Neural signal control parameter migration system and method based on cross-species multi-modal information
By constructing a cross-species multimodal data acquisition and spatiotemporal co-representation computation module, the differences between species were eliminated, and parameter transfer and individualized optimization of mouse neural signals to human neural signals were realized. This solved the problems of cross-species neural data feature alignment and parameter transfer, and improved the efficiency and accuracy of neural regulation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-11
- Publication Date
- 2026-03-13
AI Technical Summary
Cross-species neural data differ in sampling rate, channel mapping, and noise level, leading to inconsistent feature distributions and making it difficult to achieve effective feature alignment. Existing transfer learning frameworks are ill-suited for multimodal, non-stationary, high-dimensional neural signal scenarios with limited samples, and struggle to achieve adaptive fine-tuning and closed-loop optimization of parameters under online feedback conditions, hindering the effective translation from animal experiments to human applications.
A cross-species multimodal data acquisition module was constructed. Data was processed through standardization and unified coding. Cross-species temporal adversarial coding and spatial adversarial coding units were used to eliminate inter-species differences and form a cross-species spatiotemporal co-representation. The parameter transfer and fine-tuning module was used to realize the mapping and individualized optimization of mouse data to human neural signal control parameters.
This study achieved effective alignment of mouse and human neural signals in a unified spatiotemporal feature space, improved the consistency of cross-species neural response patterns, and enhanced the safety, efficiency, and accuracy of neural modulation experiments. It also provides technical support for the intelligent and personalized development of brain-computer interfaces and neural modulation strategies.
Smart Images
Figure CN121658902A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial intelligence and neuroscience technology, and in particular relates to a neural signal control parameter transfer system and method based on cross-species multimodal information. Background Technology
[0002] In recent years, cross-species neural information modeling and brain-computer interface control have become important directions in the field of neural engineering. Mice, as an important model organism for human nervous system research, share high similarities with humans in anatomical structure, physiological mechanisms, and gene regulation, providing a feasible basis for cross-species research. Particularly in the fields of proteomics and genomics, cross-species comparative studies have revealed significant functional conservation between humans and mice in neuronal development, synaptic transmission, and gene expression regulation. Experiments have shown that evoked potentials in mice and humans under cortical stimulation are significantly similar, especially in the thalamus-cortex circuit, where hyperpolarization and rebound mechanisms have been validated in different species. This conservation of neural responses suggests that mouse-based neural signal control and response modeling can provide a valuable reference for parameter optimization, strategy validation, and simulation in humans.
[0003] However, in practical applications, due to the limitations of human brain signal acquisition and the high cost of annotation, cross-species data exhibit significant differences at the spatial, temporal, and modal levels, severely limiting the efficiency of transferring regulatory parameters and response patterns. Specifically, existing technologies face the following key problems: differences in sampling rate, channel mapping, and noise levels among cross-species neural data lead to inconsistent feature distributions, making it difficult to achieve effective feature alignment directly; existing transfer learning frameworks are mostly derived from highly labeled data environments such as natural language or computer vision, making it difficult to directly adapt to multimodal, non-stationary, high-dimensional neural signal scenarios with limited samples; there are still significant shortcomings in cross-species spatiotemporal consistency modeling, cross-modal alignment of EEG and functional magnetic resonance imaging, and joint optimization of parameter response representations; at the same time, neural modulation experiments require strict safety boundaries, and existing methods struggle to achieve adaptive fine-tuning and closed-loop optimization of parameters under online feedback conditions. These technical deficiencies severely hinder the effective translation from animal experiments to human applications, becoming a key bottleneck in realizing intelligent and personalized neural control systems. To solve these problems, it is necessary to develop new methods that can integrate multimodal neural data, construct unified cross-species representations, and support efficient parameter transfer and fine-tuning. Summary of the Invention
[0004] To address the aforementioned technical challenges, this invention proposes a neural signal control parameter transfer system and method based on cross-species multimodal information. This system enables the transfer of multimodal neural signal representation and control parameters from mice to humans, thereby promoting the intelligent and personalized development of brain-computer interfaces and neural modulation strategies.
