A neural state decoding method and system based on energy condition coupled tensor learning

Abstract

The application discloses a neural state decoding method and system based on energy condition coupling tensor learning, and belongs to the technical field of neural signal processing and intelligent decoding, which comprises: dividing a multi-brain region local field potential signal into multiple time windows, extracting low-frequency phase and high-frequency amplitude, and calculating a frequency resolution phase amplitude coupling tensor composed of a real part, an imaginary part and a modulus value; adding phase frequency coordinates and amplitude frequency coordinates to a coupling atlas, and obtaining a coupling feature token through coordinate perception coding; extracting a frequency band energy and generating a scale parameter and a bias parameter, and calibrating the coupling feature token through bounded feature linear modulation; sequentially performing cross-brain region attention aggregation and time sequence aggregation, and outputting a decoding result of a Parkinson's disease baseline state and a levodopa-induced dyskinesia related state. The application retains a coupling frequency structure, limits the modulation amplitude of energy information, and takes into account cross-brain region and time sequence information, thereby improving the stability and interpretability of cross-subject decoding.

Description

Technical Field

[0001] This invention belongs to the field of neural signal processing and intelligent decoding technology, specifically relating to a neural state decoding method and system based on energy conditional coupling tensor learning. Background Technology

[0002] Levodopa is a commonly used drug in the treatment of motor symptoms in Parkinson's disease. With prolonged administration, some subjects may experience levodopa-induced dyskinesia. This state is associated with changes in rhythmic activity in multiple brain regions within the basal ganglia-cortical circuit. Local field potentials can reflect neural group activity and are suitable for extracting state-related bioelectrical characteristics from the perspective of spectral and transfrequency interactions.

[0003] Phase-amplitude coupling is used to describe the modulation relationship between the phase of a low-frequency oscillation and the amplitude of a high-frequency oscillation. Existing methods typically compress the coupling matrix into several statistics or classify it using only the magnitude of the coupling strength. Such processing loses the directional information of complex coupling values ​​and makes it difficult to explicitly distinguish coupling modes that are at different phase and amplitude frequencies but have similar values.

[0004] On the other hand, the coupling of frequency band energy and phase amplitude reflects different properties of neural signals. Directly concatenating the two types of features can easily make the energy features, which are larger in scale or easier to fit, the dominant input of the classifier; directly multiplying the energy as a weight onto the coupling scalar makes it difficult to implement fine-grained calibration in the already learned high-dimensional coupled representation, and may amplify noise or individual differences.

[0005] Multi-brain region neural signals also exhibit significant spatial heterogeneity and temporal dynamics. Fixed averaging across multiple brain regions or time windows fails to adaptively select brain regions and time points based on state correlation. Furthermore, dividing the training and test sets at the level of random segments may result in data from the same subject simultaneously entering both the training and testing processes, leading to an overestimation of cross-subject generalization performance.

[0006] Therefore, a neural state decoding scheme is needed that can preserve complex phase amplitude coupling information and frequency positional relationships, introduce energy conditions in a controlled manner, and jointly model brain region differences and temporal dynamics. Summary of the Invention

[0007] In view of the above, the purpose of this invention is to overcome the above deficiencies and provide a neural state decoding method and system based on energy conditional coupling tensor learning, which can be used to characterize the phase amplitude coupling structure in local field potential signals of multiple brain regions and to decode the baseline state of Parkinson's disease and the state related to levodopa-induced dyskinesia across subjects.

[0008] To achieve the above-mentioned objectives, an embodiment provides a neural state decoding system based on energy-conditionally coupled tensor learning, comprising: The coupling tensor construction module is used to extract multiple low-frequency phase components and multiple high-frequency amplitude envelope components from the local field potential signal in each time window of each brain region and calculate the complex phase amplitude coupling value. The real part, imaginary part and magnitude of the complex phase amplitude coupling value are used as complex representation channels to form a phase amplitude coupling tensor. The coordinate-aware encoding module is used to extract the phase-amplitude coupling feature token after adding normalized coordinate channels to the phase-amplitude coupling tensor to characterize the phase frequency position and amplitude frequency position. The energy condition modulation module is used to extract the frequency band energy features corresponding to the time window and brain region from the local field potential signal, generate scale parameters and bias parameters based on the frequency band energy features, and calibrate the phase amplitude coupling feature token based on the scale parameters and bias parameters and using bounded feature linear modulation. The aggregation decoding module is used to sequentially perform cross-brain region attention aggregation and cross-time window temporal aggregation on the calibrated phase amplitude coupling feature tokens to obtain the analysis fragment-level representation and decode it to obtain the neural state decoding result, which is used to distinguish between the baseline state of Parkinson's disease and the state related to levodopa-induced dyskinesia.

