Cerebral stroke patient rehabilitation exercise analysis system based on medical big data

Abstract

The invention relates to the technical field of medical data analysis and rehabilitation, in particular to a cerebral apoplexy patient rehabilitation exercise analysis system based on medical big data, which comprises a data acquisition module used for acquiring electromyographic signals, exercise data and behavior data of a patient to form a data set; the data processing and fusion module is used for preprocessing the electromyographic signals in the data set, normalizing motion and behavior data and fusing the motion and behavior data into a multi-modal feature vector; the hierarchical analysis module is used for performing motion intention recognition according to the preprocessed electromyographic signals and outputting an intention classification result, recognizing the motion intention and predicting a rehabilitation effect through a double-flow Transform encoder, and evaluating a behavior risk level in combination with a risk factor formula; the optimization module is used for dynamically adjusting the mirror image treatment intensity according to the risk level and optimizing the prediction model by adopting incremental learning; and the interaction module is used for generating a thermodynamic diagram superposition rehabilitation curve in the X-Y coordinate system, and marking doctor intervention suggestions to complete visual interaction.

Description

Technical Field

[0001] The present invention relates to the field of medical data analysis and rehabilitation technology, and in particular to a rehabilitation movement analysis system for stroke patients based on medical big data. Background Art

[0002] Stroke rehabilitation requires the integration of multi-dimensional data such as the patient's motor ability, rehabilitation participation, and lifestyle habits. Existing technologies have the following problems: relying solely on electromyographic signals or motion data, it is impossible to comprehensively assess the rehabilitation status; traditional models lack real-time feedback and cannot dynamically adjust rehabilitation strategies; rehabilitation results lack intuitive visual expression, and the efficiency of doctor-patient collaboration is low. Summary of the Invention

[0003] (1) Purpose of the invention In order to solve the technical problems existing in the background technology, the present invention proposes a rehabilitation movement analysis system for stroke patients based on medical big data. It combines electromyographic signals, movement data and behavioral status, optimizes rehabilitation strategies through dynamic models, and realizes visual interaction.

[0004] (2) Technical solution To solve the above problems, the present invention provides a rehabilitation movement analysis system for stroke patients based on medical big data, comprising: Data acquisition module, used to collect patients' electromyographic signals, motion data and behavioral data to form a data set; A data processing and fusion module is used to pre-process the electromyographic signals in the data set, normalize the motion and behavior data, and fuse them into a multimodal feature vector; A hierarchical analysis module is used to identify movement intentions based on the preprocessed electromyographic signals and output intention classification results, identify movement intentions through a dual-stream Transformer encoder, predict rehabilitation effects based on a fuzzy neural network, and assess behavioral risk levels in combination with a risk factor formula; An optimization module, which dynamically adjusts the intensity of mirror therapy based on risk level and optimizes the prediction model using incremental learning; The interactive module is used to generate a heat map and superimpose the recovery curve in the XY coordinate system, and mark the doctor's intervention suggestions to complete the visual interaction.

[0005] Preferably, the electromyographic signals are collected from the affected hand and the healthy hand via a surface electromyographic sensor; The exercise data: uses an inertial measurement unit to record the patient's exercise intensity, exercise frequency and exercise duration. The exercise intensity is calculated by fusing heart rate and acceleration data; The behavioral data: uses wearable devices to obtain the patient's work dimension data, diet data and rest index, the work dimension data includes weekly working hours and work intensity quantification value, and the diet data includes weekly drinking frequency and weekly cholesterol intake ratio.

[0006] Preferably, the data preprocessing and feature fusion process includes: Performing wavelet noise reduction processing on the electromyographic signals in the data set and filtering out power frequency interference using a blind source separation algorithm; The Z-score normalization method was used to standardize the movement and behavior data; Construct a multimodal feature vector, including the time domain and frequency domain features of the electromyographic signal, the intensity-frequency joint index of the motion data, and the risk factor of the behavioral data:

[0007] Among them, Y, E, and N are exercise, work, and diet factors respectively, and d1, d2, and d3 are preset weights.

