Brain-computer interface signal augmentation and evaluation method incorporating physical information

Abstract

The application discloses a brain-computer interface signal enhancement and evaluation method fusing physical information. First, based on individual structural magnetic resonance images, a personalized six-layer brain tissue dissection model containing the cortex, white matter, cerebrospinal fluid, dura mater, skull and scalp is constructed, and differentiated conductivity parameters are given. Second, neurophysiological prior knowledge such as neural dynamics and white matter anisotropic conduction is converted into a representation rule that can be learned by a neural network. Then, these rules are embedded in the physical information neural network in the form of differentiable physical constraints, including the residual error of the cortical neural dynamics equation, the residual error of the volume conduction equation of each layer, and the continuity constraint of the interlayer interface. Finally, the scalp electroencephalogram signal is used to train the network, and by minimizing the loss function containing the reconstruction and physical constraints, a high-fidelity intracranial electroencephalogram signal and multi-modal neurophysiological information are jointly output. The application effectively improves the fidelity, physical reasonableness and clinical interpretability of signal enhancement.

Description

Technical Field

[0001] This application relates to the field of brain-computer interface signal processing and artificial intelligence, and in particular to a method for enhancing and evaluating brain-computer interface signals that integrates physical information. Background Technology

[0002] Brain-computer interface (BCI) technology enables direct communication between the human brain and external devices by analyzing neural electrical signals, and has broad application prospects in fields such as neurological rehabilitation, diagnosis of consciousness disorders, and intelligent prosthetic control. Currently, non-invasive scalp electroencephalography (EEG) is widely used due to its safety and convenience. However, its signals must traverse multiple layers of biological tissue, including cerebrospinal fluid, dura mater, skull, and scalp, to reach the scalp. These tissues have vastly different conductivities (e.g., cerebrospinal fluid conductivity is approximately 224 times that of the skull), resulting in severe spatial ambiguity, high-frequency component attenuation, and loss of depth information in the EEG signal, leading to unsatisfactory signal-to-noise ratio and spatial resolution. While invasive intracranial electroencephalography (iEEG) can provide high-quality signals, it is accompanied by surgical risks, high costs, and ethical restrictions.

[0003] To improve the quality of non-invasive signals, existing technologies have the following limitations: 1) Simplified modeling: Homogeneous models or simplified three-layer spherical models are often used, ignoring anatomical differences such as skull thickness and cortical folds between individuals, as well as key electrophysiological characteristics such as white matter anisotropy and high cerebrospinal fluid conductivity, leading to distortion of the forward conduction model; 2) Lack of physical constraints: Purely data-driven deep learning models (such as end-to-end mapping networks) lack clear physical constraints, and their reconstruction results may violate the generation and propagation mechanisms of neural electrical activity, resulting in poor generalization ability and low interpretability; 3) Single constraints: Even when introducing physical information neural networks (PINN), they often only embed a unified volume conduction equation, failing to systematically model cortical neural dynamics, interlayer interface continuity conditions, etc., as differentiable constraints, resulting in insufficient physical self-consistency of the model; 4) Single output information: Existing methods usually only output the reconstructed iEEG signal, failing to provide deep information such as the signal propagation process in each layer of tissue, cortical source current distribution, and neural dynamic state, limiting their auxiliary decision-making value in clinical diagnosis. Therefore, how to achieve high-fidelity, highly interpretable, and biophysical-compliant EEG signal enhancement under non-invasive conditions remains a technical challenge that urgently needs to be overcome in this field. Summary of the Invention

[0004] In view of this, this application provides a brain-computer interface signal enhancement and evaluation method that integrates physical information, aiming to solve the problems of signal propagation distortion caused by oversimplification of the head physical model, insufficient generalization ability due to high dependence on training data, and poor interpretability of results due to lack of physiological mechanism constraints in the existing technology.

[0005] To address the aforementioned technical problems, this application adopts the following technical solution: a method for enhancing and evaluating brain-computer interface signals that integrate physical information, comprising:

[0006] Acquire structural magnetic resonance images of an individual;

[0007] Based on structural magnetic resonance imaging, a personalized layered anatomical model of multi-layered brain tissue, including the cortex, white matter, cerebrospinal fluid, dura mater, skull, and scalp, was constructed.

[0008] Transforming prior knowledge of neuroelectrophysiology of multi-layered brain tissue into representation rules that can be learned by neural networks;

[0009] The representation rules are embedded in the physical information neural network in the form of differentiable physical constraints. The physical constraints include at least the physical equation constraints within each tissue layer and the coupling constraints at the interface between adjacent tissue layers.

[0010] The collected scalp EEG signals are input into a physical information neural network for training. By minimizing the total loss function, which includes signal reconstruction loss and multiple physical constraint losses, the enhanced intracranial EEG signals and multimodal neurophysiological information are jointly output.

[0011] To address the aforementioned technical problems, another technical solution adopted in this application is: providing a brain-computer interface signal enhancement and evaluation device that integrates physical information, comprising:

[0012] The acquisition module is used to acquire structural magnetic resonance images of an individual.

[0013] The module is used to construct a personalized layered anatomical model of the brain tissue, including the cortex, white matter, cerebrospinal fluid, dura mater, skull, and scalp, based on structural magnetic resonance imaging.

[0014] The conversion module is used to convert prior knowledge of neuroelectrophysiology of multi-layered brain tissue into representation rules that can be learned by neural networks.

[0015] The embedding module is used to embed the representation rules into the physical information neural network in the form of differentiable physical constraints. The physical constraints include at least the physical equation constraints within each layer of the organization and the coupling constraints at the interface between adjacent organizational layers.

[0016] The output module is used to input the collected scalp EEG signals into the physical information neural network for training. By minimizing the total loss function, which includes signal reconstruction loss and multiple physical constraint losses, it jointly outputs the enhanced intracranial EEG signals and multimodal neurophysiological information.

[0017] To solve the above-mentioned technical problems, another technical solution adopted in this application is: to provide a computer device, the computer device including a processor and a memory coupled to the processor, the memory storing program instructions, and when the program instructions are executed by the processor, causing the processor to perform the steps of the brain-computer interface signal enhancement and evaluation method that integrates physical information as described above.

[0018] To solve the above-mentioned technical problems, another technical solution adopted in this application is to provide a storage medium storing program instructions capable of implementing the brain-computer interface signal enhancement and evaluation method for fusing physical information as described above.

[0019] Compared to existing technologies, this invention achieves a breakthrough in EEG signal enhancement technology by deeply integrating individualized neuroanatomical structures, multi-level bioelectrophysiological prior knowledge, and physical information neural networks. The personalized six-layer brain tissue anatomical model constructed by this method meticulously depicts the significant differences in conductivity among different layers and the anisotropy of white matter. This fundamentally corrects the signal propagation distortion caused by traditional homogeneous or simplified models, significantly improving the fidelity and spatial resolution of reconstructed intracranial EEG signals, and particularly effectively restoring high-frequency neural oscillations severely attenuated by the skull. By simultaneously embedding the dynamic equations describing neural activity and the volumetric conduction equations describing signal propagation as differentiable constraints into the network training, and strictly ensuring the continuity of potential and current at interlayer interfaces, the enhancement results strictly adhere to the fundamental physical laws of neural electrical activity, greatly improving the model's physical rationality and generalization ability under data scarcity. Furthermore, the network can simultaneously output multimodal information such as cortical source current distribution, neurodynamic state parameters, and signal propagation paths across layers, providing unprecedented visual insights and quantitative evidence for clinical lesion localization and neural mechanism analysis, greatly enhancing the interpretability and clinical applicability of the technology. Simultaneously, the network supports individualized fine-tuning of key biophysical parameters and balances multiple constraints through adaptive optimization strategies, achieving precise and robust signal enhancement for different individuals. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating an embodiment of the brain-computer interface signal enhancement and evaluation method integrating physical information according to the present invention;

[0021] Figure 2 This is a functional module diagram of an embodiment of the brain-computer interface signal enhancement and evaluation device that integrates physical information according to the present invention;

[0022] Figure 3 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention;

[0023] Figure 4This is a schematic diagram of the structure of the storage medium according to an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0025] The terms "first," "second," and "third" in this application are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationships and movements between components in a specific orientation (as shown in the figures). If the specific orientation changes, the directional indications also change accordingly. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0026] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0027] Figure 1 This is a flowchart illustrating the brain-computer interface signal enhancement and evaluation method integrating physical information according to an embodiment of the present invention. It should be noted that if substantially the same result is obtained, the method of this embodiment is not necessarily identical. Figure 1 The illustrated process sequence is limited. For example... Figure 1 As shown, the brain-computer interface signal enhancement and evaluation method that integrates physical information includes the following steps:

[0028] Step S1: Acquire the individual's structural magnetic resonance imaging.

[0029] Specifically, structural magnetic resonance imaging (MRI) images of the individual are acquired, typically high-resolution T1-weighted images. The purpose of this step is to provide source data for all subsequent personalized modeling. Unlike existing techniques that use population average templates, this embodiment emphasizes individual specificity. Individuals exhibit significant differences in anatomical structures such as skull thickness, cortical fold morphology, and ventricular size, which profoundly affect the propagation path and attenuation of EEG signals. Therefore, modeling based on the individual's own MRI images is a fundamental prerequisite for overcoming the homogeneity assumption and achieving high-precision signal enhancement. This image will serve as input, driving subsequent automatic segmentation, geometric reconstruction, and parameter personalization processes.