[0009] Preferably, the complex phase amplitude coupling value is calculated according to the following formula: in, Indicates the first The first low-frequency phase component and the second Complex phase amplitude coupling values ​​between the high-frequency amplitude envelope components This indicates the number of sampling points within the time window. Indicates the first Each high-frequency component at the sampling time The instantaneous amplitude, Indicates the first The low-frequency component at the sampling time The instantaneous phase, The imaginary unit is represented; the instantaneous amplitude and the instantaneous phase are obtained by bandpass filtering and Hilbert transform.

[0010] Preferably, the phase amplitude coupling tensor is a tensor formed by arranging the complex characterization channels of complex phase amplitude coupling values ​​according to the time window dimension, brain region dimension, phase frequency band dimension, and amplitude frequency band dimension.

[0011] Preferably, the normalized coordinate channels are generated based on the center frequency of the phase band and the center frequency of the amplitude band, respectively, and are spliced ​​with the complex representation channels corresponding to the real part, imaginary part and modulus value in the channel dimension to realize the addition of normalized coordinate channels to the phase amplitude coupling tensor; The phase amplitude coupling feature tokens are extracted by a coordinate-aware coding network based on the phase amplitude coupling tensor with added normalized coordinate channels. The coordinate-aware coding network includes at least two convolutional layers, a nonlinear activation layer, a normalization layer, and a linear projection layer to preserve the positional relationship of the coupling mode on the phase frequency axis and the amplitude frequency axis.

[0012] Preferably, the frequency band energy characteristics include the relative logarithmic band power within each preset frequency band, wherein the relative logarithmic band power is obtained by taking the logarithm of the corresponding frequency band power and normalizing it relative to the reference power of the same time window and the same brain region; The conditional parameter generation network generates scale parameters and bias parameters based on the frequency band energy characteristics. The conditional parameter generation network adopts a multilayer perceptron and outputs scale parameters and bias parameters with the same dimension as the phase amplitude coupling feature token.

[0013] Preferably, the bounded characteristic linear modulation is used to calibrate the phase amplitude coupling characteristic token according to the following formula: in, and These represent the phase amplitude coupling characteristic tokens before and after calibration, respectively. Indicates the scale parameter. Indicates the bias parameter. Indicates modulation intensity and , for function.

[0014] Preferably, the cross-brain region attention aggregation includes: for each time window, calculating normalized brain region weights based on the calibrated phase amplitude coupling feature tokens corresponding to each brain region, and performing a weighted summation of the calibrated phase amplitude coupling feature tokens corresponding to each brain region to obtain the brain region fusion representation for each time window.

[0015] Preferably, the cross-time window temporal aggregation includes: inputting the brain region fusion representations of each time window into a bidirectional gated recurrent unit in chronological order to obtain time window-level hidden states; calculating normalized time weights based on each time window-level hidden state and performing weighted summation to obtain analysis segment-level representations.

[0016] To achieve the above-mentioned objectives, embodiments of the present invention provide a neural state decoding method based on energy-conditionally coupled tensor learning, which employs the aforementioned system and includes the following steps: Multiple low-frequency phase components and multiple high-frequency amplitude envelope components are extracted from the local field potential signal in each time window of each brain region, and complex phase amplitude coupling values ​​are calculated. The real part, imaginary part, and magnitude of the complex phase amplitude coupling values ​​are used as complex characterization channels to form a phase amplitude coupling tensor. After adding normalized coordinate channels to the phase-amplitude coupling tensor to characterize the phase frequency position and amplitude frequency position, the phase-amplitude coupling feature token is extracted. Frequency band energy features corresponding to time windows and brain regions are extracted from local field potential signals. Scale parameters and bias parameters are generated based on the frequency band energy features. The phase amplitude coupling feature token is calibrated based on the scale parameters and bias parameters and using bounded feature linear modulation. After calibration, the phase amplitude coupling feature tokens are sequentially subjected to cross-brain region attention aggregation and cross-time window temporal aggregation to obtain the analysis fragment level representation. After decoding, the neural state decoding result is obtained, which is used to distinguish between the baseline state of Parkinson's disease and the state related to levodopa-induced dyskinesia.

[0017] To achieve the above-mentioned objectives, the embodiments also provide a computing device, including a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, it is used to implement the above-mentioned neural state decoding method based on energy conditional coupling tensor learning.

