Cerebral stroke patient home rehabilitation training content dynamic optimization system based on GRPO

Abstract

The technical scheme of the invention discloses a dynamic optimization system for the household rehabilitation training content of a stroke patient based on GRPO. The exercise performance and physiological feedback of the patient are collected in real time, the system can adjust the training content in a closed-loop mode, the scientificity, safety and efficiency of rehabilitation training are remarkably improved, and the purposes of progressive training difficulty and individualization are both achieved. The system disclosed by the invention not only breaks through the limitations of fixed templates, manual adjustment, feedback lag and the like of traditional rehabilitation training, but also has good expansibility, can be integrated with a rehabilitation robot, intelligent wearable equipment and a mobile health management platform, and promotes collaborative development of rehabilitation medicine, intelligent medical treatment and health industries. Therefore, a new personalized rehabilitation mode based on frontier artificial intelligence decision is provided for patients with cerebral apoplexy and related diseases, theory and engineering systems in the field of intelligent rehabilitation are greatly enriched, and the method has important clinical application value and wide industrial popularization prospects.

Description

TECHNICAL FIELD

[0001] The present application relates to an application of reinforcement learning, especially a group relative policy optimization (GRPO) algorithm, to dynamic optimization of home rehabilitation training content for stroke patients, to achieve real-time adaptation and autonomous evolution of rehabilitation programs to individual patient ability changes. The present application relates to artificial intelligence medical treatment, rehabilitation engineering, intelligent human-computer interaction, wearable remote health monitoring, and other fields, and belongs to the field of high-tech artificial intelligence and medical health, focusing on intelligent rehabilitation auxiliary decision-making, personalized training content adaptive optimization, and rehabilitation medical treatment innovation driven by big data. BACKGROUND

[0002] With the continuous progress of medical artificial intelligence technology and rehabilitation engineering, the rehabilitation mode of stroke patients with functional disorders is undergoing profound changes. Modern rehabilitation concepts emphasize "full-process, scientific, and individualized" rehabilitation management, especially in the home rehabilitation stage after patients are discharged from the hospital. How to ensure the continuity and pertinence of rehabilitation training has become a key factor affecting the effectiveness of functional recovery. Stroke patients often need to adhere to rehabilitation training for a long time. Only a continuous, scientific, and dynamically adjusted training program can effectively promote neural remodeling and reconstruction of motor function. In the home scenario, the rehabilitation effect of patients largely depends on whether the training content can be optimized in real time according to individual ability and recovery progress, otherwise, problems such as poor rehabilitation effectiveness and prolonged recovery period are likely to occur.

[0003] Currently, most home rehabilitation training systems or auxiliary tools on the market still mainly use "fixed plans" or "phased action arrangements". Such systems usually rely on the clinical experience of rehabilitation therapists, use standardized templates, and push pre-set training actions or programs for patients in different groups according to the functional classification of patients at the time of initial assessment. Some products support coarse-grained hierarchical recommendations based on static characteristics such as age and gender, and individual high-end devices can set weekly and monthly training periods and regularly adjust action content through a schedule. However, this design based on artificial rules or static templates, while standardizing the home rehabilitation process of patients to some extent and providing basic training guidance for patients, lacks real-time collection and dynamic analysis of actual training feedback from patients, resulting in a lag in training content adjustment relative to changes in patient ability, making it difficult to adapt to the unique rehabilitation trajectory of each patient. Patients often encounter action difficulty mismatches during training, such as repeated failures due to overly difficult tasks or limited training effectiveness due to overly easy content, which directly affects the rehabilitation compliance and initiative of patients.

[0004] For further improvement of training content recommendation, some existing systems attempt to introduce rule-based grouping push methods. For example, according to the patient's age, gender or initial motor function, etc. Static features, different rehabilitation training plans are designed for different groups. This kind of grouping logic can to some extent avoid the drawbacks of one-size-fits-all solutions and provide basic adaptive training guidance for patients at different levels. However, due to the lack of automatic collection and analysis of daily and each training results of patients, the granularity of its content recommendation is still coarse, and it cannot respond sensitively to the changes in patient's ability. When the patient makes progress or regresses, the system cannot make dynamic adjustments, the rehabilitation program adjustment lags behind, and it cannot achieve truly continuous individualized dynamic rehabilitation management.

[0005] In traditional home rehabilitation practice, the individualization of most rehabilitation training still depends on regular offline follow-up and periodic manual assessment. Rehabilitation therapists usually manually adjust the patient's training plan according to the results of monthly or quarterly face-to-face consultations. This method can achieve some individualized optimization, but it has obvious limitations such as long evaluation period, slow response, many subjective factors, and untimely adjustment. For the majority of stroke patients distributed in the home environment, this mode of relying on manual assessment not only leads to low efficiency, but also seriously restricts the accessibility and popularity of rehabilitation services, which is not conducive to the realization of large-scale promotion and patient self-management.

[0006] With the development of artificial intelligence and big data technology, more and more AI-assisted rehabilitation recommendation systems have emerged, but overall, the current market mainstream products are mainly based on static models, grouping recommendations or simple prediction algorithms based on historical data. Most systems still rely on initial stratification in training content push and difficulty adjustment, and lack a closed-loop mechanism that can learn, evolve and optimize content based on real-time training feedback of patients. In order to achieve personalized recommendation, some systems require a large amount of manually annotated data and tedious parameter settings, which not only increases the actual operation burden, but also makes it difficult to achieve intelligent, adaptive and efficient automatic learning in the whole process.