[0030] Step S2: Based on structural magnetic resonance imaging, construct a personalized layered anatomical model of the brain tissue, including the cortex, white matter, cerebrospinal fluid, dura mater, skull, and scalp.

[0031] It should be noted that existing physical information neural network methods, when processing EEG signals, typically treat the head as a homogeneous medium or use a simplified three-layer spherical model (brain tissue-skull-scalp). This simplification ignores the unique electrical properties of key tissue layers such as cerebrospinal fluid and dura mater, resulting in physical constraints failing to accurately reflect the real bioelectrical propagation process. This embodiment proposes a six-layer physiological domain spatial analysis method based on neuroanatomy, providing a precise spatial basis for subsequent layered physical constraint learning.

[0032] Furthermore, the construction of a personalized layered anatomical model of the brain, comprising multiple layers of brain tissue including the cortex, white matter, cerebrospinal fluid, dura mater, skull, and scalp, specifically includes:

[0033] 1. A deep learning segmentation network was used to process structural magnetic resonance images, automatically identifying and segmenting the boundaries of six layers of brain tissue.

[0034] First, in this embodiment, the three-dimensional space of the head is divided into six non-overlapping subdomains according to the neuroanatomical structure, and each subdomain corresponds to a tissue layer with a clear neuroelectrophysiological function:

[0035] The cortical region corresponds to the gray matter of the brain and is the aggregation area of ​​pyramidal neuron cell bodies, serving as the primary source of neural electrical signals. This region is isotropic in its electrical properties, with a conductivity of approximately 0.33 S / m. The cortical thickness is approximately 2-4 mm, exhibiting a highly folded geometric shape.

[0036] The white matter domain corresponds to the white matter region of the brain and is composed of myelinated axonal fiber bundles. It is the main pathway for the conduction of neural electrical signals in the brain. The electrical properties of this domain are anisotropic, with the conductivity of current conducted along the fiber direction (approximately 0.14 S / m) being significantly higher than that perpendicular to the fiber direction (approximately 0.065 S / m), and the anisotropy ratio is approximately 2.15.

[0037] The cerebrospinal fluid domain corresponds to the cerebrospinal fluid region in the subarachnoid space. It is a highly conductive ionic solution (approximately 1.79 S / m), with a conductivity about 224 times that of the skull. Due to its extremely high conductivity, the potential distribution within this domain tends to be spatially uniform, resulting in significant current convergence and short-circuit effects, leading to a loss of spatial resolution for deep source signals.

[0038] The dura mater corresponds to the dura mater layer that is closely attached to the inner side of the skull. It is a thin layer structure composed of dense connective tissue, with a thickness of about 0.5-1 mm and low electrical conductivity (about 0.02 S / m). Due to its thin-layer characteristics, a thin-layer impedance model is used to describe its potential transition characteristics.

[0039] The cranial region corresponds to the bone tissue area of ​​the skull, consisting of three layers: outer cortical bone, cancellous bone, and inner cortical bone. Its overall electrical conductivity is extremely low (approximately 0.008 S / m), making it a major attenuation barrier for EEG signal propagation. The thickness of the skull varies spatially, approximately 5 mm in the temporal region and 10-15 mm in the apex. This variation in thickness results in different degrees of signal attenuation at different locations on the scalp.

[0040] The scalp domain corresponds to the soft tissue area of ​​the scalp and is the measurement layer for the EEG electrodes. It has a conductivity of approximately 0.43 S / m and a thickness of approximately 5-8 mm. The above six subdomains satisfy the condition of spatial non-intersection, and their union covers the entire head space.

[0041] This embodiment defines five interlayer physiological interfaces, each of which is the common boundary between two adjacent tissue layers: gray matter-white matter interface, white matter-subarachnoid space interface, cerebrospinal fluid-dura mater interface, dura mater-skull interface, and skull-scalp interface. These interfaces are the locations where subsequent interface coupling constraints apply.

[0042] To achieve automatic identification and differentiable processing of the six physiological domains of brain tissue, this embodiment provides a spatial domain classification network for brain-computer interface signal enhancement and evaluation centered on fused physical information, namely a deep learning segmentation network. This network receives three-dimensional spatial coordinates of the head as input and outputs the probability distribution of the brain tissue layer to which the coordinate point belongs, enabling the physical information neural network to automatically identify different brain tissue regions and apply corresponding physical constraints. The network structure adopts a multilayer perceptron architecture, including an input layer, multiple hidden layers, and an output layer. The input layer receives the three-dimensional spatial coordinates of the head; the hidden layers adopt a fully connected structure, with each layer containing 64 to 128 neurons, using the ReLU activation function; the output layer contains 6 neurons, corresponding to the six brain tissue layers of cortex, white matter, cerebrospinal fluid, dura mater, skull, and scalp, respectively, and outputs the probability distribution using the Softmax activation function. To enhance the network's ability to express high-frequency spatial changes in brain tissue boundaries, this embodiment uses Fourier feature encoding for the input coordinates, mapping the low-dimensional coordinates to a high-dimensional feature space through a series of sine and cosine functions of different frequencies, enabling the network to accurately capture complex anatomical boundaries such as cortical folds.

[0043] It should be noted that, to achieve end-to-end differentiable training, this embodiment uses a soft masking mechanism instead of hard classification. For any spatial coordinate, the mask value belonging to the k-th layer is calculated using a temperature-controlled Softmax function:

[0044] ;

[0045] in, Indicates spatial location Belongs to the The soft mask value of the layered brain tissue has a range of values. ; Indicates spatial domain classification network for location The logit output of the organization at level k; The temperature parameter controls the softness or hardness of the soft mask. The time approaches hard classification, The distribution tends to be uniform. These correspond to six tissue layers: the cortex, white matter, cerebrospinal fluid, dura mater, skull, and scalp.

[0046] The temperature parameter gradually decreases during training using an exponential decay strategy:

[0047] ;

[0048] in The initial temperature. Here, t represents the decay rate, and t represents the number of training steps. This allows the soft mask to gradually approach hard classification, achieving a smooth transition from soft constraints to hard constraints.

[0049] It is important to understand that, based on the soft masking mechanism, this embodiment selectively applies physical constraints corresponding to different brain tissue layers to different spatial locations of the head. For any spatial coordinate, the physical constraint loss is a weighted sum of the physical constraints of the six brain tissue layers, with the weights determined by the soft mask value. This design enables the brain-computer interface signal enhancement and evaluation network that integrates physical information to automatically identify the brain tissue layer to which each spatial point belongs and apply the corresponding neurophysiological physical constraints (such as the cortical Poisson equation, the white matter anisotropic Laplace equation, cerebrospinal fluid isopotential constraints, etc.) without the need for manual specification of tissue boundaries.

[0050] 2. Set initial conductivity parameters for each tissue layer based on neurophysiological measurements, wherein the conductivity of the white matter layer is represented by an anisotropic conductivity tensor constructed based on diffusion tensor imaging data.

[0051] Specifically, existing methods typically use fixed conductivity values ​​from literature, which cannot adapt to individual differences and the need for parameter optimization during training. Therefore, this embodiment proposes a learnable conductivity parameterization method based on prior knowledge, which allows the network to fine-tune conductivity parameters during training while maintaining physical rationality.

[0052] (a) Setting the a priori values ​​for conductivity in six-layer tissues:

[0053] This embodiment, based on neuroelectrophysiology literature, sets differentiated a priori conductivity values ​​for the six tissue layers. It is particularly noteworthy that the conductivity ratio of cerebrospinal fluid to that of the skull is approximately 224 times. This significant difference is the primary physiological reason for the severe attenuation of deep-source signals and is a key factor emphasized in the modeling of this embodiment.

[0054] (b) Design of learnable conductivity parameters:

[0055] This embodiment designs learnable conductivity correction parameters for each tissue layer. The actual conductivity is calculated exponentially by multiplying the prior value by the correction factor.

[0056] ;

[0057] in, Indicates the first Actual electrical conductivity of layered brain tissue (unit: S / m). This indicates the first [number] based on neuroelectrophysiology literature. The first a priori value of the electrical conductivity of the layered tissue; For learnable conductivity correction parameters; This embodiment sets the maximum allowable correction factor. That is, the actual conductivity is within the prior value. The conductivity can vary within a range of several times. This design allows for a degree of individualized adjustment while preventing the conductivity from deviating from its physically reasonable range.

[0058] (c) Construction of the anisotropic conductivity tensor of white matter:

[0059] The conductivity of the white matter domain is described using an anisotropic tensor to reflect the directional conduction characteristics of myelinated axon fibers. This embodiment proposes a formula for constructing anisotropic conductivity tensors based on fiber orientation:

[0060] ;

[0061] in, Indicates the spatial location of white matter place Electrical conductivity tensor (unit: S / m); This represents the longitudinal conductivity along the direction of the myelinated axonal fibers, approximately 0.14 S / m; This represents the transverse conductivity perpendicular to the fiber direction, approximately 0.065 S / m; for identity matrix; For position The fiber orientation unit vector at that location is extracted from diffusion tensor imaging data. The anisotropy ratio of this tensor is... This reflects the physiological characteristic that current is conducted much more easily along the fiber direction than in the perpendicular direction.

