Deep learning-based personalized rehabilitation nursing scheme generation method and intelligent chip

Abstract

This invention relates to the field of rehabilitation data processing technology, and in particular to a method for generating personalized rehabilitation care plans based on deep learning and an intelligent chip. First, multimodal data including patient physiological signals, movement data, subjective feedback, rehabilitation therapist assessments, and environmental information is loaded. State modeling is then used to obtain a patient implicit state vector that integrates physiological, motor, and psychological data. Next, an initial personalized reward function is learned using inverse reinforcement learning. A meta-learner dynamically adjusts the reward function weights based on the patient's state and rehabilitation effect. The updated reward function and implicit state vector are input into a reinforcement learning policy network to optimize and obtain the optimal rehabilitation action, which is then sent to the rehabilitation execution device. Simultaneously, new data is collected for closed-loop iteration, achieving long-term adaptive rehabilitation plan generation.

Description

Technical Field

[0001] This invention relates to the field of rehabilitation data processing technology, and in particular to a method for generating personalized rehabilitation care plans based on deep learning and an intelligent chip. Background Technology

[0002] In the field of personalized rehabilitation program generation based on reinforcement learning, existing technologies typically pre-define fixed reward functions to guide strategy optimization. For example, CN120221132A discloses a multimodal data-driven rehabilitation nursing assistance system, whose reward function is composed of a weighted sum of functional recovery increments and psychological improvement increments. However, the weight coefficients α and β are preset constants and cannot be dynamically adjusted according to changes in the patient's state. CN121331347A proposes a Parkinson's disease treatment strategy generation system based on hierarchical and preference learning. It learns patient preferences and updates the reward function through periodic trajectory comparisons. However, its updates rely on offline interaction and lack immediate response to real-time physiological and psychological states, making it difficult to adapt to continuous changes in patient motivation and abilities during long-term rehabilitation. CN120600220A uses meta-learning to optimize the parameters of a large multimodal model, but the optimization object is the model weights rather than the reward function itself, failing to solve the problem of personalized adaptation of decision-making goals. All of the above methods, due to the static or lagging update of the reward function, result in a disconnect between the rehabilitation program and the patient's current needs, affecting rehabilitation effectiveness and compliance. To this end, the present invention provides a personalized dynamic reward function learning method for long-term rehabilitation. It initializes individual preferences through inverse reinforcement learning and uses a meta-learner to dynamically adjust the reward function weights according to the patient's real-time status and rehabilitation effect, thereby achieving online adaptive optimization of rehabilitation goals. Summary of the Invention

[0003] This invention addresses the problem in existing technologies where rehabilitation plans are out of sync with patients' current needs, affecting rehabilitation outcomes and compliance. It provides a method for generating personalized rehabilitation care plans based on deep learning and an intelligent chip to solve this problem.

[0004] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: In a first aspect, the present invention provides a method for generating personalized rehabilitation care plans based on deep learning, including: Load multimodal data of patients during the rehabilitation process, wherein the multimodal data includes physiological signal data, motion data, patient subjective feedback data, rehabilitation therapist assessment data, and environmental information data; Based on the multimodal data, patient state modeling is performed to obtain the patient's implicit state vector, wherein the patient's implicit state vector integrates the representations of physiological load state, motor function state and psychological motivation state. Based on the patient's implicit state vector and a preset set of sub-reward functions, an initial personalized reward function is learned to obtain an initial reward function. The initial reward function is a weighted combination of multiple sub-reward functions, and the set of sub-reward functions includes at least a safety reward function, an effort reward function, and a movement quality reward function. The patient's implicit state vector and recent rehabilitation effect indicators are input into a pre-trained meta-learner, and dynamic reward function adjustment is performed to obtain an updated reward function. The meta-learner is used to output the weight adjustment amount of the reward function according to the current patient state and rehabilitation effect. The patient's implicit state vector and the updated reward function are input into a reinforcement learning policy network to perform rehabilitation policy optimization and obtain the optimal rehabilitation action. The optimal rehabilitation action is sent to the rehabilitation execution device, and new multimodal data after the optimal rehabilitation action is executed is collected for the next round of rehabilitation strategy optimization.

[0005] Optionally, load multimodal data of the patient during the recovery process, including: The patient's wearable sensors collect electromyography (EMG) signals, heart rate signals, and electrodermal signals, which are then added to the physiological signal data. The joint angles, motion trajectories, and speeds collected by the built-in encoder of the rehabilitation execution device, as well as the magnitude of the assistive force output by the rehabilitation execution device, are added to the motion data. The results of the fatigue measurement scale, pain score and motivation level questionnaire filled out by the patient before and after training were obtained through the interactive interface and added to the patient's subjective feedback data. Obtain functional scores and movement quality scores from rehabilitation therapists' periodic assessments and add them to the aforementioned rehabilitation therapist assessment data; Obtain the operating status and training environment parameters of the rehabilitation execution equipment, and add them to the environmental information data.

[0006] Optionally, based on the multimodal data, patient state modeling is performed to obtain the patient's implicit state vector, including: The multimodal data is time-aligned and features are extracted to obtain multimodal temporal features; The multimodal temporal features are input into a temporal neural network to extract a low-dimensional implicit state vector of the patient. The temporal neural network is composed of a convolutional neural network layer and a long short-term memory network layer connected in series.

[0007] Optionally, based on the patient's implicit state vector and a preset set of sub-reward functions, initial personalized reward function learning is performed to obtain an initial reward function, which includes the following steps: Construct the sub-reward function set, where: The safety reward function is calculated by monitoring whether the physiological signal data exceeds a preset safety threshold range. When the monitored value exceeds the preset safety threshold range, a negative reward value is output. The absolute value of the negative reward value is proportional to the extent of exceeding the limit. The effort reward function is calculated based on the ratio of the magnitude of the patient's active force exertion in the motion data to the total torque output by the rehabilitation execution device, and the ratio is proportional to the positive reward value. The motion quality reward function is calculated based on the dynamic time warping distance between the real-time joint motion trajectory in the motion data and the preset standard rehabilitation motion trajectory. The dynamic time warping distance is inversely proportional to the positive reward value.

[0008] Specifically, based on the patient's implicit state vector and a preset set of sub-reward functions, initial personalized reward function learning is performed to obtain an initial reward function, including: Collect the patient's preference selection data in the early stage of rehabilitation, wherein the preference selection data includes the patient's preferred option selected from multiple candidate training options; Collect initial guidance and demonstration data from rehabilitation therapists for the patients; An inverse reinforcement learning algorithm is used, with the preference selection data and the initial guidance demonstration data as supervision, and the goal of maximizing the log likelihood of the observed behavior, to solve for the initial weight vectors corresponding to the safety reward function, the effort level reward function and the action quality reward function; The initial reward function is obtained by weighting and combining the sub-reward function set according to the initial weight vector.

[0009] Optionally, the patient's implicit state vector and recent rehabilitation effect indicators are input into a pre-trained meta-learner, and dynamic reward function adjustment is performed to obtain an updated reward function, including: Obtain the recent rehabilitation effect indicators, wherein the rehabilitation effect indicators include the change in functional scores and training compliance rate within a preset time window; The patient's implicit state vector and the rehabilitation effect index are input into the meta-learner, and the meta-learner outputs the adjustment amount of the current reward function weight; The adjustment amount is applied to the current reward function weights to generate an updated reward function, and the updated weights are normalized to ensure that the updated weights are non-negative and sum to 1. The meta-learner is a pre-trained neural network, and the training data for the meta-learner comes from the complete rehabilitation trajectories of multiple patients in the historical rehabilitation database. The training objective is to maximize the cumulative rehabilitation effect of patients in subsequent rehabilitation stages.

[0010] Optional, also includes: When a preset change in the patient's condition is detected, the dynamic reward function adjustment step is triggered.

