Intelligent guidance system for rehabilitation training of burn patients based on deep learning

Abstract

The present invention relates to the technical field of rehabilitation training for burn patients, and specifically to an intelligent guidance system for rehabilitation training for burn patients based on deep learning. Multimodal data integration technology is used to collect video records, physiological signals and skin elasticity measurement data of patient rehabilitation training, and the data is cleaned, standardized and classified. Data enhancement technology is applied to control and improve the generalization ability of the model; a deep learning classification model is developed, including a convolutional neural network CNN, which is used to analyze image data of skin recovery, and a recurrent neural network RNN ​​or a long short-term memory network LSTM is applied to predict the patient's future rehabilitation trend; a generative adversarial network GAN is used to generate personalized training movements or treatment plans for different rehabilitation stages, and a reinforcement learning method is used to optimize the rehabilitation training plan, and the training intensity and type are dynamically adjusted according to the patient's real-time feedback; finally, natural language processing NLP technology is integrated to realize speech recognition and generation, and the training is guided by communicating with the patient through natural language.

Description

Technical Field

[0001] The present invention relates to the technical field of rehabilitation training for burn patients, and in particular to an intelligent guidance system for rehabilitation training for burn patients based on deep learning. Background Art

[0002] In the current rehabilitation training guidance system, there is a more traditional method, such as the system described in Chinese invention patent publication number CN2024101049979, which includes multiple patient clients, intermediate clients and decision-making clients. Although this method can provide certain rehabilitation training guidance, it has shortcomings and disadvantages, especially in processing a large amount of patient data, the accuracy of real-time feedback, the formulation of personalized rehabilitation plans, and the improvement of rehabilitation efficiency. First, the data circulation method in the system may cause delays in information processing. In this structure, the patient client first collects the feedback data of the rehabilitation patients and then sends it to the intermediate client, which then filters the data and sends the target feedback data to the decision-making client. This multi-level data transmission and processing process not only increases the complexity of the system, but also may slow down the processing and response speed of key information. For patients who need to adjust their rehabilitation plans in a timely manner, any delay may affect the rehabilitation effect. Secondly, the system's personalized processing capabilities are limited. Although the intermediate client can filter out feedback data that requires special attention, this screening process depends on the experience and judgment of the intermediate client operator, which may lead to individual differences being ignored. Rehabilitation training needs to adjust the training plan according to the specific situation of the patient, including the type, frequency and intensity of training. If it is not possible to make quick and precise adjustments based on each patient's feedback, the best rehabilitation effect may not be achieved. In addition, the system is insufficient in real-time data analysis and instant feedback.

[0003] The existing system architecture requires data to be transmitted between multiple clients, which is not conducive to the implementation of real-time data analysis and feedback based on complex algorithms. For example, using deep learning models to analyze patient behavior and provide immediate training adjustment suggestions requires a large amount of computing resources and high-speed data processing capabilities. In traditional multi-client systems, data processing and transmission delays at each step may affect the timeliness and effectiveness of rehabilitation training. Furthermore, regarding the evaluation and feedback of rehabilitation effects, traditional systems usually lack comprehensive evaluation tools and mechanisms. Rehabilitation training needs to consider not only short-term effects, but also progress in the long-term rehabilitation process. An effective evaluation mechanism should include continuous monitoring of the patient's rehabilitation status, adjustment feedback of the training process, and long-term tracking of rehabilitation results, but existing systems often find it difficult to achieve the organic integration of these functions. Summary of the invention

[0004] The purpose of the present invention is to provide an intelligent guidance system for rehabilitation training of burn patients based on deep learning, so as to solve some of the drawbacks and shortcomings pointed out in the background technology.

[0005] The present invention solves the above-mentioned technical problems by adopting the following technical solutions: an intelligent guidance system for rehabilitation training of burn patients based on deep learning, comprising: firstly, using multimodal data integration technology to collect video records of patient rehabilitation training, physiological signals and skin elasticity measurement data, cleaning, standardizing and classifying the data, and applying data enhancement technology to control and improve the generalization ability of the model;

[0006] Next, we developed deep learning classification models, including convolutional neural networks (CNNs), to analyze image data of skin recovery, and applied recurrent neural networks (RNNs) or long short-term memory networks (LSTMs) to predict the patient's future recovery trend.

[0007] Then, the generative adversarial network (GAN) is used to generate personalized training actions or treatment plans for different rehabilitation stages, and the rehabilitation training plan is optimized through reinforcement learning methods, dynamically adjusting the training intensity and type based on the patient's real-time feedback;

[0008] Finally, natural language processing (NLP) technology is integrated to realize speech recognition and generation, and natural language is used to communicate with patients to guide training. At the same time, computer vision technology is used to analyze the patient's training movements in real time, build a system to guide the correctness of movements and provide instant feedback.

[0009] Furthermore, the multimodal data integration technology is implemented in the following steps:

[0010] S1. First, a burn-specific feature fusion network is used, using a multi-dimensional weighted fusion formula:

[0011]

[0012] Where x represents the image feature vector of the burn area, y represents the physiological signal vector of the pain response, z is the change in skin elasticity, γ is the nonlinear index that adjusts the influence of data features, and α(x) is the weight function that is dynamically adjusted according to the input features;

[0013] S2. Then use the conditional generative adversarial network to enhance the burn area data and apply the random process formula:

[0014]

[0015] Where x represents the original rehabilitation data, and the data generated by Fourier series representation retains the rehabilitation characteristics while introducing controlled randomness to simulate the variation at different recovery stages;

[0016] S3. Then the data is classified and labeled through self-supervised learning method, using calculus formula:

[0017]

[0018] Where x is the burn recovery data to be classified, and the higher-level data features are extracted by combining differential and integral operations to control high-precision automatic classification;

[0019] S4. Finally, a joint rehabilitation status assessment model was developed using high-order polynomials and partial derivative formulas:

[0020]

[0021] Where x and y represent the feature vectors of visual data and physiological data, respectively. The visual and physiological data are fused through the divergence operation of the vector field to control the comprehensiveness of the rehabilitation guidance program.

[0022] Furthermore, the deep learning classification model specifically includes:

[0023] First, a multi-scale convolutional neural network (CNN) was developed, which was designed to analyze the image data of burn patients’ skin, using a nonlinear fusion formula:

[0024]

[0025] Where x represents the pixel value or feature of the skin restoration image; α is the width of the adjustment function, which affects the local or global analysis of the feature, making the network adaptable to the analysis of burn images of different scales; β k is a frequency parameter that adjusts the response to different frequencies of the image and helps the network capture recovery patterns from subtle textures to wide areas;

[0026] At the same time, the bidirectional long short-term memory network Bi-LSTM is combined to predict the patient's future recovery trend, using the time series analysis formula:

[0027]

[0028] Among them, x(t) represents the quantitative indicator of the rehabilitation status at time t, including pain level, inflammatory response or skin elasticity; σ is the attenuation parameter that is dynamically adjusted according to the rehabilitation progress, and predicts the future rehabilitation path based on the patient's current and past rehabilitation data;

[0029] Finally, through the feature fusion strategy, use the formula:

[0030]

[0031] Among them, x and y are the features of CNN and Bi-LSTM outputs, respectively, reflecting the visual recovery characteristics of the skin and the temporal dynamics of the recovery state; γ is used as a regulation parameter to balance the influence of the two data sources.

