Cross-species ECoG signal decoding method and system based on ensemble learning, terminal and storage medium

Abstract

The invention discloses a cross-species ECoG signal decoding method and system based on ensemble learning, a terminal and a storage medium, and the method comprises the steps: collecting brain ECoG signals of different species during the execution of a motion task, and carrying out the data preprocessing, and obtaining a preprocessed target brain ECoG signal; according to a target brain ECoG signal, frequency band features are adaptively selected for different species; constructing a decoding model, and training the decoding model by using the frequency band features to obtain a trained decoding model; and obtaining a current brain ECoG signal of a new target species, inputting the current brain ECoG signal into the trained decoding model for decoding, and outputting motion intention information of the target species. According to the method, explainable frequency band power features are used as a link, a frequency band feature analysis and transfer learning strategy is combined, a universal cross-species neural motion decoding framework is constructed, unified decoding of motion intentions of different species is achieved, and the cross-species adaptive capacity and the transfer performance of a decoding model are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of brain-computer interface signal decoding technology, and in particular to a cross-species ECoG signal decoding method, system, terminal and computer-readable storage medium based on ensemble learning. Background Art

[0002] Existing ECoG (Electrocorticography) signal decoding systems are mostly built for a single species and lack decoding models that are universal across species, making it difficult to effectively transfer decoding models between different species. Because brain motor signals from different species differ significantly in terms of spectral characteristics, decoding models trained on one species are difficult to directly apply to other species, limiting their transferability. Existing technologies generally suffer from insufficient model stability and low transfer efficiency when addressing differences in neural signatures across species.

[0003] Therefore, the existing technology still needs to be improved and developed. Summary of the Invention

[0004] The main purpose of the present invention is to provide a cross-species ECoG signal decoding method, system, terminal and computer-readable storage medium based on ensemble learning, aiming to solve the problem in the existing technology of lacking effective countermeasures when facing cross-species brain signal differences and being unable to ensure the universal adaptability of the decoding model.

[0005] To achieve the above objectives, the present invention provides a cross-species ECoG signal decoding method based on ensemble learning, which comprises the following steps:

[0006] collecting brain ECoG signals of different species when performing motor tasks, performing data preprocessing on the brain ECoG signals, and obtaining preprocessed target brain ECoG signals;

[0007] Adaptively selecting frequency band features for different species based on the target brain ECoG signal;

[0008] Constructing a decoding model, and training the decoding model using the frequency band features to obtain a trained decoding model;

[0009] The current brain ECoG signal of the new target species is obtained, the current target brain ECoG signal is input into the trained decoding model for decoding, and the movement intention information of the target species is output.

[0010] Optionally, the cross-species ECoG signal decoding method based on ensemble learning, wherein the collecting of brain ECoG signals of different species when performing motor tasks, and performing data preprocessing on the brain ECoG signals to obtain preprocessed target brain ECoG signals, specifically includes:

[0011] Using implantable microelectrode arrays to collect brain ECoG signals from different species while they perform similar tasks or behavioral conditions;

[0012] The brain ECoG signal is subjected to band-pass filtering, artifact removal, and adaptive noise template denoising to obtain a preprocessed target brain ECoG signal.

[0013] Optionally, the cross-species ECoG signal decoding method based on ensemble learning, wherein the data preprocessing of the brain ECoG signal to obtain a preprocessed target brain ECoG signal, further comprises:

[0014] The pre-processed target brain ECoG signals of different species are standardized and synchronously segmented according to a unified format.

[0015] Optionally, the cross-species ECoG signal decoding method based on ensemble learning, wherein the adaptive selection of frequency band features for different species based on the target brain ECoG signal, specifically includes:

[0016] Calculate the power spectral density of each target brain ECoG signal segment, perform Fourier transform on the segmented signals using the Welch algorithm, and average to obtain spectrum estimation;

[0017] The spectrum is integrated according to typical EEG bands, the average power value of each frequency band is extracted, different frequency bands are defined, and the area under the power spectrum density curve in each frequency band is calculated as the energy feature of the corresponding frequency band. Each target brain ECoG signal is converted into a multidimensional frequency band feature vector.

[0018] The frequency band characteristics of different species were compared and analyzed, and the distribution of energy in each frequency band of each species under different motion states was statistically analyzed to obtain the frequency range that meets the requirements.

[0019] Optionally, the cross-species ECoG signal decoding method based on ensemble learning, wherein the constructing of a decoding model and the use of the frequency band features to train the decoding model to obtain a trained decoding model specifically include:

[0020] A decoding model is constructed using an ensemble learning type of machine learning model, wherein the ensemble learning type of machine learning model includes a random forest model and an XGBoost gradient boosting decision tree model;

[0021] The multi-dimensional frequency band feature vector is used as input, and a supervised learning method is adopted to train the decoding model using the existing data labels of each species. The output is the predicted movement state category, thereby obtaining a trained decoding model.

[0022] Optionally, in the cross-species ECoG signal decoding method based on ensemble learning, the training of the decoding model adopts a hierarchical training strategy, first training a sub-model for each species to ensure a good fit to its own data, and then considering cross-species fusion.

[0023] Optionally, the cross-species ECoG signal decoding method based on ensemble learning, wherein the step of constructing a decoding model and training the decoding model using the frequency band features to obtain a trained decoding model, further comprises:

[0024] aligning the feature spaces of different species using a cross-species alignment strategy and adjusting the decoding model;

[0025] The cross-species alignment strategy includes feature space reconstruction and transfer learning fine-tuning. The feature space reconstruction is used to map the frequency band features of the source species and the target species into a common feature space. The transfer learning fine-tuning is used to use the model parameters pre-trained on the source species and retrain and adjust them on a preset amount of target species data.