[0005] To achieve the above objectives, this invention provides a neural signal control parameter transfer system based on cross-species multimodal information, comprising: A cross-species multimodal data acquisition module is used to collect and preprocess multimodal neural data from mice and humans to form samples in a unified format. A cross-species spatiotemporal co-representation computation module is used to extract temporal features and represent spatial patterns from the multimodal neural data, construct a cross-species spatiotemporal feature representation space, and eliminate distribution differences between species. The cross-species parameter transfer and fine-tuning module is used to map mouse experimental data into parameters suitable for human neural signal control and response modeling based on the cross-species spatiotemporal feature representation, and to perform individualized optimization and strategy recommendation.
[0006] Optionally, the cross-species multimodal data acquisition module includes a standardization and unified coding unit, used to perform filtering, artifact removal, resampling, normalization, time alignment, and unified format coding on the acquired data.
[0007] Optionally, the cross-species spatiotemporal co-representation computing module includes a cross-species temporal adversarial coding unit, a cross-species spatial adversarial coding unit, and a spatiotemporal information fusion unit; The cross-species temporal adversarial coding unit is used to receive EEG data from mice and humans, extract the temporal-spectral features of both for adversarial training, and obtain cross-species temporal representations. The cross-species spatial adversarial coding unit is used to receive MRI / fMRI data from mice and humans, extract the anatomical-functional spatial patterns of both for adversarial training, and obtain cross-species spatial representations. The spatiotemporal information fusion unit is used to align and fuse cross-species temporal representations and cross-species spatial representations to output a cross-species spatiotemporal co-representation.
[0008] Optionally, the cross-species temporal adversarial coding unit includes a temporal feature extractor and a first domain discriminator; The time-series feature extractor is used to extract time-series features from EEG data; The first domain discriminator is used to erase species information through adversarial training.
[0009] Optionally, the cross-species spatial adversarial coding unit includes a spatial pattern encoder and a second domain discriminator; The spatial pattern encoder is used to extract the spatial pattern of MRI / fMRI data; The second domain discriminator is used to erase species information through adversarial training.
[0010] Optionally, the cross-species parameter transfer and fine-tuning module includes a mouse neural signal control parameter recommendation unit and a human neural signal control parameter fine-tuning unit; The mouse neural signal control parameter recommendation unit is used to receive preprocessed mouse neural signal control experimental data, learn parameter-response mapping relationships in a cross-species spatiotemporal co-representation space, and infer candidate sets of human neural signal control parameters. The human neural signal control parameter fine-tuning unit is used to update the model weights using an efficient parameter fine-tuning method after acquiring feedback data from human neural signal control experiments, so as to form an individualized human neural signal control model.
[0011] Optionally, the human neural signal control parameters output by the mouse neural signal control parameter recommendation unit include stimulation frequency, stimulation amplitude, pulse width, duty cycle, stimulation duration, stimulation site / channel combination, and stimulation sequence.
[0012] On the other hand, to achieve the above objectives, the present invention also provides a method for transferring neural signal control parameters based on cross-species multimodal information, including: A cross-species spatiotemporal co-representation computation module was trained using publicly available cross-species multimodal neural data to construct cross-species spatiotemporal co-representations; A neural signal control experiment was conducted on mice to collect multimodal neural signals and neural signal control parameters. After preprocessing and standardization, a neural signal control parameter recommendation unit was trained in mice to generate candidate human neural signal control parameters. Human neural signal control experiments were conducted based on the candidate's neural signal control parameters. Feedback data was collected, and after preprocessing and standardization, the model parameters were finely tuned to form a human neural signal control parameter fine-tuning unit. Human multimodal data is input into the human neural signal control parameter fine-tuning unit to obtain recommended human neural signal control parameters for individualized optimization and strategy recommendation.
[0013] Technical Effects of this Invention: This invention discloses a neural signal control parameter transfer system and method based on cross-species multimodal information. By constructing a system for cross-species multimodal data acquisition, spatiotemporal co-representation calculation, and parameter transfer and fine-tuning, it achieves effective alignment of mouse and human neural signals in a unified spatiotemporal feature space, significantly improving the consistency of cross-species neural response patterns. This invention utilizes adversarial training and distribution distance constraints to eliminate the distribution differences of temporal and spatial features between species, and forms a unified cross-species spatiotemporal co-representation through attention fusion. Based on this representation, the system can transfer neural signal control parameters learned from mouse experiments to human applications, and achieve individualized optimization with limited human experimental data using efficient parameter fine-tuning strategies. This method effectively solves the problems of difficult cross-species parameter transfer, low experimental efficiency, and insufficient accuracy of individualized modeling, significantly improving the safety, efficiency, and accuracy of neural modulation experiments, and providing reliable technical support for the intelligent and individualized development of brain-computer interfaces and neural modulation strategies. Attached Figure Description
[0014] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the structure of a neural signal control parameter transfer system based on cross-species multimodal information according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating a neural signal control parameter transfer method based on cross-species multimodal information, according to an embodiment of the present invention. Detailed Implementation
[0015] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.
