Multi-mode nerve physiological signal fusion-based magic limb pain function evaluation system and method

Abstract

The invention provides a magic limb pain function evaluation system and method based on multi-mode nerve physiological signal fusion, and relates to the technical field of medical instrument biomedical engineering and nerve regulation and control. According to the system, central nervous function signals, peripheral electromyographic signals and autonomic nerve signals of a patient suffering from phantom limb pain are synchronously collected and preprocessed, and then multi-modal indexes such as a function connection index, an activation asymmetry index, a function recovery index and skin electric response characteristics are extracted. A dynamic Bayesian network is utilized to construct a model containing three layers of hidden states of central nerves, peripheral muscles and autonomic nerves, the nerve-physiological state of a patient is evaluated in real time, and personalized transcranial magnetic stimulation treatment scheme suggestions are generated. The system realizes a closed-loop control process of data acquisition, fusion analysis and treatment feedback, effectively overcomes the limitation of a traditional single evaluation mode, and provides a reliable quantitative basis for accurate diagnosis and treatment of phantom limb pain.

Description

Technical Field

[0001] This invention relates to the fields of biomedical engineering and neuromodulation technology, specifically to a phantom limb pain functional assessment system and method based on multimodal neurophysiological signal fusion. Background Technology

[0002] Phantom limb pain (PLP) is a common chronic pain syndrome in amputees. Its pathological mechanism is complex, involving multiple factors such as central nervous system reorganization, autonomic dysfunction, and abnormal electromyographic activity of the residual limb. Currently, clinical assessment of PLP mainly relies on patient-reported pain intensity scales (such as the Visual Analogue Scale, VAS), lacking objective physiological indicators.

[0003] Traditional assessment methods cannot simultaneously reflect the multidimensional physiological state of the central nervous system, peripheral muscle activity, and autonomic nervous system, nor can they obtain real-time dynamic changes in the patient's pain status, making it difficult to optimize diagnostic and treatment plans according to the individual characteristics of each patient. Furthermore, existing technologies often lack closed-loop feedback mechanisms when providing treatment plans, making it difficult to dynamically adjust treatment parameters based on the patient's real-time condition.

[0004] Therefore, there is an urgent need for a method that can integrate and analyze multiple physiological signals in real time to provide more accurate and quantitative evidence for the diagnosis and treatment of phantom limb pain patients.

[0005] Therefore, those skilled in the art provide a phantom limb pain functional assessment system and method based on multimodal neurophysiological signal fusion to solve the problems mentioned in the background art. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a phantom limb pain functional assessment system and method based on multimodal neurophysiological signal fusion. It can simultaneously collect multiple physiological signals from patients and perform joint analysis of the signals through advanced algorithm models to achieve real-time assessment of phantom limb pain status and generation of personalized treatment suggestions. This solves the problem that existing technologies often lack a closed-loop feedback mechanism when providing treatment plans, making it difficult to dynamically adjust treatment parameters according to the patient's real-time status.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A functional assessment system for phantom limb pain based on multimodal neurophysiological signal fusion includes: a multimodal data acquisition module, a signal preprocessing and feature extraction module, a multimodal data fusion analysis module, an adaptive learning and personalized suggestion generation module, and a data storage and visualization feedback module;

[0009] The multimodal data acquisition module is used to simultaneously acquire the patient's fNIRS, EMG and EDA signals. The fNIRS sensor is arranged in the bilateral primary motor cortex (M1), dorsolateral prefrontal cortex (DLPFC) and primary sensory cortex (S1). The EMG surface electrode is attached to the proximal muscle group of the residual limb and the corresponding muscle group on the opposite side. The EDA electrode is attached to the thenar eminence of the palm.

[0010] The signal preprocessing and feature extraction module is used to denoise and calculate features of the acquired multimodal signals. It uses a combination of improved time derivative distribution recovery (TDDR) and Hampel filtering to remove motion artifacts from the fNIRS signal, wavelet threshold denoising filtering to process the EMG signal, and polynomial fitting correction to remove baseline drift of the EDA signal. The extracted features include: activation asymmetry index of fNIRS, functional connectivity index based on wavelet analysis, functional recovery index (FRI) of EMG, and skin conductance response (SCR) amplitude of EDA.