[0026] The transfer learning fine-tuning includes parameter transfer and multi-task learning. The parameter transfer is used to use the model parameters trained for the source species as the initial values of the target species model, and then iteratively fine-tune with the data of the target species. The multi-task learning is used to construct a unified model that simultaneously inputs multi-species data, including the source species task and the target species task in the loss function. Through joint training, the hidden layer of the unified model serves the decoding of the two species at the same time.

[0027] In addition, to achieve the above-mentioned purpose, the present invention further provides a cross-species ECoG signal decoding system based on ensemble learning, wherein the cross-species ECoG signal decoding system based on ensemble learning comprises:

[0028] A data acquisition and preprocessing module is used to collect brain ECoG signals of different species when performing motor tasks, perform data preprocessing on the brain ECoG signals, and obtain preprocessed target brain ECoG signals;

[0029] A frequency band feature extraction and analysis module, configured to adaptively select frequency band features for different species based on the target brain ECoG signal;

[0030] A decoding model construction and training module is used to construct a decoding model and train the decoding model using the frequency band features to obtain a trained decoding model;

[0031] The movement intention information recognition module is used to obtain the current brain ECoG signal of a new target species, input the current brain ECoG signal into a trained decoding model for decoding, and output the movement intention information of the target species.

[0032] In addition, to achieve the above-mentioned purpose, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and an integrated learning-based cross-species ECoG signal decoding program stored in the memory and runnable on the processor, wherein the integrated learning-based cross-species ECoG signal decoding program, when executed by the processor, implements the steps of the integrated learning-based cross-species ECoG signal decoding method described above.

[0033] In addition, to achieve the above-mentioned purpose, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a cross-species ECoG signal decoding program based on ensemble learning, and when the cross-species ECoG signal decoding program based on ensemble learning is executed by a processor, the steps of the cross-species ECoG signal decoding method based on ensemble learning as described above are implemented.

[0034] In the present invention, brain ECoG signals of different species are collected when performing motor tasks, and data preprocessing is performed on the brain ECoG signals to obtain preprocessed target brain ECoG signals; based on the target brain ECoG signals, frequency band features are adaptively selected for different species; a decoding model is constructed, and the decoding model is trained using the frequency band features to obtain a trained decoding model; the current brain ECoG signal of the new target species is obtained, and the current brain ECoG signal is input into the trained decoding model for decoding, and the movement intention information of the target species is output. The present invention uses interpretable frequency band power features as a link, combines frequency band feature analysis with transfer learning strategies, and constructs a universal cross-species neural motion decoding framework to achieve unified decoding of movement intentions of different species, thereby improving the cross-species adaptability and transfer performance of the decoding model. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is a flow chart of a preferred embodiment of the cross-species ECoG signal decoding method based on ensemble learning of the present invention;

[0036] Figure 2 Schematic diagram of the overall framework process of a preferred embodiment of the cross-species ECoG signal decoding method based on ensemble learning of the present invention;

[0037] Figure 3 2 is a schematic diagram showing the pre-experimental frequency band characteristics in a preferred embodiment of the cross-species ECoG signal decoding method based on ensemble learning of the present invention;

[0038] Figure 4 Schematic diagram showing the difference in frequency bands of movement states of rabbits and macaques in a preferred embodiment of the cross-species ECoG signal decoding method based on ensemble learning of the present invention;

[0039] Figure 5 2 is a schematic diagram of the rabbit frequency band classification results in a preferred embodiment of the cross-species ECoG signal decoding method based on ensemble learning of the present invention;

[0040] Figure 6 2 is a schematic diagram of the frequency band classification results of macaque monkeys in a preferred embodiment of the cross-species ECoG signal decoding method based on ensemble learning of the present invention;

[0041] Figure 7 2 is a structural diagram of a preferred embodiment of a cross-species ECoG signal decoding system based on ensemble learning according to the present invention;

[0042] Figure 8 FIG. 4 is a structural diagram of a preferred embodiment of the terminal of the present invention. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solutions and advantages of the present invention more clear and distinct, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0044] Brain-computer interface technology controls external devices by analyzing neural signals and is widely used in motor rehabilitation and prosthetic control. ECoG signals are cortical electrical signals between non-invasive EEG (Electroencephalography) and single-neuron recordings, with high temporal and spatial resolution and good stability. They are widely used in BCI (Brain-Computer Interface) tasks such as motor intention, language, and emotion. With the development of brain-computer interface (BCI) research, the importance of animal experiments in algorithm verification continues to increase. However, the signal characteristics of EEG signals (especially cortical potential ECoG) vary greatly between different species, including spectral distribution, signal amplitude, spatial topology, etc., which makes model training often highly species-dependent. Some studies have used methods such as transfer learning and canonical correlation analysis to try to transfer models from non-human primates to humans, but there are still problems with the model's high dependence on data and poor generalization ability.

[0045] Existing technical solutions include: decoding model solutions and cross-species research solutions.

[0046] A. Decoding model solution:

[0047] (1) Currently, commonly used ECoG decoding models include traditional machine learning methods and deep learning methods. Most use linear classifiers such as SVM, LDA, and Softmax to process specific frequency band features, such as β and γ waves. These traditional methods are widely used due to their simplicity and easy interpretation, but they often have difficulty capturing nonlinear features and have limited effectiveness in complex brain signals.

[0048] (2) Research on using ensemble learning (such as Random Forest, XGBoost) or Convolutional Neural Network (CNN) to process local time window features. Ensemble learning methods such as Random Forest and Gradient Boosting Decision Tree (such as XGBoost) can better handle nonlinear relationships and excel in improving the decoding performance of a single species. Deep learning methods (such as RNN, Transformer, etc.) can automatically extract multi-level features from raw data and have shown great potential in tasks such as ECoG speech decoding.

[0049] However, whether traditional or deep methods, current research mostly focuses on data from humans or a certain type of experimental animal, and lacks consideration of the universality of decoding across species.