[0018] To achieve the above-mentioned objectives, the embodiments also provide a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements the above-mentioned neural state decoding method based on energy conditional coupling tensor learning.

[0019] Compared with the prior art, the beneficial effects of the present invention include at least the following: (1) At the same time, the real part, imaginary part and magnitude of the complex phase amplitude coupling value are retained to reduce the information loss caused by using only the coupling strength; (2) The position of the coupling unit is explicitly marked by the normalized coordinates of the phase frequency and the normalized coordinates of the amplitude frequency, so that the convolutional coding when extracting the phase amplitude coupling feature token can distinguish the local patterns at different frequency positions. (3) Introducing band energy using a bounded linear modulation method allows the energy to perform conditional calibration on the high-dimensional coupling characterization with a limited amplitude, reducing the risk of energy characteristics suppressing coupling characteristics or amplifying noise; (4) By utilizing brain region attention and temporal aggregation, information from different brain regions and different time windows is adaptively fused to adapt to the spatial heterogeneity and temporal dynamics of neural states in multiple brain regions. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the neural state decoding method based on energy conditional coupling tensor learning according to the present invention; Figure 2 This is a schematic diagram of the overall process of the present invention, illustrating the offline training, classification performance evaluation, and online state decoding processes; Figure 3 This is a schematic diagram of the frequency coordinate sensing encoding and energy conditional bounded modulation process according to an embodiment of the present invention; Figure 4 This is an explanatory diagram showing the overall phase amplitude coupling difference in one embodiment of the present invention; Figure 5 This is an explanatory diagram showing the difference in partitioned phase amplitude coupling in one embodiment of the present invention; Figure 6 This is an explanatory diagram of brain region hotspots in one embodiment of the present invention; Figure 7 This is an explanatory diagram of the energy condition modulation parameters in one embodiment of the present invention; Figure 8 This is a schematic diagram of the neural state decoding system based on energy conditional coupling tensor learning according to the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of this invention.

[0023] This invention provides a neural state decoding scheme based on energy-conditional coupling tensor learning. Instead of simply weighting the original phase-amplitude coupling values ​​with energy after obtaining the phase-amplitude coupling tensor, it generates constrained scale and bias parameters from the energy features after obtaining high-dimensional coupling feature tokens in a coordinate-aware network. These parameters are then bounded and calibrated. Finally, through cross-brain region attention aggregation and cross-time window temporal aggregation, the scheme decodes and distinguishes between the baseline state of Parkinson's disease and the state related to levodopa-induced dyskinesia. This invention preserves the coupling frequency structure, limits the modulation amplitude of energy information, and considers both cross-brain region and temporal information, improving the stability and interpretability of decoding across subjects.

[0024] like Figure 2As shown, the overall process of the neural state decoding scheme of this invention includes two stages: offline training and online decoding. In the offline stage, local field potential signals from multiple brain regions are preprocessed and time-windowed. A complex PAC tensor is constructed sequentially, and relative logarithmic band power is extracted. Through frequency coordinate perceptual encoding, conditional parameter generation, energy conditionally bounded feature linear modulation, cross-brain region attention aggregation, and temporal encoding and attention, a classification model is trained and its parameters are saved. The classification performance is then evaluated. In the online stage, the pre-trained model is loaded according to the same data processing flow for state decoding, and the neural state decoding results are output. Each stage is described in detail below.

[0025] like Figure 1 As shown in the embodiment, a neural state decoding method based on energy-conditionally coupled tensor learning is provided, including the following steps: S1 extracts multiple low-frequency phase components and multiple high-frequency amplitude envelope components from the local field potential signal in each time window of each brain region and calculates the complex phase amplitude coupling value. The real part, imaginary part and magnitude of the complex phase amplitude coupling value are used as complex characterization channels to form the phase amplitude coupling tensor.

[0026] In this embodiment, firstly, local field potential signals of multiple brain regions of the subject are acquired, the local field potential signals are preprocessed, and each analysis segment of each brain region is divided into multiple time windows.

[0027] Then, bandpass filtering was applied to the local field potential signals for each time window and each brain region. The low-frequency filtering result was then subjected to Hilbert transform to obtain the instantaneous phase. The instantaneous amplitude is obtained by Hilbert transform of the high-frequency filtering result. Then the first The first low-frequency phase component and the second Complex phase amplitude coupling value between each high-frequency amplitude envelope component for: in, This indicates the number of sampling points within the time window. Indicates the first Each high-frequency component at the sampling time The instantaneous amplitude, Indicates the first The low-frequency component at the sampling time The instantaneous phase, It represents the imaginary unit. The magnitude of the coupling vector reflects the coupling strength, and the real and imaginary parts together preserve the directional information of the coupling vector in the complex plane.