[0011] The multimodal data fusion analysis module employs a DBN model for joint inference. The model defines three latent states: neural, muscular, and autonomic. The observation vector consists of the functional connectivity index, activation asymmetry index, functional recovery index (FRI), and SCR amplitude. A forward-backward algorithm is used to calculate the posterior probability of the latent states in real time, yielding assessment results of the patient's neurophysiological state at each time point. A comprehensive assessment index can also be calculated, corresponding to the traditional VAS classification (e.g., none, mild, moderate, severe).

[0012] The adaptive learning and personalized suggestion generation module updates the evaluation model and generates treatment suggestions online based on real-time evaluation results and historical data. When the model prediction error continues to exceed the set threshold, the system automatically triggers the incremental learning algorithm to update the DBN model parameters to improve prediction accuracy. At the same time, combined with clinical experience rules and treatment plan library, it automatically generates personalized neuromodulation treatment parameter suggestions (such as stimulation frequency, intensity and target adjustment) and rehabilitation training or other intervention plans based on the comprehensive evaluation results, and feeds them back to medical staff through a visual interface.

[0013] The data storage and visualization feedback module stores and manages the collected data and analysis results in a hierarchical manner, and displays the assessment results and treatment suggestions to medical staff through a visual interface.

[0014] Furthermore, the signal preprocessing and feature extraction module includes a multimodal signal preprocessing unit and a time-frequency domain feature extraction unit, used to perform noise reduction processing on the fNIRS, EMG and EDA signals, and extract the corresponding time-domain and frequency-domain features.

[0015] Furthermore, the multimodal data fusion analysis module uses DBN for modeling. The observation vector of the DBN model includes four features: functional connectivity index, activation asymmetry index, functional recovery index, and skin conductance response amplitude. The posterior probabilities of neural state, muscle state, and autonomic nervous state are calculated in real time using a forward-backward algorithm.

[0016] Furthermore, the data storage and visualization feedback module includes a visualization display module and a structured report generation module, used to display the analysis results in the form of charts or reports and generate treatment reports.

[0017] Furthermore, it also includes a data storage and management module for storing and managing the multimodal acquired data and analysis results.

[0018] Furthermore, a functional assessment method for phantom limb pain based on multimodal neurophysiological signal fusion includes the following steps:

[0019] S1. Simultaneously acquire fNIRS, EMG, and EDA signals of PLP patients before, during, and after treatment using a multimodal data acquisition module;

[0020] S2. Preprocess and extract features from the signals acquired in each modality to obtain the functional connectivity index and activation asymmetry index of fNIRS, functional recovery index, and EDA skin electroreactivity amplitude.

[0021] S3. Input the modal features into the DBN model and obtain the posterior probability distribution of the PLP state at each time step through forward-backward algorithm reasoning;

[0022] S4. Use the visualization module to present the above analysis results to medical staff in the form of multi-indicator dynamic charts and generate a structured treatment report;

[0023] S5. The collected multimodal signal data and analysis results are stored according to a hierarchical storage strategy and managed through the patient database management module.

[0024] This invention provides a functional assessment system and method for phantom limb pain based on multimodal neurophysiological signal fusion. It has the following beneficial effects:

[0025] 1. This invention provides a functional assessment system and method for phantom limb pain based on multimodal neurophysiological signal fusion. The system uses multiple sensors such as functional near-infrared spectroscopy (fNIRS), electromyography (EMG), and electrical skin conductance (EDA) to simultaneously collect data and extract multidimensional features such as central nervous system hemodynamics, peripheral muscle activity, and autonomic nerve response to comprehensively reflect the physiological state of PLP patients.

[0026] 2. This invention provides a functional assessment system and method for phantom limb pain based on the fusion of multimodal neurophysiological signals. It uses DBN to construct a three-layer hidden state model of the central nervous system, muscles and autonomic nervous system. Through joint inference, the posterior probability of each hidden state is calculated in real time to realize the fusion analysis of multimodal signals. This model can reveal the coupling relationship between different physiological pathways and improve the objectivity and accuracy of PLP state determination.