[0050] B. Cross-species research plan:

[0051] (1) There are cross-subject decoding studies based on Transformer and GNN, but few involve cross-species ECoG.

[0052] (2) In cross-species research, some have proposed developing a cross-species cell type classification model based on high-density Neuropixels probes. A deep learning classifier has been constructed to achieve cross-species identification of neuronal cell types based on data recorded by high-density electrode arrays, successfully classifying different cerebellar neurons in mice and macaques. These research results indicate that through adequate training and feature extraction strategies, algorithms can potentially decode neural information across species differences.

[0053] Taking the above two aspects into consideration, existing technical solutions lack effective countermeasures when facing cross-species differences in brain signals and cannot guarantee the universal adaptability of the decoding model.

[0054] Existing brain-computer interface decoding technologies are primarily optimized for a single species. When applied to other species, performance often degrades, exhibiting the following deficiencies:

[0055] (1) Frequency band differences affect decoding migration: Different species have different optimal decoding frequency bands. If fixed frequency band features are used, the accuracy of the model on new species will be significantly reduced. Traditional methods fail to adjust feature extraction strategies based on species physiological differences, resulting in poor cross-species adaptability of the model.

[0056] (2) Models lack a universal structure: Many decoding models fine-tune parameters or structures for specific datasets, such as being designed only for human data, and lack a model framework that is universal across different species.

[0057] (3) Limited interpretability and scalability: Although deep learning methods can be applied across domains, their internal decision-making processes are difficult to explain, which is not conducive to understanding the differences in neural mechanisms between different species. Traditional methods that rely too much on specific frequency bands may not be applicable when extended to new species. Existing technologies lack a cross-species decoding solution that can both explain the role of frequency bands and have universal scalability.

[0058] To address these technical issues, the present invention provides an ECoG signal decoding framework that can be effectively generalized to multiple species, thereby improving the cross-species adaptability and transfer performance of the decoding model. The present invention's technical solution encompasses the entire process, from data acquisition to model training and cross-species adaptation. The overall concept is to utilize interpretable frequency band power features as a link, combined with transfer learning strategies, to achieve unified decoding of movement intentions across different species.

[0059] The cross-species ECoG signal decoding method based on ensemble learning described in the preferred embodiment of the present invention is as follows: Figure 1 and Figure 2 As shown, the cross-species ECoG signal decoding method based on ensemble learning includes the following steps:

[0060] Step S10: collecting brain ECoG signals of different species when performing motor tasks, performing data preprocessing on the brain ECoG signals, and obtaining preprocessed target brain ECoG signals.

[0061] Specifically, in terms of data acquisition, brain ECoG signals of two or more species are obtained when they perform similar motor tasks. For example, brain signals of rabbits in free motion and macaques in autonomous motion can be selected as sources of training data. The signals of each species are collected through an implantable microelectrode array, and the corresponding behavioral labels (such as movement type or state) are synchronized. In other words, the data involved in the present invention include brain movement-related neural signals of at least two species. The ideal data source is the ECoG signals recorded under similar tasks or behavioral conditions to ensure that the decoding tasks of different species are comparable. Example data can come from: motor cortex ECoG recordings of rabbits walking freely, and cortical ECoG recordings of macaques when performing grasping tasks. In the experimental setting, different species may use their own adapted high-density cortical electrode arrays (such as μECoG electrodes), and synchronously collect EEG signals and motor behavior videos through wired or wireless devices.

[0062] Subsequently, strict data preprocessing is performed (i.e., the collected raw ECoG data passes through a unified preprocessing pipeline). Bandpass filtering: Bandpass filtering (for example, a 0.5-300Hz bandpass filter) is uniformly applied to the data of all species to remove high-frequency noise interference such as DC drift and power frequency, and to remove non-target signal components to the greatest extent; Artifact Removal: Check and remove abnormal high-amplitude segments caused by loose electrodes, myoelectric bursts, etc.; Adaptive Noise Template (ANT) Denoising: For the periodic interference common in implantable recordings, an adaptive noise template algorithm is introduced for denoising. This is a major feature of the preprocessing step. For the periodic noise that often appears in long-term implanted recordings (such as low-frequency modulation noise caused by heartbeat pulsation, respiratory rhythm, etc.), the present invention introduces the ANT algorithm to eliminate it. The specific approach is: collect background ECoG signals for a certain period of time in a resting state, calculate its mean value to obtain a noise template; use this template to subtract the noise component from the formal data (original ECoG data or data also mixed with the noise) through adaptive filtering, and update the template based on the residual, iterating repeatedly until the noise signal converges. This method can effectively remove cyclostationary interference while retaining neural signals, so that the processed data has a higher signal-to-noise ratio and stability. At the end of preprocessing, the data of different species can be standardized and synchronously segmented according to a unified format to prepare for feature extraction. That is, the ANT method uses a noise template that is updated over time to identify and subtract cyclostationary noise, effectively improving the signal-to-noise ratio of ECoG signals of different species. The high-quality signal after preprocessing (i.e., the target brain ECoG signal) lays the foundation for subsequent feature extraction.

[0063] Step S20: Adaptively selecting frequency band features for different species based on the target brain ECoG signal.

