A rehabilitation state evaluation method based on pattern recognition

Abstract

The application discloses a rehabilitation state evaluation method based on pattern recognition, comprising the following steps: acquiring rehabilitation training images and medical record information of a patient; inputting a PoseNet model to extract joint coordinates and generate a posture key point sequence; inputting the medical record information and the posture sequence into a rehabilitation knowledge graph to initialize an individualized rehabilitation training plan; based on a Dyna-Q reinforcement learning model, combining action normative scoring and subjective feedback to adjust training intensity and frequency; updating a path in the rehabilitation knowledge graph and optimizing the training plan through an incremental learning mechanism; dynamically pruning a network layer of the PoseNet model; and constructing a safety rule library to detect pain feedback in real time, triggering an artificial intervention process to ensure the safety of the patient's rehabilitation process. The application realizes multi-modal data fusion and intelligent learning, and improves the rehabilitation effect of the patient and the work efficiency of a therapist.

Description

Technical Field

[0001] This invention relates to the field of rehabilitation medicine technology, and in particular to a rehabilitation status assessment method based on pattern recognition. Background Technology

[0002] In modern rehabilitation medicine, traditional rehabilitation training relies on face-to-face guidance and supervision from therapists. While this approach provides personalized training, it also faces numerous challenges. Due to the limited number of therapists, patients often do not receive sufficient attention, especially in home or community rehabilitation centers, where training effectiveness and adherence can be limited. Furthermore, it is often difficult to effectively monitor whether patients are performing rehabilitation exercises correctly, and many patients may experience poor rehabilitation outcomes or secondary injuries due to improper form.

[0003] Some AI-based rehabilitation assessment methods are beginning to be applied in clinical practice. For example, posture estimation technology uses cameras to capture patients' movements to help assess the accuracy of those movements; reinforcement learning algorithms attempt to adjust training plans based on patient feedback to achieve better rehabilitation outcomes. A problem with existing technologies is that while these methods can assess movement standardization in real time, most systems rely on a single data source, neglecting the comprehensive consideration of multi-dimensional information such as patient medical records, subjective feedback, and training history. Existing intelligent rehabilitation systems still lack effective knowledge graph support, making it difficult for the system to adaptively optimize when facing different patients and personalized rehabilitation needs.

[0004] Therefore, how to provide a pattern recognition-based method for assessing rehabilitation status is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a rehabilitation status assessment method based on pattern recognition. This invention fully integrates posture estimation, reinforcement learning, and knowledge graph technologies to assess the standardization of patients' rehabilitation movements in real time and dynamically adjust the rehabilitation plan based on patients' feedback data. It has the advantages of strong real-time performance, high personalization, and improved rehabilitation effects.

[0006] A rehabilitation status assessment method based on pattern recognition according to an embodiment of the present invention includes the following steps:

[0007] Acquire patient rehabilitation training images and medical record information;

[0008] Input rehabilitation training images into the PoseNet pose estimation model to extract joint coordinates and generate a sequence of pose key points.

[0009] Input medical record information and posture key point sequences into the rehabilitation knowledge graph, and initialize the rehabilitation training plan based on the path relationships in the rehabilitation knowledge graph;

[0010] The subjective feedback data of patients during the training process is obtained. The posture key point sequence and subjective feedback data are input into the Dyna-Q reinforcement learning model. The Dyna-Q reinforcement learning model evaluates the effect of the current training plan based on the action standardization score and subjective feedback, and dynamically adjusts the training intensity and frequency, and outputs the adjusted training plan parameters.

[0011] The paths in the rehabilitation knowledge graph are updated based on the adjusted training plan parameters, and the training plan is iteratively optimized based on the daily execution feedback data using the incremental learning mechanism of the Dyna-Q reinforcement learning model.

[0012] During each pose estimation, dynamic pruning is performed on the PoseNet pose estimation model based on the complexity of the training task, selecting the network layers to be activated.

[0013] A medical safety boundary rule base is constructed to detect the sequence of key posture points and subjective feedback data in real time during rehabilitation training. When a patient reports severe pain, the training is forcibly paused and a manual intervention process is triggered. The manual correction instruction is received, and the manually corrected training plan is applied to subsequent rehabilitation training until the rehabilitation assessment indicators reach the preset standards.

