Systems and methods for seizure prediction and detection

Inventors

Kamousi, Baharan • Hajinoroozi, Mehdi • Karunakaran, Suganya • Grant, Alexander • Yi, Jianchun • Woo, Raymond • Parvizi, Josef • Chao, Xingjuan

Assignees

Ceribell Inc

Abstract

The present disclosure provides systems and methods for seizure detection. The method for seizure detection may include receiving a plurality of electroencephalography (EEG) signals over a plurality of channels for a subject, preprocessing the plurality of EEG signals by segmenting the plurality of EEG signals for each channel into a plurality of temporal data segments, extracting a plurality of features from each temporal data segment for each channel, and applying a machine learning algorithm to the plurality of features to perform a seizure binary classification for each temporal data segment for each channel. A control policy may be employed to determine a seizure burden on the aggregated seizure binary classifications. When the seizure burden is equal to or exceeds a threshold, a notification may be generated. The notification may be usable by a healthcare practitioner to assess whether the subject may be at risk of having a seizure.

Core Innovation

A method for seizure detection receives a plurality of electroencephalography (EEG) signals over a plurality of channels for a subject. The method preprocesses the plurality of EEG signals by segmenting the plurality of EEG signals for each channel into a plurality of temporal data segments, where each temporal data segment is associated with a time epoch, and extracts a plurality of features from each temporal data segment for each channel.

The method applies a machine learning algorithm to the plurality of features to perform a seizure binary classification for each temporal data segment for each channel, generating a plurality of classifications. Each seizure binary classification classifies each temporal data segment for each channel as seizure-positive or seizure-negative, and the plurality of classifications are compared sequentially across a plurality of time epochs on each channel.

A subset of the classifications is discarded if the subset comprises fewer than three seizure-positive classifications in a row. The remaining non-discarded classifications are aggregated over a moving time window, and a seizure burden is determined based on the aggregated classifications, where the seizure burden comprises a percentage of the temporal data segments classified as seizure-positive and provides a measure of a degree of severity or likelihood of a seizure.

One or more notifications are generated when the seizure burden is equal to or exceeds one or more thresholds. The notifications are indicative of different seizure activities and are useable for assessing whether the subject is at risk of having a seizure.

Claims Coverage

The coverage centers on one independent method claim defining four inventive features: multi-channel EEG epoch feature extraction, per-epoch seizure-positive/seizure-negative classification, moving time window seizure burden percentage, and threshold-based notifications for different seizure activities.

The inventive combination links multi-channel EEG segmentation and feature extraction, per-epoch seizure-positive/seizure-negative classification with discarding of short positive runs, computation of a seizure burden as a percentage of seizure-positive segments over a moving time window, and threshold-triggered notifications indicative of different seizure activities for seizure risk assessment.

Stated Advantages

Documented Applications