[0032] Furthermore, the generation of the personalized training action or treatment plan includes:

[0033] S1. First, the conditional generative adversarial network cGAN is used to generate personalized training actions or treatment plans for different rehabilitation stages based on the type, depth and recovery stage of the burn.

[0034] S2, then the multi-armed bandit MAB model is used to select rehabilitation actions based on the patient's real-time feedback, including pain intensity and comfort assessment, to control the rehabilitation training program in response to the patient's changes;

[0035] S3. Finally, the rehabilitation training plan is further adjusted and optimized through the deep Q network DQN, using the data obtained from the skin elasticity monitor sensor and the patient's self-reported feedback to dynamically adjust the training intensity and type.

[0036] Furthermore, the conditional generative adversarial network cGAN is used to generate the input burn type t, depth d, and recovery stage s, combined with random noise z, through the formula:

[0037]

[0038] Specifically, z represents the random noise introduced from the prior distribution; t represents the numerical representation of the burn type, including chemical burns, thermal burns, etc., which is used to generate rehabilitation plans for specific types of burns; d represents the numerical representation of the burn depth, including first-degree, second-degree, and third-degree burns, which affects the intensity and sensitivity of the generated actions; s represents the numerical representation of the recovery stage, including the acute stage, healing stage, and reconstruction stage, and the treatment plan is adjusted according to the different rehabilitation stages; σ represents the standard deviation of the sensitivity of the generated samples; ω represents the periodic adjustment factor, which is used to simulate the periodic adjustment during rehabilitation training.

[0039] Then, based on the patient’s real-time feedback data x, the feedback adjustment formula is used:

[0040]

[0041] Where x represents the data extracted from the patient's real-time feedback, including the pain level or the quantitative index of rehabilitation progress; α and β are model adjustment parameters, which are combined through nonlinear and integral operations to control the actual effect of training actions on the patient's rehabilitation progress; tanh(αx ′ ) represents the feedback function, which is provided by the hyperbolic tangent function;

[0042] Finally, a dynamic adjustment mechanism is adopted by adjusting the formula:

[0043]

[0044] Here, x represents the rehabilitation progress data obtained from physiological monitoring devices including skin elasticity monitors or pain rating systems.

[0045] Furthermore, the multi-armed bandit MAB model construction includes:

[0046] S1. First, define and classify different rehabilitation actions as arms of a multi-armed slot machine. Each arm represents a rehabilitation training action using the formula:

[0047]

[0048] Where t is time, γ, β k and δ k is an adjustment parameter specific to action k, s(t) represents the rehabilitation status indicator at time t, which is used to adjust the selection probability of each action;

[0049] S2. Secondly, collect the patient's pain intensity and comfort feedback when performing each rehabilitation action, and adopt an update strategy through the multi-armed bandit model:

[0050]

[0051] where r i is the reward feedback at time i, δ ki is the indicator function, θ k and ν k is the learning rate and exploration parameter of action k, adjusting the weights of history and new feedback;

[0052] S3, finally through dynamic adjustment algorithm:

[0053]

[0054] Where s(u) is the rehabilitation state score at time u, λ and ρ are parameters for adjusting the sensitivity of rehabilitation movements, and the integral represents the cumulative effect of the time-attenuated rehabilitation state score on the adjustment of training movements from the beginning to the current moment.

[0055] Furthermore, the adjustment and optimization of the rehabilitation training plan covers the following technical implementation steps:

[0056] First, through the deep Q network DQN model, use the function:

[0057]

[0058] For training and optimization, s trepresents the data obtained from the skin elasticity monitor sensor and the patient's self-reported pain intensity or comfort rating, a t represents the potential training action, θ and φ are the parameters adjusted during the learning process;

[0059] Then collect sensor and self-reported data in real time, and use the integral calculation formula:

[0060]

[0061] where α and β are coefficients that adjust for the effects of time decay.

[0062] Furthermore, the natural language processing (NLP) technology uses the function:

[0063]

[0064] Realize speech recognition and generation, where s(t) represents the change of speech feature variables over time, r(t) represents the emotional state variables fed back by the patient, and α, β, and κ are model parameters used to process speech data and generate appropriate training guidance feedback;

[0065] Secondly, combined with computer vision technology, through the function:

[0066]

[0067] Real-time analysis of the correctness of the patient's training movements, where d(t) represents the movement deviation under time ·, ω is the periodic adjustment factor, and γ is the time decay parameter, ensuring that the movement analysis accurately reflects the actual performance of the patient's movements;

[0068] Finally, the output of vision and language processing is adjusted by the algorithm:

[0069]

[0070] Dynamic integration, where x and y represent quantitative data of visual and language feedback, respectively, to provide instant training adjustments and feedback.

[0071] Beneficial effects of the present invention:

[0072] 1. Personalized rehabilitation program: By using deep learning models to analyze patients' real-time data, including skin elasticity, pain intensity, comfort, etc., the system can customize a personalized rehabilitation program for each patient. This personalization ensures the relevance and effectiveness of rehabilitation training and adapts to the specific recovery needs of each patient.

[0073] 2. Dynamic training adjustment: The system can collect and analyze patient feedback data in real time, and dynamically adjust the intensity and type of training through advanced algorithms. This not only reduces the patient's discomfort, but also prevents potential injuries caused by overtraining.

[0074] 3. Enhanced interactivity and participation: By integrating natural language processing and computer vision technology, the system provides real-time interaction with patients, enhancing their participation in the rehabilitation process. Patients can give feedback on their feelings through voice, and the system adjusts the training content accordingly, making patients feel more cared for and satisfied.

[0075] 4. Instant feedback and evaluation: The system can instantly evaluate the effectiveness of rehabilitation training and provide feedback. This helps physical therapists or medical teams monitor the patient's progress and make necessary adjustments in a timely manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 This is a flow chart of the intelligent guidance system for rehabilitation training of burn patients based on deep learning in the present invention.

[0077] Figure 2 This is a flowchart of the implementation steps of the multimodal data integration technology of the present invention.

[0078] Figure 3 This is a flow chart of the degree learning classification model of the present invention.

[0079] Figure 4 This is a flow chart of the personalized training movements or treatment plans of the present invention. DETAILED DESCRIPTION

[0080] The specific implementation modes of the present invention will be described in detail below in conjunction with the accompanying drawings.

[0081] Intelligent guidance system for rehabilitation training of burn patients based on deep learning, as shown in the attached Figure 1 S1 first uses multimodal data integration technology, including collecting video records of patients during rehabilitation training, physiological signals such as heart rate and blood pressure, and skin elasticity measurement data. These data can provide comprehensive information to evaluate the patient's rehabilitation status; then, all collected data need to be cleaned to remove noise and irrelevant information, and then standardized to ensure the comparability of data from different sources and at different times. The classification processing is to divide the data into different categories based on the characteristics of the data, so that the subsequent deep learning model can better learn and recognize; finally, by applying data enhancement techniques such as rotation, scaling, shearing or artificially increasing the changes in data samples through algorithms, the generalization ability of the model when processing unseen data can be significantly improved.