[0016] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.
[0017] like Figure 1 As shown, this embodiment provides a neural signal control parameter transfer system based on cross-species multimodal information, including: A cross-species multimodal data acquisition module is used to collect and preprocess multimodal neural data from mice and humans to form samples in a unified format. A cross-species spatiotemporal co-representation computation module is used to extract temporal features and represent spatial patterns from the multimodal neural data, construct a cross-species spatiotemporal feature representation space, and eliminate distribution differences between species. The cross-species parameter transfer and fine-tuning module is used to map mouse experimental data into parameters suitable for human neural signal control and response modeling based on the cross-species spatiotemporal feature representation, and to perform individualized optimization and strategy recommendation.
[0018] Furthermore, the cross-species multimodal data acquisition module includes a standardization and unified coding unit, which is used to perform filtering, artifact removal, resampling, normalization, time alignment and unified format coding on the acquired data.
[0019] Specifically, the cross-species multimodal data acquisition module uniformly completes the acquisition, cleaning, and clock alignment of publicly available and experimental data from mice and humans, outputting standardized data frames. This module is configured with EEG channels (at least 32 channels, sampling rate at least 500Hz) for signal acquisition. The system supports impedance monitoring (Hz) and MRI / fMRI image import channels (compatible with DICOM / NIfTI), and can be expanded to include EMG, EOG, and auxiliary measurement channels such as heart rate / respiration. For preprocessing and synchronization, EEG undergoes bandpass filtering, power frequency notch filtering, ICA / robust PCA artifact removal, segmented cropping, and normalization. Images undergo head motion correction, temporal alignment, spatial registration, denoising regression, and standardization. Synchronization pulses or trigger events are used to eliminate cross-device / cross-species clock drift to construct a unified cross-modal time base. For data standardization and annotation, a unified sampling rate, channel / ROI mapping, and coordinate system are implemented. A cross-species channel mapping table based on homologous cortical regions or networks is established, and parameters such as stimulus intensity, frequency, waveform, and target points, as well as labels such as experimental status, task, and behavioral scores, are written into metadata. For quality monitoring, real-time / offline signal-to-noise ratio, artifact rate, and channel health monitoring, along with a full-process audit log, are provided to ensure the availability and traceability of input data.
[0020] Furthermore, the cross-species spatiotemporal co-representation computing module includes a cross-species temporal adversarial coding unit, a cross-species spatial adversarial coding unit, and a spatiotemporal information fusion unit; The cross-species temporal adversarial coding unit is used to receive EEG data from mice and humans, extract the temporal-spectral features of both for adversarial training, and obtain cross-species temporal representations. The cross-species spatial adversarial coding unit is used to receive MRI / fMRI data from mice and humans, extract the anatomical-functional spatial patterns of both for adversarial training, and obtain cross-species spatial representations. The spatiotemporal information fusion unit is used to align and fuse cross-species temporal representations and cross-species spatial representations to output a cross-species spatiotemporal co-representation.