[0008] Preferably, the preprocessed EMG signals are fed into a two-stream Transformer encoder, with a 200ms time window for the long sequence branch and a 50ms time window for the short sequence branch, and the intention classification results, including grasp, stretch, and relax, are output through a cross-attention module; A dynamic prediction model was constructed based on a fuzzy neural network. The motion data and intention classification results were input, and the Levenberg-Marquardt algorithm was used to optimize the network parameters and output the rehabilitation score. The risk level is generated based on the risk factor XW and the preset threshold, including high risk, medium risk, and low risk.

[0009] Preferably, if the risk level is high, the intensity of mirror therapy is increased, and grip training of the affected hand is enforced through rehabilitation equipment; If the rehabilitation score is lower than the preset value, the incremental learning mechanism is triggered to update the fuzzy neural network weights based on the new data of this week and optimize the prediction model for the next cycle.

[0010] Preferably, a patient behavior status curve is drawn in an XY coordinate system, with the horizontal axis being time and the vertical axis being the risk factor XW, and a rehabilitation score curve is superimposed; a heat map is used to color-code high-risk periods and annotate doctor intervention recommendations.

[0011] Preferably, the incremental learning mechanism is specifically: When the proportion of newly added data in the original training set exceeds the preset proportion each week, the entire network parameter update is triggered; When the ratio is lower than the preset ratio, only the weight of the last fully connected layer is updated.

[0012] Preferably, the method for generating the heat map includes: Divide the XY coordinate system into an a×a pixel grid; According to the risk level distribution of data points in the grid, the bilinear interpolation algorithm is used to fill the color, and the color transparency is positively correlated with the risk density.

[0013] Preferably, it also includes abnormal behavior warning, specifically: If the exercise duration is lower than the preset threshold for a set number of consecutive days, a reminder message will be sent to the patient; If the weekly cholesterol intake ratio exceeds the preset value, a dietary adjustment suggestion will be generated and pushed to the user terminal.

[0014] Preferably, the logic for generating the doctor's intervention suggestion is as follows: High-risk patients: Mandatory appointment for offline rehabilitation assessment this week; Moderate-risk patients: provide a link to a customized video guidance course; Low-risk patients: Generate weekly recovery reports and send them by email.

[0015] The above technical solution of the present invention has the following beneficial technical effects: This invention breaks the limitations of a single data source by fusing electromyographic signals, motion data, and behavioral status (work, diet, rest), and comprehensively assesses the patient's rehabilitation status. Based on the incremental learning mechanism of a fuzzy neural network, the model parameters are dynamically optimized according to new data each week, allowing the rehabilitation plan to adapt to the patient's progress in real time. The dual-stream Transformer encoder is used, combined with a cross-attention mechanism for long and short temporal features, to significantly improve the reliability of motion control of rehabilitation assistive devices. High-risk periods are marked in real time through risk factor formulas and heat map visualization. The visual interactive interface integrates heat maps, rehabilitation curves, and doctor recommendations to convert abstract data into intuitive charts. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a system block diagram of the present invention; Figure 2 This is a spatiotemporal feature pyramid diagram in one embodiment of the present invention. DETAILED DESCRIPTION

[0017] To make the objectives, technical solutions, and advantages of the present invention more clearly understood, the present invention will be further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings. It should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present invention. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessary confusion of the concepts of the present invention.

[0018] refer to Figure 1-2 .

[0019] The present invention provides a stroke patient rehabilitation movement analysis system based on medical big data, comprising: The data acquisition module (electromyographic sensors, motion monitoring equipment, behavioral state sensors) is used to collect patients' electromyographic signals, motion data, and behavioral data to form a data set. The data processing and fusion module preprocesses the electromyographic signals in the data set, normalizes the motion and behavioral data, and fuses them into a multimodal feature vector. The hierarchical analysis module (noise elimination, feature extraction, multimodal fusion) identifies movement intent based on the preprocessed electromyographic signals and outputs an intention classification result. It uses a two-stream Transformer encoder to identify movement intent, predicts rehabilitation outcomes based on a fuzzy neural network, and assesses behavioral risk levels using a risk factor formula. (Motion intent identification, rehabilitation outcome prediction, behavioral risk assessment) The optimization module dynamically adjusts the intensity of mirror therapy based on risk levels and uses incremental learning to optimize the prediction model. (Incremental learning model updates, rehabilitation strategy adjustment) The interaction module generates a heat map in the XY coordinate system and overlays the rehabilitation curve, annotating physician intervention recommendations for visual interaction. (Heat map, rehabilitation curve, physician intervention interface)