[0014] Optionally, acquiring patient rehabilitation training images and medical record information specifically includes:

[0015] The system terminal obtains the patient's medical record information and configures initial training parameters for the patient based on the medical record information. The initial training parameters include training actions, training intensity and training frequency, and generate a personalized initial rehabilitation plan.

[0016] Patients use mobile terminal camera devices to capture real-time training video streams and extract rehabilitation training images from the real-time training video streams.

[0017] Rehabilitation training images are image frames continuously captured by a mobile terminal camera device when patients perform training movements under the guidance of a personalized initial rehabilitation plan.

[0018] Optionally, the generation of the pose keypoint sequence specifically includes:

[0019] Input the rehabilitation training images into the PoseNet pose estimation model to obtain the initial pose key point sequence and the confidence level of each joint coordinate;

[0020] Obtain information about the surgical site or the site of functional impairment from the medical record to identify the joints to focus on;

[0021] Joints with a confidence level below a preset threshold in the initial pose keypoint sequence are identified as joints to be completed.

[0022] When the joint to be completed is a key focus joint, the coordinates of the visible joints adjacent to the joint to be completed are obtained. Based on the personalized bone length ratio and joint range of motion constraints in the medical record information, and combined with the historical coordinates of the joint to be completed in the preceding image frame, the completion coordinates are calculated using an interpolation algorithm based on human kinematic constraints.

[0023] Replace the corresponding joint coordinates in the initial pose keypoint sequence with the completed coordinates to generate a complete pose keypoint sequence that conforms to the individual anatomical characteristics of the patient.

[0024] Optionally, the rehabilitation knowledge graph specifically includes:

[0025] Obtain knowledge sources in rehabilitation medicine, extract entity types and relation types from these sources, and construct an ontology model for the rehabilitation domain.

[0026] Collect historical rehabilitation case data, extract entity instances and inter-entity relationship instances from the historical rehabilitation case data, map the entity instances and relationship instances to the rehabilitation domain ontology model, and generate a basic rehabilitation knowledge graph;

[0027] The individual characteristic parameters contained in the medical record information are obtained, the personalized weight of each entity instance is calculated based on the individual characteristic parameters, and the personalized weight is added as an attribute to the corresponding entity instance in the basic rehabilitation knowledge graph to generate a rehabilitation knowledge graph containing personalized weights.

[0028] Obtain the patient's medical record information, input the medical record information into a rehabilitation knowledge graph containing personalized weights, retrieve rehabilitation paths that match the medical record information, and extract recommended training parameters from the retrieval results;

[0029] The recommended training parameters are presented to the therapist's terminal, and the therapist's confirmation or adjustment instructions for the recommended training parameters are received. A personalized initial rehabilitation plan is generated based on the confirmation or adjustment instructions.

[0030] Optionally, the Dyna-Q reinforcement learning model specifically includes:

[0031] The posture key point sequence, subjective feedback data, and training parameters corresponding to the current rehabilitation path in the rehabilitation knowledge graph are jointly constructed into a state vector;

[0032] An environmental model unit is constructed to predict the state changes of patients after implementing the adjusted training plan, and simulated experience samples are generated based on the prediction results. The real training experience samples and simulated experience samples are stored together in the experience memory set.

[0033] Q-value updates are performed based on the experience memory set, and the value function is iteratively updated using an incremental learning approach. Newly acquired daily execution feedback data covers the value estimates corresponding to historically low-relevance samples during the update process.

[0034] Upon receiving a manual correction instruction, the manually corrected training plan parameters are converted into external reward signals and input into the value update unit to perform reinforcement correction on the Q value of the corresponding state-action pair.

[0035] After completing the value function update, the output training plan adjustment parameters are synchronously written into the rehabilitation knowledge graph to update the entity weights of the corresponding rehabilitation path.

[0036] Optionally, the dynamic pruning of the PoseNet pose estimation model specifically includes:

[0037] Obtain the image complexity features of the current rehabilitation training images. The image complexity features include image resolution, number of human targets, and degree of joint occlusion.

[0038] Based on the image complexity features, the current pose estimation task is mapped to one of several preset computational accuracy levels;

[0039] Obtain the average confidence level of the action prescriptive scores output by the Dyna-Q reinforcement learning model for historical tasks of the same accuracy level.

[0040] If the average confidence level is lower than the preset accuracy threshold and the system resource utilization rate does not exceed the load limit, the current task will be automatically upgraded to a higher accuracy level and the entire network layer will be activated to participate in the calculation.