[0021] Specifically, the cross-species spatiotemporal co-representation computation module performs domain alignment and feature extraction for both temporal (EEG) and spatial (fMRI) dimensions, and forms a unified spatiotemporal representation space in the fusion unit to achieve alignment of human-mouse cross-species neural response features. The temporal adversarial coding unit employs a hybrid structure of temporal Transformer and gated convolution to capture short-to-medium-to-long-range dependencies. A domain discriminator is introduced to align the human-mouse temporal distribution. The loss function is composed of reconstruction terms such as temporal mean square error and frequency domain spectral consistency, as well as adversarial / distribution distance terms such as domain cross-entropy and MMD. The spatial adversarial coding unit extracts cortical / network-level spatial activation maps based on a hybrid backbone of Swin-Transformer or CNN-Transformer and aligns the human-mouse latent spatial distribution with an adversarial domain adapter. It is also supplemented with structural consistency regularization (such as graph Laplacian constraints and functional connectivity consistency terms for adjacent ROIs) to maintain spatial topological stability. The spatiotemporal information fusion unit uses cross-modal attention and gated residual fusion to jointly encode temporal-spectral features, spatial-network features, and stimulus context (parameters, target points, electric fields / BOLD priors, etc.) and outputs a cross-species spatiotemporal co-representation vector Z, which serves as a unified input for subsequent transfer and parameter recommendation.
[0022] Furthermore, the cross-species temporal adversarial coding unit includes a temporal feature extractor and a first domain discriminator; The time-series feature extractor is used to extract time-series features from EEG data; The first domain discriminator is used to erase species information through adversarial training.
[0023] Specifically, the cross-species temporal adversarial coding unit consists of a temporal feature extractor. Domain discriminator Composition, in which For temporal feature extractors (encoders), the parameters are: ; For the domain discriminator, the parameters are Given an input EEG fragment and its domain tags (For example Represents the mouse domain. (Representing the human domain), firstly, temporal feature representations are obtained through a feature extractor. Suppose there are a total of [number] items in the current mini-batch. The nth sample, the nth The original input and reconstruction result of each sample are denoted as follows: and The corresponding features are The domain tag is The cross-species temporal adversarial coding unit is optimized under the following joint loss.
[0024] (1) Reconstruction Loss (MSE): ; in, The number of samples; L2 norm, used to measure the first norm of the second norm. Original EEG fragments of each sample With reconstructed fragments The Euclidean distance between them To characterize the overall reconstruction error, the reconstruction loss is used to ensure that the extracted temporal features retain sufficient original signal information.
[0025] (2) Counter-loss (BCE): ; in, For the domain discriminator to the first One sample belongs to the human domain ( The predicted probability of ) The standard binary cross-entropy loss is used to drive the domain discriminator to distinguish different species. This is used to train the feature extractor. At that time A gradient inversion layer is added before the data to remove species-related information by maximizing the domain discrimination error, thereby obtaining more domain-invariant temporal features.
[0026] (3) Distribution Distance Loss (MMD): To further align cross-species characteristic distributions, a distribution distance constraint based on Maximum Mean Discrepancy (MMD) is introduced: ; in, For distribution distance loss; For the first The representation of a mouse sample in the output space of the feature extractor; For the first The representation of a human sample in this space; and These represent the number of samples from the mouse domain and the human domain in the current batch, respectively. and The first The mouse sample and the first The representation of a human sample in the output space of the feature extractor; A kernel function (such as a Gaussian kernel) is used to measure the similarity between two features. This is achieved by minimizing... This allows the distribution of mouse characteristics to be as close as possible to the distribution of human characteristics in the regenerative nucleus Hilbert space, achieving cross-species distribution alignment at the temporal characteristic level.
[0027] Furthermore, the cross-species spatial adversarial coding unit includes a spatial pattern encoder and a second domain discriminator; The spatial pattern encoder is used to extract the spatial pattern of MRI / fMRI data; The second domain discriminator is used to erase species information through adversarial training.
[0028] Specifically, the cross-species spatial adversarial coding unit comprises a spatial pattern encoder and a domain discriminator, and encodes voxel / cortical region maps or functional connectivity maps constructed for MRI / fMRI. It uses the reconstruction loss, adversarial loss, and distribution distance loss calculated in the cross-species temporal adversarial coding unit to achieve cross-species spatial representation alignment.
[0029] Furthermore, the cross-species parameter transfer and fine-tuning module includes a mouse neural signal control parameter recommendation unit and a human neural signal control parameter fine-tuning unit; The mouse neural signal control parameter recommendation unit is used to receive preprocessed mouse neural signal control experimental data, learn parameter-response mapping relationships in a cross-species spatiotemporal co-representation space, and infer candidate sets of human neural signal control parameters. The human neural signal control parameter fine-tuning unit is used to update the model weights using an efficient parameter fine-tuning method after acquiring feedback data from human neural signal control experiments, so as to form an individualized human neural signal control model.