[0007] In the field of intelligent optimization of home rehabilitation training content for stroke patients, hierarchical adaptive recommendation is a traditional but widely used technical approach. This method usually divides patients into different levels or stages based on their basic abilities, stage assessments, or key physiological indicators. Each level is equipped with a pre-set training action library and recommendation logic. When a patient completes the tasks specified in the current level or reaches a certain threshold, the system automatically adjusts to a higher level and recommends higher intensity or difficulty training content. This method has certain feasibility in practical application and can achieve progressive training tasks and coarse-grained individual adaptation. However, the core limitation of hierarchical adaptive recommendation is the coarse adjustment granularity and lack of real-time performance. The improvement and recovery curve of patient ability often does not completely match the grading settings, leading to phenomena such as "stuck level" or "level skipping," making it difficult to accurately match the actual needs in a timely manner. In addition, traditional hierarchical methods still rely on manually developed rules, with weak system self-learning and optimization capabilities, making it difficult to automatically adjust the recommended content based on small changes in patients, affecting the personalized experience and rehabilitation efficiency.

[0008] Fuzzy logic control is an optimization method that uses fuzzy set theory to describe and make decisions on patient status. This method dynamically adjusts rehabilitation tasks by defining a set of fuzzy rules (such as "if the patient's subjective fatigue is high and the action completion rate is low, then reduce the training difficulty") and combining continuous indicators collected by sensors. Compared with hierarchical methods, fuzzy logic can be more flexible in handling state boundaries, enabling multi-variable and multi-interval comprehensive decision-making. However, in practical applications, fuzzy logic control still faces several technical bottlenecks: first, the rule base design highly depends on expert experience and manual tuning, making it difficult to cover all actual clinical situations; second, the self-learning capability is limited, and the rule system cannot evolve automatically based on patient historical feedback; third, it performs poorly in sequential decision-making and long-term optimal path exploration for rehabilitation training, often only achieving short-term and local training content adjustments, making it difficult to achieve long-term dynamic global optimization.

[0009] The decision tree-based push mechanism refers to establishing a training content decision tree, selecting a path based on the patient's current physiological state, past training results, and other feature nodes, and finally generating personalized training action recommendations. This method can achieve a certain degree of personalized recommendation and explicitly express the decision-making process through tree structure, facilitating interpretation and clinical verification. However, decision tree models generally have insufficient generalization ability and rigid decision boundaries. When the patient's recovery trajectory exhibits new changes beyond the sample space, the decision tree may make incorrect branch selections. Additionally, once the structure of the decision tree model is fixed, subsequent self-adaptive learning and dynamic expansion are difficult, relying on manual updates of nodes and branches to maintain applicability, resulting in high engineering maintenance and expansion costs.

[0010] In summary, existing home rehabilitation training recommendation systems have the following main defects and shortcomings:

[0011] Firstly, the lack of real-time dynamic individualized adjustment based on the training results of each patient, the content and difficulty cannot be automatically adjusted according to the daily or step training feedback, resulting in the disconnection between the training plan and the actual ability of the patient.

[0012] Secondly, the system design and content recommendation rely too much on the subjective experience of the rehabilitation therapist and the static stratification, the intelligent level is limited, the automatic learning and self-evolution mechanism is insufficient, and it is difficult to realize high-level automatic individualized management.

[0013] Third, the content adjustment response is slow, and it is difficult to adapt to the dynamic changes of the patient's ability and the pace of rehabilitation progress, and the training content often lags behind the actual recovery state of the patient, affecting the rehabilitation effect.

[0014] Fourth, there is a lack of effective use of training feedback data, which cannot continuously optimize the rehabilitation content based on long-term interaction data of patients, and it is difficult to establish a self-improving content push closed loop, and the level of individualization and precision is limited.

[0015] Fifth, the patient's daily training compliance and initiative are difficult to be continuously guaranteed, and some systems need frequent manual intervention, the operation is complex, and the patient's use experience is not good, which further affects the rehabilitation effect. SUMMARY

[0016] The purpose of the present application is to provide a new intelligent rehabilitation training recommendation system, which should be able to collect and analyze the training feedback of patients in real time, have the ability of automatic learning and dynamic optimization of content, and truly realize individualized, dynamic and scientific home rehabilitation management. Only by realizing the whole process closed loop self-adaptation, the rehabilitation efficiency and intelligent level can be greatly improved, so that every stroke patient can obtain the best rehabilitation plan that is accurately matched with their own functional state and recovery progress, and promote the deep integration of rehabilitation engineering technology and intelligent medical treatment and industrial upgrading.

[0017] In order to achieve the above purpose, the technical scheme of the present application discloses a stroke patient home rehabilitation training content dynamic optimization system based on GRPO, characterized in that it comprises:

[0018] A patient information acquisition module for collecting and processing real-time data of patients throughout the process;

[0019] The strategy optimization module: on the one hand, the strategy optimization module generates the personalized strategy of the current round of training according to the real-time data collected by the patient information acquisition module by applying the group relative strategy optimization algorithm; on the other hand, the strategy optimization module automatically analyzes the performance advantages and disadvantages of the patient in the historical training based on the real-time data and the task completion results fed back by the rehabilitation feedback acquisition and evaluation module, and dynamically updates the individual training recommendation strategy of the current round of training, adjusts the action type, difficulty and training load, and obtains the personalized strategy of the current round of training;

[0020] The hierarchical rehabilitation action library;

[0021] The recommendation engine is used to realize the dynamic content push of the patient terminal based on the rehabilitation training content library relying on the personalized strategy output by the strategy optimization module;

[0022] The rehabilitation feedback acquisition and evaluation module is used to automatically and comprehensively collect the completion quality, action performance and feedback feeling of the patient in each training, quantitatively and qualitatively evaluate the completion quality and target achievement of each round of training task based on the built-in judgment rules and algorithm model, judge whether the training task meets the standard, quickly assess the training success or failure, and obtain the task completion result data.