[0028] Finally, for each Take the real part respectively virtual part And modulus This forms three complex representation channels. These channels are arranged according to the time window dimension, brain region dimension, phase frequency band dimension, and amplitude frequency band dimension to form a batch-input phase amplitude coupling tensor, which can be represented as: in, Indicates batch size. Indicates the number of time windows. The number represents the number of brain regions, K and Q represent the number of phase bands and amplitude bands, respectively, and 3 represents the number of channels.

[0029] S2, after adding normalized coordinate channels to the phase amplitude coupling tensor to characterize the phase frequency position and amplitude frequency position, extract the phase amplitude coupling feature token.

[0030] In the embodiments, such as Figure 3 As shown in the upper half, to enable the network to distinguish the meaning of the same local shape appearing at different frequency positions, a phase frequency coordinate map is generated based on the center frequency of each phase band, and an amplitude frequency coordinate map is generated based on the center frequency of each amplitude band, forming phase frequency coordinate channels and amplitude frequency coordinate channels respectively. The center frequency is linearly normalized to the interval [-1, 1] and copied in the corresponding dimension so that the two coordinate maps have the same spatial size as the K×Q phase amplitude coupling tensor.

[0031] The phase-frequency coordinate graph and the amplitude-frequency coordinate graph are used as two additional channels, along with... , and The three channels are concatenated to obtain a five-channel input for the coordinate-aware coding network. This five-channel input is then sequentially processed through two layers of coordinate-aware convolution, non-linear activation, and normalization. It is then flattened, linearly projected, and normalized to obtain the corresponding input for each time window and each brain region. Phase-Amplitude Coupling (PAC) Feature Token .

[0032] S3 extracts the frequency band energy features corresponding to the time window and brain region from the local field potential signal, generates scale parameters and bias parameters based on the frequency band energy features, and calibrates the phase amplitude coupling feature token based on the scale parameters and bias parameters and using bounded feature linear modulation.

[0033] In this embodiment, for each time window and each brain region, the band power corresponding to a preset frequency band is calculated. To reduce individual gain and electrode impedance differences, the band power is first logarithmically calculated, and then normalized relative to the reference power of that time window and brain region to obtain a relative logarithmic band power vector. The frequency band energy characteristics are composed of the reference power, which can be the total power of a preset wideband, the sum of the power of each frequency band, or the reference power determined in the training data.

[0034] like Figure 3 As shown in the lower half, the relative logarithmic power vector Input conditional parameter generation network, where the conditional parameter generation network can be a multilayer perceptron (MLP), generates PAC feature tokens. Scale parameters with the same dimensions and bias parameters Then, the following formula is used to apply the PAC feature token. Perform bounded characteristic linear modulation: in, and These represent the phase amplitude coupling characteristic tokens before and after calibration, respectively. Indicates modulation intensity and , This is the hyperbolic tangent function. The hyperbolic tangent function restricts scale and bias variations to a state determined by... Within a defined range, preferably, This allows the energy branch to primarily serve as a conditional calibration function, rather than as a high-weight classification input parallel to the coupling branch.

[0035] S4 sequentially performs cross-brain region attention aggregation and cross-time window temporal aggregation on the calibrated phase amplitude coupling feature tokens to obtain the analysis fragment level representation and decode it to obtain the neural state decoding result, which is used to distinguish between the baseline state of Parkinson's disease and the state related to levodopa-induced dyskinesia.

[0036] In this embodiment, for each time window, normalized brain region weights are calculated based on the calibrated phase amplitude coupling feature tokens corresponding to each brain region, and the calibrated phase amplitude coupling feature tokens corresponding to each brain region are weighted and summed to obtain the brain region fusion representation for each time window, thereby achieving cross-brain region attention aggregation. Then, the brain region fusion representations of each time window are input into a bidirectional gated loop unit in chronological order to obtain the time window-level hidden state. Normalized time weights are calculated based on the time window-level hidden state and weighted and summed to achieve cross-time window temporal aggregation, thereby obtaining the analysis segment-level representation.