[0011] Optional, also includes: After performing the dynamic reward function adjustment step, the updated reward function, the patient's implicit state vector, and the corresponding rehabilitation effect indicators are sent to the manual review terminal to receive confirmation or correction instructions from the rehabilitation therapist based on the review results. If the correction instruction is received, the updated reward function is corrected according to the weight adjustment value in the correction instruction to obtain a manually corrected reward function, which is then used for subsequent rehabilitation strategy optimization.

[0012] Optionally, the patient's implicit state vector and the updated reward function are input into a reinforcement learning policy network to perform rehabilitation policy optimization and obtain the optimal rehabilitation action, including: The patient's implicit state vector is used as the environment state for reinforcement learning, and the updated reward function is used as the current optimization objective. The reinforcement learning policy network performs policy reasoning and outputs a rehabilitation action that maximizes the expected cumulative reward. The rehabilitation action includes training intensity, assist force magnitude, and training mode.

[0013] In a second aspect, the present invention provides a smart chip, comprising: A multimodal data loading module is used to load multimodal data of patients during the rehabilitation process, wherein the multimodal data includes physiological signal data, motion data, patient subjective feedback data, rehabilitation therapist assessment data, and environmental information data; The patient state modeling module is used to perform patient state modeling based on the multimodal data to obtain a patient implicit state vector, wherein the patient implicit state vector integrates the representations of physiological load state, motor function state and psychological motivation state. The reward function acquisition module is used to perform initial personalized reward function learning based on the patient's implicit state vector and a preset set of sub-reward functions to obtain an initial reward function. The initial reward function is a weighted combination of multiple sub-reward functions, and the set of sub-reward functions includes at least a safety reward function, an effort reward function, and a movement quality reward function. The reward function update module is used to input the patient's implicit state vector and recent rehabilitation effect indicators into a pre-trained meta-learner, perform dynamic reward function adjustment, and obtain an updated reward function. The meta-learner is used to output the weight adjustment amount of the reward function according to the current patient state and rehabilitation effect. The rehabilitation strategy optimization module is used to input the patient's implicit state vector and the updated reward function into the reinforcement learning policy network, perform rehabilitation strategy optimization, and obtain the optimal rehabilitation action; The strategy loop optimization module is used to send the optimal rehabilitation action to the rehabilitation execution device and collect new multimodal data after the optimal rehabilitation action is executed, so as to be used for the next round of rehabilitation strategy optimization.

[0014] By implementing this invention, it is possible to load multimodal data of patients during the rehabilitation process. The multimodal data includes physiological signal data, motion data, patient subjective feedback data, rehabilitation therapist assessment data, and environmental information data. This comprehensively covers objective physiology, motor performance and subjective feelings, professional assessment, and environmental influences, avoiding information bias caused by single data points. It allows rehabilitation analysis to be based on the real and complete patient state, laying a solid data foundation for accurate and personalized modeling and ensuring the reliability of subsequent calculations and decisions.

[0015] By implementing this invention, patient state modeling can be performed based on the multimodal data to obtain a patient implicit state vector, wherein the patient implicit state vector integrates representations of physiological load state, motor function state, and psychological motivation state; highly redundant and multidimensional raw data are transformed into a compact and unified state representation, efficiently condensing core rehabilitation state information, reducing subsequent computational complexity, and fully preserving key state features, enabling the system to accurately perceive the patient's comprehensive state and providing precise input for reward function and strategy optimization.

[0016] By implementing this invention, it is possible to perform initial personalized reward function learning based on the patient's implicit state vector and a preset set of sub-reward functions to obtain an initial reward function. The initial reward function is a weighted combination of multiple sub-reward functions, and the set of sub-reward functions includes at least a safety reward function, an effort reward function, and a movement quality reward function. Based on the patient's initial preferences and the therapist's demonstration, initial reward rules tailored to individual needs are quickly generated, balancing safety, active participation, and movement standardization. This ensures that rehabilitation optimization goals are personalized from the outset, avoiding the poor adaptability problems caused by generic reward functions.

[0017] By implementing this invention, the patient's implicit state vector and recent rehabilitation effect indicators can be input into a pre-trained meta-learner to perform dynamic reward function adjustment and obtain an updated reward function. The meta-learner is used to output the weight adjustment amount of the reward function based on the current patient state and rehabilitation effect. This breaks through the limitations of traditional static reward functions by dynamically adjusting and optimizing the target weights based on the patient's real-time state and rehabilitation effect, adapting in real-time to the continuous changes in motivation, ability, and recovery status during the rehabilitation process. This ensures that the reward function always matches the current rehabilitation needs, improving long-term rehabilitation adaptability.

[0018] By implementing this invention, the patient's implicit state vector and the updated reward function can be input into a reinforcement learning policy network to perform rehabilitation strategy optimization and obtain the optimal rehabilitation action. Guided by dynamic personalized rewards, the system can accurately generate training intensity, assistance force, training mode, and other actions that maximize the expected rehabilitation effect, ensuring that the rehabilitation plan always iterates around the current optimal goal, achieving a high degree of matching between the strategy and the patient's state and rehabilitation goals.

[0019] By implementing this invention, the optimal rehabilitation action can be sent to the rehabilitation execution device, and new multimodal data after the optimal rehabilitation action is executed can be collected for the next round of rehabilitation strategy optimization; a closed-loop iterative mechanism of decision-making-execution-feedback-optimization can be formed, so that the rehabilitation plan can be continuously updated with the results of each round of execution, constantly correcting the optimization direction, achieving long-term dynamic self-adaptation, and ensuring that the rehabilitation strategy continuously fits the patient's real-time status.

[0020] In summary, by implementing this invention, online adaptive generation of rehabilitation nursing plans can be achieved, solving the problems of lagging updates and insufficient personalization in traditional static reward functions. This allows rehabilitation plans to adapt to changes in the patient's physiological, motor, and psychological states in real time, improving rehabilitation outcomes and patient compliance, and supporting long-term personalized rehabilitation nursing. Attached Figure Description

[0021] Figure 1 A flowchart illustrating the method for generating personalized rehabilitation and nursing care plans based on deep learning provided by this invention; Figure 2 This is a schematic diagram of the structure of the smart chip provided by the present invention.

[0022] In the attached diagram, the components represented by each number are as follows: Multimodal data loading module 11, patient status modeling module 12, reward function acquisition module 13, reward function update module 14, rehabilitation strategy optimization module 15, and strategy loop optimization module 16. Detailed Implementation

[0023] Example 1, as Figure 1 As shown, this embodiment of the invention provides a method for generating personalized rehabilitation care plans based on deep learning, including: S100: Load the patient's multimodal data during the rehabilitation process, wherein the multimodal data includes physiological signal data, motion data, patient subjective feedback data, rehabilitation therapist assessment data, and environmental information data; S200: Based on the multimodal data, perform patient state modeling to obtain a patient implicit state vector, wherein the patient implicit state vector integrates representations of physiological load state, motor function state and psychological motivation state. S300: Based on the patient's implicit state vector and a preset set of sub-reward functions, perform initial personalized reward function learning to obtain an initial reward function, wherein the initial reward function is a weighted combination of multiple sub-reward functions, and the set of sub-reward functions includes at least a safety reward function, an effort reward function, and a movement quality reward function; S400: Input the patient's implicit state vector and recent rehabilitation effect indicators into a pre-trained meta-learner, perform dynamic reward function adjustment, and obtain an updated reward function, wherein the meta-learner is used to output the weight adjustment amount of the reward function according to the current patient state and rehabilitation effect; S500: Input the patient's implicit state vector and the updated reward function into the reinforcement learning policy network, perform rehabilitation policy optimization, and obtain the optimal rehabilitation action; S600: Send the optimal rehabilitation action to the rehabilitation execution device and collect new multimodal data after executing the optimal rehabilitation action for use in the next round of rehabilitation strategy optimization.