[0027] 3. This invention provides a functional assessment system and method for phantom limb pain based on multimodal neurophysiological signal fusion. The system forms a closed-loop workflow of "data acquisition - fusion analysis - treatment feedback". Based on the comprehensive assessment results, the system automatically generates personalized neuromodulation intervention plan suggestions and provides feedback to medical staff through a visual interface, realizing real-time monitoring and dynamic adjustment of the treatment process.

[0028] 4. This invention provides a functional assessment system and method for phantom limb pain based on multimodal neurophysiological signal fusion. The system is designed with a hierarchical storage and visualization feedback mechanism to manage the collected raw data, feature data and assessment results in a structured manner, and to display the analysis results in the form of charts and reports, providing data support for subsequent clinical decision-making and research. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the overall framework of the phantom limb pain functional assessment system based on multimodal neurophysiological signal fusion of the present invention, showing the data flow and workflow between the functional modules;

[0030] Figure 2 The flowchart of the signal preprocessing and feature extraction module of the present invention illustrates the denoising and feature extraction steps of fNIRS, EMG and EDA signals.

[0031] Figure 3 This is a flowchart of the multimodal data fusion and evaluation module of the present invention, illustrating the process of jointly modeling and inferring multimodal features using a dynamic Bayesian network;

[0032] Figure 4 This is a schematic diagram of the adaptive learning and feedback module of the present invention, illustrating the process of online model updating and personalized treatment suggestion generation;

[0033] Figure 5 This is a schematic diagram of the visualization report interface of the present invention, illustrating how the system presents assessment results and treatment recommendations in a graphical and reporting manner. Detailed Implementation

[0034] The technical solutions of the specific embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described specific embodiments are only a part of the specific embodiments of the present invention, and not all of them. Based on the specific embodiments of the present invention, all other specific embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] Implementation method 1:

[0036] like Figure 1-5 As shown, this embodiment of the invention provides a phantom limb pain functional assessment system based on multimodal neurophysiological signal fusion, including: a multimodal data acquisition module, a signal preprocessing and feature extraction module, a multimodal data fusion analysis module, an adaptive learning and personalized suggestion generation module, and a data storage and visualization feedback module;

[0037] The multimodal data acquisition module is used to simultaneously acquire the patient's fNIRS, EMG and EDA signals. The fNIRS sensor is arranged in the bilateral primary motor cortex (M1), dorsolateral prefrontal cortex (DLPFC) and primary sensory cortex (S1). The EMG surface electrode is attached to the proximal muscle group of the residual limb and the corresponding muscle group on the opposite side. The EDA electrode is attached to the thenar eminence of the palm.

[0038] The signal preprocessing and feature extraction module is used to denoise and calculate features of the acquired multimodal signals. It uses a combination of improved time derivative distribution recovery (TDDR) and Hampel filtering to remove motion artifacts from the fNIRS signal, wavelet threshold denoising filtering to process the EMG signal, and polynomial fitting correction to remove baseline drift of the EDA signal. The extracted features include: activation asymmetry index of fNIRS, functional connectivity index based on wavelet analysis, functional recovery index (FRI) of EMG, and skin conductance response (SCR) amplitude of EDA.

[0039] The multimodal data fusion analysis module employs a DBN model for joint inference. The model defines three latent states: neural, muscular, and autonomic. The observation vector consists of the functional connectivity index, activation asymmetry index, functional recovery index (FRI), and SCR amplitude. A forward-backward algorithm is used to calculate the posterior probability of the latent states in real time, yielding assessment results of the patient's neurophysiological state at each time point. A comprehensive assessment index can also be calculated, corresponding to the traditional VAS classification (e.g., none, mild, moderate, severe).

[0040] The adaptive learning and personalized suggestion generation module updates the evaluation model and generates treatment suggestions online based on real-time evaluation results and historical data. When the model prediction error continues to exceed the set threshold, the system automatically triggers the incremental learning algorithm to update the DBN model parameters to improve prediction accuracy. At the same time, combined with clinical experience rules and treatment plan library, it automatically generates personalized neuromodulation treatment parameter suggestions (such as stimulation frequency, intensity and target adjustment) and rehabilitation training or other intervention plans based on the comprehensive evaluation results, and feeds them back to medical staff through a visual interface.

[0041] The data storage and visualization feedback module stores and manages the collected data and analysis results in a hierarchical manner, and displays the assessment results and treatment suggestions to medical staff through a visual interface.