[0062] (d) Conductivity prior regularization loss:

[0063] To prevent the conductivity parameter from deviating excessively from the prior value during training, this embodiment designs a priori regularization loss for conductivity:

[0064] ;

[0065] in, This indicates the conductance prior regularization loss; The conductivity correction parameter for the k-th layer of tissue; The regularization weights for the k-th layer of tissue are set as follows: larger weights are assigned to tissues with reliable measurements (such as cerebrospinal fluid and skull), and smaller weights are assigned to tissues with large individual differences (such as cortex and white matter).

[0066] 3. Based on the segmentation results, extract the three-dimensional boundary surfaces of each tissue layer and calculate the local thickness distribution of each tissue layer.

[0067] Specifically, individual differences in neuroanatomy significantly affect the propagation characteristics of electroencephalogram (EEG) signals. Existing methods typically use population-averaged templates, which fail to reflect individual-specific key anatomical features such as skull thickness distribution and cortical fold morphology. Therefore, this embodiment proposes an automated neuroanatomical model reconstruction method based on individual structural MRI, providing an individualized anatomical basis for signal enhancement and evaluation of brain-computer interfaces that fuse physical information with brain-computer interfaces based on brain-layered prior physical information neural networks.

[0068] (a) Deep learning-based six-layer MRI tissue segmentation network:

[0069] This embodiment designs a six-layer tissue segmentation network for signal enhancement and evaluation in brain-computer interfaces that integrates physical information. A three-dimensional U-Net architecture is used to achieve automatic mapping from T1-weighted MRI images to six-layer tissue segmentation. This embodiment proposes a combined segmentation loss function:

[0070] ;

[0071] in, Indicates the total partition loss; This indicates Dice's loss. and These are the predicted probability and the true label of the $k$-th layer, respectively; , representing the cross-entropy loss; This represents the boundary enhancement loss, emphasizing the segmentation accuracy of interlayer interfaces; , , These are the weighting coefficients for each loss term.

[0072] (b) Implicit representation and extraction of interlayer boundary surfaces:

[0073] This embodiment proposes an implicit representation method for interlayer interfaces based on the symbolic distance field. For the interface between the i-th layer and the j-th layer... Define implicit functions:

[0074] ;

[0075] in, Indicates position The implicit function value at that location; and These represent the segmentation probabilities of this position belonging to the i-th and j-th layers, respectively. (Interface) Defined as the zero isosurface of an implicit function The Marching Cubes algorithm is used to extract the triangular mesh representation.

[0076] (c) Calculation of individualized tissue thickness field:

[0077] This embodiment proposes a tissue thickness field calculation method based on bidirectional distance transformation:

[0078] ;

[0079] in, Indicates the position of the k-th layer of organization. Local thickness at the location (unit: mm); Indicates position To the interface The shortest Euclidean distance; and These represent the interfaces between the k-th layer and its adjacent inner and outer layers, respectively. This thickness field reflects individual-specific anatomical variations, such as the spatial distribution of skull thickness.

[0080] (d) Conditional encoding of individualized geometric features:

[0081] This embodiment proposes a personalized geometric feature conditional coding method that embeds anatomical features into a brain-computer interface signal enhancement and evaluation network that integrates physical information.

[0082] ;

[0083] in, Represents an individualized geometric condition vector; This represents the average thickness of the k-th layer of tissue; Indicates brain volume; Represents the surface area of ​​the cortex; Represents the morphological feature vectors of the cortex (including gyral index, cortical curvature distribution, etc.); This is a geometric feature encoding network. The conditional vector is concatenated with EEG features and then input into the main network, enabling the network to adaptively adjust the signal enhancement strategy based on individual anatomical characteristics.

[0084] Step S3: Transform the prior knowledge of neurophysiology of multi-layered brain tissue into representation rules that can be learned by the neural network.

[0085] Specifically, this step is the core of the knowledge representation of the entire brain-computer interface signal enhancement and evaluation system that integrates physiological hierarchical prior physical information neural network with physical information. Its core function is to convert the prior knowledge of multi-layered brain tissue into a rule form that the neural network can learn, including key physiological characteristics such as cortical neurodynamics, white matter anisotropic conductivity, cerebrospinal fluid isoelectric effect, dura mater thin layer impedance, and skull frequency-dependent attenuation. This fundamentally solves the problem of existing methods lacking prior knowledge of neuroelectrophysiology.

[0086] Furthermore, the conversion of prior knowledge of neurophysiological processes in multi-layered brain tissue into learnable representation rules for neural networks specifically includes:

[0087] 1. Construct a spatially extended neurodynamic model, extending the ordinary differential equations describing the activity of neuronal populations into spatiotemporally distributed partial differential equations defined on the cortical surface.

[0088] Specifically, existing methods lack modeling of the spatiotemporal evolution of cortical neural activity. Classical neurodynamic models only describe the temporal dynamics of a single cortical column and cannot reflect the spatial distribution characteristics of neural activity on the cortical surface. Therefore, this embodiment extends the neurodynamic model to a spatially distributed dynamic system and uses it as a physical constraint to embed into the training of a brain-computer interface signal enhancement and evaluation network that integrates physical information, so that the reconstructed neural signals conform to the principles of neurophysiology.

[0089] (a) Classical neurodynamic model:

[0090] This embodiment innovatively extends the neurodynamic model. This model describes the activity of cortical neuronal populations, comprising three interacting neuronal populations: pyramidal cell populations (the main source of EEG signals), excitatory interneuron populations, and inhibitory interneuron populations. The model's state variables include the mean membrane potential of pyramidal cells, excitatory postsynaptic potentials, inhibitory postsynaptic potentials, and their time derivatives, totaling six state variables. These state variables directly determine the source current density of the brain-computer interface signal. The model's dynamics are described by a set of ordinary differential equations, with the core equation being:

[0091] ;

[0092] ;

[0093] ;

[0094] in, This represents the excitatory postsynaptic potential (unit: mV), reflecting the input from the excitatory interneuron to the pyramidal cell; This represents the inhibitory postsynaptic potential (unit: mV), reflecting the input from the inhibitory interneuron to the pyramidal cell; The value represents the mean membrane potential of the pyramidal cell; A and B are the excitatory and inhibitory synaptic gains, respectively (unit: mV); a and b are the reciprocals of the excitatory and inhibitory time constants, respectively (unit: mV). ); For Sigmoid activation function, it describes the nonlinear relationship between the average firing rate of a neuron population and the membrane potential; Half of the maximum discharge rate (unit: Hz); The slope parameter of the Sigmoid function; The membrane potential threshold for generating half-maximum discharge rate; The strength of the connection from the inhibitory interneuron to the pyramidal cell.

[0095] (b) Spatial distributed expansion:

[0096] This embodiment extends the state variables of the classical neurodynamic model from time functions to spatiotemporal functions; that is, the state variables are not only functions of time but also functions of cortical spatial location. This embodiment proposes a spatially extended neurodynamic equation:

[0097] ;

[0098] in, Indicates cortical location Excitatory postsynaptic potential at time t (unit: mV); This represents the average membrane potential of the cone cells at that location; Lateral diffusion coefficient (unit: ), describes the electrotense coupling strength between adjacent cortical columns; To define in the cortical manifold The Laplace-Beltrami operator describes the spatial diffusion of neural activity on the cortical surface; The co-activation intensity coefficient describes the active co-activation between functional columns via horizontal fiber synaptic connections. Indicates position Neighborhood The spatial mean of the excitatory postsynaptic potentials. This extension allows the neurodynamic state to become a spatial field defined on the cortical surface, capable of describing the spatial distribution and temporal evolution of neural activity in different cortical regions.

[0099] (c) Neurodynamic state estimation network:

[0100] This embodiment designs a neurodynamic state estimation network for brain-computer interface (BCI) signal enhancement and evaluation based on fused physical information. This network estimates the neurodynamic state at each location on the scalp from scalp EEG features. The network input includes cortical spatial coordinates, time, and feature vectors extracted from the scalp EEG signal. The network first performs positional encoding on the cortical spatial coordinates and time, employing a sine-cosine encoding scheme to map low-dimensional coordinates to a high-dimensional feature space, enhancing the network's ability to represent complex anatomical structures such as cortical folds. The encoded spatiotemporal features are concatenated with the scalp EEG features and input into a multilayer perceptron, outputting a six-dimensional neurodynamic state vector for subsequent source current density calculation and BCI signal reconstruction.

[0101] (d) Neural dynamics parameter estimation network

[0102] This embodiment designs a neurodynamic parameter estimation network to estimate the physiological parameters of a neurodynamic model from EEG features. This embodiment proposes a parameter constraint transformation formula:

[0103] ;

[0104] in, This represents the k-th neurodynamic parameter (such as synaptic gain A, B, reciprocal of time constant a, b, connection strength). wait); and These are the lower and upper physiological bounds of the parameter, respectively. For the Sigmoid function; For parameter estimation networks; EEG feature vectors; These are the network parameters. This design ensures that the estimated parameters are always within a physiologically reasonable range.

[0105] (e) Mapping of neurodynamic states to source current density:

[0106] This embodiment establishes a mapping relationship between neural dynamics and source current density, realizing a complete causal chain of neural activity → source current → potential distribution. This embodiment proposes a formula for generating source current density:

[0107] ;

[0108] in, Indicates cortical location The source current density vector at time t (unit: ); Current density amplitude coefficient (unit: ), which are learnable parameters; This represents the excitation-inhibition balance potential difference, reflecting the potential difference between the apical dendrite and the cell body of a pyramidal cell; For the dermal surface at position The outward normal unit vector at a location reflects the spatial orientation of the apical dendrites of the pyramidal cell; For learnable nonlinear modulation functions:

[0109] ;

[0110] in, This is a nonlinear modulation factor, with a value range of [value range missing]. ; These are learnable parameters. This nonlinear mapping can capture the differences in current generation efficiency of neuronal populations under different excitation-inhibition states.