[0024] In step S100 of this application embodiment, loading the patient's multimodal data during the rehabilitation process includes: The patient's wearable sensors collect electromyography (EMG) signals, heart rate signals, and electrodermal signals, which are then added to the physiological signal data. The joint angles, motion trajectories, and speeds collected by the built-in encoder of the rehabilitation execution device, as well as the magnitude of the assistive force output by the rehabilitation execution device, are added to the motion data. The results of the fatigue measurement scale, pain score and motivation level questionnaire filled out by the patient before and after training were obtained through the interactive interface and added to the patient's subjective feedback data. Obtain functional scores and movement quality scores from rehabilitation therapists' periodic assessments and add them to the aforementioned rehabilitation therapist assessment data; Obtain the operating status and training environment parameters of the rehabilitation execution equipment, and add them to the environmental information data.

[0025] In step S100 of this application embodiment, the purpose of the above steps is to comprehensively collect all dimensions of patient rehabilitation information, so as to provide complete and realistic input basis for subsequent patient status modeling, reward function learning and rehabilitation strategy optimization, and ensure that the rehabilitation plan fits the actual situation of the patient.

[0026] To achieve the above objectives, it is first necessary to obtain the electromyography signals, heart rate signals, and skin conductance signals collected by the patient's wearable sensors and add them to the physiological signal data; This involves collecting electromyography (EMG), heart rate, and ductal skin signals using wearable sensors and incorporating them into the physiological signal data. For example, if a wearable sensor detects an EMG signal intensity of 35 μV, a heart rate of 85 beats per minute, and a ductal skin signal of 12 μS, these three sets of data are added to the physiological signal data.

[0027] Next, it is necessary to obtain the joint angles, motion trajectories, and speeds collected by the built-in encoder of the rehabilitation execution device, as well as the magnitude of the assistive force output by the rehabilitation execution device, and add them to the motion data; This involves using the built-in encoder of the rehabilitation device to collect joint angles, movement trajectories, and speeds, as well as the magnitude of the auxiliary force output by the device, and incorporating these data into the motion data. For example, if the built-in encoder of the rehabilitation device detects that the patient's knee joint angle is 90 degrees, the movement trajectory is a standard arc, the speed is 15 cm / s, and the device outputs an auxiliary force of 20 N, these data are added to the motion data.

[0028] The built-in encoder in the rehabilitation execution device is a position and motion sensing component installed inside the device. It is used to collect motion data such as joint angles, movement trajectories, and movement speeds during the patient's rehabilitation movements in real time. It can also work with the device system to obtain the magnitude of the output assist force. This data is added to the motion data for subsequent patient status modeling, movement quality assessment, and rehabilitation strategy optimization.

[0029] Then, the fatigue measurement scale, pain score and motivation level questionnaire results filled out by the patient through the interactive interface before and after training are obtained and added to the patient's subjective feedback data. This involves collecting the results of fatigue measurement scales, pain scores, and motivation level questionnaires filled out by patients before and after training via an interactive interface, and incorporating them into the patient's subjective feedback data. For example, if a patient scores 2 points on the fatigue measurement scale, 1 point on the pain score, and 9 points on the motivation level questionnaire before training, and scores 5 points on the fatigue measurement scale, 2 points on the pain score, and 8 points on the motivation level questionnaire after training, these results are added to the patient's subjective feedback data.

[0030] Furthermore, functional scores and movement quality scores obtained from regular assessments by rehabilitation therapists are added to the rehabilitation therapist assessment data; This involves collecting the functional scores and movement quality scores regularly assessed by rehabilitation therapists and incorporating them into the therapist's evaluation data. For example, if a rehabilitation therapist regularly assesses a patient's functional score as 75 and their movement quality score as 80, these two scores will be added to the therapist's evaluation data.

[0031] The functional score is a quantitative score regularly assessed by rehabilitation therapists based on rehabilitation function indicators such as the patient's limb mobility and self-care ability. It reflects the patient's overall rehabilitation function level and is an objective indicator for measuring rehabilitation effectiveness. The movement quality score is a quantitative score assessed by rehabilitation therapists based on dimensions such as the standardization, completion, and stability of the patient's rehabilitation movements. It is used to evaluate the effectiveness of rehabilitation exercise execution.

[0032] Finally, the operating status and training environment parameters of the rehabilitation execution equipment are obtained and added to the environmental information data.

[0033] This involves collecting the operating status of the rehabilitation equipment and the parameters of the training environment, and then incorporating them into the environmental information data. For example, if the rehabilitation equipment is operating normally and the training environment parameters are a temperature of 25°C and a humidity of 50%, this information is added to the environmental information data.

[0034] In step S200 of this embodiment, based on the multimodal data, patient state modeling is performed to obtain the patient's implicit state vector, including: The multimodal data is time-aligned and features are extracted to obtain multimodal temporal features; The multimodal temporal features are input into a temporal neural network to extract a low-dimensional implicit state vector of the patient. The temporal neural network is composed of a convolutional neural network layer and a long short-term memory network layer connected in series.

[0035] In step S200 of this application embodiment, the purpose of the above steps is to fuse and reduce the dimensionality of multimodal data to generate a patient implicit state vector that can uniformly reflect the patient's physiological load state, motor function state, and psychological motivation state, so as to provide standardized state input for subsequent reward function learning and rehabilitation strategy optimization.

[0036] To achieve the above objectives, the multimodal data first needs to be time-aligned and feature-extracted to obtain multimodal temporal features. This involves calibrating and matching physiological signal data, motion data, patient subjective feedback data, rehabilitation therapist assessment data, and environmental information data along a unified timeline, and then extracting effective features from each modality. For example, heart rate signals, joint angles, and fatigue scale scores are aligned with timestamps once per second to extract heart rate change features, joint angle change features, and fatigue trend features, forming multimodal temporal features.

[0037] The extraction of effective features for each modality involves extracting key information reflecting the patient's state from physiological signal data, motion data, patient subjective feedback data, therapist assessment data, and environmental information data. For example, physiological signals include mean heart rate, electromyography amplitude, and skin conductance fluctuations. Motion data includes joint angle range, mean movement speed, and assist force ratio. Subjective feedback includes fatigue level, pain level, and motivation level score. Therapist assessment includes functional score and movement quality score. Environmental information includes equipment operating status and ambient temperature and humidity values. These features are used for subsequent time alignment and state modeling.

[0038] Then, the multimodal temporal features are input into a temporal neural network to extract a low-dimensional patient implicit state vector, wherein the temporal neural network is composed of a convolutional neural network layer and a long short-term memory network layer connected in series.

[0039] Multimodal temporal features are input into a temporal neural network to extract low-dimensional patient implicit state vectors. The temporal neural network consists of convolutional neural network layers and long short-term memory network layers connected in series. The convolutional neural network layers extract local spatial features, while the long short-term memory network layers capture temporal dependencies.

[0040] For example, in terms of model structure, the temporal neural network is constructed using a concatenated structure of convolutional neural network layers and long short-term memory network layers, specifically including convolutional neural network layers, long short-term memory network layers, and fully connected layers. The convolutional neural network layers have 32 convolutional kernels, a kernel size of 3, a stride of 1, and uniform padding. The ReLU activation function is used to extract local spatial features. The long short-term memory network layers have 64 hidden units and an output dimension of 64, used to capture temporal dependencies. The fully connected layers have a dimension of 64, use the Tanh activation function, and output a low-dimensional patient implicit state vector.

[0041] In terms of parameter settings, the learning rate of the temporal neural network was set to 0.001, the batch size was set to 16, the Adam optimizer was used, and the mean squared error was used as the loss function.