[0030] Specifically, the cross-species parameter transfer and fine-tuning module completes the transfer, recommendation, and individualized optimization of mouse experimental knowledge to human regulatory parameters. The mouse neural signal control parameter recommendation unit is based on the multimodal Z-vectors of the mouse side and corresponding stimulus-response pairs. It uses a hybrid head of metric / contrast learning and regression to learn the mapping of "representation → parameter" and outputs uncertainty and safety confidence intervals. The human neural signal control parameter fine-tuning unit uses efficient parameter fine-tuning strategies (such as LoRA, Adapter, or Prefix-Tuning) on the common representation backbone and recommender under data-scarce conditions. Combined with safety constraints such as dose / frequency / current density upper limits and MRI environment compatibility, it generates an individualized human neural signal control model. The online closed-loop optimization subunit continuously collects feedback from behavioral, clinical scales, and physiological indicators during human trials. It uses small-step online updates (such as cumulative or weighted sliding windows based on Reptile) for robust iteration, so that the recommended parameters adaptively converge with individual status and stage results.
[0031] Furthermore, the human neural signal control parameters output by the mouse neural signal control parameter recommendation unit include stimulation frequency, stimulation amplitude, pulse width, duty cycle, stimulation duration, stimulation site / channel combination, and stimulation sequence.
[0032] Furthermore, the system also includes a data management and security compliance module that runs through the above processes. This module provides data tiered anonymization, fine-grained access control, and access auditing for multi-institutional collaboration scenarios. It supports cross-domain collaboration paradigms such as federated learning and knowledge distillation, and complies with ethical approval, privacy protection, and relevant compliance requirements. The experiment execution and recording module, along with the visualization monitoring and maintenance module, are responsible for distributing control signal schemes and coordinating device linkage, managing session-level logs and early warnings, and presenting key indicators such as time-frequency graphs, functional connection heatmaps, ROI-level field distribution / electric field simulations, and parameter trajectories in a visual format. It outputs archiveable interim and final evaluation reports, supporting the stable operation and interpretable decision-making of the system in scientific research and engineering applications.
[0033] like Figure 2 As shown, this embodiment provides a method for transferring neural signal control parameters based on cross-species multimodal information, including: A cross-species spatiotemporal co-representation computation module was trained using publicly available cross-species multimodal neural data to construct cross-species spatiotemporal co-representations; A neural signal control experiment was conducted on mice to collect multimodal neural signals and neural signal control parameters. After preprocessing and standardization, a neural signal control parameter recommendation unit was trained in mice to generate candidate human neural signal control parameters. Human neural signal control experiments were conducted based on the candidate's neural signal control parameters. Feedback data was collected, and after preprocessing and standardization, the model parameters were finely tuned to form a human neural signal control parameter fine-tuning unit. Human multimodal data is input into the human neural signal control parameter fine-tuning unit to obtain recommended human neural signal control parameters for individualized optimization and strategy recommendation.
[0034] Specifically, the implementation process of this embodiment includes: During the cross-species co-characterization pre-training phase: S1. Pre-training on public datasets: Select publicly available EEG / MRI / fMRI data of humans and mice, and complete cleaning, registration, bandwidth and coordinate system I, time axis alignment and metadata completion according to the unified specifications of the acquisition module to build a standard sample set that can be directly used for cross-modal / cross-species learning; S2, Temporal / Spatial Adversarial Coding: Temporal adversarial coding and spatial adversarial coding units are trained separately to gradually align the human-mouse potential distribution in the temporal-spectrum and spatial-network. Reconstruction loss (MSE), domain adversarial loss (domain discriminator cross-entropy), and distribution distance loss (MMD) are used in combination, and structural consistency regularization (such as ROI graph Laplacian and functional connectivity matrix consistency) is introduced to ensure spatial topological stability. S3. Spatiotemporal fusion and task self-supervision: Spatiotemporal information fusion units are jointly trained using self-supervised tasks such as occlusion reconstruction, cross-modal comparison, and spatiotemporal order discrimination to obtain cross-species spatiotemporal co-representation vector Z, and the backbone weights are solidified as the initialization parameter library for subsequent transfer.