[0111] 2. Based on diffusion tensor imaging data, an anisotropic conductivity tensor reflecting the directional conduction characteristics of white matter myelinated axonal fibers is constructed.

[0112] Specifically, this embodiment proposes a method for constructing anisotropic conductivity tensors of white matter based on diffusion tensor imaging data, providing an accurate white matter electrical model for signal enhancement and evaluation of brain-computer interfaces that fuse physical information with brain-layered prior physical information neural networks.

[0113] (a) Decoding the orientation of myelinated axonal fibers based on DTI:

[0114] This embodiment proposes a method for extracting white matter fiber orientation from diffusion tensor imaging data. The diffusion tensor is decomposed into features, and the principal feature vectors are extracted as fiber orientations.

[0115] ;

[0116] in, Indicates position The unit vector of the fiber direction at that location; Received by DTI measurement Diffusion tensor (unit: ); This represents the eigenvector corresponding to the largest eigenvalue. This eigenvector reflects the direction in which water molecules diffuse most freely, i.e., the main orientation of myelinated axonal fibers.

[0117] (b) Fiber orientation field continuum network:

[0118] This embodiment designs a fiber orientation field continuum network to interpolate discrete fiber orientations into a continuous orientation field:

[0119] ;

[0120] in, This represents the unit vector of the fiber direction after continuation; For fiber orientation field network, parameters This network is learned by fitting discrete DTI data. It enables the fiber orientation to be queried at any spatial location, supporting continuous spatial sampling of physical information neural networks.

[0121] (c) Construction of the anisotropic conductivity tensor:

[0122] This embodiment proposes a formula for constructing anisotropic conductivity tensors based on fiber orientation:

[0123] ;

[0124] in, Indicates the location of white matter place Electrical conductivity tensor (unit: S / m); This represents the longitudinal conductivity along the direction of the myelinated axonal fibers, designed as a learnable parameter with an initial value of approximately 0.14 S / m; This represents the transverse conductivity perpendicular to the fiber direction, designed as a learnable parameter with an initial value of approximately 0.065 S / m; for Identity matrix.

[0125] (d) Anisotropy ratio constraint loss:

[0126] This embodiment designs anisotropy ratio constraint loss to ensure that the anisotropy ratio is within a physiologically reasonable range:

[0127] ;

[0128] in, This represents the anisotropy ratio constraint loss; and These represent the physiological lower and upper bounds of the anisotropy ratio, respectively. This loss penalizes deviations from the physiologically reasonable range of the anisotropy ratio, ensuring that the conductivity tensor conforms to neurophysiological priors.

[0129] 3. Establish an approximate model of the equipotential field to describe the current convergence effect in the cerebrospinal fluid domain due to its high conductivity.

[0130] Specifically, cerebrospinal fluid (CSF) has extremely high conductivity (approximately 1.79 S / m), about 224 times that of the skull. This high conductivity leads to a spatially uniform potential distribution within the subarachnoid space, resulting in significant current convergence and short-circuit effects. This embodiment proposes an equipotential field approximation and current convergence effect modeling method based on the high conductivity characteristics of CSF, providing an accurate CSF electrical model for signal enhancement and evaluation of brain-computer interfaces that fuse physical information with brain-based layered prior physical information neural networks.

[0131] (a) Analysis of the physical effects of high conductivity:

[0132] According to Ohm's Law In highly conductive media, even a small potential gradient can generate a large current. To satisfy the current continuity condition... In a high-conductivity medium, the potential gradient must be very small, meaning the potential distribution tends to be spatially uniform. This embodiment establishes an approximate model of the equipotential field of cerebrospinal fluid based on this physical principle.

[0133] (b) Isopotential field decomposition model:

[0134] This embodiment proposes an isoelectric field decomposition model for the potential within the cerebrospinal fluid domain:

[0135] ;

[0136] in, Indicates location within the cerebrospinal fluid domain Potential at time t (unit: ); It represents the spatial average potential within the cerebrospinal fluid domain, and is only a function of time, reflecting the overall potential level of the cerebrospinal fluid as an equipotential body; This represents the potential disturbance term, which should approach zero under the equipotential approximation. It indicates the spatial extent of the cerebrospinal fluid domain.

[0137] (c) Cerebrospinal fluid potential field estimation network:

[0138] This embodiment designs a dedicated cerebrospinal fluid potential field estimation network, which includes an average potential subnetwork and a perturbation subnetwork:

[0139] ;

[0140] ;

[0141] in, For the average potential subnetwork, from EEG characteristics Estimate the spatial mean potential of the cerebrospinal fluid domain; To estimate the spatial perturbation of the potential in the perturbed subnetwork; The disturbance amplitude coefficient is designed as a learnable parameter, and its initial value is set to a small value to strengthen the equipotential constraint.

[0142] (d) Isopotential constraint loss:

[0143] This embodiment employs an equipotential constraint loss design to force the potential within the cerebrospinal fluid domain to tend towards spatial uniformity.

[0144] ;

[0145] in, This represents the loss due to equipotential constraint. This represents the spatial variance within the cerebrospinal fluid domain; This represents the spatial expectation within the cerebrospinal fluid domain; and This is the weighting coefficient. This loss forces the potential distribution within the cerebrospinal fluid domain to tend towards spatial uniformity, reflecting the equipotential characteristics of a high-conductivity medium.

[0146] 4. Establish a thin-layer impedance model to describe the relationship between the potential jumps on the inner and outer surfaces of the dura mater and the normal current density.

[0147] Specifically, the dura mater is a thin, dense layer of connective tissue closely adhering to the inner side of the skull, approximately 0.5-1 mm thick, with low electrical conductivity (approximately 0.02 S / m). Due to its thin layer characteristics, it is unsuitable for description using volumetric conduction equations, and existing methods typically ignore the dura mater or treat it as part of the skull. This embodiment proposes a potential jump modeling method based on the impedance characteristics of the thin dura mater, providing a complete six-layer physical model for signal enhancement and evaluation of brain-computer interfaces that fuse physical information with neural networks based on brain layered priors.

[0148] (a) The principle of thin-layer impedance approximation:

[0149] For thin-layer structures with thicknesses much smaller than other geometric scales, the thin-layer impedance approximation can be used to simplify the three-dimensional volume conduction problem into two-dimensional interface conditions. Based on this principle, this embodiment models the dura mater as a thin-layer barrier with a certain surface resistance, describing the relationship between the potential jump and the through current between its inner and outer surfaces.

[0150] (b) Dura mater thin-layer impedance model:

[0151] This embodiment proposes a formula for the thin-layer impedance potential jump of the dura mater:

[0152] ;

[0153] in, and These respectively indicate the position of the outer surface (closer to the skull) and inner surface (closer to the cerebrospinal fluid) of the dura mater. Potential at time t (unit: V); Indicates the location of the dura mater Thin film resistance at (unit: ), For local thickness, Dura mater conductivity; This represents the normal current density passing through the dura mater (unit: ); This refers to the dura mater interface.

[0154] (c) Learnable thin-film resistance parameters:

[0155] This embodiment is designed with learnable thin-film resistance parameters to accommodate individual differences:

[0156] ;

[0157] in, These are prior values ​​of thin-film resistance based on literature. For location-dependent learnable correction parameters, a dura mater electrical resistance correction network is used. Output; These are network parameters. This design allows for spatial variation of the thin-layer resistance to reflect spatial non-uniformity in dura mater thickness.

[0158] (d) Thin-layer impedance constraint loss:

[0159] This embodiment designs a thin-layer impedance constraint loss to force the network output to satisfy the thin-layer impedance relationship:

[0160] ;

[0161] in, This represents the thin-layer impedance constraint loss. This loss forces the potential jumps on the inner and outer surfaces of the dura mater to satisfy the thin-layer impedance relationship with the normal current density.

[0162] 5. Establish a transfer function model describing the frequency-dependent attenuation characteristics of the skull to EEG signals, and design an optimal filtering regularization framework for recovering high-frequency components.

[0163] Specifically, the skull exhibits a significant low-pass filtering effect on high-frequency EEG signals, with high-frequency components experiencing severe attenuation as they pass through it. During inverse reconstruction, simple inverse filtering can lead to explosive amplification of high-frequency noise. Existing methods lack effective mechanisms to address skull frequency-dependent attenuation. This embodiment proposes a skull frequency-dependent inverse attenuation regularization method based on optimal filtering, providing an optimal balance between high-frequency component recovery and noise suppression for brain-computer interface signal enhancement and evaluation based on brain-layered prior physical information neural networks fusing physical information.

[0164] (a) Skull frequency-dependent decay model:

[0165] This embodiment establishes a frequency-dependent attenuation transfer function model of the skull:

[0166] ;

[0167] in, Indicates the frequency of the skull The attenuation transfer function at that point; Indicates skull thickness (unit: mm); Indicates skin depth (unit: mm), which decreases with increasing frequency; Angular frequency; Permeability; Let be the electrical conductivity of the skull. This model describes the exponential attenuation characteristics of high-frequency signals passing through the skull.