[0082] A deep learning classification model was then developed to further analyze and predict the recovery process. Figure 1 S2 uses a convolutional neural network (CNN) to analyze image data of skin recovery. CNN is suitable for processing image data, and it can effectively identify and analyze local patterns and features in images, such as texture, color, and edge changes of burned skin. These features are essential for evaluating the recovery status of the skin and monitoring the progress of rehabilitation. In addition, the system also applies a recurrent neural network (RNN) or a long short-term memory network (LSTM) to predict the patient's future rehabilitation trend. RNN and LSTM are particularly suitable for processing time series data and can capture dynamic changes in the rehabilitation process, such as the frequency and intensity of rehabilitation activities and the changing trend of patient feedback. Such time series analysis helps the model understand the long-term trends and periodic patterns of rehabilitation progress, and provide patients with more personalized and precise rehabilitation advice. By combining the analysis results of CNN and RNN or LSTM, the system can comprehensively evaluate the patient's current rehabilitation status and predict his future recovery trajectory, thereby effectively guiding and adjusting the rehabilitation training plan.

[0083] Generative adversarial networks (GANs) are further used to generate personalized training movements or treatment plans for different rehabilitation stages. Figure 1 In S3, GAN works by setting two neural networks - a generator and a discriminator - to compete with each other, where the generator learns to create new data that looks similar to real data, while the discriminator tries to distinguish between generated data and real data. In this system, the generator generates effective training actions or treatment plans based on the patient's current rehabilitation status and rehabilitation stage, while the discriminator evaluates whether these plans are practical and feasible. In addition, the system also uses reinforcement learning methods to optimize rehabilitation training plans. The reinforcement learning model observes the patient's response to the training actions (such as pain level, comfort, and recovery speed) and adjusts its strategy based on this feedback to maximize the patient's rehabilitation effect. The model continuously learns and adjusts, making the training plan more effective over time. In this process, the system is able to dynamically adjust the intensity and type of training based on the patient's real-time feedback, ensuring that the training plan always responds best to the patient's specific needs and rehabilitation progress.

[0084] Finally, natural language processing (NLP) technology is integrated to achieve speech recognition and generation, as shown in the attached Figure 1S4 enables the system to communicate with patients through natural language and directly guide training. This method of communication enables patients to ask questions, express their feelings or feedback discomfort during training through voice, and the system can understand and respond to the patient's needs in real time, provide specific guidance on how to perform specific rehabilitation movements, explain the purpose of the movements, or adjust the intensity and type of training. At the same time, the system also uses computer vision technology to analyze the training movements performed by patients in real time to ensure the accuracy of the movements. This includes monitoring every detail of the movement, such as the angle of the arm or leg, as well as the speed and amplitude of the movement, to ensure that all movements are performed in accordance with the standards of the physical therapist, thereby maximizing the rehabilitation effect. When the system detects any deviation or error through visual analysis, it will immediately provide feedback through NLP technology to tell the patient how to adjust the movement to correct the error. This instant feedback greatly improves the effectiveness of the training and the safety of the patient.

[0085] Embodiment 1:

[0086] Combined with Figure 2 In this embodiment, the system uses multimodal data integration technology and a burn-specific feature fusion network to help burn patients recover more effectively. The multi-dimensional weighted fusion formula used is as follows:

[0087]

[0088] In this equation:

[0089] x represents the image feature vector of the burn area, which may be texture, color, and boundary features extracted from the image of the burn area;

[0090] y represents the physiological signal vector of pain response, such as data extracted from electromyography (EMG) or heart rate variability (HRV);

[0091] z is the change in skin elasticity, usually obtained by a skin elasticity measurement device, indicating the degree of skin recovery.

[0092] γ is a nonlinear index that adjusts the influence of data features. It is used to control the influence of skin elasticity data in the overall model. The value of γ can be set to 0.5, which is a commonly used value for balancing the influence of linear and nonlinearity.

[0093] α(x) is a weight function that is dynamically adjusted according to the input features. It is set as a function based on the complexity of the image features, for example, α(x) = 1-norm(x), where normx is the normalized value of x.

[0094] Suppose the following data is collected from a burn patient:

[0095] The image feature vector of x is extracted through image processing technology, and its normalized value is 0.8.

[0096] The pain signal of y, measured by a physiological sensor, has a value of 20 (this can be a pain score).

[0097] The skin elasticity of z was measured to be 0.7 (this represents the elasticity ratio relative to normal skin).

[0098] Substituting into the formula, first calculate the weight function: α(x) = 1-0.8 = 0.2

[0099] Then calculate the integral part:

[0100]

[0101] Next, we calculate log(y+1):

[0102] log(20+1)=log(21)≈3.04

[0103] Finally calculate z γ :

[0104] z γ =0.7 0.5 ≈0.836

[0105] Putting it all together:

[0106] f(x,y,z)=0.1492·3.04+0.836≈0.453568+0.836=1.289568

[0107] The calculation result f(x,y,z)=1.289568 represents a comprehensive rehabilitation score that combines image features, pain signals, and skin elasticity. This score can be used to guide rehabilitation training, such as adjusting the intensity, type, and duration of training to optimize the patient's rehabilitation effect.

[0108] Then, the conditional generative adversarial network (cGAN) was used to enhance the burn area data to generate more training data to improve the system's adaptability and prediction accuracy for different burn recovery stages. By applying a random process formula based on Fourier series, the system can introduce controlled randomness while retaining the recovery characteristics to simulate various variations in the burn recovery process.

[0109] The specific formula is:

[0110]

[0111] Here, x represents the original rehabilitation data, such as skin elasticity measurements, pain scores, etc., and periodic changes are introduced through the Fourier series method to generate a dataset with similar rehabilitation characteristics but containing random small changes.

[0112] Assume that the original pain score x obtained from a specific case is 3. Substitute this value into the above formula and calculate the first few terms of the formula to estimate its result (in practical applications, the number of terms in the series is usually limited to ensure the feasibility of the calculation). Considering the rapid convergence of the series, the first three terms can be calculated as an example:

[0113] When n=1:

[0114]

[0115] When n=2:

[0116]

[0117] When n=3:

[0118]

[0119] Summing the values ​​of these three terms gives an approximate estimate:

[0120] g(3)≈1.3365+1.944+0.395≈3.676

[0121] This result indicates that by introducing randomness and Fourier series methods, new pain score data can be generated that statistically maintains similar characteristics to the original data while containing a certain degree of variation, simulating the natural fluctuations in patients' pain perception in the real world.

[0122] Finally, the burn recovery related data is classified and labeled through self-supervised learning methods to support the adaptive and personalized guidance of the system. The composite calculus formula is used to process and analyze the data:

[0123]

[0124] Here, x represents the burn recovery data to be classified, such as skin elasticity index, pain level or recovery status score. By combining differentiation and integration operations to extract deeper features in the data, high-precision classification can be controlled.

[0125] First, the differential term Used to highlight the rate of change and trend of each data point within a small range, which helps to identify local characteristics and subtle changes in the data. For example, if x is a pain score, this operation will help identify sensitive areas of pain change.

[0126] Secondly, the integral term ∫x 2ln(x)dx is used to accumulate global features of the entire dataset, providing a holistic view of the data characteristics. For example, for skin elasticity data, this part can help assess the overall recovery trend and the long-term improvement or degradation that occurs during the recovery process.