[0041] If the average confidence level is higher than the preset accuracy threshold or the system resource utilization rate exceeds the load limit, the current task will be automatically downgraded by one level of calculation accuracy and some network layers will be skipped.

[0042] Based on the final determined target accuracy level, the corresponding set of network layers is activated from the PoseNet pose estimation model, and joint coordinates are extracted to generate a sequence of pose key points.

[0043] Optionally, the medical security boundary rule base specifically includes:

[0044] Real-time acquisition of pain index from posture key point sequences and subjective feedback data;

[0045] The sequence of posture key points is compared with a pre-stored high-risk motion library, which stores joint angle thresholds and trajectory features.

[0046] When the system detects that the posture key point sequence matches high-risk action features or that the pain index exceeds a preset safety threshold, it immediately interrupts the execution of the current training plan and pauses the decision output of the Dyna-Q reinforcement learning model.

[0047] Generate an alert message and send it to the therapist's terminal, mark it as a triggered safety event, and record it in the safety log;

[0048] The system receives intervention instructions from the therapist's terminal, adjusts the subsequent training plan according to the intervention instructions, and writes the adjusted training plan parameters as safety constraints into the rehabilitation knowledge graph, while simultaneously updating the personalized weights of relevant entity instances on the corresponding rehabilitation path.

[0049] Upon receiving the training resumption instruction, the decision output of the Dyna-Q reinforcement learning model is restarted to continue the rehabilitation training.

[0050] Optionally, triggering a security event and recording it in the security log specifically includes:

[0051] When a security event is triggered by the medical safety boundary rule base, the system automatically collects the continuous attitude key point sequence before the trigger time, the decision trajectory of the Dyna-Q reinforcement learning model during that period, and the state prediction results of the environmental model unit to construct the causal chain data of the security event.

[0052] The causal chain data of safety incidents is stored as a structured event record and a visual playback interface is provided. The visual playback interface allows therapists to annotate key problem nodes in the event process.

[0053] Obtain the key problem nodes marked by the therapist, extract new high-risk action feature templates from the posture key point sequence corresponding to the key problem nodes, and add the new high-risk action feature templates to the high-risk action library of the medical safety boundary rule base.

[0054] Based on the causal chain data of safety events and the annotation results of therapists, the prediction bias of the environmental model unit for high-risk actions is calculated, and the prediction parameters of the environmental model unit are corrected using the prediction bias.

[0055] The beneficial effects of this invention are:

[0056] (1) By combining posture estimation, reinforcement learning, and knowledge graphs, the normalization of patients' movements is assessed in real time, and the rehabilitation plan is dynamically adjusted based on the patients' medical records and subjective feedback. This greatly improves the personalization of rehabilitation training and ensures that each patient can receive the training program most suitable for their recovery progress and needs.

[0057] (2) The system employs real-time motion monitoring and patient feedback mechanisms, which can promptly detect problems such as non-standard patient movements, pain, or fatigue, and automatically adjust the training plan according to the patient's condition, avoiding the risk of readmission or secondary injury due to non-standard movements. The system's built-in medical safety boundary rules further ensure that patients do not exceed the safe range during training, thus guaranteeing the patient's rehabilitation safety.

[0058] (3) By combining therapist intervention with an intelligent system, therapists only need to review the rehabilitation pathways and adjustment plans recommended by the system, avoiding a large amount of repetitive and inefficient manual guidance in traditional rehabilitation. This method reduces the workload of therapists and allows them to focus on more complex and urgent cases, thereby improving overall treatment efficiency. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0060] Figure 1 This is a flowchart of a rehabilitation status assessment method based on pattern recognition proposed in this invention;

[0061] Figure 2 This is a schematic diagram of the overall architecture of a pattern recognition-based rehabilitation status assessment method proposed in this invention.

[0062] Figure 3 This is a diagram of the improved Dyna-Q reinforcement learning structure of a pattern recognition-based rehabilitation status assessment method proposed in this invention. Detailed Implementation

[0063] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0064] refer to Figures 1-3 A pattern recognition-based method for assessing rehabilitation status includes the following steps:

[0065] Acquire patient rehabilitation training images and medical record information;

[0066] Input rehabilitation training images into the PoseNet pose estimation model to extract joint coordinates and generate a sequence of pose key points.