[0042] This phantom limb pain functional assessment method based on multimodal neurophysiological signal fusion includes the following steps:

[0043] S1. Simultaneously acquire fNIRS, EMG, and EDA signals of PLP patients before, during, and after treatment using a multimodal data acquisition module;

[0044] S2. Preprocess and extract features from the signals acquired in each modality to obtain the functional connectivity index and activation asymmetry index of fNIRS, functional recovery index, and EDA skin electroreactivity amplitude.

[0045] S3. Input the modal features into the DBN model and obtain the posterior probability distribution of the PLP state at each time step through forward-backward algorithm reasoning;

[0046] S4. Use the visualization module to present the above analysis results to medical staff in the form of multi-indicator dynamic charts and generate a structured treatment report;

[0047] S5. The collected multimodal signal data and analysis results are stored according to a hierarchical storage strategy and managed through the patient database management module.

[0048] Example 1: System Setup and Data Acquisition

[0049] In a quiet and comfortable environment, physiological signal data of PLP patients are synchronously acquired through a multimodal data acquisition module.

[0050] First, the sensors were deployed: the fNIRS sensors covered the bilateral M1 area, DLPFC, and S1 area to collect brain blood oxygenation signals;

[0051] EMG surface electrodes are attached to the target muscle group proximal to the residual limb and the corresponding muscle group on the opposite side to collect electromyographic signals in real time.

[0052] EDA electrodes are attached to the thenar eminence of the palm to collect signals of skin electrical activity.

[0053] The methods for data collection tasks include:

[0054] 1) Resting state: Instruct the patient to remain awake and relaxed, and continuously collect fNIRS, EMG and EDA signals for 10 minutes to obtain baseline physiological activity data;

[0055] 2) Mirror movement rehabilitation state: Under guidance, perform mirror movement training (such as bending and straightening the leg, and observe the movement of the healthy leg with the help of a mirror to stimulate phantom limb sensation). During the training, the system continuously and synchronously collects fNIRS, EMG and EDA signals for 10 minutes and records the patient's training cooperation level.

[0056] 3) Recovery state: After the training, the patient is instructed to stop exercising and rest quietly for 10 minutes. During this stage, the system continues to collect fNIRS signals to monitor the recovery process of cerebral blood oxygenation level.

[0057] The above data acquisition scheme ensures continuous monitoring of multimodal signals in the three stages of pre-treatment, treatment, and post-treatment, providing a rich data foundation for subsequent analysis.

[0058] Example 2: Signal Preprocessing and Feature Extraction

[0059] For the acquired multimodal signals, preprocessing and feature extraction are performed on each signal, specifically as follows:

[0060] 1) fNIRS signal processing: First, the raw fNIRS data is processed using an improved TDDR algorithm combined with Hampel filtering to remove motion artifacts and improve signal quality;

[0061] Then, the denoised fNIRS signal is subjected to continuous wavelet transform in the 0.01 to 0.08 Hz frequency band. The amplitude of the wavelet coefficients in this frequency band is calculated and the mean value is taken to quantify the degree of activation asymmetry between the left and right hemispheres.

[0062] Simultaneously, wavelet coherence coefficients of signals from corresponding regions of the left and right hemispheres are calculated to quantify the functional connectivity index characteristics, ultimately yielding the functional connectivity index and activation asymmetry index of fNIRS.

[0063] ①Signal preprocessing

[0064] An improved TDDR algorithm (combined iterative TDDR and Hampel filtering) is applied to the fNIRS signal to remove motion artifacts, including the following process:

[0065] Step 1 (TDDR): Calculate the first derivative of the signal, detect outliers (threshold: ±3 standard deviations), and perform linear interpolation repair;

[0066] Step 2 (Hampel filtering): Calculate the local median M and absolute deviation AD using a sliding window (window length = 5s), and replace outliers (threshold: |x(t)-M|>3×AD);

[0067] Step 3: Iterate and optimize until the residual energy of artifacts is less than 1%.

[0068] Linear interpolation repair is performed when |x'(t)|>3σ;

[0069] M = median(x) t-L:t+L AD = median(|x) i -M∣), if |x i -M∣>k×AD,x i =M.