[0127] Assume that x = 2, which represents the recovery index of a specific burn patient. Each part of the calculation formula is as follows:

[0128] 1. Calculation of the differential part:

[0129]

[0130] Assume sin(2)≈0.9093 (obtained from standard trigonometric tables),

[0131] Using approximate differentiation methods (such as numerical differentiation), the derivative of this expression is approximately 0.2 (this is an example value and the exact calculation depends on the accuracy of the differentiation step).

[0132] 2. Calculation of the integral part: ∫ 2 ln(2)dx=4ln(2)·x

[0133] When x=2, 4ln(2)·2=8ln(2)≈5.545 (using the approximation of ln(2)≈0.693).

[0134] Adding these two parts together:

[0135] h(2)≈0.2+5.545=5.745

[0136] This result h(2)=5.745 can be used to help classify and label data, where the calculated value can be used as a component of the feature vector and input into the machine learning model to predict and analyze the recovery status of burn patients.

[0137] Finally, a joint rehabilitation status assessment model was developed, using high-order polynomials and partial derivative formulas to fuse and analyze information from different data sources, thereby providing patients with a comprehensive rehabilitation guidance program. The specific calculation formula is as follows:

[0138]

[0139] In this formula:

[0140] x and y represent feature vectors obtained from computer vision systems and physiological monitoring devices, respectively. x is the motion accuracy feature extracted from the patient's rehabilitation training video, such as motion amplitude, speed, etc., while y is the value of physiological signals such as heart rate, blood pressure or electromyography.

[0141] By applying the divergence operation of the vector field, this formula analyzes the information of each independent data source and also analyzes the information of each independent data source through the interaction terms x sin(y) and y 3 e x To consider the mutual influence of these data, making the rehabilitation plan more comprehensive and accurate.

[0142] Suppose the following data is obtained during a patient's rehabilitation training:

[0143] Visual data x=2, which indicates the deviation of the patient's movement range from the ideal range in a certain training movement.

[0144] Physiological data y=3, which is the average heart rate of the patient when performing the training.

[0145] Substituting these values ​​into the formula:

[0146] 1. Calculate x sin(y)+y 3 e x :

[0147] x sin(y) = 2 sin(3) ≈ 2 × 0.1411 = 0.2822 (approximate value obtained using a standard trigonometric table or a calculator)

[0148] y 3 e x =27e 2 ≈27×7.389=199.503 (same, using e 2 ≈7.389 approximation)

[0149] The total is: 0.2822+199.503=199.7852

[0150] 2. Calculate the divergence

[0151] This step needs to take into account the spatial distribution and variation of the data, which is usually calculated in multidimensional data space by numerical methods. The divergence calculated by numerical methods is set to 0.5 (this is an example value, depending on the distribution and detailed processing of the data).

[0152] The calculation result k(2,3)=0.5 provides a comprehensive score based on visual and physiological data, which can be used to judge the adaptability and effectiveness of the patient's current rehabilitation training.

[0153] Embodiment 2:

[0154] In this embodiment, combined with the attached Figure 3, developed a multi-scale convolutional neural network (CNN) for accurate analysis of image data of burn patients' skin. A nonlinear fusion formula was applied to process image features to enhance the model's ability to recognize skin changes during burn recovery. The formula is as follows:

[0155]

[0156] in:

[0157] x represents the pixel value or feature vector extracted from the burn area image.

[0158] α is the width parameter of the adjustment function, which controls the Gaussian weight function The decay rate of , thus affecting the local or global resolution of the feature. In practical applications, α is selected in the range of 0.1 to 1.0 to flexibly adjust the sensitivity of the model to image details.

[0159] β k is a frequency parameter, each k value corresponds to a different frequency, these parameters help the network capture image features at different scales, such as subtle textures or recovery patterns in wide areas. These parameters are usually selected in the range of 0.5 to 5.0 to adapt to different texture and pattern recognition needs.

[0160] Assume that a set of image data is obtained from the recovery area of ​​a burn patient, and the extracted pixel value feature x is ready to be input into the model after specific processing. Select α=0.5 and β1=1.0, β2=2.0, β3=3.0 as the instance values ​​of the parameters, and calculate the fusion effect of the first three items:

[0161] 1. Calculate each β k For the corresponding parts, the contribution of each part is estimated using the numerical approximation method of integration (such as the Monte Carlo method). The calculation results of each item are as follows:

[0162] For k=1:

[0163] For k = 2:

[0164] For k=3:

[0165] 2. Combine these results and multiply them by the Gaussian weight function Finally, it is integrated into the model for training and analysis.

[0166] Through this method, CNN is able to deeply analyze and understand the various stages of skin recovery in burn patients and provide patients with customized rehabilitation guidance.

[0167] At the same time, the bidirectional long short-term memory network (Bi-LSTM) is combined to predict the patient's future recovery trend, and the time series analysis formula is used to enhance the depth processing and prediction accuracy of the rehabilitation data. The formula is as follows:

[0168]

[0169] in:

[0170] x(t) represents a quantitative indicator of the recovery status at time t, including pain level, inflammatory response, or skin elasticity, which is obtained from medical monitoring devices or patient feedback.

[0171] σ is a decay parameter, which is used to adjust the weight of the data in the time window T, mainly reflecting the impact of time distance on the prediction. In theory, this parameter can be dynamically adjusted according to the changes in the patient's rehabilitation progress. For example, it can be set to a higher value (such as σ = 0.5) in the early stage of rehabilitation to capture more subtle changes, and lowered in the stable period (such as σ = 0.1) to reduce the impact of transitional noise.

[0172] T is the length of the time window, which can be days, weeks or even months, depending on the frequency of monitoring of the rehabilitation process and the required prediction horizon.

[0173] Regarding the recovery of skin elasticity of a patient, the monitoring data show that the skin elasticity data x(t) in the past week (T = 7 days) are as follows: 1.2, 1.3, 1.5, 1.4, 1.6, 1.7, 1.8 (the unit is a certain skin elasticity index). Select σ = 0.1 and t0 = 0 (calculated from the beginning of the week), and calculate g(t):

[0174]

[0175] Through the numerical integration method, the value of this integral can be estimated to predict the skin elasticity trend in the next week. For example, if the calculation result shows g(t)≈1.5, it indicates that the patient's skin elasticity is steadily improving. Based on this, the system can recommend continuing the current rehabilitation training or making appropriate adjustments.

[0176] Finally, the system's comprehensive analysis and prediction capabilities are enhanced by effectively integrating the features of different data sources. The feature fusion strategy is adopted, and the following formula is used to integrate the features output by the CNN and Bi-LSTM models:

[0177]

[0178] In this formula: x and y represent the output features obtained from the multi-scale convolutional neural network (CNN) and the bidirectional long short-term memory network (Bi-LSTM), respectively. Among them, x refers to the visual recovery characteristics of the skin during the burn recovery process (such as skin color change, texture recovery, etc.), and y refers to the recovery status characteristics obtained by time series data analysis (such as changes in skin elasticity index, trends in pain levels, etc.).

[0179] γ is a tuning parameter used to balance the impact of the two data sources to avoid excessive influence of either data source on the results. Typically, γ can be selected from 0.1 to 1.0, depending on the characteristics of the actual data and the specific requirements of the system.

