Multi-Level DNN for EEG Mental State Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional devices for recognizing mental behavioral attributes based on bio-signals, such as EEG, face limitations in identifying complex features and dependencies among brain regions due to lateralization effects and are biased by gender or handedness, making them unsuitable for accurate mental state recognition.
Innovation Solution
An electronic device employing a multi-level deep neural network (DNN) model that processes EEG and other bio-signals to identify invariant mental behavioral attributes, accounting for lateralization, gender, and handedness, and tracks changes in mental states, providing diagnostic and predictive reports.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional supervised machine learning models are used for mental state recognition, then the device can identify simple features from limited inputs, but it fails to identify complex features and relationships among different brain regions due to lateralization effects
Solution Approach 1:
The patent segments the brain signal processing into multiple independent feature extraction channels corresponding to different brain regions, allowing each channel to process signals independently before combining results. This segmentation approach enables the system to handle complex features and relationships among brain regions while maintaining simplicity in individual processing streams.
Solution Approach 2:
The patent transitions from traditional 2D signal processing to a multi-dimensional approach by incorporating spatial information from multiple EEG electrodes arranged in specific patterns. This dimensional expansion allows the system to capture complex spatial relationships and dependencies among different brain regions that conventional 1D or 2D models cannot detect.
2Ease of manufacture
If conventional machine learning models are trained with hand-crafted features, then the model can learn from bio-signals, but it introduces bias due to lateralization of brain functions related to gender or handedness
Solution Approach 1:
The patent extracts and removes bias-related components from the feature representation by identifying and eliminating features that correlate with gender or handedness. This extraction approach allows the model to focus on task-relevant features while discarding confounding variables that introduce bias in mental state recognition.
Solution Approach 2:
The patent applies different processing strategies to different brain regions based on their functional characteristics. By tailoring the feature extraction and processing to local regional properties rather than applying uniform processing across all electrodes, the system reduces the impact of lateralization effects while maintaining sensitivity to genuine neural patterns.
3Device complexity
If conventional devices use limited input sets for training, then the model training is simpler, but the device cannot accurately identify complex relationships and dependencies among brain regions
Solution Approach 1:
The patent performs preliminary feature extraction and dimensionality reduction on the raw EEG signals before feeding them to the main classification model. This preliminary processing step simplifies the input data structure while preserving critical information about complex relationships among brain regions, making the subsequent model training more efficient and accurate.
Solution Approach 2:
The patent introduces intermediate processing layers that act as mediators between raw signal inputs and final classification outputs. These intermediate layers extract and represent complex relationships in a simplified form, enabling the model to learn from limited training data while still capturing intricate dependencies among different brain regions.
Data Source
AI summary
An electronic device that handles recognition of mental behavioral, affect, emotional, mental states, mental health, or mood-based attributes based on deep neural networks (DNNs), stores a set of EEG signals and a set of bio-signals associated with a subject. The electronic device trains a plurality of first recognition models on a training set of EEG signals and a training set of bio-signals associated with different training subjects. The electronic device trains a second recognition model on a feature vector from output layers of the plurality of first recognition models. The electronic device estimates a plurality of dependency or relationship data by application of the trained plurality of first recognition models on the set of EEG signals and bio-signals. The electronic device identifies a mental behavioral attribute of the subject by application of the trained second recognition model on the plurality of signals and their relationship data.