[0020] The electromyographic signal is collected from the affected and unaffected hands using a surface electromyographic sensor. The motion data is recorded using an inertial measurement unit to record the patient's exercise intensity, frequency, and duration. The exercise intensity is calculated by fusing heart rate and acceleration data. The behavioral data is obtained using a wearable device to obtain the patient's work dimension data, diet data, and rest index. The work dimension data includes weekly work hours and work intensity quantification values, and the diet data includes weekly alcohol consumption and weekly cholesterol intake ratio. Examples are as follows: EMG signal acquisition: An 8-channel surface EMG sensor is placed on the extensor digitorum, flexor digitorum, and thenar muscles of the affected hand, as well as corresponding locations on the unaffected hand, to synchronously collect EMG signals. For example, the sensor model is the Delsys Trigno Wireless EMG, which supports real-time data transmission.

[0021] Motion data collection: The inertial measurement unit (IMU) is fixed to the patient's wrist to record triaxial acceleration, angular velocity and heart rate data. Intensity = 0.6 x heart rate ratio + 0.4 x acceleration value; heart rate ratio = real-time heart rate / maximum heart rate.

[0022] Behavioral data collection: Weekly working hours (hours), work intensity (wage / legal working hours ratio), weekly alcohol consumption, and weekly cholesterol intake ratio (high-fat food quality / low-fat food quality) were obtained through smart wristbands (such as Fitbit Charge 5).

[0023] The data preprocessing and feature fusion process includes: performing wavelet denoising on the EMG signals in the dataset. Specifically, wavelet transform (Daubechies 4 wavelet basis) is used to remove baseline drift, and FastICA blind source separation is combined to filter out power frequency interference. Power frequency interference is also filtered out using a blind source separation algorithm; Z-score normalization is used to standardize the motion and behavioral data; and a multimodal feature vector is constructed, including the time domain and frequency domain features of the EMG signals, the intensity-frequency joint index of the motion data, and the risk factor of the behavioral data.

[0024] Among them, Y, E, and N are exercise, work, and diet factors respectively, and d1, d2, and d3 are preset weights.

[0025] The preprocessed EMG signals are fed into a two-stream Transformer encoder. The long sequence branch uses a 200ms time window, and the short sequence branch uses a 50ms time window. The cross-attention module outputs the intent classification results, including grasp, stretch, and relax. A dynamic prediction model is constructed based on a fuzzy neural network. Motion data and intention classification results are input, and the Levenberg-Marquardt algorithm is used to optimize network parameters to output a rehabilitation score (e.g., 0-100 points). The algorithm training error threshold is 0.01 (for example only).

[0026] The risk level is generated based on the risk factor XW and the preset threshold, including high risk, medium risk and low risk. For example: high risk setting (XW<50) red alert; medium risk (50 <XW<80)黄色预警;低危(XW> 80) Green and safe.

[0027] If the risk level is high, the intensity of mirror therapy will be increased, and grip training of the affected hand will be enforced using rehabilitation equipment; For example, the intensity of mirror therapy has been increased to 150%, and the mandatory daily training time has been increased by half an hour.

[0028] If the rehabilitation score is lower than the preset value (e.g. 60), the incremental learning mechanism is triggered to update the fuzzy neural network weights based on the new data of this week and optimize the prediction model for the next cycle. The incremental learning mechanism is specifically as follows: When the proportion of newly added data in the original training set exceeds the preset proportion each week, the entire network parameter update is triggered; When the ratio is lower than the preset ratio, only the weight of the last fully connected layer is updated.

[0029] In one embodiment, the learning rule is: if the proportion of new data is greater than 20%, all network parameters are updated (learning rate 0.001); if the proportion of new data is less than or equal to 20%, only the output layer weights are updated (learning rate 0.01).

[0030] Draw a patient behavior status curve in an XY coordinate system, with time on the horizontal axis and risk factor XW on the vertical axis, and superimpose the rehabilitation score curve; use a heat map to color-code high-risk periods and mark doctor intervention recommendations. The heat map generation method includes: Divide the XY coordinate system into a 50×50 pixel grid; Based on the risk level distribution of the data points within the grid, a bilinear interpolation algorithm is used to fill in the color. The color transparency is positively correlated with the risk density. The transparency formula is: a = 0.3 + 0.7 x number of high-risk points / total number of points.