[0035] In the mouse end parameter learning and transfer preparation stage: S4. Mouse Batch Experiments and Standardization: Conduct multi-formula control signal experiments on a large number of mice, and fully record the control parameters (frequency, amplitude, waveform, duty cycle, channel / target, timing) and their EEG / fMRI responses and behavioral / physiological indicators; complete artifact removal, normalization, session-level quality control and data auditing according to unified procedures; S5. Mouse neural signal control parameter recommendation unit training: The parameter recommender is trained with the co-representation Z and stimulus-response paired samples. The regression head directly outputs candidate parameter vectors (including frequency / amplitude / duty cycle / waveform / target point), and the metric head learns the comparison space of "parameter-response" to support similar trial retrieval and cold start. In addition, the parameter confidence interval and safety boundary mask are estimated through uncertainty modeling (such as MC Dropout or deep ensemble). S6. Cross-species migration preparation: Conduct domain invariance testing (to examine the alignment of human-mouse Z distribution), sensitivity analysis (to assess robustness to modality loss and noise), and simulation evaluation (consistency between electric field and network hierarchical response). Based on this, a set of transferable priors and safety thresholds is formed to provide boundaries and initiation parameters for target domain experiments.
[0036] In the small-sample fine-tuning and individualized modeling stage on the human end: S7. Small-sample human trials and pretreatment: Under ethical approval and safety threshold constraints, exploratory trials were conducted on a small number of human subjects using the initial parameters recommended for mice, and feedback such as EEG / fMRI and scales was collected and uniform pretreatment and quality control were completed. S8. Efficient fine-tuning of human neural signal control parameters: Freeze the main body of the co-characterization backbone, introduce LoRA or Adapter only in the bottleneck / adapter layer for low-rank parameter tuning, and perform small-step updates to the end head of the parameter recommender to reduce computational power consumption and overfitting risk, while continuously adhering to safety boundaries such as dose, frequency, current density and cumulative duration. S9. Personalized Human Model Generation: Human feedback data is precipitated into individual profiles (including threshold sensitivity, frequency preference region, and network resonance window, etc.). Based on this, a personalized human neural signal control parameter fine-tuning unit is output, enabling parameter recommendations to achieve adaptive optimization and accurate push based on individual physiological characteristics and historical performance.
[0037] In the online closed-loop and continuous optimization phase: S10. Online Recommendation and Constraint Monitoring: Before each session begins, the system generates a set of parameter candidates based on real-time common representation and historical records; the monitoring subprocess dynamically constrains energy / amplitude / frequency band / session duration, etc., and triggers degradation strategies or termination mechanisms when necessary (engineering and experimental safety level).
[0038] S11, Incremental Updates and Backtracking Learning: After the session ends, new data is written into the human fine-tuning buffer, and small incremental updates or federated distillation backtracking learning are performed as planned to continuously improve model stability, generalization ability and adaptability to population heterogeneity.
[0039] S12. Evaluation and Interpretable Output: The core evaluation focuses on task / behavioral metrics, signal quality, and responsiveness, outputting interpretable parameter-effect mappings (such as ROI-level gain, network connectivity changes, and time-frequency power band response) for parameter revision, scheme verification, and engineering optimization. These four stages are interconnected, ensuring parameter transferability and process controllability from mice to humans, while also achieving individualized continuous optimization through closed-loop feedback.
[0040] This invention discloses a neural signal control parameter transfer system and method based on cross-species multimodal information. By constructing a system for cross-species multimodal data acquisition, spatiotemporal co-representation calculation, and parameter transfer and fine-tuning, it achieves effective alignment of mouse and human neural signals in a unified spatiotemporal feature space, significantly improving the consistency of cross-species neural response patterns. This invention utilizes adversarial training and distribution distance constraints to eliminate the distribution differences of temporal and spatial features between species, and forms a unified cross-species spatiotemporal co-representation through attention fusion. Based on this representation, the system can transfer neural signal control parameters learned from mouse experiments to human applications, and achieve individualized optimization with limited human experimental data using efficient parameter fine-tuning strategies. This method effectively solves the problems of difficult cross-species parameter transfer, low experimental efficiency, and insufficient accuracy of individualized modeling, significantly improving the safety, efficiency, and accuracy of neural modulation experiments, and providing reliable technical support for the intelligent and individualized development of brain-computer interfaces and neural modulation strategies.
