Description:FIELD OF THE INVENTION

[0001] The present invention relates to an AI-powered stacking ensemble model for EEG signal classification in BCI system that combines multiple machine learning classifiers with a meta-classifier to improve signal interpretation accuracy, thereby supporting real-time applications through enhanced generalization.

BACKGROUND OF THE INVENTION

[0002] Brain-Computer Interface (BCI) systems allow direct communication between the human brain and external devices, relying heavily on the accurate interpretation of Electroencephalogram (EEG) signals. These signals are inherently noisy, subject-specific, and difficult to classify due to their low signal-to-noise ratio and variability across users.

[0003] Traditional BCI systems include, single classifiers such as Support Vector Machines (SVM), k-Nearest Neighbors (k-NN), or Decision Trees for EEG signal classification. While these methods offer reasonable accuracy, these methods often fail to generalize well across different subjects and are sensitive to noise and signal distortion.

[0004] US2024104377A1 discloses the field of Electroencephalogram (EEG) classification, and, more particularly, to method and system for EEG motor imagery classification. Existing deep learning works employ the sensor-space for EEG graph representations wherein the channels of the EEG are considered as nodes and connection between the nodes are either predefined or are based on certain heuristics. However, these representations are ineffective and fail to accurately capture the underlying brain's functional networks. Embodiments of present disclosure provide a method of training a weighted adjacency matrix and a Graph Neural Network (GNN) to accurately represent the EEG signals. The method also trains a graph, a node, and an edge classifier to perform graph classification (i.e. motor imagery classification), node and edge classification. Thus, representations generated by the GNN can be additionally used for node and edge classification unlike state of the art methods.

[0005] WO2021090331A1 discloses a method for predicting a level of autism spectrum disorders (ASD) using extended reality platform. In one embodiment, the method comprising: displaying, by a Virtual Reality Display (VRD) / Head Mounted Display (HMD) a one or more interactive animations to a user; capturing, by a pair of eye tracking device and a combination of BCI, EEG and fNIR (Functional near infrared spectroscopy) sensors of the VRD/HMD, the user's eye gazing coordinates display on a screen to predict the attention span and the user's cognitive functions, behavioral functions and linguistic functions during one or more specific interactive animation incidents occurring within said interactive animations, computing, by a processing unit of the VRD, an Autism Rating score of the user based on the captured eye gazing coordinates, cognitive functions, behavioral functions and linguistic functions, validating, by the processing unit of the VRD/HMD, the computed Autism Rating score of the user with a pre-stored Autism Rating score and predicting the level of autism spectrum disorders (ASD) of the user based on the validated Autism Rating score.

[0006] Conventionally, many systems have not leveraged ensemble learning approaches, which are known to reduce classification error by combining the strengths of diverse models. Additionally, cross-subject variability and the need for real-time classification demand techniques that adapt to different users without extensive retraining.

[0007] In order to overcome the aforementioned drawbacks, there exists a need in the art to develop system that requires to be capable of generalization using data augmentation, extract relevant features from EEG signals, and deliver real-time predictions for robust performance in BCI systems.

OBJECTS OF THE INVENTION

[0008] The principal object of the present invention is to overcome the disadvantages of the prior art.

[0009] An object of the present invention is to develop a system that is capable of integrating multiple machine learning classifiers to improve the accuracy and reliability of EEG signal classification for BCI applications.

[0010] Another object of the present invention is to develop a system that is capable of improving model generalization and robustness against signal variability.

[0011] Another object of the present invention is to develop a system that is capable including dimensionality reduction techniques to meet the latency requirements of practical BCI implementations.

[0012] Another object of the present invention is to develop a system that is capable of providing robust and interpretable classification results by combining the outputs of base classifiers thereby improving decision-making in BCI systems.

[0013] Yet another object of the present invention is to develop a system that is capable of enhancing signal detection in event-related potential-based interfaces.

[0014] The foregoing and other objects, features, and advantages of the present invention will become readily apparent upon further review of the following detailed description of the preferred embodiment as illustrated in the accompanying drawings.