[0180] Assume that the following data is obtained from the actual rehabilitation training of a burn patient:

[0181] From the CNN model, a series of output features x after image data processing are obtained, which represent the visual characteristics of skin restoration, such as a texture clarity index of 0.8.

[0182] At the same time, from the Bi-LSTM model, the output y of the time series analysis was obtained, which represents the dynamics of the recovery status, such as the change of the pain index over time was recorded as 0.6.

[0183] Selecting γ = 0.5, we can calculate the integral part of feature fusion:

[0184]

[0185] You can use numerical integration to estimate the value of this expression. For example, set the integral result to about 0.25 (this is an example value, which needs to be calculated numerically). Then, substitute this value into the fusion strategy formula:

[0186]

[0187] This result h(0.8,0.6)=0.63 indicates that the comprehensive index obtained by feature fusion can provide rehabilitation guidance for burn patients based on comprehensive evaluation of visual and physiological data.

[0188] Embodiment 3:

[0189] In this embodiment, a conditional generative adversarial network (cGAN) is used to generate personalized training actions or treatment plans based on the type, depth and recovery stage of the burn. This method can not only generate customized plans adapted to the conditions of individual patients, but also improve the pertinence and effectiveness of treatment.

[0190] cGAN consists of a generator and a discriminator. The generator is used to generate seemingly real rehabilitation training movements, while the discriminator determines whether these movements are realistic enough and meet the given rehabilitation conditions.

[0191] The conditions input to the generator include burn type t (eg, chemical burn, thermal burn), burn depth d (eg, first degree, second degree, third degree burn), and recovery stage s (eg, acute phase, healing phase, reconstructive phase).

[0192] Used to generate training action formula:

[0193]

[0194] in,

[0195] z is the random noise introduced from the prior distribution to increase the diversity of the generation process.

[0196] σ controls the width of the Gaussian decay function and affects the sensitivity of the generated action. The general value range is 0.1 to 1.0.

[0197] ω is the periodic adjustment factor, which generally ranges from 0.5 to 5.0 and is used to simulate the periodic changes during rehabilitation training.

[0198] Assume that a patient has experienced second-degree thermal burns (t=2, d=2) and is currently in the healing period (s=2). The system needs to generate a rehabilitation plan suitable for the patient's current stage.

[0199] Setting σ = 0.5 and ω = 1.0 controls the shape and periodicity of the function.

[0200] Set the random noise z = 0.8 and substitute it into the formula to calculate:

[0201]

[0202] because Some values ​​are very small (close to zero), which means that the generated motion will be very conservative, ensuring that no excessive stress is placed on the patient.

[0203] In this way, cGAN can generate appropriate rehabilitation movements or treatment plans based on the specific burn type, depth, and recovery stage, thereby greatly improving the personalization and effectiveness of rehabilitation training.

[0204] In order to dynamically adjust the rehabilitation training plan to adapt to the patient's real-time status and feedback, a feedback adjustment formula is introduced. The formula adjusts the training plan based on the patient's immediate feedback and a sophisticated mathematical model to maximize the rehabilitation effect. The formula is defined as follows:

[0205]

[0206] in:

[0207] x represents data extracted from real-time feedback from patients, including quantitative indicators such as pain level or recovery progress.

[0208] α and β are model tuning parameters that control the sensitivity of the feedback function and the smoothness of the integral, respectively. Usually, α ranges from 0.1 to 10, depending on the sensitivity of the pain data; and β is usually between 1 and 10 to ensure that the denominator does not vary due to x. ′ Instability caused by approaching zero.

[0209] tanh(αx ′ ) is the hyperbolic tangent function, which is used to provide a smooth and bounded response, allowing the model to maintain feedback stability when the pain sensation varies greatly.

[0210] Assume that the pain level of a patient during rehabilitation training changes from mild to moderate, and the data x changes from 0 to 5. Select α = 2 and β = 5, and the feedback adjustment formula can be calculated by numerical integration method:

[0211] 1. Calculate each x ′ The corresponding tanh(2x ′ ), which is at 0 (corresponding to x ′ =0) to close to 1 (corresponding to x ′ =5).

[0212] 2. Substitute each value into the integral formula to calculate the corresponding

[0213] 3. Estimate the overall feedback value R(x) by integrating from 0 to 5 using a numerical integration method (such as the trapezoidal rule or Simpson's rule).

[0214] For example, if the numerical integration result is 0.8, it means that under the current pain level and rehabilitation progress, the rehabilitation training program should be adjusted accordingly to reduce pain or better adapt to the patient's rehabilitation status. This value can be used to dynamically adjust the training intensity, such as reducing the number of repetitions or adjusting the training intensity, so as to optimize the patient's rehabilitation experience and effect.

[0215] Finally, a dynamic adjustment mechanism is used to adjust the rehabilitation training plan in real time to adapt to the patient's current recovery status. The following adjustment formula is used:

[0216]

[0217] in:

[0218] x represents the recovery progress data obtained from physiological monitoring devices such as skin elasticity monitors or pain rating systems. These data can provide real-time, quantitative information about the patient's recovery status, such as the elasticity of the skin or the intensity of pain.

[0219] In the formula is an expression involving logarithmic and square terms designed to capture the nonlinear characteristics and growth trend of rehabilitation data.

[0220] By deriving the expression, the change rate of the rehabilitation parameter is calculated and used to dynamically adjust the intensity and type of training.

[0221] Assume that a data point x=3 is obtained from the skin elasticity monitoring of a specific burn patient, which represents the skin elasticity index or pain score at a specific time point. Substitute this value into the formula to calculate:

[0222] 1. First calculate the original expression:

[0223]

[0224] 2. Then calculate its derivative:

[0225]

[0226] The calculation result D(3) = 12.024 indicates that under the current rehabilitation progress and conditions, the rehabilitation training plan needs to be adjusted accordingly to adapt to the patient's recovery needs. This adjustment includes increasing or decreasing the intensity of certain training activities, or adjusting the frequency and duration of training. In this way, the dynamic adjustment mechanism ensures that each training step can be optimized according to the patient's actual recovery situation, thereby effectively improving rehabilitation efficiency and patient satisfaction.

[0227] Embodiment 4:

[0228] In this embodiment, in order to ensure the real-time and personalization of the rehabilitation training program, the system adopts a multi-armed bandit (MAB) model, and uses real-time feedback to adjust the rehabilitation movements to adapt to the changes in the patient's pain intensity and comfort, thereby optimizing the rehabilitation process.

[0229] 1. Define rehabilitation movements:

[0230] Each rehabilitation exercise is considered an arm of a multi-armed bandit, where, for example, the exercise may include gentle stretching, more intense strength training, or a specific physical therapy technique.

[0231] The movements are categorised according to the different stages and needs of burn recovery to ensure they cover the full range from initial treatment to post-operative rehabilitation.

[0232] 2. Implementation of feedback function:

[0233] Using the formula:

[0234]

[0235] t represents time, which represents the length of time from the start of rehabilitation to the present.

[0236] γ, β k and δ k are tuning parameters specific to action k, used to control the sensitivity and adaptability of action selection. The values ​​of these parameters are usually determined based on prior clinical data and experience. For example, γ is between 0.1 and 1.0, β k Between 0.5 and 2.0, δ k Between 0.1 and 1.0, depending on the type of action and the desired effect.