[0023] Preferably, the real-time data includes objective indicators of patient motor function recorded by wearable sensors and subjective information actively input by the patient through the rehabilitation APP.

[0024] Preferably, the patient information acquisition module forms a structured feature vector that can reflect the current ability and recovery stage of the patient after data fusion and standardization preprocessing of all raw data.

[0025] Preferably, the strategy optimization module uses a reward mechanism to realize the iterative update of the strategy, including the following steps:

[0026] Step 1, obtaining the training task completion data of the historical training through the rehabilitation feedback acquisition and evaluation module;

[0027] Step 2, rewarding all rehabilitation training actions or contents in the historical training, when the patient successfully completes the jth rehabilitation training action or content, the reward is "1", otherwise, the reward is "0", and the reward function of the jth rehabilitation training action or content is defined as r(q, o j ), wherein q represents the state or context information of the patient, o j represents the jth rehabilitation training action or content recommended for the state;

[0028] Step 3, based on the reward obtained in the previous step, the advantage value of all rehabilitation training actions or contents in the historical training is calculated, wherein the advantage value of the jth rehabilitation training action or content is represented as Then Wherein B is the group baseline;

[0029] Step 4, normalize all the advantage values obtained in the previous step, as shown in the following formula:

[0030]

[0031] In the formula, is the normalized advantage value, and σ r is the standard deviation of the reward;

[0032] Step 5, the group relative strategy optimization algorithm is used to optimize and update the strategy, and the parameters of the new strategy are θ, the old strategy is π old (·|q), and the reference strategy is π ref (·|q), then for a given state q, the expected optimization objective function J(θ) of the group relative strategy optimization algorithm is defined as:

[0033]

[0034] In the formula: r θ (q,o) represents the ratio of the probability of the new strategy to the old strategy on action o; ∈ is the strategy clipping coefficient; clip(r θ (q,o),1-∈,1+∈) represents that the ratio r θ (q,o) is truncated in the interval [1-∈,1+∈]; D KL (π θ (·|q)‖π ref (·|q)) is the KL divergence regular term, which is used to measure the deviation of the new strategy π θ (·|q) from the reference strategy π ref (·|q), and β is the regularization coefficient;

[0035] The strategy parameters are updated by maximizing J(θ) each time, and the new strategy π θ (·|q) is obtained, and the optimal objective J(θ) is gradually approached through multiple rounds of small-step gradient ascent iteration, and finally the converged strategy is obtained, wherein after each strategy parameter update, the system enters the next cycle, and a new training scheme set is generated again for the updated strategy, patient feedback is obtained, the advantage is calculated, and the strategy is continuously optimized accordingly, and the iteration is repeated, forming a closed-loop strategy optimization process, until the convergence condition is met or the predetermined training iteration number is reached.

[0036] Preferably, the rehabilitation action library covers various training contents from basic to high difficulty, and details the operation standards, difficulty levels and applicable stages of each action.

[0037] Preferably, when the rehabilitation action library is constructed, the rehabilitation action library covering various rehabilitation stages is preset according to clinical norms in the rehabilitation field and expert experience, and when the patient trains at home through the intelligent device, the multi-dimensional indexes are continuously collected, and the preset rehabilitation action library is updated.

[0038] The system can adjust the training content in a closed loop by collecting the motion performance and physiological feedback of the patient in real time, significantly improving the scientificity, safety and efficiency of the rehabilitation training, and balancing the gradual progress of the training difficulty and the achievement of the individualized goal. The system disclosed in the present application not only breaks through the limitations of the traditional rehabilitation training fixed template, manual adjustment and feedback lag, but also has good expansibility, can be integrated with rehabilitation robots, intelligent wearable devices and mobile health management platforms, and promotes the coordinated development of rehabilitation medicine, smart medical treatment and health industry. Therefore, the present application provides a new mode of personalized rehabilitation based on cutting-edge artificial intelligence decision-making for patients with stroke and related diseases, greatly enriches the theory and engineering system of the field of intelligent rehabilitation, and has important clinical application value and broad industrial promotion prospects.

[0039] The short board of the hierarchical adaptive recommendation method, the fuzzy logic control strategy and the dynamic pushing mechanism based on the decision tree pointed out in the background technology is that the self-learning ability is limited, the dynamic response to the patient feedback is not accurate enough, and the sequential decision and the whole cycle personalized optimization of the training content cannot be efficiently realized. The intelligent optimization technology based on the group relative strategy optimization (GRPO) reinforcement learning proposed in the present application has the following outstanding advantages:

[0040] 1) High self-adaptation and self-evolution ability: the GRPO algorithm can collect patient feedback data in real time, automatically adjust the training content distribution based on the principle of reinforcement learning, without human intervention, and optimize the recommendation strategy independently, overcoming the disadvantages of the traditional method relying on expert rules or fixed model structure;

[0041] 2) Strong dynamic individualized response ability: the system can adjust the type, difficulty and frequency of the training action according to the actual training feedback of each patient, ensure the continuous and accurate adaptation of the training content to the patient's functional state, and significantly improve the rehabilitation efficiency.

[0042] 3) Good engineering expansibility and industrialization potential: the GRPO technology takes automation and self-learning as the core, can support the large-scale deployment of multiple patients and multiple scenes, reduce the engineering maintenance cost, and has good commercial promotion prospects and clinical landing value.