[0037] Specifically, regarding the first A time window, The calibrated feature tokens corresponding to each brain region are input into a shared attention scoring network, and the brain region weights are obtained by Softmax normalization. . No. Brain region fusion representation within a time window for: in, For the first Within the first time window The calibrated feature tokens corresponding to each brain region; Arranged in chronological order The input is a bidirectional gated recurrent unit to simultaneously encode forward and backward temporal dependencies. Temporal attention weights are calculated for the hidden states of each time window and then summed to obtain the segment-level representation of the analysis.

[0038] Finally, the fragment-level representations are input into a linear classifier, which outputs decoding results based on the fragment-level representations. These results contain logical values ​​for two types of neural states, which are then processed by Softmax to obtain class probabilities. These probabilities are used to distinguish between the baseline state of Parkinson's disease and the state associated with levodopa-induced dyskinesia.

[0039] All network parameters used in the above method, including those of the coordinate-aware coding network, conditional parameter generation network, attention scoring network, bidirectional gated recurrent unit, and linear classifier, have been trained and optimized. During training, the training, validation, and test sets are divided according to the subjects, ensuring that all signals from the test subjects are excluded from parameter training, model selection, and normalized statistics calculation. All normalized statistics are calculated solely from the training subjects' data and are consistently used during the validation and testing phases. This isolation of training, validation, and test data according to the subjects, and the calculation of normalized statistics solely from the training data, allows for a more realistic evaluation and improvement of stability in cross-subject applications.

[0040] Specific experimental examples are also provided to illustrate the invention. Specifically, local field potential signals from multiple brain regions in hemilateral Parkinson's disease rats before and after levodopa administration were used. The recorded brain regions included the caudate putamen (CPu), primary motor cortex (MOp), and secondary motor cortex (MOs). The raw signal sampling rate was 2000 Hz, which was downsampled to 500 Hz during preprocessing. Each ten-minute recording was divided into non-overlapping 60-second analysis segments; each analysis segment was further divided into... A time window.

[0041] Recordings from 20 to 10 minutes before drug administration were defined as the Parkinson's disease baseline state, and recordings from 45 to 55 minutes after drug administration were defined as the levodopa-induced dyskinesia-related state, used as true labels. In this example, the latter label represents the post-drug administration state, rather than a clinical diagnosis or severity grading based on segment-by-segment behavioral scores. A total of 10 rats and 59 ten-minute recordings were recorded, including 30 baseline state recordings and 29 dyskinesia-related state recordings.

[0042] Then, proceed with steps S1-S4, setting 7 phase frequency bands and 4 amplitude frequency bands. The phase frequency bands can be set to 2–4 Hz, 3–5 Hz, 4–6 Hz, 5–7 Hz, 6–8 Hz, 8–10 Hz, and 10–12 Hz; the amplitude frequency bands can be set to 50–85 Hz, 85–120 Hz, 120–160 Hz, and 160–205 Hz. Time window. Batch size PAC Feature Token Dimensions Modulation intensity .

[0043] Leave-one-out cross-validation was employed. Each time, all data from one rat were used as the test set, while the remaining rats were used for training and validation. Then, 20% of the training rats were used as validation subjects. Data from the test subjects were not used for gradient updates, early stopping, model selection, or normalization statistic calculations. The mean and standard deviation calculated from the training set remained constant during both the validation and testing phases.

[0044] During training, the cross-entropy loss function between the decoded results and the ground truth labels, along with the AdamW optimizer, are used. An optional configuration is an initial learning rate of 5×10⁻⁶. -4 Batch size 5, weight decay 1×10 -4 The maximum training duration is 80 epochs; training should be stopped early if the validation metric fails to improve for 10 consecutive epochs, but at least 15 epochs should be trained. Experiments were repeated using three random seeds (42, 43, and 44), and results were reported as mean and standard deviation. The above parameters are merely examples and do not constitute limitations.

[0045] Table 1. Experimental results across subjects under three random initializations (mean ± standard deviation) As shown in Table 1, the macro-F1 score of the method of the present invention is 0.829±0.008, and the area under the receiver operating characteristic curve (ROC-AUC) is 0.894±0.011. The macro-F1 score of the phase amplitude coupling (PAC) statistic compared to the radial basis function kernel support vector machine (RBF-SVM) method is 0.767±0.009, and the ROC-AUC is 0.831±0.001. Compared to this control method, the method of the present invention shows improvements in both macro-F1 and ROC-AUC in the data of this embodiment.