[0070] ②Feature 1 extraction: Activation of asymmetric index

[0071] Wavelet basis function selection: The complex Morlet wavelet basis function is used, and its mathematical expression is:

[0072]

[0073] Wherein, the center frequency w0 = 1.

[0074] Continuous wavelet transform: Time-frequency decomposition of the fNIRS signal, the formula is as follows:

[0075]

[0076] Where s is the wavelet scale (inversely proportional to the frequency), τ is the translation variable (i.e., the current time point), x(t) is the original HbO2 concentration signal, and * denotes complex conjugate.

[0077] Wavelet amplitude extraction: Within the 0.01-0.08Hz frequency band, calculate the wavelet coefficient amplitude at each time point:

[0078] A(s,τ)=∣W(s,τ)∣

[0079] Finally, the average amplitude of the frequency band is extracted to obtain the feature values ​​of each channel in resting, task, and recovery states.

[0080]

[0081] Where N is the number of wavelet scales in the 0.01-0.08Hz frequency band, with a value of 1.05.

[0082] The Activation Asymmetry Index (ASI) is defined as a normalized index of the difference in eigenvalues ​​between corresponding channels in the left and right hemispheres. The specific calculation formula is as follows:

[0083]

[0084] When ASI is positive, it indicates that the activation of the healthy side of the brain is stronger than that of the affected side; when ASI is negative, it indicates that the activation of the affected side of the brain is stronger than that of the healthy side; the larger the absolute value of ASI, the higher the degree of asymmetry in activation between the two sides, which may indicate functional imbalance or abnormality.

[0085] ③ Feature 2 extraction: Functional connectivity strength between the left and right hemispheres (wavelet coherence coefficient)

[0086] First, continuous wavelet transforms were performed on the fNIRS signals x(t) and y(t) from the target channels of the left and right hemispheres, respectively. Complex Morlet wavelets were used as the basis functions, with a center frequency set to 0.1 Hz to ensure sensitivity to low-frequency changes in cerebral blood oxygenation signals. The wavelet transform formula is as follows:

[0087]

[0088] W x (s,t), W y (s,t) represent the wavelet transform results of signals x(t) and y(t), respectively. s: wavelet scale (inversely proportional to frequency); t: time; ψ * : Complex conjugate of Morlet wavelets.

[0089] Calculate the cross-wavelet spectrum and coherence coefficient of two signals.

[0090]

[0091] Subsequently, the wavelet coherence coefficients are calculated:

[0092]

[0093] Where S represents the local smoothing operator (using standard deviation σ = 0.6) s Gaussian windows are used for joint time-frequency domain smoothing to improve the robustness of wavelet spectral calculations.

[0094] FCI→1 indicates that the two channels have extremely high coherence at this scale and time point, i.e., there is a strong functional connection; FCI→0 indicates that the two channels have almost no correlation at this scale and time point, i.e., the functional connection is weak or lost.

[0095] 2) EMG signal processing: The raw EMG signal is first denoised using wavelet thresholding. Then, the temporal RMS of the EMG signal is calculated as an electromyographic feature within a set time window. In addition, the Functional Recovery Index (FRI) can be calculated based on the ratio of the difference between the muscle activity of the healthy side and the muscle activity of the residual limb to the electromyographic signal of the healthy side, in order to assess the degree of muscle function recovery.

[0096] ① Wavelet thresholding denoising

[0097] Step 1: Wavelet decomposition coefficients

[0098] Let the original electromyographic signal be x(t), and the wavelet coefficients of the k-th wavelet in the j-th layer after wavelet decomposition be:

[0099]

[0100] Step 2: Wavelet coefficients after thresholding

[0101] The wavelet coefficients are denoised using a soft thresholding function, and the processed coefficients are:

[0102]

[0103] Step 3: Signal Reconstruction Formula

[0104] The denoised signal x'(t) is reconstructed using inverse wavelet transform:

[0105]

[0106] ② Extraction of root mean square (RMS) value in the time domain:

[0107] Extract N sampling points x'(1), x'(2), ..., x'(N) from the filtered signal x'(t), and calculate the RMS value:

[0108]

[0109] ③ Calculate the Functional Recovery Index (FRI):

[0110]

[0111] An FRI close to 0% indicates that the residual limb muscle function is close to that of the healthy side, and the functional recovery is good; the higher the FRI value, the lower the level of residual limb electromyographic activity is compared to the healthy side, and the degree of functional recovery is poor.