[0168] (b) Optimal filtering regularization framework:

[0169] This embodiment proposes an optimal filtering regularization framework for inverse attenuation recovery in the frequency domain:

[0170] ;

[0171] in, Indicates the restored intracranial signal spectrum; This indicates the spectrum of EEG signals from the scalp. Represents the complex conjugate of the attenuation transfer function; The frequency-dependent regularization parameter is determined by the signal-to-noise ratio (SNR). This formula achieves full recovery in high SNR frequency bands and suppression in low SNR frequency bands.

[0172] (c) Adaptive regularization parameter learning:

[0173] This embodiment designs an adaptive regularization parameter learning network to learn the optimal frequency-dependent regularization parameters from the data:

[0174] ;

[0175] in, For learning the regularization parameters of the network; EEG feature vectors; For network parameters; The function ensures that the regularization parameter is positive. This network can adaptively adjust the regularization strength of each frequency band according to the signal characteristics.

[0176] (d) Balance loss between high-frequency reconstruction and noise suppression:

[0177] This embodiment designs a loss that balances high-frequency reconstruction and noise suppression:

[0178] ;

[0179] ;

[0180] in, Indicates high-frequency balance loss; This indicates high-frequency fidelity loss, penalizing the recovery error of high-frequency components; This indicates a loss in frequency spectrum smoothing, which suppresses high-frequency noise amplification; High-frequency threshold; and These are the weighting coefficients.

[0181] Step S4: Embed the representation rules into the physical information neural network in the form of differentiable physical constraints. The physical constraints include at least the physical equation constraints within each tissue layer and the coupling constraints at the interface between adjacent tissue layers.

[0182] Specifically, this step is the core of the physical constraints of the entire physiological hierarchical prior physical information neural network fusion physical information brain-computer interface signal enhancement and evaluation system. Its core function is to convert the multi-layer brain tissue prior physical knowledge in Module 2 into differentiable partial differential equation constraints, and design its embedding method and multi-constraint coordination optimization strategy in the fusion physical information brain-computer interface signal enhancement and evaluation network to ensure that the reconstructed brain-computer interface signal simultaneously satisfies neurodynamic constraints and volume conduction constraints.

[0183] Furthermore, embedding representation rules into the physical information neural network in the form of differentiable physical constraints specifically includes:

[0184] 1. Define the residual loss of the neurodynamic equation for cortical tissue as a physical constraint.

[0185] Specifically, existing methods lack physical constraints on cortical neural activity, potentially leading to reconstructed signals that violate neurophysiological principles. The cortex is the primary source of brain-computer interface (BCI) signals, and its neural activity should satisfy neurodynamic equations. Therefore, this embodiment embeds neurodynamic equations as physical constraints into the training of a BCI signal enhancement and evaluation network that integrates physical information, ensuring that the reconstructed cortical activity conforms to neurophysiological principles.

[0186] This embodiment proposes a neurodynamic residual constraint formula for signal enhancement and evaluation of brain-computer interfaces that integrate physical information:

[0187] ;

[0188] ;

[0189] in, Indicates cortical location The residual of the neurodynamic equation at time t; This represents the excitatory postsynaptic potential estimated by the network; This represents the mean membrane potential of the cone cells estimated by the network. The neurodynamic constraint loss is represented by the squared residual in the cortical domain. And the expected value over time. The smaller the residual, the better the network output conforms to the neurodynamic equation.

[0190] 2. Define the residual loss of the corresponding volume transfer equation for each tissue layer as a residual constraint.

[0191] Specifically, existing methods typically employ homogeneous volumetric conduction models, neglecting the differentiated electrical characteristics of the six brain tissue layers. Therefore, this embodiment designs differentiated volumetric conduction equation constraints for each of the six brain tissue layers—cortex, white matter, cerebrospinal fluid, dura mater, skull, and scalp—to ensure that the brain-computer interface signal enhancement and evaluation results, which integrate physical information, satisfy the physical laws of each tissue layer.

[0192] This embodiment designs volumetric transfer equation residual constraint mechanisms for each of the six tissue layers. The physical constraint equations for each layer are as follows:

[0193] Within the cortical domain, the potential field and the source current density satisfy the Poisson equation:

[0194] ;

[0195] in, Represents the potential within the cortical region (unit: ); Indicates cortical conductivity (unit: S / m); Source current density (unit: ).

[0196] Within the white matter domain, the potential field satisfies the anisotropic Laplace equation (source-free condition):

[0197] ;

[0198] in, This represents the potential within the white matter region; This represents the anisotropic conductivity tensor of white matter.

[0199] The dura mater is described using a thin-layer impedance model to depict the potential transitions between its inner and outer surfaces:

[0200] ;

[0201] in, and These represent the electrical potentials of the outer and inner surfaces of the dura mater, respectively. The resistance of the dura mater (unit: ), The thickness of the dura mater. Dura mater conductivity; This represents the normal current density passing through the dura mater.

[0202] Within the skull and scalp region, the potential field satisfies the Laplace equation (without source conditions):

[0203] ;

[0204] At the scalp-air interface, the insulation boundary conditions are satisfied:

[0205]

[0206] in, Indicates the outer boundary of the scalp; This represents the directional derivative along the outward normal.

[0207] 3. At the physiological interface between adjacent tissue layers, both potential continuity constraints and normal current continuity constraints are applied simultaneously.

[0208] Specifically, existing methods often neglect the physical continuity of interfaces between layers when processing multilayered brain tissue. The potential and current continuity of the five interfaces—gray matter-white matter, white matter-cerebrospinal fluid, cerebrospinal fluid-dura mater, dura mater-skull, and skull-scalp—are crucial for integrating the layered physical model into a unified system. Therefore, this embodiment establishes coupling constraints on all brain tissue interfaces for signal enhancement and evaluation of brain-computer interfaces that integrate physical information, ensuring the physical consistency of the reconstructed signal across each layer.

[0209] At any brain tissue interface, the fundamental conservation law of bioelectricity requires that both potential continuity and current continuity be satisfied. This embodiment defines a unified coupling constraint framework for the interfaces between five brain tissue layers:

[0210] Potential continuity constraint: at the interface At this point, the potential between two adjacent layers should be continuous:

[0211] ;

[0212] Current continuity constraint: at the interface At this point, the normal current density should be continuous:

[0213] ;

[0214] in, This represents the interface between the i-th layer and the j-th layer; and These represent the electrical potentials on both sides of the interface; and Represent the conductivity on both sides of the interface (for white matter, the conductivity tensor). ); Let be the unit vector normal to the interface.

[0215] Total interface coupling loss:

[0216] ;

[0217] in, This represents the total interface coupling loss; This represents a set of five inter-layer interfaces; and These represent the weights for the potential continuity and current continuity constraints, respectively. The gray-white matter interface needs to handle isotropic-anisotropic current continuity matching; the dura mater-skull interface needs to handle the connection between the thin-layer model and the volumetric model.

[0218] 4. Use a differentiable neural network to learn the residual correction terms for the boundary conditions of each interface.

[0219] Specifically, due to factors such as brain tissue model simplification, numerical errors, and individual anatomical differences, strict brain tissue interface conditions may not be fully met. Therefore, this embodiment designs a learnable brain tissue boundary condition residual correction mechanism to improve the generalization performance of brain-computer interface signal enhancement and evaluation models that integrate physical information.

[0220] This embodiment designs a residual correction network for each interface and learns the correction terms at the interface. This embodiment proposes a learnable boundary residual correction formula:

[0221] ;

[0222] ;

[0223] in, Display Interface Learnable potential correction term at (unit: ); The parameter is A multilayer perceptron; input features include the potentials on both sides of the interface. Potential gradient Conductivity parameters Interface normal vector and interface curvature When the correction item At that time, it degenerates into a strict potential continuity condition.

[0224] To prevent the correction term from becoming too large, this embodiment designs a regularization loss for the correction term:

[0225] ;

[0226] in, This represents the regularization loss of the modification term; For the first The regularization weights of the interface. This loss penalty corrects for excessively large errors, ensuring that the correction is only used to compensate for small biases caused by model simplification or numerical errors.

[0227] 5. A phased training strategy and an adaptive weight scheduling mechanism are adopted to coordinate and optimize the signal reconstruction loss and various physical constraint losses.

[0228] Specifically, the enhancement and evaluation of brain-computer interface signals incorporating physical information involves various physical constraints, such as neurodynamic constraints, six-layer volume conduction constraints, and interface coupling constraints. Therefore, this embodiment designs a multi-task joint optimization strategy and an adaptive constraint weight scheduling strategy for the enhancement and evaluation of brain-computer interface signals incorporating physical information.

[0229] This embodiment proposes a multi-task joint optimization method for brain-computer interface signal enhancement and evaluation total loss function that integrates physical information:

[0230] ;

[0231] in, Indicates the total loss; This indicates the loss of signal reconstruction in the brain-computer interface; This represents the loss due to the kth type of physical constraint (including neurodynamic constraints, volume conduction constraints at each layer, interface coupling constraints, etc.). This represents the weight of the k-th type of constraint at training step t.

[0232] Training is divided into three phases, and weights are scheduled according to the following strategy:

[0233] ;

[0234] in, The maximum weight of the k-th type of constraint; and The threshold number of training steps for phase switching. Phase 1 ( The main optimization is the reconstruction loss; the second stage ( Gradually increase the weight of physical constraints; the third stage ( All constraints reach their maximum weight.