[0046] Ablation results showed that the changes in various evaluation metrics were not entirely consistent after removing different modules. For example, when the frequency coordinate channel was removed, the mean Macro-F1 was 0.840, but the standard deviation increased to 0.018 and the ROC-AUC decreased to 0.884; when the energy conditional modulation was removed, the Macro-F1 remained basically unchanged, but the ROC-AUC decreased to 0.879. Therefore, this embodiment does not interpret each module as generating independent monotonic gains for all individual metrics, but rather uses the complete structure to improve probabilistic discrimination ability and maintain relatively stable overall performance across subject settings.

[0047] like Figure 4 and Figure 5 As shown, the phase-amplitude coupling plots of levodopa-induced dyskinesia-related states (LID-related states) and Parkinson's disease baseline states (PD baselines) were compared, and the state difference was represented by the PAC amplitude difference ΔPAC between the "LID-related states and PD baselines". A significant positive hotspot was observed near the combination of 3–7 Hz phase and 160–205 Hz amplitude, a trend observed in the CCu, MOp, and MOs brain regions, with MOp being particularly prominent. A negative trend was observed near the combination of 8–10 Hz phase and 50–85 Hz amplitude. These results indicate that preserving two-dimensional frequency locations helps in locating state-related transfrequency patterns.

[0048] like Figure 6 As shown, the hotspot summary by brain region gives the hotspot ΔPAC amplitude corresponding to the combination of 3-7 Hz phase and 160-205 Hz amplitude; as Figure 7 As shown, the scale term, bias term, and total modulation amount generated by the energy conditional branch have relatively small overall amplitudes, which aligns with the bounded modulation design objective of small-amplitude, state-dependent calibration. The bias term and total modulation amount are improved relative to the PD baseline under LID-dependent states, indicating that energy information can perform fine-grained calibration of the PAC feature tokens based on the state, without replacing the coupled representation as an independent dominant input.

[0049] The above results are derived from time-varying data of 10 animals before and after drug administration. The sample size and state label definitions are limited. In practical applications, the number of subjects can be increased, synchronous behavioral scores or clinical scales can be introduced, and adjustments can be made to the frequency band division, number of time windows, token dimension, and modulation intensity. Any combination of techniques employing complex coupling tensors, frequency coordinate sensing coding, energy conditionally bounded modulation, and brain region-temporal aggregation can fall within the scope of this invention.

[0050] like Figure 8As shown, the embodiment also provides a neural state decoding system 80 based on energy conditional coupling tensor learning, including a coupling tensor construction module 81, a coordinate-aware encoding module 82, an energy conditional modulation module 83, and an aggregation decoding module 84; The coupling tensor construction module 81 extracts multiple low-frequency phase components and multiple high-frequency amplitude envelope components from the local field potential signal within each time window of each brain region and calculates complex phase amplitude coupling values. The real part, imaginary part, and magnitude of the complex phase amplitude coupling values ​​are used as complex representation channels to form a phase amplitude coupling tensor. The coordinate sensing encoding module 82 adds normalized coordinate channels to the phase amplitude coupling tensor to represent the phase frequency position and amplitude frequency position, and then extracts the phase amplitude coupling feature tokens. The energy condition modulation module 83 extracts the frequency band energy features corresponding to the time window and brain region from the local field potential signal, generates scale parameters and bias parameters based on the frequency band energy features, and calibrates the phase amplitude coupling feature tokens based on the scale parameters and bias parameters and using bounded feature linear modulation. The aggregation decoding module 84 performs cross-brain region attention aggregation and cross-time window temporal aggregation on the calibrated phase amplitude coupling feature tokens to obtain analysis segment-level representations and decodes them to obtain neural state decoding results, which are used to distinguish between the baseline state of Parkinson's disease and the state related to levodopa-induced dyskinesia.

[0051] It should be noted that the neural state decoding system based on energy conditional coupling tensor learning provided in the above embodiments should be illustrated using the above-described functional module division as an example when performing neural state decoding. The functions described above can be assigned to different functional modules as needed, that is, the internal structure of the terminal or server can be divided into different functional modules to complete all or part of the functions described above. Furthermore, the neural state decoding system based on energy conditional coupling tensor learning provided in the above embodiments and the neural state decoding method embodiments based on energy conditional coupling tensor learning belong to the same concept. For details of their specific implementation, please refer to the neural state decoding method embodiments based on energy conditional coupling tensor learning, which will not be repeated here.