[0064] Specifically, during the feature engineering extraction phase, the present invention highlights the important role of frequency band power spectrum features. By performing power spectral density analysis (PSD) on the preprocessed ECoG signal (i.e., the target brain ECoG signal), a common method is to use the Welch algorithm to perform a Fourier transform on the segmented signal and average it to obtain a stable spectrum estimate. The spectrum is then integrated according to typical EEG bands, and the average power value of each band is extracted. For example, frequency bands such as δ: 1-4Hz, θ: 4-8Hz, α: 8-13Hz, β: 13-30Hz, low γ: 30-70Hz, and high γ: >70-150Hz are defined. The area under the PSD curve in each frequency band is calculated as the energy feature of the frequency band. In this way, each ECoG data is converted into a multidimensional frequency band feature vector. For each species' data, the correlation between each frequency band feature and the movement state is evaluated separately, and the frequency range that contributes most to decoding is selected. For example, preliminary experiments showed that the PSD of rabbit ECoG in the low-frequency beta band showed the greatest difference between movement and rest, while the PSD of macaques showed the most significant change in the high-frequency gamma band. Therefore, the present invention adaptively selects frequency band features for different species, giving higher weight or focusing analysis on the most effective frequency bands for each species in the model, thereby fully exploring the movement intention information in the brain signals of each animal.

[0065] Next, we conducted a comparative analysis of the frequency band characteristics of different species: by statistically analyzing the energy distribution of each frequency band under different motion states of each species, we found the frequency range that plays a major role in distinguishing motion states.

[0066] For each segment of ECoG signal x(t), the Welch method is used to estimate the power spectrum:

[0067]

[0068] Among them, P xx (f) represents the estimated power spectrum density, L represents the number of segments, represents Fourier transform, w(t) represents window function (such as Hamming window), x l (t) represents the first segment of the signal, t represents time, and f represents frequency.

[0069] Then, the power characteristics of each frequency band are:

[0070]

[0071] According to the experimental results, the ECoG signal power spectral density (PSD) response of rabbits and macaques in motion showed significant species differences. Figure 3As shown, when rabbits are in motion, their ECoG signals exhibit significant increases in power in the low-frequency beta band (13-30 Hz) and theta band (4-8 Hz), while changes in the high-frequency gamma band (above 70 Hz) are not significant. In contrast, macaques exhibit a significant increase in power in the high-frequency gamma band during motor activation, while changes in the low-frequency bands (including the delta, theta, and beta bands) are relatively small. This difference in spectral characteristics suggests that rabbits' motor intentions are primarily encoded in low-frequency rhythms, and their neural regulation may rely on cortical-wide synchronization and rhythmic oscillations. In contrast, macaques' movement-related neural activity is primarily manifested by increases in the high-frequency gamma band, reflecting the synchronized firing of local neuronal populations, which is generally associated with more direct motor output signals. These results reveal essential differences in the cortical motor regulation mechanisms of different species from a neurophysiological perspective, reflecting differences in neural network structure, neuronal type ratios, and motor execution methods. Therefore, in the construction of cross-species ECoG signal decoding models, it is necessary to introduce a species-specific frequency band modeling mechanism and select the most informative frequency range as the decoding feature for each species, thereby effectively improving the generalization performance of the model across different species. Figure 4 As shown, the spectrum curve has been normalized, and the shaded area represents the standard deviation, which clearly shows the species differences in the β and γ bands.

[0072] Based on this type of analysis, the present invention focuses on species-specific frequency bands during feature extraction: for rabbits, beta-band power is used as a core feature; for macaques, high gamma-band power is emphasized. This mechanism ensures that data from each species is represented using the frequency band characteristics that best represent its motor neural activity, laying a solid foundation for subsequent model training.

[0073] Step S30: construct a decoding model, and use the frequency band features to train the decoding model to obtain a trained decoding model.

[0074] Specifically, after completing the frequency band feature extraction, it is necessary to design a suitable decoding model (i.e., classification model) to map these features to specific motion intentions (such as different motion types or states). The decoding model is used to identify motion states, including stillness in motion (simple binary classification), and multiple classifications of motion states (mainly distinguished according to the intensity of the motion). The present invention gives priority to machine learning models of the ensemble learning category, typically represented by random forest (RF) and XGBoost gradient boosting decision tree models. These two models are based on a set of decision trees, have good fitting capabilities for nonlinear relationships, and can provide feature importance evaluations, which facilitate the interpretation of model decisions. In the previous study, the present invention compared the performance of RF, XGBoost and traditional Softmax regression models. The results showed that RF and XGBoost have higher accuracy on complex ECoG frequency band features and have stronger capture capabilities for nonlinear patterns. Especially for situations containing multiple frequency band features, RF and the like can automatically determine which frequency bands contribute most to the classification, while linear models such as Softmax have limited discriminative power on low frequency bands. Therefore, in implementation, RF or XGBoost models are used, with the input being a multi-band power feature vector and the output being the predicted motion state category. Model training can adopt supervised learning, using existing data labels of various species for training. It should be noted that this framework is not limited to the above two classification algorithms.

[0075] That is to say, the present invention preferably uses an ensemble learning classifier (such as random forest, XGBoost gradient boosting decision tree) as the basic model to fully utilize the nonlinear discriminative ability of frequency band features. Compared with linear models, ensemble learning methods show higher accuracy and stability when processing complex brain signal relationships. In comparative experiments, random forest and XGBoost outperformed the Softmax linear classifier in different frequency bands, especially when the data pattern was complex. Therefore, in the model training phase, the frequency band power extracted from each species was used as input feature to train the random forest / XGBoost model to classify the motion state. It is worth emphasizing that the model structure of the present invention is versatile: in addition to the above-mentioned traditional models, it also supports replacement with deep learning models such as convolutional neural networks (CNN) or Transformer for end-to-end training. When there is large-scale cross-species data, deep models can further improve performance. But no matter what classifier is used, its input is represented by frequency band features, ensuring the comparability and interpretability of the output results between different models.