[0042] For model training, the training samples for the temporal neural network were derived from the complete rehabilitation trajectories of multiple patients in a historical rehabilitation database. The training sample size consisted of 10,000 sets of multimodal temporal feature samples. The training rounds numbered 100. The convergence criterion was that the loss value decreased by less than 0.001 for 20 consecutive training rounds.

[0043] For example, the aforementioned multimodal temporal features are first input into a convolutional neural network layer to extract local features, and then input into a long short-term memory network layer to learn temporal patterns. Finally, a 64-dimensional implicit state vector of the patient containing physiological, motor, and psychological information is output.

[0044] In step S300 of this embodiment, based on the patient's implicit state vector and a preset sub-reward function set, initial personalized reward function learning is performed to obtain the initial reward function. This includes the following prior steps: Construct the sub-reward function set, wherein; The safety reward function is calculated by monitoring whether the physiological signal data exceeds a preset safety threshold range. When the monitored value exceeds the preset safety threshold range, a negative reward value is output. The absolute value of the negative reward value is proportional to the extent of exceeding the limit. The effort reward function is calculated based on the ratio of the magnitude of the patient's active force exertion in the motion data to the total torque output by the rehabilitation execution device, and the ratio is proportional to the positive reward value. The motion quality reward function is calculated based on the dynamic time warping distance between the real-time joint motion trajectory in the motion data and the preset standard rehabilitation motion trajectory. The dynamic time warping distance is inversely proportional to the positive reward value.

[0045] In step S300 of this application embodiment, the purpose of the above steps is to construct a sub-reward function set that includes a safety reward function, an effort reward function, and a movement quality reward function, so as to provide a basic calculation unit for subsequent initial personalized reward function learning and ensure that the initial reward function conforms to the three core standards of safety, effort, and quality of rehabilitation training.

[0046] To achieve the above objectives, it is first necessary to construct the aforementioned set of sub-reward functions, wherein; The safety reward function is calculated by monitoring whether the physiological signal data exceeds a preset safety threshold range. When the monitored value exceeds the preset safety threshold range, a negative reward value is output. The absolute value of the negative reward value is proportional to the extent of exceeding the limit. This system monitors physiological signal data to see if it exceeds a preset safety threshold range. Based on the monitoring results, a reward value is calculated; if it exceeds the threshold, a negative reward value is output. The absolute value of the negative reward value is proportional to the extent of the exceedance. For example, if the preset heart rate signal safety threshold range is 60 to 100 beats per minute, and the patient's real-time heart rate is 120 beats per minute, exceeding the upper limit by 20 beats per minute, a negative reward value is output. The greater the exceedance, the larger the absolute value of the negative reward value.

[0047] The safety threshold ranges are set based on clinical rehabilitation guidelines and the patient's baseline physiological indicators. First, electromyography (EMG), heart rate, and skin conductance (SCAD) signals are collected from the patient at rest during the initial rehabilitation phase to obtain the individual's baseline physiological range. Then, the rehabilitation therapist, in conjunction with general rehabilitation safety standards, determines the upper and lower limits for each physiological signal.

[0048] The effort reward function is calculated based on the ratio of the magnitude of the patient's active force exertion in the motion data to the total torque output by the rehabilitation execution device, and the ratio is proportional to the positive reward value. This involves taking the patient's active force from the motion data and calculating the ratio between the active force and the total torque output by the rehabilitation device. The positive reward value is then calculated based on this ratio, which is directly proportional to the reward value. For example, if the patient's active force is 30 Newtons and the total torque output by the rehabilitation device is 50 Newtons, the ratio is 0.6. The higher this ratio, the greater the corresponding positive reward value.

[0049] The magnitude of the patient's active force is collected by the built-in sensors and force sensors of the rehabilitation execution device, and calculated by combining electromyography (EMG) signals, reflecting the level of effort the patient exerts to complete the movement voluntarily. The total torque output by the rehabilitation execution device is directly collected by the built-in encoder and torque sensor, and is the total torque formed by the device's auxiliary force and the patient's active force. Both are inherent data acquisition items in the motion data and can be directly read and used to calculate the effort reward function.

[0050] The motion quality reward function is calculated based on the dynamic time warping distance between the real-time joint motion trajectory in the motion data and the preset standard rehabilitation motion trajectory. The dynamic time warping distance is inversely proportional to the positive reward value.

[0051] This involves calculating the dynamic time warping distance between the real-time joint movement trajectory in the motion data and the preset standard rehabilitation movement trajectory. A positive reward value is then calculated based on this distance, and the dynamic time warping distance is inversely proportional to the positive reward value. For example, if the dynamic time warping distance between the patient's real-time knee joint movement trajectory and the standard trajectory is 1.2, the smaller the distance, the more standard the movement, and the greater the corresponding positive reward value.

[0052] The dynamic time warping distance is calculated using a dynamic time warping algorithm. First, real-time joint motion trajectories are extracted from the motion data. Then, a preset standard rehabilitation motion trajectory is retrieved. The two unequal-length time-series trajectories are aligned along their time axes based on shape similarity. The cumulative distance at each point is calculated to obtain the final dynamic time warping distance.

[0053] In step S300 of this embodiment, based on the patient's implicit state vector and a preset set of sub-reward functions, initial personalized reward function learning is performed to obtain an initial reward function, including: Collect the patient's preference selection data in the early stage of rehabilitation, wherein the preference selection data includes the patient's preferred option selected from multiple candidate training options; Collect initial guidance and demonstration data from rehabilitation therapists for the patients; An inverse reinforcement learning algorithm is used, with the preference selection data and the initial guidance demonstration data as supervision, and the goal of maximizing the log likelihood of the observed behavior, to solve for the initial weight vectors corresponding to the safety reward function, the effort level reward function and the action quality reward function; The initial reward function is obtained by weighting and combining the sub-reward function set according to the initial weight vector.

[0054] In step S300 of this application embodiment, the purpose of the above step is to rely on the patient's implicit state vector and the constructed set of sub-reward functions, combined with the patient's initial preferences and the rehabilitation therapist's demonstration information, to obtain the initial weight vector of each sub-reward function through the inverse reinforcement learning algorithm, and generate an initial reward function that fits the individual patient situation after weighted combination, so as to provide a basis for subsequent dynamic reward function adjustment.

[0055] To achieve the above objectives, it is first necessary to collect the patient's preference selection data in the early stage of rehabilitation, wherein the preference selection data includes the preferred scheme selected by the patient from multiple candidate training schemes; This involves presenting patients with multiple candidate training programs and recording their preferred program, thus generating preference selection data that can be used for learning. For example, if a patient is given three candidate training programs—low intensity, medium intensity, and high intensity—and chooses the medium intensity program, this selection result constitutes preference selection data.

[0056] Next, data on the initial guidance and demonstration provided by the rehabilitation therapist to the patient were collected; This involves recording the initial training guidance and standard movement demonstrations provided by the rehabilitation therapist based on the patient's condition and physical state, forming professional monitoring data. For example, if the rehabilitation therapist instructs the patient to train twice a day for fifteen minutes each time initially, with the range of motion controlled within a safe range, this guidance constitutes the initial guidance and demonstration data.

[0057] Then, using an inverse reinforcement learning algorithm, with the preference selection data and the initial guidance demonstration data as supervision, and with the objective of maximizing the log likelihood of the observed behavior, the initial weight vectors corresponding to the safety reward function, the effort level reward function, and the action quality reward function are solved. For example, through iterative calculation using the inverse reinforcement learning algorithm, the initial weights of the safety reward function, effort reward function, and action quality reward function are obtained as follows: 0.3, 0.4, and 0.3 respectively. These three elements constitute the initial weight vector.