[0052] Based on the same inventive concept, the embodiment also provides a computing device, including a memory and one or more processors. The memory stores executable code, and when the one or more processors execute the executable code, it is used to implement the above-described neural state decoding method based on energy conditional coupling tensor learning, specifically including the following steps: S1. Extract multiple low-frequency phase components and multiple high-frequency amplitude envelope components from the local field potential signal in each time window of each brain region and calculate the complex phase amplitude coupling value. Use the real part, imaginary part and magnitude of the complex phase amplitude coupling value as complex characterization channels to form the phase amplitude coupling tensor. S2, after adding normalized coordinate channels to the phase amplitude coupling tensor to characterize the phase frequency position and amplitude frequency position, the phase amplitude coupling feature token is extracted. S3, extract the frequency band energy features corresponding to the time window and brain region from the local field potential signal, generate scale parameters and bias parameters based on the frequency band energy features, and calibrate the phase amplitude coupling feature token based on the scale parameters and bias parameters and using bounded feature linear modulation; S4 sequentially performs cross-brain region attention aggregation and cross-time window temporal aggregation on the calibrated phase amplitude coupling feature tokens to obtain the analysis fragment level representation and decode it to obtain the neural state decoding result, which is used to distinguish between the baseline state of Parkinson's disease and the state related to levodopa-induced dyskinesia.

[0053] The computing device provided in this embodiment, at the hardware level, includes not only a processor and memory, but also internal buses, network interfaces, memory, and other hardware required for business operations. The memory is non-volatile memory. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to implement the neural state decoding method based on energy conditional coupling tensor learning described in S1-S4 above. Of course, besides software implementation, this invention does not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution entity of the following processing flow is not limited to individual logic units, but can also be hardware or logic devices.

[0054] Based on the same inventive concept, the embodiments also provide a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements the above-described neural state decoding method based on energy conditional coupling tensor learning, specifically including the following steps: S1. Extract multiple low-frequency phase components and multiple high-frequency amplitude envelope components from the local field potential signal in each time window of each brain region and calculate the complex phase amplitude coupling value. Use the real part, imaginary part and magnitude of the complex phase amplitude coupling value as complex characterization channels to form the phase amplitude coupling tensor. S2, after adding normalized coordinate channels to the phase amplitude coupling tensor to characterize the phase frequency position and amplitude frequency position, the phase amplitude coupling feature token is extracted. S3, extract the frequency band energy features corresponding to the time window and brain region from the local field potential signal, generate scale parameters and bias parameters based on the frequency band energy features, and calibrate the phase amplitude coupling feature token based on the scale parameters and bias parameters and using bounded feature linear modulation; S4 sequentially performs cross-brain region attention aggregation and cross-time window temporal aggregation on the calibrated phase amplitude coupling feature tokens to obtain the analysis fragment level representation and decode it to obtain the neural state decoding result, which is used to distinguish between the baseline state of Parkinson's disease and the state related to levodopa-induced dyskinesia.

[0055] In this embodiment, the computer-readable medium includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data.

[0056] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A neural state decoding system based on energy-conditionally coupled tensor learning, characterized in that, include: The coupling tensor construction module is used to extract multiple low-frequency phase components and multiple high-frequency amplitude envelope components from the local field potential signal in each time window of each brain region and calculate the complex phase amplitude coupling value. The real part, imaginary part and magnitude of the complex phase amplitude coupling value are used as complex representation channels to form a phase amplitude coupling tensor. The coordinate-aware encoding module is used to extract the phase-amplitude coupling feature token after adding normalized coordinate channels to the phase-amplitude coupling tensor to characterize the phase frequency position and amplitude frequency position. The energy condition modulation module is used to extract the frequency band energy features corresponding to the time window and brain region from the local field potential signal, generate scale parameters and bias parameters based on the frequency band energy features, and calibrate the phase amplitude coupling feature token based on the scale parameters and bias parameters and using bounded feature linear modulation. The aggregation decoding module is used to sequentially perform cross-brain region attention aggregation and cross-time window temporal aggregation on the calibrated phase amplitude coupling feature tokens to obtain the analysis fragment-level representation and decode it to obtain the neural state decoding result, which is used to distinguish between the baseline state of Parkinson's disease and the state related to levodopa-induced dyskinesia.

2. The neural state decoding system based on energy-conditionally coupled tensor learning according to claim 1, characterized in that, The complex phase amplitude coupling value is calculated according to the following formula: in, Indicates the first The first low-frequency phase component and the first Complex phase amplitude coupling values ​​between the high-frequency amplitude envelope components This indicates the number of sampling points within the time window. Indicates the first Each high-frequency component at the sampling time instantaneous amplitude, Indicates the first The low-frequency component at the sampling time The instantaneous phase, The imaginary unit is represented; the instantaneous amplitude and the instantaneous phase are obtained by bandpass filtering and Hilbert transform.