[0043] The technical scheme of the present application relies on a reinforcement learning framework, fully excavates and utilizes the individualized training performance and recovery trajectory of patients. The system automatically adjusts the training content and stage goal according to the state change and advancement speed of each patient, realizes the precise individualized management of 'thousand faces', and flexibly upgrades or adjusts the training difficulty, frequency and content according to the specific ability and recovery rhythm of the patient, which not only guarantees the scientificity of training, but also improves the compliance and rehabilitation enthusiasm. Compared with the traditional method of relying on expert experience to set a general scheme, the present application greatly improves the precision and effect of individualized rehabilitation.

[0044] I) The present application is based on the group relative policy optimization (GRPO) reinforcement learning algorithm, which can automatically adjust and dynamically optimize the subsequent training content according to the real-time feedback of each training of the patient. The system adjusts the training difficulty, type and intensity in real time through precise analysis of the patient's action completion, success and failure mode, ensures that the training tasks pushed are always accurately matched with the current ability level of the patient. This dynamic closed-loop mechanism effectively avoids the failure frustration caused by too difficult training content, or the rehabilitation bottleneck phenomenon caused by too easy. A large number of experimental and practical application results show that after adopting the technical scheme of the present application, the overall success rate of patient rehabilitation training is significantly improved, the rehabilitation process is smoother, and the training results and functional recovery efficiency are better than the traditional static recommendation and rule grading scheme.

[0045] II) The present application realizes the automatic generation, real-time adjustment and autonomous evolution of the whole process of rehabilitation training content through reinforcement learning and multi-modal feedback mechanism. The system does not need human intervention during operation, does not need frequent intervention or manual setting of stage goals by doctors, and can realize all-weather, continuous and self-adaptive rehabilitation content pushing and adjustment. The highly automated intelligent decision-making mechanism greatly reduces the management burden of doctors, and also simplifies the operation experience of patients, so that patients can independently and efficiently complete rehabilitation training, and promote the popularization and standardized implementation of rehabilitation work.

[0046] III) The technical scheme of the present application relies on a reinforcement learning framework, fully excavates and utilizes the individualized training performance and recovery trajectory of patients. The system automatically adjusts the training content and stage goal according to the state change and advancement speed of each patient, realizes the precise individualized management of 'thousand faces', and flexibly upgrades or adjusts the training difficulty, frequency and content according to the specific ability and recovery rhythm of the patient, which not only guarantees the scientificity of training, but also improves the compliance and rehabilitation enthusiasm. Compared with the traditional method of relying on expert experience to set a general scheme, the present application greatly improves the precision and effect of individualized rehabilitation.

[0047] IV. In the implementation of the present application, stroke patients carry out home training in the family scene through the intelligent rehabilitation training system. Basic health data and functional status are input by the patient or the rehabilitation team when the system is first deployed, and then combined with wearable multi-modal sensors and rehabilitation App, real-time collection of multi-dimensional indicators such as limb activity amplitude, muscle strength, training action completion, subjective fatigue and pain is carried out. All raw data is converted into a feature vector after fusion and normalization, which is input into the strategy optimization module. The system dynamically matches the multi-level training content in the rehabilitation action library with the patient's state, and automatically optimizes the training content push through the GRPO reinforcement learning algorithm, with "group sampling-reward discrimination-advantage normalization-KL regularization convergence" as the core process. In each round of training recommendation, the system generates multiple candidate actions based on the last training result and the current ability state, collects the completion of each action as a binary reward (such as 1 for completion and 0 for failure), uses the group average reward as the baseline, calculates the advantage value of each action and normalizes it. The clipping of the probability ratio of new and old strategies and the KL regularization term jointly limit the strategy update amplitude, ensuring that the training task progresses at a moderate and stable pace. The system automatically increases the probability of actions with positive advantage values, gradually eliminates training content that is not suitable for the patient, and forms a rehabilitation path that gradually increases in difficulty and highly matches individual ability. In multiple iterations, the strategy automatically captures changes in the patient's recovery trajectory, dynamically optimizes the rehabilitation content from the basic to the advanced, and significantly improves the training success rate and functional recovery speed. After each training, the system generates a stage feedback report and personalized training suggestions, providing timely motivation and guidance for patients to adjust their training pace, effectively improving compliance and initiative. A large number of practical application results show that the intelligent rehabilitation system not only shortens the average rehabilitation period and improves the ability to take care of oneself, but also greatly reduces the burden on doctors through automated data analysis and strategy evolution, supporting large-scale deployment in multiple centers, communities and remote scenarios, providing a scientific and efficient intelligent home rehabilitation solution for stroke patients, and promoting the digitalization and intelligentization upgrade of the rehabilitation medical field.

[0048] V. The core algorithm and system architecture design of the present application have high universality and can be flexibly expanded to various rehabilitation training scenarios and different types of chronic disease management platforms. The system is not only suitable for stroke patients, but also can be smoothly transplanted to orthopedic postoperative, sports injury, and functional decline in the elderly, etc. Relying on the highly automated, large-scale remote deployment and cloud data support architecture, the technical solution of the present application has a significant industrialization landing prospect, can support community medical care, home rehabilitation, specialized hospitals and Internet medical care, etc. multi-level, multi-scene application, and promote the intelligentization, standardization and high efficiency development of the rehabilitation industry. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 schematically illustrates the overall system architecture of the present application;

[0050] Figure 2 The strategy iteration update process is illustrated;

[0051] Figure 3 The process of dynamically adjusting recommended content of the application is illustrated;

[0052] Figure 4 The application flow of the system of the application is illustrated. DETAILED DESCRIPTION

[0053] The application will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the application and not used to limit the scope of the application. Furthermore, it should be understood that after reading the content taught by the application, those skilled in the art can make various modifications or changes to the application, and these equivalent forms also fall within the scope defined by the appended claims of the application.