[0031] Abnormal behavior warning, specifically: If the exercise duration is lower than the preset threshold for a set number of consecutive days, a reminder message will be sent to the patient; If the weekly cholesterol intake ratio exceeds the preset value, a dietary adjustment suggestion will be generated and pushed to the user terminal.

[0032] Doctors intervene, automatically popping up an offline assessment appointment window; pushing customized rehabilitation video links; and generating PDF weekly reports. These correspond to different risk levels.

[0033] The logic for generating the doctor's intervention recommendations is as follows: high-risk patients: mandatory appointment for offline rehabilitation assessment this week; medium-risk patients: provide customized video guidance course links; low-risk patients: generate weekly rehabilitation reports and send them by email.

[0034] In order to better understand the present invention, an actual case is used for illustration below.

[0035] The patient, Mr. Zhang, a 58-year-old male, suffered from right hemiplegia after stroke (Brunnstrom stage III) and underwent 4 weeks of rehabilitation training.

[0036] Data collection (Day 1) The electromyographic sensor collected electromyographic signals of the affected hand (sampling rate 1000 Hz), the IMU recorded movement data (average intensity 45, frequency 5 times / day, duration 20 minutes), and the smart bracelet obtained behavioral data (weekly working hours 50 hours, weekly cholesterol intake ratio 1.8).

[0037] Data processing and fusion After wavelet denoising and FastICA processing, the EMG signals were extracted with RMS (root mean square value) = 120 μV, iEMG (integrated EMG) = 3500 μV·s, and MF9 (median frequency) = 90 Hz.

[0038] Calculating risk factors , marked as medium risk.

[0039] Analysis: Motion Intent Recognition: Two-stream Transformer outputs the “grasp” intention (92% confidence level); Rehabilitation score: Fuzzy neural network output 68 points (out of 100); Risk level: Medium (XW=52).

[0040] Adjust the rehabilitation plan, increase the intensity of mirror therapy to 120%, and add 10 minutes of grip training per day; Incremental learning: New data accounts for 15%, and only the output layer weights are updated.

[0041] Visual feedback: Doctors review the heat map and find that patients have a higher risk of developing symptoms during workdays. They then provide a "work break stretching tutorial"; The system automatically sends dietary advice: "The weekly cholesterol intake ratio needs to be reduced to below 1.5."

[0042] Experimental results: The rehabilitation score increased to 82 points (↑20%), the risk factor XW decreased to 65 (medium risk → low risk), and the grip strength of the affected hand recovered to 70% of the healthy side (initial 40%).

[0043] This invention forms a closed loop of "acquisition-processing-analysis-optimization-feedback" through (data acquisition) → (data cleaning and fusion) → (intelligent analysis) → (strategy optimization) → (visualized interaction). Multimodal data fusion (electromyography, motion, and behavioral data) enhances comprehensive assessments, dynamic models (fuzzy neural networks + incremental learning) implement personalized rehabilitation strategies, and visual interaction reduces the cost of doctor-patient communication.

[0044] In certain embodiments, reference Figure 2 In the fusion feature, multi-scale convolution kernel is used to extract spatiotemporal features. Local feature layer: 3×3×3 convolution kernel extracts single joint motion features, regional feature layer: 5×5×5 convolution kernel captures multi-joint collaborative motion patterns, global feature layer: 1×1×1 convolution kernel integrates whole body motion trajectory features, temporal feature enhancement layer: bidirectional LSTM network captures long-term motion trends.

[0045] The basic principles, main features, and advantages of the present invention are shown and described above. It should be understood by those skilled in the art that the present invention is not limited to the above-described embodiments. The above-described embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention, and such changes and modifications fall within the scope of the invention claimed.

Claims

1. A stroke patient rehabilitation exercise analysis system based on medical big data, characterized by: include: Data acquisition module, used to collect patients' electromyographic signals, motion data and behavioral data to form a data set; A data processing and fusion module is used to pre-process the electromyographic signals in the data set, normalize the motion and behavior data, and fuse them into a multimodal feature vector; A hierarchical analysis module is used to identify movement intentions based on the preprocessed electromyographic signals and output intention classification results, identify movement intentions through a dual-stream Transformer encoder, predict rehabilitation effects based on a fuzzy neural network, and assess behavioral risk levels in combination with a risk factor formula; An optimization module, which dynamically adjusts the intensity of mirror therapy based on risk level and optimizes the prediction model using incremental learning; The interactive module is used to generate a heat map and superimpose the recovery curve in the XY coordinate system, and mark the doctor's intervention suggestions to complete the visual interaction.