[0058] The inverse reinforcement learning algorithm is an algorithm model directly deployed in the rehabilitation system. Its core function is to deduce the reward function weights from observable behaviors such as patient preference selection data and initial guidance demonstration data from rehabilitation therapists.

[0059] For example, the inverse reinforcement learning algorithm can be constructed using the maximum entropy inverse reinforcement learning algorithm, with the feature mapping dimension set to 64 dimensions to match the dimension of the patient's implicit state vector; the policy network consists of two fully connected layers, with 128 neurons in each layer; the reward function fitter uses a linear regression model with a fitting coefficient ranging from 0 to 1. The learning rate is set to 0.001, the batch size to 32, the entropy regularization coefficient to 0.1, the optimizer uses the Adam optimizer, and the loss function is cross-entropy loss.

[0060] The training samples were sourced from patient preference selection data during the initial rehabilitation phase and initial guidance demonstration data from rehabilitation therapists. The training sample size was set at 500 sets of standardized behavioral sequence samples. The training rounds were set at 80 rounds. The convergence criterion was determined by observing a log-likelihood value fluctuation of less than 0.001 for the observed behavior within 15 consecutive training rounds, and a change in the weights of each sub-reward function of less than 0.0005.

[0061] Finally, the sub-reward function set is weighted and combined according to the initial weight vector to obtain the initial reward function.

[0062] The initial reward function is obtained by multiplying the results of each sub-reward function by its corresponding initial weight and then summing the products. For example, the initial reward function is obtained by multiplying the results of the safety reward function, effort reward function, and action quality reward function by 0.3, 0.4, and 0.3 respectively and then summing them.

[0063] In step S400 of this embodiment, the patient's implicit state vector and recent rehabilitation effect indicators are input into a pre-trained meta-learner, and dynamic reward function adjustment is performed to obtain an updated reward function, including: Obtain the recent rehabilitation effect indicators, wherein the rehabilitation effect indicators include the change in functional scores and training compliance rate within a preset time window; The patient's implicit state vector and the rehabilitation effect index are input into the meta-learner, and the meta-learner outputs the adjustment amount of the current reward function weight; The adjustment amount is applied to the current reward function weights to generate an updated reward function, and the updated weights are normalized to ensure that the updated weights are non-negative and sum to 1. The meta-learner is a pre-trained neural network, and the training data for the meta-learner comes from the complete rehabilitation trajectories of multiple patients in the historical rehabilitation database. The training objective is to maximize the cumulative rehabilitation effect of patients in subsequent rehabilitation stages.

[0064] In step S400 of this application embodiment, the purpose of the above steps is to adjust the weights of the reward function online adaptively through a pre-trained meta-learner based on the patient's real-time status and recent rehabilitation effects, so as to obtain an updated reward function that is more in line with the current patient situation, and to make the rehabilitation decision goal continuously match the patient's rehabilitation process.

[0065] To achieve the above objectives, it is first necessary to obtain the recent rehabilitation effect indicators, which include the change in functional scores and training compliance rate within a preset time window. This involves statistically analyzing the change in functional score and training adherence rate within a preset time window. The change in functional score is the difference between the current functional score and the functional score of the previous cycle, and the training adherence rate is the ratio of the number of training sessions actually completed to the number of training sessions planned to be completed. For example, if the preset time window is 7 days, the patient's functional score in the previous cycle was 70 points, the current functional score is 75 points, the change in functional score is 5 points, the planned training sessions are 14, and the patient actually completed 13 sessions, resulting in a training adherence rate of 0.93.

[0066] Next, the patient's implicit state vector and the rehabilitation effect index are input into the meta-learner, and the meta-learner outputs the adjustment amount of the current reward function weight; The patient's implicit state vector and rehabilitation effect indicators are input into the meta-learner. The meta-learner performs inference calculations through its internal neural network and outputs adjustments to the weights of the current reward function. For example, if the 64-dimensional patient implicit state vector, along with a functional score change of 5 points and a training compliance rate of 0.93, are input into the meta-learner, the meta-learner will output an adjustment of 0.05 for the safety reward function weight, -0.03 for the effort reward function weight, and -0.02 for the movement quality reward function weight.

[0067] The meta-learner is a pre-trained neural network, and the training data for the meta-learner comes from the complete rehabilitation trajectories of multiple patients in the historical rehabilitation database. The training objective is to maximize the cumulative rehabilitation effect of patients in subsequent rehabilitation stages.

[0068] For example, the meta-learner can be built based on a model-independent meta-learning algorithm. It employs a four-layer fully connected neural network structure: 128 neurons in the first layer, 64 in the second, 32 in the third, and a third output layer. The output layer has 3 neurons, corresponding to the weight adjustments for the safety reward function, effort reward function, and action quality reward function, respectively. The activation function is uniformly ReLU. The meta-learning rate is set to 0.001, the batch size to 32, the optimizer is Adam, and the loss function is mean squared error.

[0069] The training samples for the meta-learner are the complete rehabilitation trajectories of multiple patients from a historical rehabilitation database. The training sample size is 10,000 sets. The training rounds are 100. The convergence criterion is that the loss decrease is less than 0.001 for 20 consecutive rounds, and the prediction error of the weight adjustment is consistently below 0.005, indicating that the model has converged.

[0070] The cumulative rehabilitation effect refers to the combined sum of the total improvement in functional scores, the average training compliance rate, and the improvement in movement quality scores during a continuous rehabilitation phase. It measures the overall degree of rehabilitation improvement over a period of time. It is obtained by weighting and summing the changes in functional scores, movement quality scores, and training compliance rate within a preset time window according to cumulative weights. This is the core optimization objective of the meta-learner; a higher value indicates better patient rehabilitation progress. For example, if a patient's functional score improves by 5 points, movement quality by 4 points, and training compliance rate by 0.95 within 7 days, the weighted sum yields the cumulative rehabilitation effect value for that period.

[0071] In calculating the cumulative rehabilitation effect, the weighted cumulative weights are first set based on the experience of experts in the rehabilitation field, with a weight of 0.5 for the total improvement of functional scores, a weight of 0.2 for the average training compliance rate, and a weight of 0.3 for the improvement of movement quality scores. Then, based on the rehabilitation trajectory data of patients in the historical rehabilitation database, the cumulative weights are fine-tuned with the final rehabilitation achievement rate as the target, so that the cumulative weights are adapted to the characteristics of different patient groups. The final determined cumulative weights must satisfy the constraint that they are non-negative and sum to 1.

[0072] Then, the adjustment amount is applied to the current reward function weights to generate an updated reward function, and the updated weights are normalized to ensure that the updated weights are non-negative and sum to 1. First, the weight adjustment amounts of each sub-reward function output by the meta-learner are added to the current weights of their respective sub-reward functions to obtain the initially adjusted weights. Then, the initially adjusted weights are normalized by dividing them by the sum of all weights, ensuring that the final weights are non-negative and sum to 1. For example, if the safety weight is currently 0.3 with an adjustment of 0.05, the initial adjustment will be 0.35. After normalization, the sum of the safety, effort, and action quality weights is ensured to be 1 and without negative values.

[0073] In step S400 of the embodiments of this application, the following is also included: When a preset change in the patient's condition is detected, the dynamic reward function adjustment step is triggered.

[0074] In step S400 of this application embodiment, the purpose of the above steps is to realize the intelligent triggering of dynamic reward function adjustment, avoid ineffective and frequent calculations, update the reward function in a timely manner when the patient's state changes significantly, and ensure that the rehabilitation strategy is highly matched with the patient's current state.