[0054] The application focuses on the series of technical pain points and deficiencies existing in the current home rehabilitation training content recommendation system for stroke patients, and positively proposes the following innovation goals:

[0055] Firstly, the traditional home rehabilitation recommendation system generally lacks a dynamic collection and response mechanism for patient training feedback, and it is difficult to make real-time personalized adjustments to the training content according to the performance of each training of the patient. Existing solutions often only formulate training plans according to initial stratification or static indicators, ignoring the changes and fluctuations in the patient's rehabilitation ability in a short period. To solve this problem, the application is dedicated to developing an intelligent decision-making system that can continuously collect real-time rehabilitation performance and historical training feedback of patients, automatically optimize training content based on data-driven, and realize the automatic evolution of training difficulty and content with the progress of patient ability, thereby improving the scientific nature and individual adaptation of rehabilitation.

[0056] Secondly, the widely used training content recommendation system highly depends on the manual experience and subjective judgment of rehabilitation therapists, with low adjustment frequency and lagging response, making it difficult to adapt to the nonlinear progress rhythm of patient rehabilitation in a timely manner. This traditional mode of relying on subjective judgment not only leads to the inability of the rehabilitation scheme to evolve dynamically, but also makes the training content obtained by the patient often lag behind the actual needs, affecting the efficiency of functional recovery. Therefore, the application proposes an intelligent algorithm path that can realize self-learning and self-evolution of training strategies without human intervention, thereby fundamentally improving the individualization level and decision-making efficiency of rehabilitation.

[0057] Third, the existing training content recommendation system is mainly based on static grouping or simple rules, with coarse granularity, which is difficult to realize accurate identification and real-time dynamic adjustment of individual differences of patients. Such recommendation mechanism lacks adaptive learning ability when facing the wide heterogeneity and complex rehabilitation trajectory of stroke patients, and cannot balance efficiency and accuracy. The present application realizes the closed-loop linkage of recommended content and patient rehabilitation progress through the introduction of reinforcement learning and sequential optimization mechanism, so that the training content can accurately adapt to the individual rehabilitation trajectory of each patient.

[0058] In addition, the utilization rate of rehabilitation training feedback data is generally low, and the system is difficult to correct the training scheme in time through the immediate information of training success or failure, resulting in slow rehabilitation path adjustment and limited improvement of patient rehabilitation effect. The present application provides a dynamic content optimization technology method which can efficiently collect and fully utilize training feedback information, greatly improving patient training compliance and rehabilitation efficiency.

[0059] In summary, the present application aims to systematically solve the problem of lack of real-time personalization, intelligent dynamic optimization and closed-loop adaptation in existing home rehabilitation systems, and to develop a highly intelligent and closed-loop self-learning rehabilitation recommendation system that can continuously optimize training content based on patient training feedback, promoting the scientific and precise new stage of stroke home rehabilitation management.

[0060] Based on the above objectives, as shown in Figure 1 The embodiment of the present application discloses a stroke patient home rehabilitation training content dynamic optimization system based on group relative strategy optimization (Group Relative Policy Optimization, GRPO) algorithm, mainly including the following contents.

[0061] I) Patient information acquisition module

[0062] The patient information acquisition module is responsible for collecting and processing the real-time data of the patient throughout the process. In the embodiment of the present application, the real-time data of the patient can include: using wearable sensors to record the objective indicators of the patient's motor function (such as: joint range of motion, power output, action completion time, etc.); subjective information such as training experience, functional self-evaluation and rehabilitation confusion actively input by the patient through the rehabilitation APP. After data fusion and standardized preprocessing of all raw data, a structured feature vector reflecting the current ability and recovery stage of the patient is formed, providing a data basis for intelligent decision-making.

[0063] II) Strategy optimization module

[0064] The strategy optimization module generates the personalized strategy of the current round of training by applying a group relative strategy optimization (GRPO) algorithm based on the real-time data collected by the patient information acquisition module. On the other hand, the strategy optimization module also applies the group relative strategy optimization (GRPO) algorithm based on the real-time data and task completion results fed back by the rehabilitation feedback acquisition and evaluation module, automatically analyzes the performance advantages and disadvantages of the patient in the historical training, and dynamically updates the individual training recommended strategy of the current round of training to adjust the action type, difficulty and training load, so as to obtain the personalized strategy of the current round of training.

[0065] In the embodiment of the application, the strategy optimization module models the recommendation process of the rehabilitation training content as a sequential decision strategy, and the sequential decision strategy of the i th round of training is represented as p i (o j |q), wherein q represents the state or context information of the patient, o j represents the j th rehabilitation training action or content recommended for the state, and as Figure 2 shown, the strategy optimization module realizes the iterative update of the strategy by using a reward mechanism, and the reward mechanism adopts a binary judgment based on objective standards:

[0066] Step 1: Obtain the training task completion data of the historical training through the rehabilitation feedback acquisition and evaluation module, including the motion performance, action completion data and the like collected by the sensor.

[0067] Step 2: Perform reward judgment on all rehabilitation training actions or contents in the historical training, and obtain a reward of “1” when the patient successfully completes the j th rehabilitation training action or content, otherwise, obtain a reward of “0”. In the embodiment of the application, the reward function of the j th rehabilitation training action or content is defined as r(q,o j ).

[0068] Step 3: Based on the reward obtained in the above step, the advantage value of all rehabilitation training actions or contents in the historical training is calculated, wherein the advantage value of the j th rehabilitation training action or content is represented as . The meaning is: if the reward r(q,o j ) of the j th rehabilitation training action or content o i is 1 and the group baseline B is less than 1, the j th rehabilitation training action or content o j performs better than the average level, and the advantage value is positive; otherwise, if r(q,o i ) is 0 and the group baseline B is greater than 0, the j th rehabilitation training action or content o j is lower than the average level, and the advantage value is negative.