[0235] This embodiment is designed based on adaptive weight adjustment according to constraint satisfaction:

[0236] ;

[0237] in, This represents the average value of all physical constraint losses; To adjust the rate parameters, constraints with poor satisfaction (loss higher than average) are given greater weight, promoting balanced optimization among all constraints.

[0238] Step S5: Input the collected scalp EEG signals into the physical information neural network for training. By minimizing the total loss function that includes signal reconstruction loss and multiple physical constraint losses, the enhanced intracranial EEG signals and multimodal neurophysiological information are jointly output.

[0239] Specifically, this embodiment is the core output of the entire brain-computer interface signal enhancement and evaluation system that integrates physiological hierarchical prior physical information neural network with physical information, solving the fundamental defects of existing methods such as single output information and lack of interpretability. This embodiment realizes the joint estimation and output of multiple clinically valuable neuroelectrophysiological signals, significantly improving the interpretability and clinical application value of brain-computer interface signal enhancement and evaluation results that integrate physical information.

[0240] The multimodal neuroelectrophysiological information jointly output includes: the spatiotemporal distribution of potentials in each layer of brain tissue, the three-dimensional vector field of cortical source current density, the time series of neurodynamic state variables, the individualized set of neurodynamic parameters, and the distribution of normal current density at each layer interface.

[0241] The physiological hierarchical prior physical information neural network in this embodiment simultaneously outputs multiple clinically valuable brain-computer interface signals, providing rich diagnostic information for clinical applications such as non-invasive brain-computer interface control. The mathematical expressions of each output signal are as follows:

[0242] (a) Reconstructed intracranial electroencephalogram (EEG) signals:

[0243] This embodiment proposes a formula for intracranial EEG reconstruction based on potential field sampling:

[0244] ;

[0245] in, Indicates the location of the m-th intracranial electrode The reconstruction potential at time t (unit: M represents the total number of intracranial electrodes; The potential field estimated by the network; It is a learnable electrode-specific noise correction term used to compensate for factors such as electrode-tissue contact impedance.

[0246] (b) Spatiotemporal distribution of tissue potentials in each layer:

[0247] This embodiment proposes an analytical formula for layered potential fields:

[0248] ;

[0249] in, Indicates the position of the k-th layer of organization. The potential at time t; Let be the soft mask value for the k-th layer. This formula allows clinicians to visualize the spatiotemporal distribution of potentials in different tissue layers and understand the signal propagation path.

[0250] (c) Three-dimensional vector field of cortical source current density:

[0251] ;

[0252] in, Indicates cortical location Source current density vector (unit: This output can be used for clinical applications such as epileptic focus localization and functional area mapping.

[0253] (d) Time series of neurodynamic state variables:

[0254] ;

[0255] in, Indicates cortical location The neurodynamic state vector at time t contains excitatory and inhibitory postsynaptic potentials and their time derivatives. This output reflects the excitation-inhibition balance of cortical neural activity.

[0256] (e) Individualized neurodynamic parameter set:

[0257] ;

[0258] in, This represents an individualized set of neurodynamic parameters, including synaptic gain (A, B), time constant (a, b), and connection strength (…). - ), diffusion coefficient ( , ), Co-activation intensity ( , ), and the conductivity of each layer ( - Parameters such as ) have clear neurophysiological significance.

[0259] (f) Distribution of normal current density at each interface:

[0260] This embodiment proposes a formula for calculating interface current density:

[0261]

[0262] in, Represents the interface between layer i and layer j. Normal current density at (unit: ); This represents the set of five interlayer interfaces. The output reflects the current transfer process between the layers.

[0263] Furthermore, the method also includes a systematic evaluation of the enhanced brain-computer interface signals, specifically including:

[0264] 1. Based on the qualitative assessment by professional physicians, the qualitative assessment should include at least the electroencephalogram (EEG) fidelity score and the clinical decision support effectiveness score.

[0265] 2. Quantitative assessment based on objective indicators, including at least inter-layer signal attenuation recovery, high-frequency oscillation recovery index, and similarity of neurodynamic state trajectories.

[0266] Furthermore, quantitative assessments based on objective indicators also include:

[0267] In at least one downstream brain-computer interface task, performance metrics are compared between those obtained using enhanced intracranial EEG signals and those obtained using raw scalp EEG signals. The performance metrics include at least one of classification accuracy, information transfer rate, and epileptiform focus lateralization concordance rate.

[0268] Specifically, this embodiment establishes a multi-dimensional systematic evaluation framework to verify the effectiveness of brain-computer interface signal enhancement and evaluation that integrates physical information from three dimensions: qualitative evaluation by professional physicians, quantitative evaluation of multiple indicators, and comprehensive evaluation of downstream tasks, providing sufficient evidence support for clinical translation.

[0269] (1) Qualitative evaluation module for the signal enhancement and assessment effect of brain-computer interface based on professional physicians and fusion of physical information:

[0270] Existing methods lack qualitative assessments based on clinical expert experience, and purely numerical indicators cannot reflect the clinical applicability and physiological rationality of enhanced signals. This embodiment proposes a qualitative assessment method for the enhancement and evaluation of brain-computer interface signals incorporating physical information, based on the subjective evaluation of professional neurologists, to verify whether the enhanced brain-computer interface signals meet the high-quality expectations of clinicians.

[0271] This embodiment proposes a qualitative assessment framework based on blinded evaluation by professional physicians, inviting neurologists with extensive experience in interpreting intracranial EEG signals to conduct a blinded evaluation of the enhanced brain-computer interface signals. This embodiment proposes the following qualitative assessment indicators:

[0272] (a) Fidelity score of electroencephalogram (EEG) characteristics:

[0273] ;

[0274] in, Indicates the fidelity score of electroencephalographic features; Indicates the number of physiological characteristics to be evaluated; This represents the clinical importance weight of the f-th physiological characteristic; This is an indicator function. Physiological characteristics include: morphological integrity of sleep spindle waves, identifiability of K-complexes, fidelity of epileptiform spike-slow wave complexes, sensorimotor rhythm μ-wave modulation characteristics, and high-frequency detail retention of gamma oscillations. This indicator specifically assesses the enhancement of brain-computer interface signals incorporating physical information and evaluates whether characteristic EEG waveforms with clinical diagnostic value are preserved.

[0275] (b) Cross-modal consistency blind evaluation index:

[0276] ;

[0277] in, Indicates the cross-modal consistency blind evaluation index; Indicates the number of paired assessments; physicians determine the enhanced signal without knowing its source. Compared with real intracranial EEG Whether it comes from the same subject. This indicator assesses whether the enhanced signal reaches a quality level that is indistinguishable from real intracranial EEG.

[0278] (c) Clinical decision support effectiveness score:

[0279] ;

[0280] in, This score indicates the effectiveness of clinical decision support; physicians make clinical decisions (such as epileptic focus lateralization and surgical resection recommendations) based on both enhanced signals and actual intracranial EEG data, and the consistency between these two decisions is assessed. This indicator directly evaluates whether enhanced signals can support correct clinical decisions.

[0281] (d) Neurophysiological rationality score:

[0282] ;

[0283] in, R represents the neurophysiological rationality score; R represents the number of neurophysiological assessment dimensions; the assessment dimensions include: whether the spatial distribution of cortical source activity conforms to functional anatomy (r=1), whether the signal propagation delay conforms to the nerve conduction velocity (r=2), whether the interlaminar potential attenuation conforms to the volume conduction law (r=3), and whether the cross-frequency coupling of neural oscillations is maintained (r=4). This index specifically assesses whether the enhanced signal conforms to the basic principles of neurophysiology.

[0284] (2) Quantitative evaluation module for signal enhancement and evaluation effect of brain-computer interface based on objective indicators and fusion of physical information

[0285] Existing technologies struggle to comprehensively quantify signal enhancement effects from multiple perspectives. This embodiment designs a multi-dimensional objective indicator system to comprehensively and quantitatively evaluate the signal enhancement and assessment effects of brain-computer interfaces that integrate physical information from four dimensions: signal quality, time-frequency characteristics, spatial distribution, and physical consistency. This embodiment establishes a multi-dimensional quantitative evaluation indicator system, proposing the following evaluation indicators for brain-computer interface signal enhancement and assessment that integrate physical information, in addition to conventional basic indicators such as correlation coefficient and mean square error:

[0286] (a) Enhancement and assessment of brain-computer interface signals incorporating physical information:

[0287] i. Interlayer signal attenuation recovery:

[0288] ;

[0289] in, This indicates the degree of signal attenuation recovery between layers; this indicator assesses whether the enhancement method correctly recovered the degree of attenuation of the EEG signal as it passed through six layers of brain tissue. This indicates that the attenuation recovery is accurate. This metric is a specific evaluation indicator for the layered physical model in this embodiment.

[0290] ii. High-frequency oscillation recovery index:

[0291] ;

[0292] in, Indicates the high-frequency brain-computer interface signal recovery index; This indicates the power of the high-frequency band (80-500Hz) of the brain-computer interface signal; This indicates the total power. The skull severely attenuates high-frequency signals; this indicator is specifically used to assess the recovery capability of high-frequency components in this embodiment.