[0075] To achieve the above objectives, it is necessary to continuously collect and calculate the patient's implicit state vector. The Euclidean distance between the current vector and the vector saved in the previous cycle is calculated, and the result is compared with a preset threshold. When the distance exceeds the preset threshold, it is determined that the patient's state has changed to a preset degree, and the system automatically triggers the dynamic reward function adjustment step. For example, if the system presets the patient state change threshold to 0.2, and the real-time calculated Euclidean distance between the current patient's implicit state vector and the vector from the previous time step is 0.32, which exceeds the set threshold, the dynamic reward function adjustment step is triggered, updating the reward function.

[0076] The preset threshold is statistically calibrated based on the patterns of patient state changes in a historical rehabilitation database. First, implicit state vectors of multiple rehabilitation patients at different stages are extracted. The Euclidean distance between vectors at adjacent time points is calculated, and the distance range corresponding to significant changes in rehabilitation effectiveness is selected. The critical value of this range is used as the initial threshold. Then, it is fine-tuned by combining the rehabilitation therapist's clinical experience with actual training scenarios to ensure that significant changes in patient state are effectively identified without inadvertently triggering adjustments due to minor fluctuations. For example, after statistical and empirical calibration, the preset threshold is ultimately set to 0.2 to determine whether to trigger dynamic reward function adjustments.

[0077] In step S400 of the embodiments of this application, the following is also included: After performing the dynamic reward function adjustment step, the updated reward function, the patient's implicit state vector, and the corresponding rehabilitation effect indicators are sent to the manual review terminal to receive confirmation or correction instructions from the rehabilitation therapist based on the review results. If the correction instruction is received, the updated reward function is corrected according to the weight adjustment value in the correction instruction to obtain a manually corrected reward function, which is then used for subsequent rehabilitation strategy optimization.

[0078] In step S400 of this application embodiment, the purpose of the above steps is to introduce a manual review mechanism by rehabilitation therapists to verify and correct the reward function after it is automatically updated by the system, so as to ensure that the weight of the reward function is both in line with the actual condition of the patient and in line with the clinical rehabilitation standards, thereby improving the safety and rationality of subsequent rehabilitation strategy optimization.

[0079] To achieve the above objectives, after performing the dynamic reward function adjustment step, the updated reward function, the patient's implicit state vector, and the corresponding rehabilitation effect indicators need to be sent to the manual review terminal to receive confirmation or correction instructions from the rehabilitation therapist based on the review results. Specifically, after adjusting the dynamic reward function, the system pushes the updated reward function, the patient's implicit state vector, and the corresponding rehabilitation effect indicators to the manual review terminal used by the rehabilitation therapist. The therapist views the relevant data on the terminal and provides the review results, and the system receives confirmation or correction instructions in real time. For example, the system sends the updated reward function with weights of 0.35, 0.37, and 0.28, the 64-dimensional patient implicit state vector, the functional score change of 5, and the training compliance rate of 0.93 to the manual review terminal, and the therapist reviews it and issues a correction instruction.

[0080] If the correction instruction is received, the updated reward function is corrected according to the weight adjustment value in the correction instruction to obtain a manually corrected reward function, which is then used for subsequent rehabilitation strategy optimization.

[0081] The system recalculates the weights of each sub-reward function in the updated reward function according to the weight adjustment values ​​contained in the instruction. After correction, a manually corrected reward function is obtained, which replaces the original function in subsequent rehabilitation strategy optimization. For example, if the correction instruction provides weight adjustment values ​​of 0.38, 0.34, and 0.28, the system corrects the reward function accordingly, obtaining a manually corrected reward function for subsequent use.

[0082] In step S500 of this embodiment, the patient's implicit state vector and the updated reward function are input into the reinforcement learning policy network to perform rehabilitation policy optimization and obtain the optimal rehabilitation action, including: The patient's implicit state vector is used as the environment state for reinforcement learning, and the updated reward function is used as the current optimization objective. The reinforcement learning policy network performs policy reasoning and outputs a rehabilitation action that maximizes the expected cumulative reward. The rehabilitation action includes training intensity, assist force magnitude, and training mode.

[0083] In step S500 of this application embodiment, the purpose of the above steps is to obtain the optimal rehabilitation action that balances safety, effort level and action quality by using reinforcement learning strategy network inference calculation based on the patient's current state and the updated reward function as the optimization target, so as to provide a precise execution plan for actual rehabilitation training.

[0084] The training intensity is a training load level set based on the patient's functional score and tolerance, and can be divided into three levels: low, medium, and high. The numerical values ​​correspond to the number of movements completed and the range of motion per unit time. The assist force is the active assistance provided by the rehabilitation device to the patient to compensate for insufficient muscle strength, and its range can be between 0N and 50N. The training mode is the movement paradigm executed by the device, including fixed modes such as continuous flexion-extension, intermittent flexion-extension, passive traction, and active resistance. The system outputs a combination of these three factors through a reinforcement learning strategy network to form the optimal rehabilitation movement adapted to the patient's current state.

[0085] To achieve the above objectives, the patient's implicit state vector should first be used as the environment state for reinforcement learning, and the updated reward function should be used as the current optimization objective. This approach uses the patient's implicit state vector as the input to the environment during reinforcement learning, while setting the updated reward function as the objective function for policy optimization, thus constructing a complete reinforcement learning optimization environment. For example, by using the 64-dimensional patient implicit state vector as the environment state and substituting it into the updated reward function after weight adjustment and normalization, the optimization environment can be built.

[0086] For example, the reinforcement learning policy network can be built using a deep deterministic policy gradient algorithm. It employs a three-layer fully connected structure: the Actor network has an input layer dimension of 64, hidden layer neurons of 128 and 64 respectively, and an output layer of 3 neurons, corresponding to training intensity, assist force, and training mode; the Critic network has the same structure as the Actor, with the output layer representing the action value evaluation, and both using ReLU activation functions. The learning rate is set to 0.0003, the discount factor to 0.95, the standard deviation of exploration noise to 0.1, the batch size to 64, and the update frequency to once every 10 steps.

[0087] The training samples for the reinforcement learning policy network are the complete rehabilitation trajectories of multiple patients from a historical rehabilitation database, with a training sample size of 15,000 sets. The training runs for 200 rounds. The convergence criterion is that the average expected cumulative reward fluctuation of the policy network over 30 consecutive rounds is less than 0.01, and the action output error is consistently below 0.005, indicating model convergence. The training objective is to maximize the expected cumulative reward.

[0088] Then, policy reasoning is performed through the reinforcement learning policy network to output the rehabilitation action that maximizes the expected cumulative reward, wherein the rehabilitation action includes training intensity, assist force magnitude, and training mode.

[0089] This involves using a reinforcement learning policy network to reason about the current environmental state and the optimization goal. It then searches the available action space to select the rehabilitation action that maximizes the expected cumulative reward. The output rehabilitation action includes three key parameters: training intensity, assist force, and training mode. For example, the policy network might determine the optimal rehabilitation action to be a combination of moderate training intensity, 20N assist force, and continuous flexion-extension training mode, which maximizes the expected cumulative reward.

[0090] The expected cumulative reward is the prediction of the total reward that the reinforcement learning policy network can obtain over multiple time steps after performing a certain set of rehabilitation actions, given the current patient's implicit state vector. This value is calculated by weighting the updated reward function, which integrates the real-time outputs of the safety reward function, effort reward function, and action quality reward function, reflecting the overall contribution of the rehabilitation action to the rehabilitation effect. Through continuous reasoning, the policy network selects the combination of training intensity, assistance level, and training mode that maximizes the expected cumulative reward, thereby determining the optimal rehabilitation action.

[0091] S600: Send the optimal rehabilitation action to the rehabilitation execution device and collect new multimodal data after executing the optimal rehabilitation action for use in the next round of rehabilitation strategy optimization.

[0092] In step S600 of this application embodiment, the purpose of the above steps is to implement the optimized rehabilitation actions and collect new multimodal data after the actions are implemented, so as to provide real-time and accurate patient status feedback for the next round of rehabilitation strategy optimization, forming a closed loop of optimization-execution-feedback-re-optimization, and continuously improving the adaptability of rehabilitation strategies.