[0076] If the amount of data is sufficient or the real-time requirements are high, the model can be replaced with a deep learning model such as a convolutional neural network (CNN), a long short-term memory network (LSTM) or a Transformer. These models can further mine time series information and complex patterns. For example, CNN can combine time and space convolution to extract local combination features of spectral features; Transformer can fuse frequency band features of different time periods to capture long-range correlations. However, no matter how the underlying classifier changes, the present invention always uses frequency band features as an intermediate representation, so the structure of the model is universal and pluggable (i.e., universal and portable), which does not affect the cross-species applicability of the overall framework. A hierarchical training strategy can be adopted for model training: first train a sub-model for each species to ensure a good fit to its own data, and then consider cross-species fusion. This step-by-step training helps stabilize model performance and prepare for transfer learning.

[0077] Step S40: Acquire the current brain ECoG signal of the new target species, input the current brain ECoG signal into the trained decoding model for decoding, and output the movement intention information of the target species.

[0078] Specifically, the most important innovation of the present invention lies in the cross-species model adaptation strategy (i.e., cross-species alignment strategy). After training is completed, in order to achieve cross-species decoding, the present invention provides two complementary adaptation mechanisms: feature space reconstruction and transfer learning fine-tuning (i.e., transfer learning adaptation). Feature space reconstruction refers to mapping the frequency band features of the source species and the target species into a common feature space. For example, the feature vectors of different species can be normalized so that the numerical distribution of the same frequency band is similar, or dimensionality reduction methods such as principal component analysis (PCA) can be used to extract the main components of the spectrum shared across species. In this way, a feature representation that is consistent with the spectrum statistics of the training species is obtained on the new species. Transfer learning fine-tuning refers to using the model parameters pre-trained on the source species to retrain and adjust on a small amount of data from the target species. For example, the initial model is first trained with rabbit data, and then the model weights are further fine-tuned with macaque data. This approach allows the model to retain its sensitivity to common features while adjusting for the characteristics of the new species. For deep learning models, methods such as adversarial domain adaptation can also be used to make the features extracted by the model insensitive to species labels, thereby achieving species-independent discriminant representation. By combining these strategies, the model can align the feature distributions of different species, reducing the impact of interspecies differences on decoding performance. In other words, by performing cross-domain adaptation at both the feature and model levels, the present invention constructs a decoding model that is robust and shared across species.

[0079] First, feature space reconstruction aims to make the frequency band features of different species statistically comparable. One specific method involves normalizing and standardizing the features: for example, the mean and variance of the frequency band features for rabbit and macaque data are calculated separately, and the respective features are normalized to zero mean and unit variance. This process makes the numerical range of the same frequency band more consistent across species, preventing bias in the model caused by systematically large or small feature values for a particular species. Alternatively, a common feature space can be constructed through feature transformation. One implementation involves combining the data from the two species and performing principal component analysis (PCA) or linear discriminant analysis (LDA) to find a low-dimensional projection subspace that can simultaneously distinguish the motion states of the two species. In this subspace, the data from different species are more closely distributed, allowing the model to learn common patterns regardless of species origin.

[0080] Among them, the feature normalization process is as follows:

[0081]

[0082] in, represents the standardized feature value, which represents the normalized feature of the i-th sample in the j-th frequency band. After the mean-variance standardization, this value has zero mean and unit standard deviation, which helps to ensure the consistency of feature scales among different species; i，j represents the frequency band characteristics of the i-th sample in the j-th frequency band, μ j and σ j represent the mean and standard deviation of the j-th frequency band in the source species, respectively.

[0083] Principal component analysis (PCA) is used for feature space alignment: Z = XW, where The original feature matrix, PCA projection matrix, Z: common subspace representation after dimensionality reduction, n represents the number of samples, d represents the feature dimension, and k represents the dimension after dimensionality reduction.

[0084] Secondly, at the model level, the present invention uses transfer learning to further optimize cross-species performance. Specifically, there are two transfer methods: one is parameter transfer, which uses the model parameters trained on the source species as the initial values of the target species model, and then uses the target species data for a small number of iterative fine-tuning. This is similar to the process of fine-tuning a pre-trained model. It can use the frequency band discrimination ability learned from a large amount of source species data and correct the parts that do not match the target species. The total loss function in transfer learning can be expressed as:

[0085]

[0086] in, represents the taxonomic loss of source species, represents the fine-tuning loss of the target species, represents the domain alignment loss (such as adversarial loss or MMD), α and β are hyperparameters that control the weight of the loss term.

[0087] The second approach is multi-task learning. This involves constructing a unified model that simultaneously inputs data from multiple species. The loss function incorporates both the source and target species tasks. Through joint training, the model's hidden layers simultaneously serve to decode both species. Within this multi-task framework, a species-discriminating branch can be added. Through adversarial training, the features extracted by the model are rendered species-independent, thereby achieving truly species-agnostic representation learning. This is similar to the adversarial training concept used in domain adaptation. This strategy allows the model to automatically correct for species biases. For example, for macaque data, the model tends to rely more on high-frequency features, while for rabbit data, it prioritizes low-frequency features. However, the overall model structure and parameters are shared across species. After feature alignment and transfer learning, a universal decoding model is ultimately obtained. This model performs well on the source species and achieves near-source performance on the target species, without requiring retraining from scratch. This cross-species model significantly improves decoding generalization and verifies the existence of common features across species.

[0088] Current experimental results show that the decoding model can decode both rabbits and monkeys (rabbit architecture framework, monkey validation model), demonstrating cross-species decoding capabilities. To switch to other species, simply search for different frequency bands in the previous steps. For example, the characteristic frequency bands of small animals like rodents are concentrated in low frequencies, while those of primates like macaques and humans require shifting to higher frequencies to complete decoding.

[0089] The feasibility and superiority of the present invention are verified based on existing pre-experiments. First, in a single-species scenario, the present invention verifies the importance of frequency band feature selection. The models are trained using the motion ECoG data of rabbits and macaques, and the decoding accuracy of different frequency bands is compared. Figure 5 and Figure 6 As shown in the results, the rabbit data achieved the highest classification accuracy of 0.89 in the beta band (13-30Hz), while the macaque data reached a maximum of approximately 0.8 in the high gamma band (70-140Hz). This result is consistent with the findings of the spectral analysis of the present invention, demonstrating the necessity and effectiveness of selecting corresponding frequency band features for each species.