[0293] iii. Source depth dependency error distribution:

[0294] ;

[0295] in, The expected reconstruction error is represented at a source depth of d; this metric assesses the reconstruction capability of enhancement methods for neural sources at different depths. Deep source signals exhibit more severe attenuation; this metric reveals the effectiveness of brain-computer interface signal enhancement and assessment for deep sources, particularly when integrating physical information.

[0296] (b) Neurodynamic consistency index:

[0297] i. Similarity of neural state trajectories:

[0298] ;

[0299] in, Indicates the similarity of neural state trajectories; This represents the neurodynamic state vector (pyramidal cell membrane potential, excitatory postsynaptic potential, inhibitory postsynaptic potential). This index assesses whether the enhanced signal maintains the correct neurodynamic evolutionary trajectory.

[0300] ii. Excitation-inhibition balance maintenance:

[0301] ;

[0302] in, Indicates the degree of maintenance of excitation-inhibition balance; This represents the ratio of excitatory to inhibitory postsynaptic potentials. The excitation-inhibition balance in the nervous system is an important physiological characteristic, and this indicator assesses whether enhanced signals maintain this crucial property.

[0303] iii. Consistency in the estimation of neurodynamic parameters:

[0304] ;

[0305] in, This indicates the correlation of the estimated neurodynamic parameters; This represents the model parameter vector. This metric evaluates whether this embodiment can accurately estimate individualized neurodynamic parameters.

[0306] (c) Cross-modal reconstruction quality indicators:

[0307] i.EEG-iEEG cross-modal information retention rate:

[0308] ;

[0309] in, Indicates the cross-modal information retention rate; This indicates mutual information. This metric assesses whether the enhanced signal maintains the correct association with the original EEG information.

[0310] ii. Spatial resolution enhancement factor:

[0311] ;

[0312] in, Indicates the spatial resolution enhancement factor; This represents the full width at half maximum (FWHM) of the point spread function. This indicates that spatial resolution has been improved. This metric quantifies the degree to which this embodiment improves spatial blur.

[0313] (3) Comprehensive evaluation module for the signal enhancement and evaluation effect of brain-computer interface based on downstream task-integrated physical information:

[0314] This embodiment designs a comprehensive evaluation method for downstream tasks of non-invasive brain-computer interfaces (BCIs). By comparing the effect with that using the original signal, it verifies the practical value of signal enhancement for clinical applications. This embodiment verifies the effectiveness of enhanced signals in actual non-invasive BCI clinical tasks. In addition to conventional basic indicators such as accuracy and F1 score, it proposes downstream task evaluation indicators for BCI signal enhancement and evaluation that incorporate physical information:

[0315] (a) Marginal benefit index of signal enhancement:

[0316] i. Gains under task difficulty conditions:

[0317] ;

[0318] in, This represents the relative performance gain when the task difficulty is d; It can be defined as the number of categories, time window length, signal-to-noise ratio, etc. This metric evaluates the marginal benefit of signal enhancement under different difficulty conditions, revealing the advantages of this embodiment in difficult tasks.

[0319] ii. Cross-subject generalization gain:

[0320] ;

[0321] in, Indicates the generalization gain across subjects; This represents the cross-subject classification accuracy on the i-th test subject. This metric assesses whether signal enhancement improves the cross-subject generalization ability of the brain-computer interface system.

[0322] iii. Few-shot learning gain:

[0323] ;

[0324] in, This represents the relative performance gain when the number of training samples is n. This metric assesses whether signal enhancement reduces the brain-computer interface system's dependence on the amount of training data.

[0325] (b) Clinical decision support indicators:

[0326] i. Concordance rate of epileptic focus localization:

[0327] ;

[0328] in, This indicates the concordance rate between epileptiform focus lateralization results based on enhanced signals and those based on actual intracranial EEG. Epileptiform focus lateralization is a crucial step in preoperative assessment, and this indicator evaluates the ability of enhanced signals to support clinical decision-making.

[0329] ii. Consistency in predicting surgical prognosis:

[0330] ;

[0331] in, The Kappa coefficient represents the agreement between the predicted surgical outcome based on the enhanced signal and the actual surgical outcome. This index assesses whether the enhanced signal can support accurate prediction of surgical outcome.

[0332] iv. Accuracy of seizure onset time detection:

[0333] ;

[0334] in, This indicates the detection error (in seconds) for the onset time of a seizure. Accurate seizure onset time detection is crucial for localizing the epileptic focus, and this indicator assesses whether the enhancement signal has maintained sufficient temporal accuracy.

[0335] (c) Real-time performance metrics of brain-computer interfaces:

[0336] i. Reduced online classification latency:

[0337] ;

[0338] in, This represents the reduction in the shortest time window length required to achieve the target accuracy (in milliseconds). This metric assesses whether signal enhancement can shorten the response time of the brain-computer interface system.

[0339] ii. Improved information transmission rate:

[0340] ;

[0341] in, The information transmission rate is expressed in bits per minute; N represents the number of classification categories; P represents the classification accuracy; and T represents the time of a single trial. It is a core indicator for measuring the practicality of brain-computer interface systems.

[0342] iii. Extended adaptive calibration cycle:

[0343] ;

[0344] in, Indicates the percentage increase in the recalibration cycle; This indicates the time required for classification performance to drop below a threshold. This indicates that enhanced signals improve the long-term stability of the brain-computer interface system.

[0345] (d) Multi-task comprehensive evaluation indicators:

[0346] i. Performance consistency across tasks:

[0347] ;

[0348] in, The coefficient of variation represents the performance improvement of different downstream tasks; K represents the number of downstream tasks. The smaller the value, the more consistent the signal enhancement effect is across different tasks.

[0349] ii. Pareto optimality assessment:

[0350] ;

[0351] in, This indicates the proportion of tasks in which the enhanced signal outperforms the original signal in a Pareto sense; This indicates Pareto dominance. It assesses whether signal enhancement simultaneously improves multiple tasks without sacrificing certain tasks.

[0352] This embodiment achieves a breakthrough in EEG signal enhancement technology by deeply integrating individualized neuroanatomical structures, multi-level bioelectrophysiological prior knowledge, and physical information neural networks. The personalized six-layer brain tissue anatomical model constructed by this method meticulously depicts the significant differences in conductivity among different layers and the anisotropy of white matter. This fundamentally corrects the signal propagation distortion caused by traditional homogeneous or simplified models, significantly improving the fidelity and spatial resolution of reconstructed intracranial EEG signals, particularly effectively restoring high-frequency neural oscillations severely attenuated by the skull. By embedding the dynamic equations describing neural activity and the volumetric conduction equations describing signal propagation as differentiable constraints into the network training, and strictly ensuring the continuity of potential and current at interlayer interfaces, the enhancement results strictly adhere to the basic physical laws of neural electrical activity, greatly improving the physical rationality of the model and its generalization ability under data scarcity. Furthermore, the network can simultaneously output multimodal information such as cortical source current distribution, neural dynamic state parameters, and signal cross-layer propagation paths, providing unprecedented visual insights and quantitative evidence for clinical lesion localization and neural mechanism analysis, greatly enhancing the interpretability and clinical practical value of the technology. Meanwhile, the network supports individualized fine-tuning of key biophysical parameters and balances multiple constraints through adaptive optimization strategies, achieving precise and robust signal enhancement for different individuals.

[0353] Figure 2This is a schematic diagram of the functional modules of the brain-computer interface signal enhancement and evaluation device that integrates physical information according to an embodiment of the present invention. Figure 2 As shown, the brain-computer interface signal enhancement and evaluation device 20 that integrates physical information includes: an acquisition module 21, a construction module 22, a conversion module 23, an embedding module 24, and an output module 25.

[0354] Acquisition module 21 is used to acquire structural magnetic resonance images of an individual;

[0355] Module 22 is used to construct a personalized layered anatomical model of multi-layered brain tissue, including the cortex, white matter, cerebrospinal fluid, dura mater, skull, and scalp, based on structural magnetic resonance imaging.

[0356] The conversion module 23 is used to convert prior knowledge of neurophysiological processes of multi-layered brain tissue into representation rules that can be learned by neural networks.

[0357] Embedding module 24 is used to embed the representation rules into the physical information neural network in the form of differentiable physical constraints. The physical constraints include at least the physical equation constraints within each layer of the organization and the coupling constraints at the interface between adjacent organizational layers.

[0358] The output module 25 is used to input the collected scalp EEG signals into the physical information neural network for training. By minimizing the total loss function, which includes signal reconstruction loss and multiple physical constraint losses, it jointly outputs the enhanced intracranial EEG signals and multimodal neurophysiological information.

[0359] Optionally, the construction module 22 performs operations to construct a personalized layered anatomical model of multiple layers of brain tissue, including the cortex, white matter, cerebrospinal fluid, dura mater, skull, and scalp. Specifically, this includes:

[0360] Deep learning segmentation networks were used to process structural magnetic resonance images, automatically identifying and segmenting the boundaries of six layers of brain tissue.

[0361] Initial conductivity parameters based on neurophysiological measurements were set for each tissue layer, where the conductivity of the white matter layer was represented by an anisotropic conductivity tensor constructed based on diffusion tensor imaging data;

[0362] Based on the segmentation results, the three-dimensional boundary surfaces of each tissue layer are extracted, and the local thickness distribution of each tissue layer is calculated.