[0093] To achieve the above objectives, the implementation method consists of two parts: The first part involves sending the optimal rehabilitation movement to the rehabilitation execution device. The system accurately sends the optimal rehabilitation movement parameters output by the reinforcement learning strategy network, including training intensity, assistive force, and training mode, to the rehabilitation execution device via the data transmission module. After receiving the parameters, the device automatically matches the corresponding execution mode and starts rehabilitation training. For example, the system sends the optimal rehabilitation movement with moderate training intensity, 20N assistive force, and continuous flexion-extension mode to the rehabilitation execution device. After receiving the data, the rehabilitation execution device automatically adjusts the parameters and starts training according to the set mode.

[0094] The second part involves collecting new multimodal data after the exercise. After the patient completes optimal rehabilitation exercise training, the rehabilitation execution device simultaneously collects multimodal data, including the patient's physiological signals, movement trajectory signals, and training completion status. The collected data undergoes standardized preprocessing to remove noise interference and ensure accuracy. The preprocessed new multimodal data is then stored in the system database for use in the next round of patient implicit state vector construction, reward function adjustment, and rehabilitation strategy optimization. For example, after the patient completes training, the rehabilitation execution device collects new multimodal data such as electromyographic signals, joint movement trajectory data, and a training compliance rate of 0.95. This data is preprocessed and stored to provide data support for the next round of rehabilitation strategy optimization.

[0095] Example 2, as Figure 2 As shown, based on the same inventive concept as the deep learning-based personalized rehabilitation care plan generation method provided in Embodiment 1, this embodiment of the invention also provides a smart chip, including: The multimodal data loading module 11 is used to load the patient's multimodal data during the rehabilitation process, wherein the multimodal data includes physiological signal data, motion data, patient subjective feedback data, rehabilitation therapist assessment data, and environmental information data; The patient state modeling module 12 is used to perform patient state modeling based on the multimodal data to obtain a patient implicit state vector, wherein the patient implicit state vector integrates the representations of physiological load state, motor function state and psychological motivation state. The reward function acquisition module 13 is used to perform initial personalized reward function learning based on the patient's implicit state vector and a preset set of sub-reward functions to obtain an initial reward function, wherein the initial reward function is a weighted combination of multiple sub-reward functions, and the set of sub-reward functions includes at least a safety reward function, an effort reward function, and a movement quality reward function; The reward function update module 14 is used to input the patient's implicit state vector and recent rehabilitation effect indicators into a pre-trained meta-learner, perform dynamic reward function adjustment, and obtain an updated reward function. The meta-learner is used to output the weight adjustment amount of the reward function according to the current patient state and rehabilitation effect. The rehabilitation strategy optimization module 15 is used to input the patient's implicit state vector and the updated reward function into the reinforcement learning strategy network, perform rehabilitation strategy optimization, and obtain the optimal rehabilitation action; The strategy loop optimization module 16 is used to send the optimal rehabilitation action to the rehabilitation execution device and collect new multimodal data after the optimal rehabilitation action is executed, so as to be used for the next round of rehabilitation strategy optimization.

[0096] Furthermore, the multimodal data loading module 11 includes the following execution steps: The patient's wearable sensors collect electromyography (EMG) signals, heart rate signals, and electrodermal signals, which are then added to the physiological signal data. The joint angles, motion trajectories, and speeds collected by the built-in encoder of the rehabilitation execution device, as well as the magnitude of the assistive force output by the rehabilitation execution device, are added to the motion data. The results of the fatigue measurement scale, pain score and motivation level questionnaire filled out by the patient before and after training were obtained through the interactive interface and added to the patient's subjective feedback data. Obtain functional scores and movement quality scores from rehabilitation therapists' periodic assessments and add them to the aforementioned rehabilitation therapist assessment data; Obtain the operating status and training environment parameters of the rehabilitation execution equipment, and add them to the environmental information data.

[0097] Furthermore, the patient status modeling module 12 includes the following execution steps: The multimodal data is time-aligned and features are extracted to obtain multimodal temporal features; The multimodal temporal features are input into a temporal neural network to extract a low-dimensional implicit state vector of the patient. The temporal neural network is composed of a convolutional neural network layer and a long short-term memory network layer connected in series.

[0098] Furthermore, the reward function acquisition module 13 includes the following execution steps: Construct the sub-reward function set, wherein; The safety reward function is calculated by monitoring whether the physiological signal data exceeds a preset safety threshold range. When the monitored value exceeds the preset safety threshold range, a negative reward value is output. The absolute value of the negative reward value is proportional to the extent of exceeding the limit. The effort reward function is calculated based on the ratio of the magnitude of the patient's active force exertion in the motion data to the total torque output by the rehabilitation execution device, and the ratio is proportional to the positive reward value. The motion quality reward function is calculated based on the dynamic time warping distance between the real-time joint motion trajectory in the motion data and the preset standard rehabilitation motion trajectory. The dynamic time warping distance is inversely proportional to the positive reward value.

[0099] Collect the patient's preference selection data in the early stage of rehabilitation, wherein the preference selection data includes the patient's preferred option selected from multiple candidate training options; Collect initial guidance and demonstration data from rehabilitation therapists for the patients; An inverse reinforcement learning algorithm is used, with the preference selection data and the initial guidance demonstration data as supervision, and the goal of maximizing the log likelihood of the observed behavior, to solve for the initial weight vectors corresponding to the safety reward function, the effort level reward function and the action quality reward function; The initial reward function is obtained by weighting and combining the sub-reward function set according to the initial weight vector.

[0100] Furthermore, the reward function update module 14 includes the following execution steps: Obtain the recent rehabilitation effect indicators, wherein the rehabilitation effect indicators include the change in functional scores and training compliance rate within a preset time window; The patient's implicit state vector and the rehabilitation effect index are input into the meta-learner, and the meta-learner outputs the adjustment amount of the current reward function weight; The adjustment amount is applied to the current reward function weights to generate an updated reward function, and the updated weights are normalized to ensure that the updated weights are non-negative and sum to 1. The meta-learner is a pre-trained neural network, and the training data for the meta-learner comes from the complete rehabilitation trajectories of multiple patients in the historical rehabilitation database. The training objective is to maximize the cumulative rehabilitation effect of patients in subsequent rehabilitation stages.

[0101] Also includes: When a preset change in the patient's condition is detected, the dynamic reward function adjustment step is triggered.

[0102] After performing the dynamic reward function adjustment step, the updated reward function, the patient's implicit state vector, and the corresponding rehabilitation effect indicators are sent to the manual review terminal to receive confirmation or correction instructions from the rehabilitation therapist based on the review results. If the correction instruction is received, the updated reward function is corrected according to the weight adjustment value in the correction instruction to obtain a manually corrected reward function, which is then used for subsequent rehabilitation strategy optimization.

[0103] Furthermore, the rehabilitation strategy optimization module 15 includes the following execution steps: The patient's implicit state vector is used as the environment state for reinforcement learning, and the updated reward function is used as the current optimization objective. The reinforcement learning policy network performs policy reasoning and outputs a rehabilitation action that maximizes the expected cumulative reward. The rehabilitation action includes training intensity, assist force magnitude, and training mode.