[0090] Next, the present invention evaluated the effectiveness of the cross-species transfer learning strategy. In a control experiment, the present invention used a model trained only with macaque data directly for rabbit data decoding. The accuracy dropped significantly to about the random level, indicating that the model cannot be generalized without adaptation. This highlights the necessity of cross-species adaptation. After adopting the transfer learning scheme proposed by the present invention, the accuracy of the model on rabbit data was significantly improved. Specifically, the present invention first trained the initial model with macaque data, and then fine-tuned it with a small amount of rabbit data. Through transfer learning, the model successfully learned the low-frequency information unique to rabbits, while retaining the useful common features in the macaque model.

[0091] The technical effects that the present invention can bring are as follows:

[0092] (1) Adapting to cross-species spectrum differences: By introducing a species-specific frequency band selection mechanism, the impact of frequency component differences between different species on decoding performance is resolved, ensuring that the model can select the most effective feature frequency band for each species, thereby improving the transfer learning effect.

[0093] (2) Improve the generalization ability of the model: Using transfer learning and feature space alignment strategies, the model can generalize the rules learned from one species to another species, and maintain a high decoding accuracy without the need for a large amount of new species data.

[0094] (3) Ensure the interpretability and universality of the solution: Utilize physiologically interpretable features such as frequency band power as a bridge to build a universal decoding framework. This framework does not rely on specific hardware, has a clear implementation principle, and can replace classification algorithms or add new species data as needed, with good modularity and scalability.

[0095] The key points of the present invention are as follows:

[0096] (1) Species-specific frequency band selection mechanism: A method is proposed to select characteristic frequency bands for decoding based on the differences in the neural signal spectra of different species. For each species, the frequency range most relevant to its movement intention is determined (e.g., rodents focus on low-frequency beta waves, primates focus on high-frequency gamma waves), and different frequency bands are adaptively given different weights in the model. This mechanism ensures that the model uses the most effective information for each species, improving the accuracy of cross-species decoding.

[0097] (2) Transfer learning strategy based on spectral differences: A model transfer learning process specifically for spectral differences was designed. This includes feature space alignment (normalizing / transforming frequency band features to make the feature distribution of different species consistent) and cross-domain model fine-tuning (using the model parameters trained on the source species to fine-tune the target species). This strategy enables the same model to be applied to decoding tasks of multiple species, achieving "one-time training, multi-species generalization", significantly improving the adaptability of the model.

[0098] (3) Cross-species feature mapping method: A method for mapping neural features of different species into a common representation space is invented. Species-specific components in the features are removed through statistical and machine learning methods (such as principal component analysis and adversarial training), and spectral features shared between species are extracted. This method ensures that the decoding model can extract motion-related information without being aware of species differences, theoretically resolving the obstacles caused by species differences and possessing broad versatility.

[0099] (4) Universal decoding framework driven by frequency band features: A universal decoding system framework based on physiological frequency band features was constructed. The framework has clear modules, including signal preprocessing, frequency band extraction, model training, and migration adaptation. Each module is independent and can be organically connected, suitable for different hardware platforms and algorithm upgrades. Without changing the overall framework, the classification model can be easily replaced (for example, from random forest to CNN) or new species data can be added for training, reflecting a high degree of scalability and compatibility.

[0100] Compared with existing similar brain-computer interface technology solutions, this invention has the following outstanding advantages:

[0101] (1) The generalization ability of the decoding model is greatly improved: Traditional models are often limited to specific species. The present invention uses frequency band alignment and transfer learning to achieve a single model applicable to multiple species. Compared with existing methods that require training models for different species separately, the cross-species model of the present invention can be quickly deployed on new species with excellent performance, significantly broadening the scope of application of brain-computer interfaces.

[0102] (2) Strong feature interpretability and clear biological significance: The present invention uses power in different frequency bands as the main feature, which directly corresponds to neural oscillation activity, so the model decision has clear physiological significance. The decoding results can be interpreted from a frequency perspective, which facilitates researchers to understand the neural mechanism of motor intention encoding in different species. This interpretability is difficult to provide with complex deep networks, providing an intuitive analytical method for cross-species research.

[0103] (3) Algorithm compatibility and modular design: The framework of the present invention adopts a modular design, with each link loosely coupled, and algorithms can be flexibly replaced or upgraded as needed. For example, the preprocessing module can be connected to different denoising algorithms, the feature extraction module can be expanded with phase information or brain network features, and the classifier module can be replaced with the latest deep learning architecture, etc., without changing the overall process. This compatible and open design makes the present invention have long-term vitality and can integrate future new technological achievements.

[0104] (4) Advantages in decoding performance and robustness: By selecting the optimal frequency band features for each species and using an integrated learning model with excellent performance, the present invention outperforms many traditional methods with fixed frequency bands in terms of decoding accuracy. At the same time, pre-processing such as ANT denoising improves signal quality and reduces the impact of noise interference on decoding. Coupled with the potential of combining multi-model methods (such as soft voting to fuse the outputs of different classifiers), the system can maintain stable and reliable performance even in complex environments.

[0105] (5) Research and application value: This invention can not only be used to compare brain functions of different species in experimental research, but also to build cross-species brain-computer interface systems in practical applications. For example, decoding models that have been proven effective in animal experiments can be transferred to humans, thereby accelerating the development of clinical brain-computer interfaces. This "cross-species migration" capability from animals to humans is rarely addressed in existing technologies, and this invention provides a feasible implementation path.