[0237] s(t) represents the rehabilitation status index at time t, such as pain intensity or skin elasticity, which are obtained in real time through rehabilitation monitoring equipment.

[0238] 3. Example application:

[0239] Assume that at a certain moment t=30 days during the patient's recovery process, his pain level s(t) is recorded as moderate, that is, s(30)=0.5.

[0240] For a specific rehabilitation action k, set γ = 0.5, β k =1.0,δ k =0.8.

[0241] Substitute into the formula to calculate the selection probability:

[0242]

[0243] Due to e -15.4 is close to zero, and the selection probability of the action is very low, indicating that the action is not the most suitable choice under the current recovery state.

[0244] In this way, the MAB model can dynamically adjust the selection of rehabilitation movements and respond to the patient's changing condition in real time, ensuring the personalization and maximum effect of rehabilitation training.

[0245] Secondly, the multi-armed bandit (MAB) model is used to process the pain intensity and comfort feedback of patients when performing each rehabilitation action. Through dynamic update strategies, the system can optimize the selection of each rehabilitation action in real time to achieve the best rehabilitation effect and patient comfort. The following are the specific update strategies:

[0246]

[0247] In this formula:

[0248] U k (t) is the utility evaluation of action k at time t.

[0249] r i is the reward feedback at time i, representing the effect of the action at that moment, such as reducing pain or improving comfort.

[0250] δ ki is an indicator function indicating whether action k is selected at time i.

[0251] θ k is the learning rate parameter of action k, which is used to adjust the balance between new feedback and historical utility. Usually, θ k The value range can be from 0.1 to 0.9, depending on the stability and importance of the action.

[0252] ν k is an exploration parameter used to increase the tendency to explore new actions to avoid premature convergence to the local optimum. Usually, ν k The value range can be 0.01 to 0.5.

[0253] Suppose a burn patient is to monitor the effects of different actions on pain during rehabilitation training. Consider action k, which the patient performed on 6 days out of the past t = 10 days, and the corresponding pain intensity reduction (reward feedback) r i They are 0.8, 0.7, 0.85, 0.75, 0.9 and 0.65 respectively.

[0254] Choosing the parameter θ k = 0.7 and ν k = 0.1, the update utility of the action can be calculated:

[0255] 1. Calculate the historical average reward:

[0256]

[0257] 2. Set U k (9) is the utility evaluation value of the previous cycle, which is set to 0.5.

[0258] 3. Calculate U k (10):

[0259]

[0260] This updated utility value U k (10) = 0.4371 indicates that over time, the effectiveness of the action in alleviating pain has been evaluated and adapted in real time.

[0261] Finally, the algorithm is dynamically adjusted to adjust the rehabilitation training program according to the real-time data of the patient's rehabilitation status. The time changes of the rehabilitation status score are taken into account to ensure that the rehabilitation program continues to adapt to the patient's current needs.

[0262] formula:

[0263]

[0264] in:

[0265] s(u) represents the recovery status score at time u, including indicators such as pain level, skin elasticity, and motor ability.

[0266] λ and ρ are adjustment parameters used to control the sensitivity and reaction speed of rehabilitation actions to changes in patient status. Generally, λ can be set between 0.1 and 1.0, while ρ is usually set between 0.01 and 0.5, depending on the requirements for the time decay speed and the impact on the rehabilitation status.

[0267] Consider a patient in a rehabilitation period, and monitor the changes in his rehabilitation status score s(u) from the beginning of rehabilitation to the current moment (set as t = 30 days). The following is a change in the rehabilitation status score, which is simplified to be recorded every 5 days, and the data are: 0.2, 0.4, 0.6, 0.7, 0.85, 0.9.

[0268] Choosing λ = 0.5 and ρ = 0.1, Δr (30) can be calculated by numerical integration method:

[0269] Calculate each element in the integral, for example, when u = 25 (the closest recorded point):

[0270] e -0.1(30-25) tanh(0.9)≈e -0.5 0.716≈0.606 0.716=0.434

[0271] Similarly, similar calculations are performed for all points from u=0 to u=30, and all values ​​are integrated.

[0272] Estimate the total value of this integral using a numerical integration method such as the trapezoidal rule or Simpson's rule.

[0273] Assuming the overall integral calculation result is 4.5, then:

[0274] Δr(30)=0.5×4.5=2.25

[0275] This result of 2.25 means that in the past 30 days, based on the change in the rehabilitation status score, the rehabilitation training plan should be adjusted by a unit value of 2.25 to adapt to the patient's current rehabilitation needs. This means increasing the intensity of rehabilitation training, adjusting the frequency of rehabilitation training, or introducing new types of rehabilitation training to ensure that the patient's rehabilitation training is in line with his recovery progress and can adapt to changes in his comfort level.

[0276] Embodiment 5:

[0277] In this embodiment, the rehabilitation training plan is further optimized and personalized through the introduction of deep Q network (DQN). DQN can dynamically adjust the intensity and type of rehabilitation training, adapting according to real-time data obtained from skin elasticity monitors and patient self-reported feedback. This approach aims to improve rehabilitation efficiency and reduce patient discomfort.

[0278] The technical implementation steps of the Deep Q Network DQN model include:

[0279] 1. Data collection and input preparation: First, the system collects data from the skin elasticity monitor, which provides quantitative indicators of the patient's skin recovery status. At the same time, the system also collects the patient's self-reported pain intensity or comfort evaluation, which provides direct information about the patient's feelings.

[0280] 2. Model training and optimization: Through the DQN model, the following specific functions are used to process this data and perform training:

[0281]

[0282] s t Represents the recovery status data at time t (such as skin elasticity index, pain intensity, etc.).

[0283] a t Indicates potential training actions, which may include different types or intensities of physical therapy activities.

[0284] θ and φ are tuning parameters that are optimized during the training process to ensure that the model outputs training recommendations that are best suited to the patient's current state. Parameter θ is usually in the range of 0.01 to 0.1, while φ is usually in the range of 0.1 to 1.0, and the specific values ​​are adjusted based on the training data and the expected model sensitivity.

[0285] 3. Selection and adjustment of training actions: The output of the DQN model is used to dynamically adjust the training plan and select the training actions that best suit the current rehabilitation stage and patient comfort. The model not only considers the current rehabilitation data, but also attempts to predict the impact of different actions on the future recovery process.

[0286] Assume that a patient has a low skin elasticity index (st =0.2), and also reported the moderate pain intensity (a t =0.5). Calculate using the above parameters and formula:

[0287]

[0288]

[0289] This calculation result can be used to evaluate the suitability of a specific training action for the patient's current rehabilitation status. If this value is relatively high, it indicates that the selected action is beneficial to the patient's rehabilitation process; if it is low, the type or intensity of the action needs to be adjusted. Through continuous iteration and optimization, the DQN model ensures that the rehabilitation training plan always meets the patient's latest rehabilitation status, thereby improving rehabilitation efficiency and patient satisfaction.

[0290] Then, precise adjustments are made using sensors and self-reported data collected in real time. Through the integral calculation formula, the system can take into account past and current rehabilitation data and dynamically adjust rehabilitation movements to optimize rehabilitation effects.