Claims

1. A method for generating personalized rehabilitation nursing plans based on deep learning, characterized in that, include: Load multimodal data of patients during the rehabilitation process, wherein the multimodal data includes physiological signal data, motion data, patient subjective feedback data, rehabilitation therapist assessment data, and environmental information data; Based on the multimodal data, patient state modeling is performed to obtain the patient's implicit state vector, wherein the patient's implicit state vector integrates the representations of physiological load state, motor function state and psychological motivation state. Based on the patient's implicit state vector and a preset set of sub-reward functions, an initial personalized reward function is learned to obtain an initial reward function. The initial reward function is a weighted combination of multiple sub-reward functions, and the set of sub-reward functions includes at least a safety reward function, an effort reward function, and a movement quality reward function. The patient's implicit state vector and recent rehabilitation effect indicators are input into a pre-trained meta-learner, and dynamic reward function adjustment is performed to obtain an updated reward function. The meta-learner is used to output the weight adjustment amount of the reward function according to the current patient state and rehabilitation effect. The patient's implicit state vector and the updated reward function are input into a reinforcement learning policy network to perform rehabilitation policy optimization and obtain the optimal rehabilitation action. The optimal rehabilitation action is sent to the rehabilitation execution device, and new multimodal data after the optimal rehabilitation action is executed is collected for the next round of rehabilitation strategy optimization.

2. The method for generating personalized rehabilitation nursing plans based on deep learning as described in claim 1, characterized in that, Load multimodal data of patients during the recovery process, including: The patient's wearable sensors collect electromyography (EMG) signals, heart rate signals, and electrodermal signals, which are then added to the physiological signal data. The joint angles, motion trajectories, and speeds collected by the built-in encoder of the rehabilitation execution device, as well as the magnitude of the assistive force output by the rehabilitation execution device, are added to the motion data. The results of the fatigue measurement scale, pain score and motivation level questionnaire filled out by the patient before and after training were obtained through the interactive interface and added to the patient's subjective feedback data. Obtain functional scores and movement quality scores from rehabilitation therapists' periodic assessments and add them to the aforementioned rehabilitation therapist assessment data; Obtain the operating status and training environment parameters of the rehabilitation execution equipment, and add them to the environmental information data.

3. The method for generating personalized rehabilitation nursing plans based on deep learning as described in claim 1, characterized in that, Based on the multimodal data, patient state modeling is performed to obtain the patient's implicit state vector, including: The multimodal data is time-aligned and features are extracted to obtain multimodal temporal features; The multimodal temporal features are input into a temporal neural network to extract a low-dimensional implicit state vector of the patient. The temporal neural network is composed of a convolutional neural network layer and a long short-term memory network layer connected in series.

4. The method for generating personalized rehabilitation nursing plans based on deep learning as described in claim 1, characterized in that, Based on the patient's implicit state vector and a preset set of sub-reward functions, initial personalized reward function learning is performed to obtain the initial reward function, which includes the following: Construct the sub-reward function set, where: The safety reward function is calculated by monitoring whether the physiological signal data exceeds a preset safety threshold range. When the monitored value exceeds the preset safety threshold range, a negative reward value is output. The absolute value of the negative reward value is proportional to the extent of exceeding the limit. The effort reward function is calculated based on the ratio of the magnitude of the patient's active force exertion in the motion data to the total torque output by the rehabilitation execution device, and the ratio is proportional to the positive reward value. The motion quality reward function is calculated based on the dynamic time warping distance between the real-time joint motion trajectory in the motion data and the preset standard rehabilitation motion trajectory. The dynamic time warping distance is inversely proportional to the positive reward value.

5. The method for generating personalized rehabilitation nursing plans based on deep learning as described in claim 4, characterized in that, Based on the patient's implicit state vector and a preset set of sub-reward functions, initial personalized reward function learning is performed to obtain an initial reward function, including: Collect the patient's preference selection data in the early stage of rehabilitation, wherein the preference selection data includes the patient's preferred option selected from multiple candidate training options; Collect initial guidance and demonstration data from rehabilitation therapists for the patients; An inverse reinforcement learning algorithm is used, with the preference selection data and the initial guidance demonstration data as supervision, and the goal of maximizing the log likelihood of the observed behavior, to solve for the initial weight vectors corresponding to the safety reward function, the effort level reward function and the action quality reward function; The initial reward function is obtained by weighting and combining the sub-reward function set according to the initial weight vector.

6. The method for generating personalized rehabilitation nursing plans based on deep learning as described in claim 1, characterized in that, The patient's implicit state vector and recent rehabilitation performance indicators are input into a pre-trained meta-learner, and a dynamic reward function adjustment is performed to obtain an updated reward function, including: Obtain the recent rehabilitation effect indicators, wherein the rehabilitation effect indicators include the change in functional scores and training compliance rate within a preset time window; The patient's implicit state vector and the rehabilitation effect index are input into the meta-learner, and the meta-learner outputs the adjustment amount of the current reward function weight; The adjustment amount is applied to the current reward function weights to generate an updated reward function, and the updated weights are normalized to ensure that the updated weights are non-negative and sum to 1. The meta-learner is a pre-trained neural network, and the training data for the meta-learner comes from the complete rehabilitation trajectories of multiple patients in the historical rehabilitation database. The training objective is to maximize the cumulative rehabilitation effect of patients in subsequent rehabilitation stages.

7. The method for generating personalized rehabilitation nursing plans based on deep learning as described in claim 1, characterized in that, Also includes: When a preset change in the patient's condition is detected, the dynamic reward function adjustment step is triggered.

8. The method for generating personalized rehabilitation nursing plans based on deep learning as described in claim 1, characterized in that, Also includes: After performing the dynamic reward function adjustment step, the updated reward function, the patient's implicit state vector, and the corresponding rehabilitation effect indicators are sent to the manual review terminal to receive confirmation or correction instructions from the rehabilitation therapist based on the review results. If the correction instruction is received, the updated reward function is corrected according to the weight adjustment value in the correction instruction to obtain a manually corrected reward function, which is then used for subsequent rehabilitation strategy optimization.

9. The method for generating personalized rehabilitation nursing plans based on deep learning as described in claim 1, characterized in that, The patient's implicit state vector and the updated reward function are input into a reinforcement learning policy network to perform rehabilitation policy optimization and obtain the optimal rehabilitation action, including: The patient's implicit state vector is used as the environment state for reinforcement learning, and the updated reward function is used as the current optimization objective. The reinforcement learning policy network performs policy reasoning and outputs a rehabilitation action that maximizes the expected cumulative reward. The rehabilitation action includes training intensity, assist force magnitude, and training mode.

10. A smart chip, said smart chip being used to implement the method for generating personalized rehabilitation care plans based on deep learning as described in any one of claims 1-9, characterized in that, The smart chip includes: A multimodal data loading module is used to load multimodal data of patients during the rehabilitation process, wherein the multimodal data includes physiological signal data, motion data, patient subjective feedback data, rehabilitation therapist assessment data, and environmental information data; The patient state modeling module is used to perform patient state modeling based on the multimodal data to obtain a patient implicit state vector, wherein the patient implicit state vector integrates the representations of physiological load state, motor function state and psychological motivation state. The reward function acquisition module is used to perform initial personalized reward function learning based on the patient's implicit state vector and a preset set of sub-reward functions to obtain an initial reward function. The initial reward function is a weighted combination of multiple sub-reward functions, and the set of sub-reward functions includes at least a safety reward function, an effort reward function, and a movement quality reward function. The reward function update module is used to input the patient's implicit state vector and recent rehabilitation effect indicators into a pre-trained meta-learner, perform dynamic reward function adjustment, and obtain an updated reward function. The meta-learner is used to output the weight adjustment amount of the reward function according to the current patient state and rehabilitation effect. The rehabilitation strategy optimization module is used to input the patient's implicit state vector and the updated reward function into the reinforcement learning policy network, perform rehabilitation strategy optimization, and obtain the optimal rehabilitation action; The strategy loop optimization module is used to send the optimal rehabilitation action to the rehabilitation execution device and collect new multimodal data after the optimal rehabilitation action is executed, so as to be used for the next round of rehabilitation strategy optimization.

Images