[0067] Input medical record information and posture key point sequences into the rehabilitation knowledge graph, and initialize the rehabilitation training plan based on the path relationships in the rehabilitation knowledge graph;

[0068] The subjective feedback data of patients during the training process is obtained. The posture key point sequence and subjective feedback data are input into the Dyna-Q reinforcement learning model. The Dyna-Q reinforcement learning model evaluates the effect of the current training plan based on the action standardization score and subjective feedback, and dynamically adjusts the training intensity and frequency, and outputs the adjusted training plan parameters.

[0069] The paths in the rehabilitation knowledge graph are updated based on the adjusted training plan parameters, and the training plan is iteratively optimized based on the daily execution feedback data using the incremental learning mechanism of the Dyna-Q reinforcement learning model.

[0070] During each pose estimation, dynamic pruning is performed on the PoseNet pose estimation model based on the complexity of the training task, selecting the network layers to be activated.

[0071] A medical safety boundary rule base is constructed to detect the sequence of key posture points and subjective feedback data in real time during rehabilitation training. When a patient reports severe pain, the training is forcibly paused and a manual intervention process is triggered. The manual correction instruction is received, and the manually corrected training plan is applied to subsequent rehabilitation training until the rehabilitation assessment indicators reach the preset standards.

[0072] In this embodiment, acquiring patient rehabilitation training images and medical record information specifically includes:

[0073] The system acquires the patient's medical record information through the system terminal and configures initial training parameters for the patient based on the medical record information. The initial training parameters include training movements, training intensity, and training frequency, generating a personalized initial rehabilitation plan; this process is completed manually by the therapist.

[0074] Patients use mobile terminal camera devices to capture real-time training video streams and extract rehabilitation training images from the real-time training video streams.

[0075] Rehabilitation training images are image frames continuously captured by a mobile terminal camera device when patients perform training movements under the guidance of a personalized initial rehabilitation plan.

[0076] In this embodiment, generating the attitude key point sequence specifically includes:

[0077] Input the rehabilitation training images into the PoseNet pose estimation model to obtain the initial pose keypoint sequence and the confidence scores of each joint coordinate; while outputting the joint coordinates, the PoseNet pose estimation model will directly output the confidence score corresponding to each joint, which is the standard output of the model.

[0078] Obtain information about the surgical site or the site of functional impairment from the medical record to identify the joints to focus on;

[0079] Joints with a confidence level below a preset threshold in the initial pose keypoint sequence are identified as joints to be completed; in this embodiment, the preset threshold is set to 0.7.

[0080] When the joint to be completed is a key focus joint, the coordinates of the visible joints adjacent to the joint to be completed are obtained. Based on the personalized bone length ratio and joint range of motion constraints in the medical record information, and combined with the historical coordinates of the joint to be completed in previous image frames, an interpolation algorithm based on human kinematic constraints is used to calculate the completed coordinates. In this embodiment, the interpolation algorithm based on human kinematic constraints predicts the initial position of the current frame using a kinematic model based on the historical coordinates of the joint to be completed in previous image frames. Using adjacent visible joints as a reference, and combining the personalized bone length ratio in the medical record information, the geometrically feasible region of the joint to be completed is determined. Based on the joint range of motion constraints in the medical record information, boundary corrections are performed on positions exceeding the physiological range. The above three constraints are weighted and fused to generate the completed coordinates.

[0081] Replace the corresponding joint coordinates in the initial pose keypoint sequence with the completed coordinates to generate a complete pose keypoint sequence that conforms to the individual anatomical characteristics of the patient.

[0082] In this embodiment, the rehabilitation knowledge graph specifically includes:

[0083] The system acquires rehabilitation medicine knowledge sources, extracts entity types and relation types from these sources, and constructs an ontology model for the rehabilitation domain. Entity types include rehabilitation actions, rehabilitation stages, indications, and surgical types. Relation types include applicable to, contraindicated to, belonging to a stage, and prerequisites. Existing rehabilitation case data is imported in batches from hospital information systems, rehabilitation databases, and electronic medical records. Each case data includes information such as the patient's surgical type, rehabilitation stage, performed actions, action parameters, and rehabilitation effect score. Entity instances and relation instances are extracted through natural language processing and rule matching. The resulting basic rehabilitation knowledge graph is stored in the form of an RDF triplet library.