SUMMARY OF THE INVENTION

[0015] The present invention relates to an AI-powered stacking ensemble model for EEG signal classification in BCI system that integrates multiple classifiers while ensuring adaptability and reliability in real-time brain-computer interface applications.

[0016] According to an embodiment of the present invention, an AI-powered stacking ensemble model for EEG signal classification in BCI system comprises a stacking ensemble classification unit connected to combine multiple classifiers using a meta-classifier to improve classification accuracy, a feature extraction unit connected to EEG Signal Input Unit to extract time-domain, frequency-domain, and power spectral features, a data augmentation unit connected to feature extraction unit to improve model generalization.

[0017] According to another embodiment of the present invention, the system further includes a validation unit connected to stacking ensemble classification unit to ensure the model works well for different people, a real-time processing unit connected to all other units to handle fast processing of EEG data to support real-time Brain-Computer Interface applications, a Logistic Regression model that merges outputs from base classifiers for final prediction, time-frequency scalogram features for P300 signal classification, model generalization tested using subject-independent data splits in LOSO-CV and real-time EEG classification achieved through optimized computation using PCA for dimensionality reduction.

[0018] While the invention has been described and shown with particular reference to the preferred embodiment, it will be apparent that variations might be possible that would fall within the scope of the present invention.

BRIEF DESCRIPTION OF THE DRAWINGS

[0019] These and other features, aspects, and advantages of the present invention will become better understood with regard to the following description, appended claims, and accompanying drawings where:
Figure 1 illustrates a flow chart of an AI-powered stacking ensemble model for EEG signal classification in BCI system.

DETAILED DESCRIPTION OF THE INVENTION

[0020] The following description includes the preferred best mode of one embodiment of the present invention. It will be clear from this description of the invention that the invention is not limited to these illustrated embodiments but that the invention also includes a variety of modifications and embodiments thereto. Therefore, the present description should be seen as illustrative and not limiting. While the invention is susceptible to various modifications and alternative constructions, it should be understood, that there is no intention to limit the invention to the specific form disclosed, but, on the contrary, the invention is to cover all modifications, alternative constructions, and equivalents falling within the spirit and scope of the invention as defined in the claims.

[0021] In any embodiment described herein, the open-ended terms "comprising," "comprises,” and the like (which are synonymous with "including," "having” and "characterized by") may be replaced by the respective partially closed phrases "consisting essentially of," consists essentially of," and the like or the respective closed phrases "consisting of," "consists of, the like.

[0022] As used herein, the singular forms “a,” “an,” and “the” designate both the singular and the plural, unless expressly stated to designate the singular only.

[0023] The present invention relates to an AI-powered stacking ensemble model for EEG signal classification in BCI system that integrates real-time processing, enabling reliable communication and control in assistive neurotechnologies.

[0024] Referring to Figure 1, a flow chart of an AI-powered stacking ensemble model for EEG signal classification in BCI system is illustrated.

[0025] The system disclosed herein includes a stacking ensemble classification unit that is configured to enhance the accuracy of EEG signal classification by combining the predictive strengths of multiple machine learning classifiers. The stacking ensemble classification unit integrates multiple base classifiers including Support Vector Machine (SVM), k-Nearest Neighbors (k-NN), Decision Tree, and Gradient Boosting models. Each of these classifiers is trained independently using the same EEG feature dataset but captures different aspects and decision boundaries due to their unique nature.

[0026] During the operation, EEG signals are first processed to extract features (as further described in a later claim), and the resulting data is fed to each of the base classifiers. These classifiers individually generate prediction outputs corresponding to different brain activity classes or command states.

[0027] These individual prediction results are then passed to a meta-classifier that resides within the stacking ensemble classification unit. The meta-classifier is trained to learn from the outputs of the base classifiers and generate a final decision by analyzing the combined prediction trends. In one embodiment, the meta-classifier may be a Logistic Regression model, although other linear or non-linear models may also be employed depending on performance requirements.

[0028] The working of the stacking ensemble classification unit allows the system to overcome the limitations of any single classifier by leveraging the complementary strengths of each. The stacking ensemble classification unit, as described previously, further includes a meta-classifier designed to merge the outputs of the base classifiers for generating a final prediction. In the preferred embodiment of the present invention, this meta-classifier is implemented using a Logistic Regression model.