[0291] Integral calculation formula:

[0292]

[0293] in:

[0294] α and β are coefficients that adjust the influence of time decay, which are used to control the weight and decay rate of each term in the integral, respectively. The typical range of values ​​is between 0.1 and 1.0 for α and between 0.01 and 0.5 for β, depending on the sensitivity of the rehabilitation activity and the degree of reliance on historical data.

[0295] s τ and a τ Represent the rehabilitation status indicators and potential training actions at time τ, respectively. These indicators include pain level, skin elasticity, activity limitation, etc.

[0296] f(s τ ,a τ ) is an evaluation function, which calculates the value according to the rehabilitation status and the effect of the training action, such as the aforementioned rehabilitation training optimization function.

[0297] The patient's rehabilitation training data is set as follows: In the past 30 days, the data of the rehabilitation state s and training action a are recorded every day. Select α = 0.5 and β = 0.1, and use the previously defined function f. Set t = 30 days, and for each day's data s τ and a τ Apply the value calculated by the f function and use a numerical integration method such as the trapezoidal rule to estimate the integral:

[0298]

[0299] If f(s τ ,a τ ) is calculated as follows:

[0300] At τ = 1 day, f(s1,a1) = 0.8

[0301] At τ = 2 days, f(s2,a2) = 0.7

[0302] And so on, until τ = 30 days.

[0303] Through numerical integration, we can accumulate f(s) from the 1st day to the 30th day. τ ,a τ ) value multiplied by the corresponding attenuation factor 0.5e -0.1(30-τ) This process will generate a cumulative utility value R(30), which reflects the overall effect of the rehabilitation training program within 30 days and provides a basis for further training adjustments.

[0304] Embodiment 6:

[0305] In this embodiment, the introduction of natural language processing (NLP) technology aims to improve the quality of interaction with patients through speech recognition and generation, making rehabilitation guidance more personalized and interactive. This technology uses complex mathematical models to analyze and generate speech feedback, which is specifically achieved through the following formula:

[0306]

[0307] in,

[0308] s(t) represents the speech feature variables extracted from the speech recognition system. These variables change over time and include pitch, volume, speaking speed, etc., which are key factors in determining speech content and emotion.

[0309] r(t) represents the emotional state variable reflected by the patient through speech input, such as pain, discomfort or satisfaction, which is quantified by characteristics such as the tone and intensity of the speech.

[0310] Parameter settings:

[0311] α, β, and κ are the adjustment parameters of the model, which are used to balance the influence of speech features and emotional state variables. For example, α and β control the influence strength of the corresponding variables, while κ adjusts the combined influence of the two.

[0312] Typical values ​​range from 0.1 to 1.0 for α and β, and from 0.5 to 2.0 for κ, depending on experimental data and desired effects.

[0313] The integral operation of this formula is essentially a comprehensive analysis of the voice and emotion data at all time points, and the accuracy and timeliness of the feedback are ensured by adjusting the parameters.

[0314] In practical applications, the calculation of this formula can be implemented by discretizing time and using numerical integration methods. For example, if data is collected every 0.1 seconds, the function can be applied to the data in each time period and the overall integral can be calculated to obtain the final training guidance feedback.

[0315] It is assumed that during rehabilitation training, the patient gave feedback on pain through voice (such as frequent mention of "pain" and tense tone of voice). The system captured this information through voice features s(t) and determined the pain level through sentiment analysis r(t). α=0.7, β=0.6, and κ=1.2 were selected, and then the data was processed and integrated to generate appropriate training guidance feedback, such as adjusting the training intensity or suggesting rest.

[0316] Secondly, the integration of computer vision technology provides a means to analyze the correctness of the patient's training movements in real time. By using mathematical functions, the system is able to evaluate the accuracy of the movements and make real-time corrections to the deviations. This approach not only improves the effectiveness of rehabilitation training, but also ensures patient safety through precise monitoring.

[0317] Function formula:

[0318]

[0319] in:

[0320] d(t) represents the movement deviation at time t, which is obtained from the computer vision system by analyzing the difference between the movement performed by the patient and the predetermined standard movement.

[0321] ω is a periodic adjustment factor, which is used to adjust the period of the sine function and affects the sensitivity of motion analysis. The typical value range can be 0.5 to 5.0.

[0322] γ is a time decay parameter that controls the time decay speed of the function and affects the length and sensitivity of the analysis window. The common value range is 0.01 to 0.1, which is adjusted according to the complexity of the action and the need for rapid changes.

[0323] Assume that a patient is undergoing rehabilitation training and performs a series of movements, including stretching and bending. The system captures the movements in real time through a camera and calculates the movement deviation d(t) for each frame. Set the total training time T = 30 seconds, ω = 2.0, γ = 0.05 for analysis:

[0324] Through computer vision technology, the action at each time point t is captured and the deviation d(t) from the standard action is calculated.

[0325] Substitute the motion deviation at each time point into the formula. For example, at t = 1 second, the calculated value of d(1) is 0.3 (example value):

[0326]

[0327] The data for the entire time period T is integrated to estimate the accuracy of the overall training action. The integral from t = 0 to t = 30 is calculated using a numerical integration method (such as the trapezoidal rule or Simpson's rule).

[0328] The score provides a quantitative way to assess the patient's overall performance quality. If the score is below a certain threshold, the system can suggest adjusting the training action or intervening with a physical therapist for guidance.

[0329] Finally, the data from vision and language processing are dynamically integrated and the rehabilitation training plan is adjusted in real time. This is achieved by adjusting the algorithm, which takes into account the interactive effects of vision and language feedback to optimize the training effect and provide accurate real-time feedback.

[0330] Adjust the algorithm formula:

[0331]

[0332] in:

[0333] x and y represent quantitative data obtained from computer vision and natural language processing techniques, respectively. For example, x is the accuracy score of the patient's movements, while y is the patient's satisfaction or comfort expressed through voice feedback.

[0334] It aims to express the combined effect of visual and language data in the form of composite functions. represents the magnitude of the combined effect of visual and language data, and the denominator log(x 2 +y 2 +1) plays a role in smoothing and regulating the impact of synthetic data.

[0335] Assume that in a rehabilitation training, the patient's movement accuracy score obtained by the visual monitoring system is x = 0.8, and the patient's comfort score analyzed by the speech recognition system is y = 0.6. The value of the adjustment algorithm can be calculated to determine whether the training plan needs to be adjusted.

[0336] First, the magnitude and regulation function of the composite influence are calculated:

[0337]

[0338] Then, apply the adjustment algorithm:

[0339]

[0340] represents the derivative of x, so if the composite amplitude of visual and language data is relatively stable with the modulation amplitude, the derivative result will be 0, suggesting that no further adjustment of the rehabilitation program is needed under the current conditions. Such a calculation result indicates that the rehabilitation activities have been well matched to the patient's current needs.