[0084] Collect historical rehabilitation case data, extract entity instances and inter-entity relationship instances from the historical rehabilitation case data, map the entity instances and relationship instances to the rehabilitation domain ontology model, and generate a basic rehabilitation knowledge graph;

[0085] The individual characteristic parameters contained in the medical record information are obtained, the personalized weight of each entity instance is calculated based on the individual characteristic parameters, and the personalized weight is added as an attribute to the corresponding entity instance in the basic rehabilitation knowledge graph to generate a rehabilitation knowledge graph containing personalized weights.

[0086] The system acquires the patient's medical record information and inputs it into a rehabilitation knowledge graph containing personalized weights. It then retrieves rehabilitation paths that match the medical record information and extracts recommended training parameters from the search results. These parameters include recommended training actions, recommended training intensity, and recommended training frequency. The personalized weights are obtained by weighted summation of three parts: static matching degree, dynamic matching degree, and feedback correction factor. The initial weights are set to 0.6, 0.3, and 0.1, and can be adjusted based on clinical experience.

[0087] Static matching score is calculated by comparing patient medical record features with entity-associated attributes. For surgical type matching, if the surgical type associated with the entity is completely consistent with the patient's surgical type, the score is 1; if it is a different surgical procedure for the same site, the score is 0.5; otherwise, it is 0. For age matching, the score is calculated based on the overlap between the patient's age range and the entity's recommended age range. For contraindication matching, if the patient has a contraindication and the entity is marked as "contraindicated for" that contraindication, the total weight is changed to 0. The final score is obtained by weighting the scores of the above feature matching, with a value ranging from 0 to 1.

[0088] The dynamic matching degree reflects the degree of matching between the patient's current rehabilitation stage and the recommended stage. The rehabilitation stage is automatically determined by the system based on the number of days after the patient's surgery. Therapists can modify and correct the preset number of days according to the actual situation. The value ranges from 0 to 1.

[0089] The feedback correction factor is dynamically adjusted based on the patient's historical performance of the action, and the average of the comprehensive score is taken. The comprehensive score is obtained by fusing the action standardization score and the patient's subjective pain and fatigue score.

[0090] The recommended training parameters are presented to the therapist's terminal, and the therapist's confirmation or adjustment instructions for the recommended training parameters are received. A personalized initial rehabilitation plan is generated based on the confirmation or adjustment instructions.

[0091] In this embodiment, the action standardization score refers to the system's quantitative assessment of the degree of conformity between the patient's current action and the standard rehabilitation action, which is automatically calculated and generated by the system based on the posture key point sequence output by PoseNet;

[0092] The system pre-builds a standard movement library, which stores the standard posture template for each rehabilitation movement. The standard posture template includes the standard joint angle range, standard movement trajectory, and standard movement duration.

[0093] During patient training, the system acquires the pose key point sequence output by PoseNet in real time and extracts the set of key joint angles; it compares the key joint angles with the standard joint angle range of the corresponding movements in the standard movement library to calculate the angle matching degree; for dynamic movements, it compares the similarity between the joint movement trajectory and the standard trajectory to calculate the trajectory matching degree; it compares the deviation between the movement completion time and the standard movement duration to calculate the rhythm matching degree; and it weights and fuses the above three matching degrees to generate a movement standardization score in the range of 0 to 1.

[0094] When a patient's actions are detected to be outside the safe range, the system directly outputs a score of 0 and triggers a safety alert process.

[0095] In this embodiment, the Dyna-Q reinforcement learning model specifically includes:

[0096] The posture keypoint sequence, subjective feedback data, and training parameters corresponding to the current rehabilitation path in the rehabilitation knowledge graph are jointly constructed into a state vector. In this embodiment, the posture keypoint sequence is converted into an action feature vector. Based on the joint coordinates output by the PoseNet posture estimation model, the set of key joint angles and joint motion amplitude parameters are calculated. The set of key joint angles is normalized to form a posture feature sub-vector. The pain score and fatigue score input by the patient through the mobile terminal are numerically standardized to generate a subjective feedback sub-vector. The training action number, training intensity level, and rehabilitation stage number corresponding to the current rehabilitation path are read from the rehabilitation knowledge graph and converted into a training parameter sub-vector using discrete encoding. The posture feature sub-vector, subjective feedback sub-vector, and training parameter sub-vector are concatenated in a fixed order to form a state vector for input to the Dyna-Q reinforcement learning model.