[0106] Furthermore, the design of the present invention is highly flexible, allowing for a variety of changes and extensions to suit different application requirements:

[0107] (1) Replacement of model structure: Although the present invention adopts a frequency band + integrated learning model, it is entirely possible to introduce a deep learning model to replace or enhance the decoding performance. For example, a neural network that integrates convolution and attention mechanisms can be designed to directly input multi-channel original ECoG signals, allowing the model to automatically learn the optimal frequency band characteristics of each species. This end-to-end approach may further improve the accuracy. However, no matter how the model is replaced, the idea of cross-species adaptation of the present invention can still be maintained by adding spectral constraints or multi-task losses in the middle layer. Similarly, if a more efficient transfer learning algorithm appears in the future (such as a cross-domain adaptation method based on meta-learning), it can also be directly integrated into the framework of the present invention to replace the current fine-tuning strategy.

[0108] (2) Expansion of frequency band features: The present invention currently focuses on the power spectrum amplitude features, and can be expanded to a richer combination of frequency domain features in the future. For example, the phase synchronization feature is added to measure the phase coupling degree of different brain regions in a specific frequency band as a feature for identifying the movement state; or the time-frequency graph (such as the wavelet transform energy graph) is introduced as the input of the deep model, allowing the model to extract the discriminant pattern from it. It is also possible to combine cross-band coupling features (CFC, Cross-FrequencyCoupling), such as the indicator of high-frequency oscillation amplitude modulating low-frequency phase, to compare the similarities and differences in neural modulation mechanisms between species. The addition of these features is expected to further enhance the ability of cross-species models to characterize complex neural dynamics.

[0109] (3) Expansion of applicable species and scenarios: In addition to rabbits and macaques, the solution of the present invention can be extended to more species and more brain regions. For example, it can be applied to comparative decoding between rodents such as mice and rats and primates; one can even try to migrate animal models to human data as a bridge for the transition from preclinical research results to clinical applications. In addition, this method is not limited to signal decoding of the motor cortex, but is also applicable to neural signals of brain regions such as the visual cortex and auditory cortex. It is only necessary to adjust the frequency band selection strategy according to different functional brain regions. For example, vision-related may focus on γ-band synchronization, spatial navigation-related may focus on θ-band oscillations, etc. By expanding the application scenarios, the present invention can become a universal cross-species brain activity pattern decoding tool.

[0110] Further, if Figure 7 As shown, based on the above-mentioned cross-species ECoG signal decoding method based on ensemble learning, the present invention also provides a cross-species ECoG signal decoding system based on ensemble learning, wherein the cross-species ECoG signal decoding system based on ensemble learning includes:

[0111] A data acquisition and preprocessing module 51 is used to acquire brain ECoG signals of different species when performing motor tasks, and perform data preprocessing on the brain ECoG signals to obtain preprocessed target brain ECoG signals;

[0112] A frequency band feature extraction and analysis module 52 is used to adaptively select frequency band features for different species based on the target brain ECoG signal;

[0113] A decoding model construction and training module 53 is used to construct a decoding model and train the decoding model using the frequency band features to obtain a trained decoding model;

[0114] The movement intention information recognition module 54 is used to obtain the current brain ECoG signal of the new target species, input the current brain ECoG signal into the trained decoding model for decoding, and output the movement intention information of the target species.

[0115] Further, if Figure 8 As shown, based on the above-mentioned cross-species ECoG signal decoding method and system based on ensemble learning, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 8 Only some of the components of the terminal are shown, but it should be understood that implementation of all of the shown components is not required, and more or fewer components may be implemented instead.

[0116] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory of the terminal. In other embodiments, the memory 20 may also be an external storage device of the terminal, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (SecureDigital, SD) card, a flash card (Flash Card), etc. equipped on the terminal. Furthermore, the memory 20 may also include both an internal storage unit of the terminal and an external storage device. The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code of the installation terminal. The memory 20 may also be used to temporarily store data that has been output or is to be output. In one embodiment, a cross-species ECoG signal decoding program 40 based on ensemble learning is stored on the memory 20, and the cross-species ECoG signal decoding program 40 based on ensemble learning can be executed by the processor 10, thereby realizing the cross-species ECoG signal decoding method based on ensemble learning in the present application.

[0117] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, configured to execute program codes or process data stored in the memory 20, such as executing the cross-species ECoG signal decoding method based on ensemble learning.

[0118] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface. The processor 10, memory 20, and display 30 of the terminal communicate with each other via a system bus.

[0119] In one embodiment, when the processor 10 executes the cross-species ECoG signal decoding program 40 based on ensemble learning in the memory 20 , the steps of the cross-species ECoG signal decoding method based on ensemble learning are implemented.

[0120] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a cross-species ECoG signal decoding program based on ensemble learning, and when the cross-species ECoG signal decoding program based on ensemble learning is executed by a processor, the steps of the cross-species ECoG signal decoding method based on ensemble learning as described above are implemented.

[0121] In summary, the present invention provides a cross-species ECoG signal decoding method, system, terminal, and computer-readable storage medium based on ensemble learning. The method includes: collecting brain ECoG signals of different species when performing motor tasks, performing data preprocessing on the brain ECoG signals, and obtaining preprocessed target brain ECoG signals; adaptively selecting frequency band features for different species based on the target brain ECoG signals; constructing a decoding model, and training the decoding model using the frequency band features to obtain a trained decoding model; obtaining the current brain ECoG signal of a new target species, inputting the current brain ECoG signal into the trained decoding model for decoding, and outputting the movement intention information of the target species. The present invention utilizes interpretable frequency band power features as a link, combines frequency band feature analysis with transfer learning strategies, and constructs a universal cross-species neural motion decoding framework to achieve unified decoding of movement intentions of different species, thereby improving the cross-species adaptability and transfer performance of the decoding model.

[0122] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or terminal comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or terminal. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or terminal comprising the element.