[0341] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. Intelligent guidance system for rehabilitation training of burn patients based on deep learning, characterized by The technical implementation steps include the following: first, multimodal data integration technology is used to collect video records of patient rehabilitation training, physiological signals and skin elasticity measurement data, the data is cleaned, standardized and classified, and data enhancement technology is applied to control and improve the generalization ability of the model; Next, we developed deep learning classification models, including convolutional neural networks (CNNs), to analyze image data of skin recovery, and applied recurrent neural networks (RNNs) or long short-term memory networks (LSTMs) to predict the patient's future recovery trend. Then, the generative adversarial network (GAN) is used to generate personalized training actions or treatment plans for different rehabilitation stages, and the rehabilitation training plan is optimized through reinforcement learning methods, dynamically adjusting the training intensity and type based on the patient's real-time feedback; Finally, natural language processing (NLP) technology is integrated to achieve speech recognition and generation, and natural language is used to communicate with patients and guide training. Computer vision technology is used to analyze patients' training movements in real time, and a system is built to guide the correctness of movements and provide instant feedback. The multimodal data integration technology implementation steps are: S1. First, a burn-specific feature fusion network is used, using a multi-dimensional weighted fusion formula: Where x represents the image feature vector of the burn area, y represents the physiological signal vector of the pain response, z is the change in skin elasticity, γ is the nonlinear index that adjusts the influence of data features, and α(x) is the weight function that is dynamically adjusted according to the input features; S2. Then use the conditional generative adversarial network to enhance the burn area data and apply the random process formula: Where x represents the original rehabilitation data, and the data generated by Fourier series representation retains the rehabilitation characteristics while introducing controlled randomness to simulate the variation at different recovery stages; S3. Then the data is classified and labeled through self-supervised learning method, using calculus formula: Where x is the burn recovery data to be classified, and the higher-level data features are extracted by combining differential and integral operations to control high-precision automatic classification; S4. Finally, a joint rehabilitation status assessment model was developed using high-order polynomials and partial derivative formulas: Where x and y represent the feature vectors of visual data and physiological data, respectively. The divergence operation of the vector field is used to fuse the visual and physiological data to control the comprehensiveness of the rehabilitation guidance program. The deep learning classification model specifically includes: First, a multi-scale convolutional neural network (CNN) was developed, which was designed to analyze the image data of burn patients’ skin, using a nonlinear fusion formula: Where x represents the pixel value or feature of the skin restoration image; α is the width of the adjustment function, which affects the local or global analysis of the feature, making the network adaptable to the analysis of burn images of different scales; β k is a frequency parameter that adjusts the response to different frequencies of the image and helps the network capture recovery patterns from subtle textures to wide areas; At the same time, the bidirectional long short-term memory network Bi-LSTM is combined to predict the patient's future recovery trend, using the time series analysis formula: Among them, x(t) represents the quantitative indicator of the rehabilitation status at time t, including pain level, inflammatory response or skin elasticity; σ is the attenuation parameter that is dynamically adjusted according to the rehabilitation progress, and predicts the future rehabilitation path based on the patient's current and past rehabilitation data; Finally, through the feature fusion strategy, use the formula: Among them, x and y are the features of CNN and Bi-LSTM outputs, respectively, reflecting the visual recovery characteristics of the skin and the temporal dynamics of the recovery state; γ is used as a regulation parameter to balance the influence of the two data sources; The generation of the personalized training action or treatment plan includes: S1. First, the conditional generative adversarial network cGAN is used to generate personalized training actions or treatment plans for different rehabilitation stages based on the type, depth and recovery stage of the burn. S2, then the multi-armed bandit MAB model is used to select rehabilitation actions based on the patient's real-time feedback, including pain intensity and comfort assessment, to control the rehabilitation training program in response to the patient's changes; S3. Finally, the rehabilitation training plan is further adjusted and optimized through the deep Q network DQN, using the data obtained from the skin elasticity monitor sensor and the patient's self-reported feedback to dynamically adjust the training intensity and type.

2. The deep learning-based intelligent guidance system for rehabilitation training of burn patients according to claim 1 is characterized in that The conditional generative adversarial network cGAN is used to input the burn type t, depth d, and recovery stage s, combined with random noise z, through the formula: Specifically, z represents the random noise introduced from the prior distribution; t represents the numerical representation of the burn type, including chemical burns, thermal burns, etc., which is used to generate rehabilitation plans for specific types of burns; d represents the numerical representation of the burn depth, including first-degree, second-degree, and third-degree burns, which affects the intensity and sensitivity of the generated actions; s represents the numerical representation of the recovery stage, including the acute stage, the healing stage, and the reconstruction stage, and the treatment plan is adjusted according to the different rehabilitation stages; σ represents the standard deviation of the sensitivity of the generated samples; ω represents the periodic adjustment factor, which is used to simulate the periodic adjustment during rehabilitation training; Then, based on the patient’s real-time feedback data x, the feedback adjustment formula is used: Where x represents the data extracted from the patient's real-time feedback, including the pain level or the quantitative index of rehabilitation progress; α and β are model adjustment parameters, which are combined through nonlinear and integral operations to control the actual effect of training actions on the patient's rehabilitation progress; tanh(αx ′ ) represents the feedback function, which is provided by the hyperbolic tangent function; Finally, a dynamic adjustment mechanism is adopted by adjusting the formula: Here, x represents the rehabilitation progress data obtained from physiological monitoring devices including skin elasticity monitors or pain rating systems.

3. The intelligent guidance system for rehabilitation training of burn patients based on deep learning according to claim 1 is characterized in that The multi-armed bandit MAB model construction includes: S1. First, define and classify different rehabilitation actions as arms of a multi-armed slot machine. Each arm represents a rehabilitation training action using the formula: Where t is time, γ, β k and δ k is an adjustment parameter specific to action k, s(t) represents the rehabilitation status indicator at time t, which is used to adjust the selection probability of each action; S2. Secondly, collect the patient's pain intensity and comfort feedback when performing each rehabilitation action, and adopt an update strategy through the multi-armed bandit model: where r i is the reward feedback at time i, δ ki is the indicator function, θ k and ν k is the learning rate and exploration parameter of action k, adjusting the weights of history and new feedback; S3, finally through dynamic adjustment algorithm: Where s(u) is the rehabilitation state score at time u, λ and ρ are parameters for adjusting the sensitivity of rehabilitation movements, and the integral represents the cumulative effect of the time-attenuated rehabilitation state score on the adjustment of training movements from the beginning to the current moment.

4. The intelligent guidance system for rehabilitation training of burn patients based on deep learning according to claim 1 is characterized in that The adjustment and optimization of the rehabilitation training plan covers the following technical implementation steps: First, through the deep Q network DQN model, use the function: For training and optimization, s t represents the data obtained from the skin elasticity monitor sensor and the patient's self-reported pain intensity or comfort rating, a t represents the potential training action, θ and φ are the parameters adjusted during the learning process; Then collect sensor and self-reported data in real time, and use the integral calculation formula: where α and β are coefficients that adjust for the effects of time decay.

5. The intelligent guidance system for rehabilitation training of burn patients based on deep learning according to claim 1 is characterized in that The natural language processing (NLP) technology uses the following functions: Realize speech recognition and generation, where s(t) represents the change of speech feature variables over time, r(t) represents the emotional state variables fed back by the patient, and α, β, and κ are model parameters used to process speech data and generate appropriate training guidance feedback; Secondly, combined with computer vision technology, through the function: Real-time analysis of the correctness of the patient's training movements, where d(t) represents the movement deviation under time ·, ω is the periodic adjustment factor, and γ is the time decay parameter, ensuring that the movement analysis accurately reflects the actual performance of the patient's movements; Finally, the output of vision and language processing is adjusted by the algorithm: Dynamic integration, where x and y represent quantitative data of visual and language feedback, respectively, to provide instant training adjustments and feedback.

Images