[0069] Step 4, in order to enhance the stability of training, all the advantage values obtained in the last step are normalized:

[0070]

[0071] wherein, is the normalized advantage value, and σ r is the standard deviation of the reward.

[0072] Step 5, after obtaining the advantage values of each candidate action, the present application uses the GRPO reinforcement learning algorithm to update the strategy. The core idea of GRPO is to introduce the advantage function calculated by the above group baseline and the KL divergence regular term of the reference strategy into the policy gradient objective of the PPO (Proximal Policy Optimization) algorithm, so as to ensure that the policy update can both improve the success rate and maintain the stability and moderate amplitude of the update. Let the parameters of the new strategy be θ, the old strategy be π old (·|q), and the reference strategy be π ref (·|q). For a given state q, the expected optimization objective function J(θ) of GRPO is defined as:

[0073]

[0074] wherein, r θ (q,o) represents the ratio (i.e. importance sampling weight) of the probability of the new strategy relative to the old strategy on action o; ∈ is a policy clipping coefficient, used to limit the change amplitude of r θ (q,o) to prevent the policy from updating too much in a single step; clip(r θ (q,o),1-∈,1+∈) represents truncating the ratio r θ (q,o) within the interval [1-∈,1+∈]; in the first bracket term of the expected value, the minimum value of and the clipped forms the surrogate loss in the PPO algorithm. This minimum value form ensures that when the difference between the new and old strategies is too large, only the limited gradient signal is used for updating, thereby limiting the change amplitude of the strategy and ensuring the stability of the training process. The second term in the objective function is the KL divergence regular term, which is used to measure the difference between the new strategy π θ (·|q) and the reference strategy π refThe deviation of (·|q) is weighted by a coefficient β and incorporated into the optimization objective, directly constraining policy updates. By adjusting the value of the regularization coefficient β, a trade-off can be struck between increasing cumulative reward and maintaining policy stability: when β is large, the new policy is strictly limited to a distribution close to the reference policy, and policy updates tend to be conservative; when β is small, the KL constraint weakens, and the policy can be explored and adjusted more boldly. With this design, the objective function of GRPO simultaneously considers improving the success rate using the dominance function and controlling the distance between the policy and the reference distribution through the KL term, making it particularly suitable for sequential decision-making problems such as rehabilitation training that require a balance between gradual adjustment and ensuring safety.

[0075] Step 6: After constructing the policy optimization objective function, in each iteration, maximize J(θ) using the policy gradient algorithm or other optimization methods to update the policy parameters and obtain the new policy π. θ (·|q). The effect of policy update is to increase the probability of selecting training content with positive dominance values ​​and decrease the probability of selecting content with negative dominance values, thereby strengthening better-than-average decisions and weakening worse-than-average decisions. Quantitatively, this can be understood as the new and old policies satisfying a relative weight adjustment after the update: for any action o, if Then π after the update θ (o|q) relative to π old (o|q) will increase; conversely, if The probability of that action in the new strategy will then be relatively reduced. Intuitively, this is equivalent to factoring the probabilities of each action in the old strategy. After weighting and renormalizing, the strategy is adjusted to have a higher success rate.

[0076] In implementation, the optimal objective J(θ) is typically approximated gradually through multiple rounds of small-step gradient ascent iterations, eventually yielding a converged policy. After each policy parameter update, the system enters the next loop: it performs group sampling again to generate a new set of training schemes, obtains patient feedback, calculates advantages, and continues to optimize the policy accordingly. This iterative process forms a closed-loop policy optimization process until convergence conditions are met (e.g., the policy success rate improvement plateaus) or a predetermined number of training iterations is reached. Through this cyclical mechanism, the policy can continuously adjust based on patient training feedback, gradually improving its decision-making quality and achieving adaptive optimization of the training content to the patient's capabilities.

[0077] In the above strategy optimization iteration process, since the training feedback reward only takes two values ​​(success = 1 or failure = 0), this invention can use the average success probability of the strategy in each iteration to describe the evolution of the strategy performance. Let p be the success rate of the strategy in the i-th round (i.e., the probability that the strategy successfully recommends training content in state q).i The success rate of the reference strategy (initial strategy) is denoted as p ref Theoretical analysis shows that, under the above-mentioned GRPO updating rule, the success rate p i of the strategy of the adjacent two rounds i+1 has the following recursive relationship:

[0078]

[0079] where p i+1 and p i represent the success rates of the (i+1)th and ith strategy iteration pushing training content respectively, p ref is the success rate of the reference strategy, β is the coefficient of the above-mentioned KL regular term, and ∈ is a stability constant (used to avoid singular cases where the denominator is zero when calculating p i (1-p i )). Analysis shows that the above-mentioned function has a fixed point in the interval [0, 1], and the final stable success rate p * of the strategy can be obtained by solving the fixed point equation. Further mathematical derivation can prove that when the initial reference success rate satisfies 0 < p ref < 1, the above-mentioned fixed point must be greater than p ref . In other words, after the GRPO strategy optimization iteration, the final stable success rate of the strategy will be improved to a level higher than the initial strategy success rate. This conclusion theoretically proves that the present application can gradually amplify the success rate of the rehabilitation training scheme and continuously improve the probability of the patient successfully completing the training. In addition, when the regular coefficient p ref is appropriately valued and the initial success rate is not 0, it can be proved that the iteration relationship satisfies the local convergence condition near the fixed point, that is, if the initial success rate p ref is already close enough to the fixed point, then the subsequent p i will monotonically converge to the fixed point. Generally, p * = 1 (100% success rate) is often a stable fixed point of the above-mentioned recursion. This means that in practice, a typical phenomenon is that as the strategy optimization iteration proceeds, the success rate of the patient completing the training task will gradually improve and tend to 1, that is, the strategy eventually tends to almost always recommend training tasks that the patient can successfully complete. Through the above-mentioned theoretical analysis, it can be seen that the GRPO strategy optimization mechanism of the present application can ensure that the rehabilitation training content recommendation strategy continuously evolves in the direction of improving the success rate, and under certain conditions, quickly converges to a stable state of high success rate, providing good performance guarantee and convergence properties for the system.