[0112] 3) EDA signal processing: The acquired skin conductance signals are first corrected for baseline drift by polynomial fitting. The least squares method is used to determine the polynomial coefficients to minimize the fitting error. After correction, the SCR (Skin Conductance Response) response is detected: a threshold T is set, and when the signal change rate exceeds T, it is recorded as the SCR start point. The maximum peak point is found after the start point. The SCR response amplitude is defined as the difference between the signal amplitude at the peak point and the baseline value before the start point. This SCR amplitude reflects the patient's autonomic nerve activation response.

[0113] ① Use polynomial fitting to correct baseline drift

[0114] Let the acquired EDA signal be s(n), (n = 0, 1, ..., N-1), and use a p-th degree polynomial to fit the baseline of the signal. The expression of the polynomial is:

[0115]

[0116] Among them, a i (i = 0, 1, ..., p) are the coefficients of the polynomial.

[0117] The coefficients 'a' of the polynomial are determined using the least squares method. i This minimizes the sum of squared errors between the fitted baseline B(n) and the original signal s(n), i.e.:

[0118]

[0119] The sum of squared errors mentioned above with respect to a i Find the partial derivatives and set them equal to zero. This gives us a system of linear equations. Solving this system of equations will give us the coefficients 'a' of the polynomial. i .

[0120] Corrected EDA signal for:

[0121] ② SCR Amplitude Feature Extraction

[0122] Detecting the starting point n of SCR onset and peak point n peak : Set the threshold T, when the corrected signal When the rate of change exceeds this threshold, record that moment as the starting point n of the SCR. onset Starting from the starting point, search for the maximum value point n in the subsequent signals. peak .

[0123] The SCR response amplitude is defined as the difference between the signal value at the peak point and the baseline value before the reaction initiation point:

[0124]

[0125] Among them, A k n represents the magnitude of the k-th SCR reaction. onset,k Indicates the starting point of the reaction, n onset,k This indicates the peak point of the reaction.

[0126] Through the above preprocessing and feature extraction methods, multiple quantitative indicators describing the activity of various pathways in the patient's neuromuscular-autonomic nervous system were obtained, providing basic data for subsequent multimodal fusion analysis.

[0127] Example 3: Multimodal Data Fusion and Evaluation

[0128] DBN is used to jointly model and evaluate the extracted multimodal features in real time, specifically including:

[0129] Model establishment: Construct a DBN model with three hidden states, setting the three hidden state variables to represent the neural state, muscle state and autonomic nervous state respectively. Each state evolves over time to form a hidden Markov chain, and its dynamic characteristics are described by the state transition probability matrix.

[0130] Observation vector definition: The observation vector is composed of four features: the FCI and ASI of fNIRS, the RMS value of EMG, and the SCR magnitude of EDA. It serves as the input for each hidden state, and a corresponding observation probability distribution (such as Gaussian distribution) is set for each hidden state.

[0131] Joint inference: The forward-backward algorithm is used to perform online inference on the DBN model and calculate the posterior probability distribution of each hidden state at each time point. Specifically, based on the current and historical observation data sequence, the posterior probabilities of the neural state, muscle state and autonomic nervous state are solved in real time, so as to obtain the comprehensive neurophysiological state assessment results of the patient at different time periods.

[0132] Assessment output: A comprehensive assessment index can be calculated based on the posterior probability to reflect the severity of phantom limb pain. This index can be correlated with the traditional VAS scale and is divided into four levels: no pain (0 points), mild (VAS 1-3 points), moderate (VAS 4-7 points), and severe (VAS 8-10 points), providing a basis for subsequent treatment decisions.

[0133] This multimodal fusion analysis module can analyze the patient's neuromuscular-autonomic nervous system status in real time and quantitatively assess the degree of phantom limb pain, which is more comprehensive and objective than traditional single-indicator assessment methods.

[0134] This module is based on a three-layer hidden-state DBN model for multimodal observation sequences {O t} and hidden state sequence {S t Joint modeling and real-time inference mainly include:

[0135] 1. Definition of observation vector

[0136] At each time t, the system synchronously collects the following feature variables to form the observation vector O. t :

[0137]

[0138] FCI tFunctional Connectivity Index;

[0139] ASI t : Activation Symmetry Index;

[0140] FRI t Functional Recovery Index;

[0141] Amp t : EDA SCR Amplitude.