[0029] The Logistic Regression model operates by taking the individual output probabilities or class predictions from each of the base classifiers, namely, the Support Vector Machine (SVM), k-Nearest Neighbors (k-NN), Decision Tree, and Gradient Boosting classifiers and using them as input features. These input features are then processed through a trained logistic function, which calculates the weighted contribution of each classifier to make the most likely class prediction.

[0030] The Logistic Regression model is particularly suitable for this task due to its ability to perform well with binary and multiclass classification problems, its robustness against overfitting when properly regularized, and its interpretability. By assigning appropriate weights to each base classifier’s output, the model effectively learns which classifiers are more reliable for different types of EEG signal patterns and adjusts its decision-making process accordingly.

[0031] During real-time operation, when EEG signals are processed and predictions are generated by the base classifiers, the Logistic Regression model receives these outputs and produces a final classification result, which is then used for BCI commands or cognitive state interpretation.

[0032] A feature extraction unit is configured with the system to receive raw EEG signals from the EEG Signal Input Unit and to transform these signals into a structured format suitable for classification. Upon receiving EEG signal data, the unit processes it through multiple parallel pipelines to extract time-domain, frequency-domain, and power spectral features, each providing unique insights into the underlying neural activity.

[0033] All extracted features are normalized and structured as input vectors, which are then transmitted to a stacking ensemble classification unit included in the system.

[0034] The system further incorporates the use of time-frequency scalogram features designed to enhance classification accuracy of P300 signals. The P300 signal, exhibits distinctive patterns that vary both in time and frequency domains, making time-frequency analysis effective.

[0035] To extract these features, the system employs methods such as Continuous Wavelet Transform (CWT) or Short-Time Fourier Transform (STFT) to generate scalograms, which are visual representations of the EEG signal’s energy distribution over both time and frequency simultaneously. The scalogram captures transient and subtle variations in the EEG data associated with the P300 response, including the onset time, duration, and spectral content.

[0036] Once generated, the scalogram features are flattened or otherwise encoded into feature vectors compatible with the stacking ensemble classification unit. Building upon the previously explained feature extraction process, the system includes a data augmentation unit connected directly to the feature extraction unit. This unit receives the extracted EEG features and applies various transformations to artificially increase the diversity of the training dataset.

[0037] The data augmentation unit introduces additive Gaussian noise to the EEG features. This noise mimics real-world measurement disturbances, allowing the model to become more resilient to variations and reduce sensitivity to minor signal fluctuations.

[0038] In addition, the data augmentation unit performs time shifting by slightly or delaying the EEG feature sequences within a defined temporal window. The data augmentation unit also applies amplitude scaling, where the magnitude of EEG features is randomly adjusted. This simulates differences in signal strength due to varying electrode placements or physiological factors, further enhancing model robustness.

[0039] By creating these varied training examples, the data augmentation unit helps prevent overfitting and improves the stacking ensemble model’s ability to generalize to new, unseen EEG data.

[0040] Building upon the previously described data augmentation techniques, the system includes the application of additive Gaussian noise and random time-shifting within a range of plus or minus 100 milliseconds. The additive Gaussian noise introduces subtle variations in the EEG signal features, simulating real-world measurement noise and improving the model's ability to handle imperfect or noisy input data.

[0041] Meanwhile, the random time-shift operation adjusts the temporal alignment of EEG features by shifting them forward or backward up to 100 milliseconds. This adjustment accounts for natural timing variations in brain signals across different trials or subjects, allowing the model to become more robust to temporal discrepancies.

[0042] Building upon the previously explained data and feature extraction techniques, the system employs a validation approach using leave-one-subject-out cross-validation (LOSO-CV) to rigorously test model generalization. In this method, data from one subject is completely held out as the test set while the model is trained on data from all other subjects.

[0043] A real-time processing unit operates as the central hub for managing the rapid flow of EEG data through the system. Connected to the feature extraction, data augmentation, stacking ensemble classification, and validation units are integrated within the system. The real-time processing unit is designed to execute high-speed computations essential for real-time Brain-Computer Interface applications. It processes incoming EEG signals with minimal latency, ensuring that feature extraction, data augmentation, and classification occur swiftly and seamlessly.