2. The method for analyzing rehabilitation movement of stroke patients based on medical big data according to claim 1, characterized in that: The electromyographic signals are collected from the affected and healthy hands via surface electromyographic sensors. The exercise data: uses an inertial measurement unit to record the patient's exercise intensity, exercise frequency and exercise duration. The exercise intensity is calculated by fusing heart rate and acceleration data; The behavioral data: uses wearable devices to obtain the patient's work dimension data, diet data and rest index, the work dimension data includes weekly working hours and work intensity quantification value, and the diet data includes weekly drinking frequency and weekly cholesterol intake ratio.

3. The method for analyzing rehabilitation movement of stroke patients based on medical big data according to claim 2, characterized in that: The process of data preprocessing and feature fusion includes: Performing wavelet noise reduction processing on the electromyographic signals in the data set and filtering out power frequency interference using a blind source separation algorithm; The Z-score normalization method was used to standardize the movement and behavior data; Construct a multimodal feature vector, including the time domain and frequency domain features of the electromyographic signal, the intensity-frequency joint index of the motion data, and the risk factor of the behavioral data: Among them, Y, E, and N are exercise, work, and diet factors respectively, and d1, d2, and d3 are preset weights.

4. The method for analyzing rehabilitation movement of stroke patients based on medical big data according to claim 3 is characterized by: The preprocessed EMG signals are fed into a two-stream Transformer encoder. The long sequence branch uses a 200ms time window, and the short sequence branch uses a 50ms time window. The cross-attention module outputs the intent classification results, including grasp, stretch, and relax. A dynamic prediction model was constructed based on a fuzzy neural network. The motion data and intention classification results were input, and the Levenberg-Marquardt algorithm was used to optimize the network parameters and output the rehabilitation score. The risk level is generated based on the risk factor XW and the preset threshold, including high risk, medium risk, and low risk.

5. The method for analyzing rehabilitation movement of stroke patients based on medical big data according to claim 4 is characterized in that: If the risk level is high, the intensity of mirror therapy will be increased, and grip training of the affected hand will be enforced using rehabilitation equipment; If the rehabilitation score is lower than the preset value, the incremental learning mechanism is triggered to update the fuzzy neural network weights based on the new data of this week and optimize the prediction model for the next cycle.

6. The method for analyzing rehabilitation movement of stroke patients based on medical big data according to claim 5, characterized in that: Draw the patient's behavioral status curve in the XY coordinate system, with time on the horizontal axis and risk factor XW on the vertical axis, and superimpose the rehabilitation score curve; use a heat map to color-code high-risk periods and mark doctor intervention recommendations.

7. The method for analyzing rehabilitation movement of stroke patients based on medical big data according to claim 5, characterized in that: The incremental learning mechanism is specifically as follows: When the proportion of newly added data in the original training set exceeds the preset proportion each week, the entire network parameter update is triggered; When the ratio is lower than the preset ratio, only the weight of the last fully connected layer is updated.

8. The method for analyzing rehabilitation movement of stroke patients based on medical big data according to claim 5, characterized in that: The method for generating the heat map includes: Divide the XY coordinate system into an a×a pixel grid; According to the risk level distribution of data points in the grid, a bilinear interpolation algorithm is used to fill the color, and the color transparency is positively correlated with the risk density.

9. The method for analyzing rehabilitation movement of stroke patients based on medical big data according to claim 1, characterized in that: It also includes abnormal behavior warnings, specifically: If the exercise duration is lower than the preset threshold for a set number of consecutive days, a reminder message will be sent to the patient; If the weekly cholesterol intake ratio exceeds the preset value, a dietary adjustment suggestion will be generated and pushed to the user terminal.

10. The method for analyzing rehabilitation movement of stroke patients based on medical big data according to claim 1, characterized in that: The logic for generating the doctor's intervention suggestion is as follows: High-risk patients: mandatory appointment for offline rehabilitation assessment this week; Moderate-risk patients: provide a link to a customized video guidance course; Low-risk patients: Generate weekly recovery reports and send them by email.

Images