[0142] 2. Definition of Hidden State

[0143] The following three types of latent state variables are constructed to describe the potential states in different physiological dimensions:

[0144]

[0145] in, Represents the neural state, which indicates the potential functional state of the motor cortex at time t.

[0146] This represents the muscle state, specifically the potential activation state of the target muscle group at time t.

[0147] This represents the autonomic state, that is, the potential activation state of the autonomic nervous system at time t.

[0148] The three components form hidden Markov chains on the time axis, and their interdependence and coupling characteristics are captured through joint probability modeling.

[0149] 3. State transition probability

[0150]

[0151] in, Let be the transition matrix for the d-th state.

[0152] 4. Observational Probability Modeling (Emission Model)

[0153] Each hidden state corresponds to the observation probability of its feature variable:

[0154]

[0155] Right now

[0156]

[0157] in:

[0158] In the hidden state S t The mean of the Gaussian distribution of the i-th observed variable.

[0159] In the hidden state S t Below, the Gaussian distribution variance of the i-th observed variable

[0160] The model can be trained using the EM algorithm.

[0161] 5. Forward-Backward Algorithm

[0162] The forward-backward algorithm is used to calculate the posterior probability of the hidden state at each time point:

[0163] Forward probability:

[0164]

[0165] Backward probability:

[0166]

[0167] Posterior probability:

[0168] 6. Baum–Welch parameter estimation

[0169] Furthermore, the decision-making level conducts joint judgments on the hidden state to achieve real-time dynamic assessment of the degree of phantom limb pain.

[0170]

[0171] in,

[0172] This represents the estimated value of the d-th dimension state transition matrix.

[0173] Example 4: Adaptive Learning and Personalized Recommendation Generation

[0174] Based on the multimodal fusion assessment results, the assessment model can be updated online and personalized treatment suggestions can be generated, specifically including:

[0175] Adaptive learning: The system uses historical patient data and real-time collected data to train the evaluation model online. When the deviation between the model's prediction results and the actual observations for multiple consecutive time windows exceeds a preset threshold, the incremental learning algorithm is triggered to adjust the parameters of the DBN and improve the accuracy and robustness of the model's prediction of the patient's PLP status.

[0176] Personalized treatment recommendations: Based on the comprehensive assessment results, combined with clinical experience rules and treatment protocol library, the system automatically generates targeted neuromodulation intervention proposals (such as TMS stimulation frequency, intensity and target adjustment plans), as well as corresponding rehabilitation training plans or other intervention measures. When the assessment results show that the patient's pain level is increasing, the system can prioritize recommending optimized neuromodulation plans and push them to medical staff in the form of charts and text reports through a visual interface to assist clinicians in formulating subsequent treatment strategies.

[0177] Example 5: Data Storage and Visual Feedback

[0178] Regarding data storage and feedback, this section explains how the system manages and displays data, specifically including:

[0179] Data storage and management: The collected raw signal data, feature data and multimodal evaluation results are stored in the database according to the preset hierarchical storage strategy. The system includes a hierarchical storage module and a patient database management module. Data security is ensured through regular backups and database maintenance, and medical staff can access historical patient data at any time for comparative analysis, providing support for long-term efficacy evaluation and research.

[0180] Visual feedback: The system presents the multimodal assessment results, probability curves of each latent state, and treatment suggestions to medical staff in the form of dynamic charts and structured reports through the visualization module. Medical staff can intuitively view the patient's current neurophysiological state and the trend of efficacy changes, understand the treatment progress in real time, and adjust the treatment plan in combination with the system's suggestions to achieve precise control of phantom limb pain treatment.

[0181] In summary, through the above embodiments, those skilled in the art can fully understand the technical solution of the present invention and implement it. The present invention is not limited to the specific embodiments described above; readers can make equivalent substitutions or modifications based on the core ideas of the present invention without departing from the scope of protection of the present invention.