[0044] Optimization techniques such as dimensionality reduction (e.g., Principal Component Analysis) are employed within this unit to maintain computational efficiency without compromising classification accuracy. By enabling fast and reliable EEG signal processing, the real-time processing unit supports immediate translation of neural activity into actionable commands, facilitating effective and responsive BCI operation.

[0045] A Principal Component Analysis (PCA) acts as a dimensionality reduction technique to optimize computation during EEG classification. By reducing the number of features extracted from the EEG signals, PCA minimizes computational load while preserving the most significant information necessary for accurate classification.

[0046] This reduction accelerates the processing speed within the real-time unit, enabling quicker decision-making without sacrificing performance. The optimized computation facilitates efficient handling of high-dimensional EEG data, which is critical for supporting real-time Brain-Computer Interface applications where low latency and high accuracy are essential.

[0047] The present invention works best in the following manner, where EEG signals are acquired through the EEG signal input unit from the user. These signals undergo preprocessing, where noise and artifacts are minimized to enhance signal quality. The preprocessed signals are then forwarded to the feature extraction unit, which extracts relevant features encompassing time-domain characteristics, frequency-domain components, and power spectral density features. For specific signal types such as P300, time-frequency scalogram features are also computed to capture detailed signal dynamics. Next, the extracted features are augmented using the data augmentation unit. This involves applying techniques such as additive Gaussian noise, random time shifting within a ±100 ms range, and amplitude scaling. These augmentations enrich the training dataset, improving the model’s ability to generalize across diverse EEG patterns and users. The augmented feature set is input into the stacking ensemble classification unit, which combines multiple base classifiers including Support Vector Machine (SVM), k-Nearest Neighbors (k-NN), Decision Tree, and Gradient Boosting classifiers.

[0048] In continuation, each base classifier independently predicts class labels, and their outputs are aggregated by a meta-classifier, specifically a Logistic Regression model, to produce the final classification decision. To ensure robustness and generalization across different subjects, the validation unit employs leave-one-subject-out cross-validation (LOSO-CV). The validation technique trains the model on data from all subjects except one, and tests on the excluded subject, iterating this process to evaluate performance comprehensively. For real-time BCI applications, the real-time processing unit facilitates fast EEG data processing by incorporating Principal Component Analysis (PCA) for dimensionality reduction. PCA reduces the complexity of feature vectors, thereby speeding up classification without compromising accuracy.

[0049] Although the field of the invention has been described herein with limited reference to specific embodiments, this description is not meant to be construed in a limiting sense. Various modifications of the disclosed embodiments, as well as alternate embodiments of the invention, will become apparent to persons skilled in the art upon reference to the description of the invention. , Claims:1) An AI-powered stacking ensemble model for EEG signal classification in BCI system, comprising:

i) a stacking ensemble classification unit connected to combine multiple classifiers (SVM, k-NN, Decision Tree, Gradient Boosting) using a meta-classifier to improve classification accuracy;
ii) a feature extraction unit connected to EEG Signal Input Unit to extract time-domain, frequency-domain, and power spectral features from EEG signals for better model input;
iii) a data augmentation unit connected to feature extraction unit to apply Gaussian noise, time shifting, and amplitude scaling to training data to improve model generalization;
iv) a validation unit connected to stacking ensemble classification unit that uses leave-one-subject-out cross-validation to ensure the model works well for different people; and
v) a real-time processing unit connected to all other units to handle fast processing of EEG data to support real-time Brain-Computer Interface applications.

2) The system as claimed in claim 1, wherein the meta-classifier is a Logistic Regression model that merges outputs from base classifiers for final prediction.

3) The system as claimed in claim 1, wherein time-frequency scalogram features are used specifically for P300 signal classification.

4) The system as claimed in claim 1, wherein data augmentation includes both additive Gaussian noise and random time-shift within ±100 ms.

5) The system as claimed in claim 1, wherein model generalization is tested using subject-independent data splits in LOSO-CV.

6) The system as claimed in claim 1, wherein real-time EEG classification is achieved through optimized computation using PCA for dimensionality reduction.