[0041] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A neural signal control parameter transfer system based on cross-species multimodal information, characterized in that, include: A cross-species multimodal data acquisition module is used to collect and preprocess multimodal neural data from mice and humans to form samples in a unified format. A cross-species spatiotemporal co-representation computation module is used to extract temporal features and represent spatial patterns from the multimodal neural data, construct a cross-species spatiotemporal feature representation space, and eliminate distribution differences between species. The cross-species parameter transfer and fine-tuning module is used to map mouse experimental data into parameters suitable for human neural signal control and response modeling based on the cross-species spatiotemporal feature representation, and to perform individualized optimization and strategy recommendation.
2. The neural signal control parameter transfer system based on cross-species multimodal information as described in claim 1, characterized in that, The cross-species multimodal data acquisition module includes a standardization and unified coding unit, which is used to perform filtering, artifact removal, resampling, normalization, time alignment and unified format coding on the acquired data.
3. The neural signal control parameter transfer system based on cross-species multimodal information as described in claim 1, wherein the cross-species spatiotemporal co-representation computation module includes a cross-species temporal adversarial coding unit, a cross-species spatial adversarial coding unit, and a spatiotemporal information fusion unit; The cross-species temporal adversarial coding unit is used to receive EEG data from mice and humans, extract the temporal-spectral features of both, and perform adversarial training to obtain cross-species temporal representations. The cross-species spatial adversarial coding unit is used to receive MRI / fMRI data from mice and humans, extract the anatomical-functional spatial patterns of both for adversarial training, and obtain cross-species spatial representations. The spatiotemporal information fusion unit is used to align and fuse cross-species temporal representations and cross-species spatial representations to output a cross-species spatiotemporal co-representation.
4. The neural signal control parameter transfer system based on cross-species multimodal information as described in claim 1, characterized in that, The cross-species temporal adversarial coding unit includes a temporal feature extractor and a first domain discriminator; The time-series feature extractor is used to extract time-series features from EEG data; The first domain discriminator is used to erase species information through adversarial training.
5. The neural signal control parameter transfer system based on cross-species multimodal information as described in claim 1, characterized in that, The cross-species spatial adversarial coding unit includes a spatial pattern encoder and a second domain discriminator; The spatial pattern encoder is used to extract the spatial pattern of MRI / fMRI data; The second domain discriminator is used to erase species information through adversarial training.
6. The neural signal control parameter transfer system based on cross-species multimodal information as described in claim 1, characterized in that, The cross-species parameter transfer and fine-tuning module includes a mouse neural signal control parameter recommendation unit and a human neural signal control parameter fine-tuning unit. The mouse neural signal control parameter recommendation unit is used to receive preprocessed mouse neural signal control experimental data, learn parameter-response mapping relationships in a cross-species spatiotemporal co-representation space, and infer candidate sets of human neural signal control parameters. The human neural signal control parameter fine-tuning unit is used to update the model weights using an efficient parameter fine-tuning method after acquiring feedback data from human neural signal control experiments, so as to form an individualized human neural signal control model.
7. The neural signal control parameter transfer system based on cross-species multimodal information as described in claim 6, characterized in that, The human neural signal control parameters output by the mouse neural signal control parameter recommendation unit include stimulation frequency, stimulation amplitude, pulse width, duty cycle, stimulation duration, stimulation site / channel combination, and stimulation sequence.
8. A method for transferring neural signal control parameters based on cross-species multimodal information, characterized in that, For implementing the neural signal control parameter transfer system based on cross-species multimodal information as described in any one of claims 1-7, the method comprises: A cross-species spatiotemporal co-representation computation module was trained using publicly available cross-species multimodal neural data to construct cross-species spatiotemporal co-representations; A neural signal control experiment was conducted on mice to collect multimodal neural signals and neural signal control parameters. After preprocessing and standardization, a neural signal control parameter recommendation unit was trained in mice to generate candidate human neural signal control parameters. Human neural signal control experiments were conducted based on the candidate's neural signal control parameters. Feedback data was collected, and after preprocessing and standardization, the model parameters were finely tuned to form a human neural signal control parameter fine-tuning unit. Human multimodal data is input into the human neural signal control parameter fine-tuning unit to obtain recommended human neural signal control parameters for individualized optimization and strategy recommendation.