[0123] Of course, those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware (such as a processor, controller, etc.) through a computer program. The program can be stored in a computer-readable storage medium that can be read by a computer. When the program is executed, it can include the processes in the above-described method embodiments. The computer-readable storage medium can be a memory, a magnetic disk, an optical disk, etc.

[0124] It should be understood that the application of the present invention is not limited to the above examples. For those skilled in the art, improvements or changes can be made based on the above description. All these improvements and changes should fall within the scope of protection of the claims attached to the present invention.

Claims

1. A cross-species ECoG signal decoding method based on ensemble learning, characterized in that: The cross-species ECoG signal decoding method based on ensemble learning includes: collecting brain ECoG signals of different species when performing motor tasks, performing data preprocessing on the brain ECoG signals, and obtaining preprocessed target brain ECoG signals; Adaptively selecting frequency band features for different species based on the target brain ECoG signal; Constructing a decoding model, and training the decoding model using the frequency band features to obtain a trained decoding model; The current brain ECoG signal of the new target species is obtained, the current brain ECoG signal is input into the trained decoding model for decoding, and the movement intention information of the target species is output.

2. The cross-species ECoG signal decoding method based on ensemble learning according to claim 1, characterized in that: The collecting of brain ECoG signals of different species when performing a motor task, and performing data preprocessing on the brain ECoG signals to obtain preprocessed target brain ECoG signals specifically include: Using implantable microelectrode arrays to collect brain ECoG signals from different species while they perform similar tasks or behavioral conditions; The brain ECoG signal is subjected to band-pass filtering, artifact removal, and adaptive noise template denoising to obtain a preprocessed target brain ECoG signal.

3. The cross-species ECoG signal decoding method based on ensemble learning according to claim 1 or 2, characterized in that: The brain ECoG signal is subjected to data preprocessing to obtain a preprocessed target brain ECoG signal, and then the method further includes: The pre-processed target brain ECoG signals of different species are standardized and synchronously segmented according to a unified format.

4. The cross-species ECoG signal decoding method based on ensemble learning according to claim 3, characterized in that: Adaptively selecting frequency band features for different species based on the target brain ECoG signal specifically includes: Calculate the power spectral density of each target brain ECoG signal segment, perform Fourier transform on the segmented signals using the Welch algorithm, and average to obtain spectrum estimation; The spectrum is integrated according to typical EEG bands, the average power value of each frequency band is extracted, different frequency bands are defined, and the area under the power spectrum density curve in each frequency band is calculated as the energy feature of the corresponding frequency band. Each target brain ECoG signal is converted into a multidimensional frequency band feature vector. The frequency band characteristics of different species were compared and analyzed, and the distribution of energy in each frequency band of each species under different motion states was statistically analyzed to obtain the frequency range that meets the requirements.

5. The cross-species ECoG signal decoding method based on ensemble learning according to claim 3, characterized in that: The constructing of the decoding model and training the decoding model using the frequency band features to obtain a trained decoding model specifically includes: A decoding model is constructed using an ensemble learning type of machine learning model, wherein the ensemble learning type of machine learning model includes a random forest model and an XGBoost gradient boosting decision tree model; The multi-dimensional frequency band feature vector is used as input, and a supervised learning method is adopted to train the decoding model using the existing data labels of each species. The output is the predicted movement state category, thereby obtaining a trained decoding model.

6. The cross-species ECoG signal decoding method based on ensemble learning according to claim 5, characterized in that: The decoding model is trained using a hierarchical training strategy, where a sub-model is first trained for each species to ensure a good fit to its own data, and then cross-species fusion is considered.

7. The cross-species ECoG signal decoding method based on ensemble learning according to claim 5, characterized in that: The decoding model is constructed, and the decoding model is trained using the frequency band features to obtain a trained decoding model, and then the following steps are further included: aligning the feature spaces of different species using a cross-species alignment strategy and adjusting the decoding model; The cross-species alignment strategy includes feature space reconstruction and transfer learning fine-tuning. The feature space reconstruction is used to map the frequency band features of the source species and the target species into a common feature space. The transfer learning fine-tuning is used to use the model parameters pre-trained on the source species and retrain and adjust them on a preset amount of target species data. The transfer learning fine-tuning includes parameter transfer and multi-task learning. The parameter transfer is used to use the model parameters trained for the source species as the initial values of the target species model, and then iteratively fine-tune with the data of the target species. The multi-task learning is used to construct a unified model that simultaneously inputs multi-species data, including the source species task and the target species task in the loss function. Through joint training, the hidden layer of the unified model serves the decoding of the two species at the same time.

8. A cross-species ECoG signal decoding system based on ensemble learning, characterized by: The cross-species ECoG signal decoding system based on ensemble learning includes: A data acquisition and preprocessing module is used to collect brain ECoG signals of different species when performing motor tasks, perform data preprocessing on the brain ECoG signals, and obtain preprocessed target brain ECoG signals; A frequency band feature extraction and analysis module, configured to adaptively select frequency band features for different species based on the target brain ECoG signal; A decoding model construction and training module is used to construct a decoding model and train the decoding model using the frequency band features to obtain a trained decoding model; The movement intention information recognition module is used to obtain the current brain ECoG signal of a new target species, input the current brain ECoG signal into a trained decoding model for decoding, and output the movement intention information of the target species.

9. A terminal, characterized in that: The terminal includes: a memory, a processor, and a cross-species ECoG signal decoding program based on ensemble learning stored in the memory and executable on the processor. When the cross-species ECoG signal decoding program based on ensemble learning is executed by the processor, the steps of the cross-species ECoG signal decoding method based on ensemble learning are implemented.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a cross-species ECoG signal decoding program based on ensemble learning. When the cross-species ECoG signal decoding program based on ensemble learning is executed by a processor, the steps of the cross-species ECoG signal decoding method based on ensemble learning are implemented.

Images