detect seizures depends on the training data available to them, but prior ictal data often are unavailable in the epilepsy monitoring unit. Mr. Ehrens and colleagues tested a novel seizure-detection system using intracranial recordings of 16 patients with focal epilepsy who underwent presurgical evaluation and had a total of 68 seizures. For each patient, the system analyzed one-hour epochs of EEG that each ended in a seizure.
The device’s algorithm performed a sliding-window analysis of each patient’s EEG to examine 11 features per channel (eg, spectral parameters, complexity, and entropy measures) that data indicate are useful in seizure detection. Features were calculated over a four-second window with a one-second sliding window. The novelty detector incorporated into the system used a one-class support vector machine (SVM) that trains on a 20-minute window of features and shifts every second. The system post-processed the SVM’s output using a Kalman filter.
“We look at the past 25 minutes of the patient’s brain waves and [at] how they are evolving through every second,” said Mr. Ehrens. The device compares past data to current data and looks for differences that could indicate the onset of a seizure.
After parameter optimization, a sensitivity of 95.6% was achieved for all patients and seizures, with a mean detection delay of 5.4 seconds and a false positive rate of 3.3/hour.
Many seizure-detection algorithms are trained using model seizures that may not resemble a given patient’s actual seizures. “We’re just saying, ‘It has to be different from the normal baseline,’” said Mr. Ehrens. “Every patient has a specific, stereotypical signature pattern, and if you only train to that, then you would potentially have better results.
“The optimization of the choice of features and the dynamic adaptation of the detector allowed us to detect previously unseen seizure events without prior training of the ictal signature,” Mr. Ehrens continued. “Although all detections were performed offline in this study, the processing time of our algorithm allows for implementation of real-time seizure-onset detection. The use of a dynamically adaptive SVM is a promising paradigm for the detection of ictal events whose dynamic characteristics are unknown at the time of patient’s admission to the epilepsy monitoring unit.”
—Erik Greb