3. The neural state decoding system based on energy-conditionally coupled tensor learning according to claim 1, characterized in that, The phase amplitude coupling tensor is a tensor formed by arranging the complex representation channels of complex phase amplitude coupling values ​​according to the time window dimension, brain region dimension, phase frequency band dimension, and amplitude frequency band dimension.

4. The neural state decoding system based on energy-conditionally coupled tensor learning according to claim 1, characterized in that, The normalized coordinate channels are generated based on the center frequency of the phase band and the center frequency of the amplitude band, respectively, and are spliced ​​with the complex representation channels corresponding to the real part, imaginary part and modulus value in the channel dimension to realize the addition of normalized coordinate channels to the phase amplitude coupling tensor. The phase amplitude coupling feature tokens are extracted by a coordinate-aware coding network based on the phase amplitude coupling tensor with added normalized coordinate channels. The coordinate-aware coding network includes at least two convolutional layers, a nonlinear activation layer, a normalization layer, and a linear projection layer to preserve the positional relationship of the coupling mode on the phase frequency axis and the amplitude frequency axis.

5. The neural state decoding system based on energy-conditionally coupled tensor learning according to claim 1, characterized in that, The frequency band energy characteristics include the relative logarithmic band power within each preset frequency band, wherein the relative logarithmic band power is obtained by taking the logarithm of the corresponding frequency band power and normalizing it relative to the reference power of the same time window and the same brain region. The conditional parameter generation network generates scale parameters and bias parameters based on the frequency band energy characteristics. The conditional parameter generation network adopts a multilayer perceptron and outputs scale parameters and bias parameters with the same dimension as the phase amplitude coupling feature token.

6. The neural state decoding system based on energy-conditionally coupled tensor learning according to claim 1, characterized in that, The bounded characteristic linear modulation calibrates the phase amplitude coupling characteristic token according to the following formula: in, and These represent the phase amplitude coupling characteristic tokens before and after calibration, respectively. Indicates the scale parameter. Indicates the bias parameter. Indicates modulation intensity and , It is the hyperbolic tangent function.

7. The neural state decoding system based on energy-conditionally coupled tensor learning according to claim 1, characterized in that, The cross-brain region attention aggregation includes: for each time window, calculating normalized brain region weights based on the calibrated phase amplitude coupling feature tokens corresponding to each brain region, and performing a weighted summation of the calibrated phase amplitude coupling feature tokens corresponding to each brain region to obtain the brain region fusion representation for each time window. The cross-time window temporal aggregation includes: inputting the brain region fusion representations of each time window into a bidirectional gated recurrent unit in chronological order to obtain the time window-level hidden state; calculating the normalized time weights based on each time window-level hidden state and performing weighted summation to obtain the analysis segment-level representation.

8. A neural state decoding method based on energy-conditionally coupled tensor learning, characterized in that, The system described in any one of claims 1-7 includes the following steps: Multiple low-frequency phase components and multiple high-frequency amplitude envelope components are extracted from the local field potential signal in each time window of each brain region, and complex phase amplitude coupling values ​​are calculated. The real part, imaginary part, and magnitude of the complex phase amplitude coupling values ​​are used as complex characterization channels to form a phase amplitude coupling tensor. After adding normalized coordinate channels to the phase-amplitude coupling tensor to characterize the phase frequency position and amplitude frequency position, the phase-amplitude coupling feature token is extracted. Frequency band energy features corresponding to time windows and brain regions are extracted from local field potential signals. Scale parameters and bias parameters are generated based on the frequency band energy features. The phase amplitude coupling feature token is calibrated based on the scale parameters and bias parameters and using bounded feature linear modulation. After calibration, the phase amplitude coupling feature tokens are sequentially subjected to cross-brain region attention aggregation and cross-time window temporal aggregation to obtain the analysis fragment level representation. After decoding, the neural state decoding result is obtained, which is used to distinguish between the baseline state of Parkinson's disease and the state related to levodopa-induced dyskinesia.

9. A computing device comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that, When the one or more processors execute the executable code, they are used to implement the neural state decoding method based on energy conditional coupling tensor learning as described in claim 8.

10. A computer-readable storage medium, characterized in that, It stores a program that, when executed by a processor, implements the neural state decoding method based on energy conditional coupling tensor learning as described in claim 8.

Images