[0182] Although specific embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these specific embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A functional assessment system for phantom limb pain based on multimodal neurophysiological signal fusion, characterized in that, include: The system includes a multimodal data acquisition module, a signal preprocessing and feature extraction module, a multimodal data fusion and analysis module, an adaptive learning and personalized suggestion generation module, and a data storage and visualization feedback module. The multimodal data acquisition module is used to simultaneously acquire the patient's fNIRS, EMG and EDA signals. The fNIRS sensor is arranged in the bilateral primary motor cortex, the dorsolateral prefrontal cortex and the primary sensory cortex. The EMG surface electrode is attached to the proximal muscle group of the residual limb and the corresponding muscle group on the opposite side. The EDA electrode is attached to the thenar eminence of the palm. The signal preprocessing and feature extraction module is used to denoise and calculate features of the acquired multimodal signals. It uses an improved method combining time derivative distribution recovery (TDDR) and Hampel filtering to remove motion artifacts from the fNIRS signal, wavelet threshold denoising filtering to process the EMG signal, and polynomial fitting correction to remove baseline drift from the EDA signal. The extracted features include: the activation asymmetry index of fNIRS, the functional connectivity index based on wavelet analysis, the functional recovery index (FRI) of EMG, and the amplitude of the skin conductance response (SCR) of EDA. The multimodal data fusion analysis module employs a DBN model for joint inference. The model defines three latent states: neural, muscular, and autonomic. The observation vector consists of the functional connectivity index, activation asymmetry index, functional recovery index (FRI), and SCR amplitude. A forward-backward algorithm is used to calculate the posterior probability of the latent states in real time, yielding assessment results of the patient's neurophysiological state at each time point. A comprehensive assessment index can also be calculated, which corresponds to the traditional VAS grading. The adaptive learning and personalized suggestion generation module updates the evaluation model and generates treatment suggestions online based on real-time evaluation results and historical data. When the model prediction error continues to exceed the set threshold, the system automatically triggers the incremental learning algorithm to update the DBN model parameters and improve the prediction accuracy. At the same time, it combines clinical experience rules and treatment plan library to automatically generate personalized neuromodulation treatment parameter suggestions and rehabilitation training or other intervention plans based on the comprehensive evaluation results, and feeds them back to medical staff through a visual interface. The data storage and visualization feedback module stores and manages the collected data and analysis results in a hierarchical manner, and displays the assessment results and treatment suggestions to medical staff through a visual interface.

2. The phantom limb pain functional assessment system according to claim 1, characterized in that, The signal preprocessing and feature extraction module includes a multimodal signal preprocessing unit and a time-frequency domain feature extraction unit, which are used to perform noise reduction processing on the fNIRS, EMG and EDA signals and extract the corresponding time-domain and frequency-domain features.

3. The phantom limb pain functional assessment system according to claim 1, characterized in that, The multimodal data fusion analysis module uses DBN for modeling. The observation vector of the DBN model includes four features: functional connectivity index, activation asymmetry index, functional recovery index, and skin conductance response amplitude. The forward-backward algorithm is used to calculate the posterior probabilities of neural state, muscle state, and autonomic nervous state in real time.

4. The phantom limb pain functional assessment system according to claim 1, characterized in that, The data storage and visualization feedback module includes a visualization display module and a structured report generation module, which are used to display the analysis results in the form of charts or reports and generate treatment reports.

5. The phantom limb pain functional assessment system according to claim 1, characterized in that, It also includes a data storage and management module for storing and managing the multimodal acquired data and analysis results.

6. A method for functional assessment of phantom limb pain based on multimodal neurophysiological signal fusion, characterized in that, Includes the following steps: S1. Simultaneously acquire fNIRS, EMG, and EDA signals of PLP patients before, during, and after treatment using a multimodal data acquisition module; S2. Preprocess and extract features from the signals acquired in each modality to obtain the functional connectivity index and activation asymmetry index of fNIRS, functional recovery index, and EDA skin electroreactivity amplitude. S3. Input the modal features into the DBN model and obtain the posterior probability distribution of the PLP state at each time step through forward-backward algorithm reasoning; S4. Use the visualization module to present the above analysis results to medical staff in the form of multi-indicator dynamic charts and generate a structured treatment report; S5. The collected multimodal signal data and analysis results are stored according to a hierarchical storage strategy and managed through the patient database management module.

Images