[0363] Optionally, the conversion module 23 performs the operation of converting prior knowledge of neurophysiological processes of multi-layered brain tissue into representational rules that can be learned by the neural network, specifically including:

[0364] A spatially extended neurodynamic model is constructed, which extends the ordinary differential equations describing the activity of neuronal populations into spatiotemporally distributed partial differential equations defined on the cortical surface.

[0365] Based on diffusion tensor imaging data, an anisotropic conductivity tensor reflecting the directional conduction characteristics of myelinated axonal fibers in white matter is constructed.

[0366] An approximate model of the equipotential field describing the current convergence effect in the cerebrospinal fluid domain due to its high conductivity was established.

[0367] A thin-layer impedance model was established to describe the relationship between the potential jumps on the inner and outer surfaces of the dura mater and the normal current density.

[0368] A transfer function model describing the frequency-dependent attenuation characteristics of the skull to EEG signals is established, and an optimal filtering regularization framework for recovering high-frequency components is designed.

[0369] Optionally, the embedding module 24 performs the operation of embedding the representation rules into the physical information neural network in the form of differentiable physical constraints, specifically including:

[0370] The residual loss of the neurodynamic equations for cortical tissue is defined as a physical constraint.

[0371] Define the residual loss of the corresponding volume transfer equation for each tissue layer as a residual constraint;

[0372] At the physiological interface between adjacent tissue layers, both potential continuity constraints and normal current continuity constraints are applied simultaneously.

[0373] Differentiable neural networks are used to learn the residual correction terms for each interface boundary condition;

[0374] A phased training strategy and an adaptive weight scheduling mechanism are adopted to coordinate and optimize the signal reconstruction loss and various physical constraint losses.

[0375] Optionally, the jointly output multimodal neuroelectrophysiological information includes: the spatiotemporal distribution of potentials in each layer of brain tissue, the three-dimensional vector field of cortical source current density, the time series of neurodynamic state variables, the individualized set of neurodynamic parameters, and the distribution of normal current density at each layer interface.

[0376] Optionally, the conductivity parameter is specifically expressed as:

[0377] ;

[0378] in, Indicates the first The actual electrical conductivity of the brain tissue; This indicates the first [number] based on neuroelectrophysiology literature. The first a priori value of the electrical conductivity of the layered tissue; For learnable conductivity correction parameters; This represents the maximum allowable correction factor.

[0379] Optionally, the output module 25 is also used to systematically evaluate the enhanced brain-computer interface signal, specifically including:

[0380] Based on the qualitative assessment by professional physicians, the qualitative assessment includes at least the electroencephalogram (EEG) fidelity score and the clinical decision support effectiveness score.

[0381] Quantitative assessment based on objective indicators includes at least inter-layer signal attenuation recovery, high-frequency oscillation recovery index, and neurodynamic state trajectory similarity.

[0382] Optionally, the quantitative evaluation based on objective indicators also includes: comparing the performance indicators obtained using enhanced intracranial EEG signals with those obtained using raw scalp EEG signals in at least one downstream brain-computer interface task, where the performance indicators include at least one of classification accuracy, information transmission rate, and epileptiform lateralization concordance rate.

[0383] For further details regarding the implementation techniques of each module in the brain-computer interface signal enhancement and evaluation device that integrates physical information in the above embodiments, please refer to the description in the brain-computer interface signal enhancement and evaluation method that integrates physical information in the above embodiments, which will not be repeated here.

[0384] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For apparatus embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0385] Please see Figure 3 , Figure 3 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention. Figure 3 As shown, the computer device 30 includes a processor 31 and a memory 32 coupled to the processor 31. The memory 32 stores program instructions. When the program instructions are executed by the processor 31, the processor 31 performs the steps of the brain-computer interface signal enhancement and evaluation method that integrates physical information as described in any of the above embodiments.

[0386] The processor 31 can also be referred to as a Central Processing Unit (CPU). The processor 31 may be an integrated circuit chip with signal processing capabilities. The processor 31 can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor can be a microprocessor or any conventional processor.

[0387] See Figure 4 , Figure 4 This is a schematic diagram of the storage medium according to an embodiment of the present invention. The storage medium of this embodiment stores program instructions 41 capable of implementing the above-described brain-computer interface signal enhancement and evaluation method for fusing physical information. These program instructions 41 can be stored in the storage medium in the form of a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, or computer devices such as computers, servers, mobile phones, and tablets.

[0388] In the several embodiments provided in this application, it should be understood that the disclosed computer devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between devices or units, and may be electrical, mechanical, or other forms.

[0389] Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for enhancing and evaluating brain-computer interface signals that integrates physical information, characterized in that, include: Acquire structural magnetic resonance images of an individual; Based on the aforementioned structural magnetic resonance images, a personalized layered anatomical model of multi-layered brain tissue, including the cortex, white matter, cerebrospinal fluid, dura mater, skull, and scalp, was constructed. The prior knowledge of the neurophysiology of the multilayered brain tissue is converted into representation rules that can be learned by the neural network. The representation rules are embedded in the physical information neural network in the form of differentiable physical constraints, which include at least the physical equation constraints within each tissue layer and the coupling constraints at the interface between adjacent tissue layers. The acquired scalp EEG signals are input into the physical information neural network for training. By minimizing the total loss function, which includes signal reconstruction loss and multiple physical constraint losses, the enhanced intracranial EEG signals and multimodal neurophysiological information are jointly output; wherein: The construction of a personalized layered anatomical model of multi-layered brain tissue, including the cortex, white matter, cerebrospinal fluid, dura mater, skull, and scalp, includes: The magnetic resonance images of the structure were processed using a deep learning segmentation network to automatically identify and segment the boundaries of the six layers of brain tissue. Initial conductivity parameters based on neurophysiological measurements were set for each tissue layer, where the conductivity of the white matter layer was represented by an anisotropic conductivity tensor constructed based on diffusion tensor imaging data; Based on the segmentation results, the three-dimensional boundary surfaces of each tissue layer are extracted, and the local thickness distribution of each tissue layer is calculated. The process of converting prior neurophysiological knowledge of the multilayered brain tissue into representational rules that can be learned by the neural network includes: A spatially extended neurodynamic model is constructed, which extends the ordinary differential equations describing the activity of neuronal populations into spatiotemporally distributed partial differential equations defined on the cortical surface. Based on diffusion tensor imaging data, an anisotropic conductivity tensor reflecting the directional conduction characteristics of myelinated axonal fibers in white matter is constructed. An approximate model of the equipotential field describing the current convergence effect in the cerebrospinal fluid domain due to its high conductivity was established. A thin-layer impedance model was established to describe the relationship between the potential jumps on the inner and outer surfaces of the dura mater and the normal current density. A transfer function model describing the frequency-dependent attenuation characteristics of the skull to EEG signals was established, and an optimal filtering regularization framework for recovering high-frequency components was designed. The feature is that embedding the representation rules into the physical information neural network in the form of differentiable physical constraints includes: The residual loss of the neurodynamic equations for cortical tissue is defined as a physical constraint. Define the residual loss of the corresponding volume transfer equation for each tissue layer as a residual constraint; At the physiological interface between adjacent tissue layers, both potential continuity constraints and normal current continuity constraints are applied simultaneously. Differentiable neural networks are used to learn the residual correction terms for each interface boundary condition; A phased training strategy and an adaptive weight scheduling mechanism are adopted to coordinate and optimize the signal reconstruction loss and various physical constraint losses.

2. The brain-computer interface signal enhancement and evaluation method integrating physical information according to claim 1, characterized in that, The jointly output multimodal neuroelectrophysiological information includes: the spatiotemporal distribution of potentials in each layer of brain tissue, the three-dimensional vector field of cortical source current density, the time series of neurodynamic state variables, the individualized set of neurodynamic parameters, and the distribution of normal current density at each layer interface.

3. The brain-computer interface signal enhancement and evaluation method integrating physical information according to claim 2, characterized in that, The conductivity parameter is specifically expressed as follows: ； in, Indicates the first The actual electrical conductivity of the brain tissue; This indicates the first [number] based on neuroelectrophysiology literature. The first a priori value of the electrical conductivity of the layered tissue; For learnable conductivity correction parameters; This represents the maximum allowable correction factor.

4. The brain-computer interface signal enhancement and evaluation method integrating physical information according to claim 1, characterized in that, It also includes a systematic evaluation of the enhanced intracranial electroencephalogram (EEG) signals, specifically including: Based on the qualitative assessment by a professional physician, the qualitative assessment includes at least the electroencephalogram (EEG) fidelity score and the clinical decision support effectiveness score. The quantitative assessment based on objective indicators includes at least the inter-layer signal attenuation recovery degree, the high-frequency oscillation recovery index, and the similarity of neurodynamic state trajectories.

5. The brain-computer interface signal enhancement and evaluation method integrating physical information according to claim 4, characterized in that, The quantitative assessment based on objective indicators also includes: In at least one downstream brain-computer interface task, performance metrics obtained using the enhanced intracranial EEG signals are compared with those obtained using the original scalp EEG signals, the performance metrics including at least one of classification accuracy, information transmission rate, and epileptiform focus lateralization concordance rate.

6. A computer device, characterized in that, It includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the brain-computer interface signal enhancement and evaluation method that integrates physical information as described in any one of claims 1 to 5.

7. A computer-readable storage medium, characterized in that, The device contains a computer program that, when executed by a processor, implements the brain-computer interface signal enhancement and evaluation method that incorporates physical information as described in any one of claims 1 to 5.

Images