[0080] Three) Rehabilitation training content library and recommendation engine

[0081] The rehabilitation action library is a hierarchical database, covering various training contents from basic to high difficulty, and refining the operation standard, difficulty level and applicable stage of each action. According to the clinical norms and expert experience in the rehabilitation field, the present application presets a rehabilitation action library covering various rehabilitation stages, including training actions of different difficulty, function and target. When the patient trains at home through the intelligent device, the system disclosed in the present application continuously collects multi-dimensional indicators including motion amplitude, action completion rate, pain self-evaluation, etc., and updates the preset rehabilitation action library, ensuring that the rehabilitation action library can meet basic rehabilitation and adapt to advanced function training.

[0082] The recommendation engine relies on the personalized strategy output by the strategy optimization module, and based on the rehabilitation training content library, realizes dynamic content pushing to the patient terminal. Each recommendation can accurately adapt to the patient's ability, and ensure the diversity and advancement of the training content.

[0083] IV) Rehabilitation feedback collection and evaluation module

[0084] The rehabilitation feedback collection and evaluation module automatically and comprehensively collects the completion quality, action performance and feedback feeling of each training of the patient, and based on the built-in judgment rules and algorithm model, quantitatively and qualitatively evaluates the completion quality and target achievement of each round of training task, judges whether the training task meets the standard, quickly assesses the success or failure of the training, and obtains the task completion result data, such as whether the patient has reached the action amplitude requirement, whether the specified group number has been completed, whether abnormal fatigue or movement disorder has been shown, etc. The rehabilitation feedback collection and evaluation module feeds back the evaluation data to the strategy optimization module in real time, forms an efficient data backflow, and supports dynamic optimization of content.

[0085] In combination with Figure 3 The above system has the ability to intelligently identify the rehabilitation progress of the patient, can dynamically adjust the recommended content according to the recovery speed and ability improvement, and realizes the gradual transition from basic training to advanced comprehensive action. Each recommendation is neither too simple nor prevents sudden difficulty increase, realizing a training path with both personalization and scientificity.

[0086] The intelligent rehabilitation training system disclosed in the present application is deployed in the patient's home environment based on a home terminal device (such as a tablet computer, a mobile phone, a rehabilitation all-in-one machine, etc.). The patient performs rehabilitation training according to the daily or regular plan suggested by the attending physician or the rehabilitation team. Before each training starts, the system automatically wakes up the user account, reads the patient's latest training state, and generates a personalized training action recommendation scheme according to the latest individual ability characteristics and rehabilitation progress. The recommended content covers core parameters such as training action type, intensity, number, duration, etc., and is dynamically adjusted in combination with real-time state, ensuring that the task difficulty and patient ability level are accurately matched.

[0087] In combination with Figure 4During the training execution, the system collects the physiological and motion performance data of the patient in real time through multi-modal sensors (such as accelerometers, gyroscopes, pressure pads, cameras, etc.) when the patient performs the action. Typical collection indicators include limb activity amplitude, joint activity, action completion time, force, balance state, and subjective training experience, fatigue and discomfort scores, etc. The data is automatically uploaded to the local or cloud analysis platform, and the system quantitatively and qualitatively evaluates the completion quality and target achievement of each training task based on the built-in judgment rules and algorithm models. For example, whether the patient has reached the action amplitude requirement, whether the specified group number has been completed, whether abnormal fatigue or movement disorders have been shown, etc.

[0088] According to the collected real-time data and task completion results, the system applies the group relative strategy optimization (GRPO) algorithm to automatically analyze the performance advantages and disadvantages of the patient in this round of training. The system dynamically updates the individual training recommendation strategy based on this, adjusts the action type, difficulty and training load. The optimized recommended content is directly pushed to the patient terminal, forming a daily / round "collection-analysis-feedback-push-retraining" adaptive closed loop. The system not only responds to the current ability of the individual, but also automatically identifies the bottleneck or plateau of the recovery progress according to the long-term training curve, and actively optimizes the training rhythm and enriches the task content to help the patient break through the functional disorder and continue to progress.

[0089] After multiple rounds of dynamic optimization, the completion rate of the patient's training task shows a significant upward trend. The system can automatically identify and exclude high-difficulty actions that the patient repeatedly fails, avoiding the frustration of training confidence due to excessive tasks. At the same time, by continuously recommending appropriate new tasks, it helps the patient gradually improve their ability and achieve an easy-to-difficult, step-by-step rehabilitation path. Clinical follow-up and periodic data analysis show that most patients assisted by the intelligent optimization training system recover faster than traditional static recommendation schemes, with an average rehabilitation period shortened, improved motor function recovery, and significantly improved self-care ability.

[0090] The system disclosed in the present application fully considers the patient experience in the interactive design. After each training is completed, the system generates intuitive periodic feedback reports for the patient, including training achievement rate, ability progress trend, next stage target, etc., helping the patient to understand their own recovery status in a timely manner. Automatically pushing training tasks of appropriate difficulty greatly reduces the patient's resistance due to excessive difficulty or monotony, improves compliance and enthusiasm. At the same time, the sense of achievement and self-confidence during the training process are constantly positively motivated, further promoting the continuous development of rehabilitation training.

