Autoencoder Neural Network for EEG Seizure Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current automated seizure detection methods in electroencephalography face challenges with non-stationary epileptic seizures, resulting in low intra- and inter-patient generalization error due to variable seizure manifestations, and limited temporal resolution, leading to incorrect association of seizures occurring within short time periods.
Innovation Solution
A system and method utilizing an autoencoder neural network to analyze electroencephalographic signals by generating an input matrix, reconstructing it, calculating a loss values vector, and detecting predefined patterns in average vectors to generate an indicator of brain activity with high temporal resolution and sensitivity, allowing for unsupervised learning without external teaching signals.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If deep learning with recurrent convolutional neural network is used for seizure detection, then generalization error is reduced, but temporal resolution is limited causing seizures within 30 seconds to be wrongly associated
Solution Approach 1:
The patent segments the EEG signal analysis into multiple frequency bands (delta, theta, alpha, beta, gamma) and processes each band separately through dedicated neural network branches. This segmentation allows the system to capture temporal characteristics at different scales, improving both generalization across patients and temporal resolution for detecting closely spaced seizures.
Solution Approach 2:
The patent transforms the temporal signal analysis into a multi-dimensional frequency domain representation by applying wavelet transform and analyzing multiple frequency bands simultaneously. This dimensional transformation enables the network to detect seizures with higher temporal precision while maintaining robust generalization across different patients and seizure types.
2Ease of manufacture
If hand-crafted features are used for seizure detection, then implementation is simpler, but robustness to variable seizure manifestations is reduced
Solution Approach 1:
The patent employs an autoencoder-based unsupervised learning framework that automatically learns robust feature representations from raw EEG signals without requiring manual feature engineering. The neural network self-adapts to capture the variable characteristics of seizures across different patients and time periods, achieving high robustness while maintaining implementation feasibility through a standardized deep learning architecture.
3Device complexity
If 30 second segments are used for seizure detection, then computational complexity is reduced, but sensitivity to multiple seizures in short time period is lost
Solution Approach 1:
The patent divides the analysis into multiple parallel frequency band segments, each processed by dedicated neural network branches. This segmentation enables the system to analyze shorter time windows with higher computational efficiency while maintaining high sensitivity through multi-scale frequency analysis, allowing detection of multiple seizures occurring within brief time periods.
Solution Approach 2:
The patent applies multiple convolutional filters with different kernel sizes and strides across frequency bands, performing partial analyses at different temporal scales. This excessive action approach ensures that even if some segments miss detections, the multi-scale analysis captures seizures accurately, improving intrinsic sensitivity without proportionally increasing computational complexity.
Data Source
AI summary
A system for calculating an indicator associated to a brain activity of a subject, the system including an acquisition module configured to acquire at least an epoch of electroencephalographic signal of a subject from a plurality of electrodes and a data processing module configured to carry out the steps of: calculating an average vector (VA) using as input of an autoencoder neural network (aNN) an electroencephalographic signals (ES) of a subject acquired from a plurality of electrodes; detecting (DET) the presence of at least a predefined pattern in the consecutive average values of the average vector (VA); and generating an indicator of brain activity (Idx) of the subject when detecting the predefined pattern.