[0091] In multicenter clinical trials and practical application, the intelligent rehabilitation training content optimization system based on GRPO disclosed by the application all show significant advantages. Compared with the traditional static rehabilitation scheme formulated manually and adjusted periodically by hand, the technical scheme of the application can dynamically adapt to individual differences and training feedback of each patient, and realize high-level precision medicine. The system has good engineering scalability and can be seamlessly integrated into hospitals, rehabilitation centers and community health service platforms, supporting large-scale remote rehabilitation scene deployment. The results of a number of experiments and user surveys show that the system not only improves the rehabilitation efficiency and functional recovery quality, but also significantly reduces the frequency of manual intervention, saving medical resources and management costs. Overall, the application has strong clinical promotion potential and significant industrial application value, bringing real well-being to the family rehabilitation of patients with chronic diseases such as stroke.

Claims

1. A dynamic optimization system for home rehabilitation training content for stroke patients based on GRPO, characterized in that, include: The patient information acquisition module is used to collect and process real-time patient data throughout the entire process; Strategy optimization module: On the one hand, the strategy optimization module generates a personalized strategy for the current round of training by applying a group relative strategy optimization algorithm based on the real-time data collected by the patient information acquisition module; on the other hand, the strategy optimization module automatically analyzes the patient's strengths and weaknesses in historical training by applying a group relative strategy optimization algorithm based on the real-time data and task completion results fed back by the rehabilitation feedback acquisition and evaluation module, and dynamically updates the individual training recommendation strategy for the current round of training accordingly, adjusting the movement type, difficulty and training load to obtain a personalized strategy for the current round of training. A tiered and graded rehabilitation exercise library; The recommendation engine is used to dynamically push content to the patient's terminal based on the personalized strategies output by the strategy optimization module and the rehabilitation training content library. The rehabilitation feedback collection and assessment module is used to automatically and comprehensively collect the patient's completion quality, movement performance, and feedback feelings for each training session. Based on the built-in judgment rules and algorithm model, it performs quantitative and qualitative assessments on the completion quality and goal achievement of each round of training tasks, determines whether the training task has been achieved, quickly evaluates the success or failure of the training, and obtains task completion result data.

2. The dynamic optimization system for home rehabilitation training content of stroke patients based on GRPO as described in claim 1, characterized in that, The real-time data includes: objective indicators of the patient's motor function recorded using wearable sensors, as well as subjective information actively input by the patient through a rehabilitation app.

3. The dynamic optimization system for home rehabilitation training content of stroke patients based on GRPO as described in claim 1, characterized in that, The patient information acquisition module performs data fusion and standardization preprocessing on all raw data to form a structured feature vector that reflects the patient's current ability and recovery stage.

4. The dynamic optimization system for home rehabilitation training content of stroke patients based on GRPO as described in claim 1, characterized in that, The strategy optimization module uses a reward mechanism to achieve iterative updates of the strategy, including the following steps: Step 1: Obtain historical training task completion data through the rehabilitation feedback collection and assessment module; Step 2: Reward all rehabilitation training actions or content from the historical training. The patient receives a reward of "1" when successfully completing the j-th rehabilitation training action or content; otherwise, a reward of "0". The reward function for the j-th rehabilitation training action or content is defined as r(q,o). j ), where q represents the patient's state or context information, o j This indicates the j-th recommended rehabilitation training action or content for this condition; Step 3: Based on the rewards obtained in the previous step, calculate the advantage value of all rehabilitation training movements or content in the historical training, where the advantage value of the j-th rehabilitation training movement or content is denoted as... but Where B is the population baseline; Step 4: Normalize all the advantage values ​​obtained in the previous step, as shown in the following formula: In the formula, The normalized dominance value, σ r The standard deviation of the reward; Step 5: Optimize and update the policy using a group relative policy optimization algorithm. Let the parameters of the new policy be θ and the old policy be π. old (·|q), with the reference strategy being σ. ref (·|q), then for a given state q, the expected optimization objective function J(θ) of the population relative policy optimization algorithm is defined as: In the formula: r θ (q,o) represents the ratio of the probability of the new policy to that of the old policy on action o; ∈ is the policy clipping coefficient; clip(r θ (q,o),1-∈,1+∈) represents the ratio r θ (q,o) is truncated within the interval [1-∈, 1+∈]; D KL (σ θ (·|q)‖σ ref (·|q)) is the KL divergence regularization term, used to measure the new policy π. θ (·|q) relative to the reference strategy π ref The degree of deviation of (·|q), where β is the regularity coefficient; Each iteration updates the policy parameters by maximizing J(θ) to obtain the new policy π. θ (·|q) is used to gradually approach the optimal target J(θ) through multiple rounds of small-step gradient ascent iterations, and finally obtain a converged policy. After each policy parameter update, the system enters the next loop, performs group sampling again for the updated policy to generate a new set of training schemes, obtains patient feedback, calculates the advantage, and continues to optimize the policy accordingly. This process is repeated to form a closed-loop policy optimization process until the convergence condition is met or the predetermined number of training iterations is reached.

5. The dynamic optimization system for home rehabilitation training content of stroke patients based on GRPO as described in claim 1, characterized in that, The rehabilitation exercise library covers a variety of training content from basic to advanced levels, and details the operation standards, difficulty levels and applicable stages of each exercise.

6. The dynamic optimization system for home rehabilitation training content of stroke patients based on GRPO as described in claim 5, characterized in that, When constructing the rehabilitation exercise library, based on clinical standards and expert experience in the rehabilitation field, a rehabilitation exercise library covering various rehabilitation stages is preset. When patients train at home using smart devices, multidimensional indicators are continuously collected, and the preset rehabilitation exercise library is updated.

Images