[0097] An environment model unit is constructed to predict the state changes of patients executing the adjusted training plan, and generates simulated experience samples based on the prediction results. Real training experience samples and simulated experience samples are stored together in an experience memory set. In this embodiment, the environment model unit learns the mapping relationship between patient training behavior and state changes through historical rehabilitation execution data, which includes the patient's historical training plan parameters, postural key point sequence changes, and corresponding subjective feedback data. The environment model retrieves similar samples from historical rehabilitation case data and previous training records of the same patient based on the current state vector and training plan adjustment parameters, and generates a predicted state based on the state transition results of similar samples. The predicted state serves as the next state in the simulated experience sample, simulating the possible state changes that the patient may experience when executing the adjusted training plan. The experience memory set is used to uniformly store real training experience samples and simulated experience samples. Each experience sample consists of a state vector, training plan adjustment action, reward value, and next state vector. The experience memory set is updated chronologically and replaces the oldest stored experience sample when it reaches a preset capacity.

[0098] Q-value updates are performed based on the experience memory set, and the value function is iteratively updated using an incremental learning approach. Newly acquired daily execution feedback data covers the value estimates corresponding to historically low-relevance samples during the update process. In this embodiment, the Dyna-Q reinforcement learning model evaluates the relevance of experience samples based on the patient's current rehabilitation stage and training time window. When newly acquired daily execution feedback data enters the experience memory set, the update priority of historical experience samples that exceed the preset time window or belong to different rehabilitation stages is reduced by comparing the rehabilitation stage number of the sample and the collection time interval. During the value function update process, the experience samples corresponding to the latest stage are selected first for Q-value updates.

[0099] Upon receiving a manual correction instruction, the manually corrected training plan parameters are converted into external reward signals and input into the value update unit to perform reinforcement correction on the Q value of the corresponding state-action pair.

[0100] After completing the value function update, the output training plan adjustment parameters are synchronously written into the rehabilitation knowledge graph to update the entity weights of the corresponding rehabilitation path.

[0101] In this embodiment, performing dynamic pruning on the PoseNet pose estimation model specifically includes:

[0102] Obtain the image complexity features of the current rehabilitation training images. The image complexity features include image resolution, number of human targets, and degree of joint occlusion.

[0103] The current pose estimation task is mapped to one of several preset computational accuracy levels based on the image complexity features. In this embodiment, the preset computational accuracy levels are divided into three levels, each corresponding to a different set of network layer activations in the PoseNet pose estimation model.

[0104] Obtain the average confidence score of the action prescriptiveness score output by the Dyna-Q reinforcement learning model for historical tasks of the same accuracy level; the confidence score is calculated using the negative correlation function of the Q-value distribution entropy.

[0105] If the average confidence level is lower than the preset accuracy threshold and the system resource utilization rate does not exceed the load limit, the current task will be automatically upgraded to a higher accuracy level, activating the entire network layer to participate in the computation; the preset accuracy threshold is set to 0.75; the load limit is set to 85%.

[0106] If the average confidence level is higher than the preset accuracy threshold or the system resource utilization rate exceeds the load limit, the current task will be automatically downgraded by one level of calculation accuracy and some network layers will be skipped; the skipped network layers are preset.

[0107] Based on the final determined target accuracy level, the corresponding set of network layers is activated from the PoseNet pose estimation model, and joint coordinates are extracted to generate a sequence of pose key points.

[0108] In this implementation, the PoseNet pose estimation model employs a pre-trained multi-layer convolutional neural network structure. Dynamic pruning is achieved through a network layer skipping mechanism during the forward propagation process.

[0109] The system sets an activation flag for each convolutional layer; during the model loading stage, it generates the corresponding layer activation mask based on the final determined target computational accuracy level; during forward propagation, it traverses each convolutional layer, and only performs the computation of that layer if the activation flag is true; if the activation flag is false, the input of that layer is passed through a skip connection.

[0110] In this embodiment, the medical safety boundary rule base specifically includes:

[0111] Real-time acquisition of pain index from posture key point sequences and subjective feedback data;

[0112] The sequence of key posture points is compared with a pre-stored high-risk movement library, which stores joint angle thresholds and trajectory features. The joint angle thresholds store the critical angles of hyperextension and excessive flexion for different surgical sites and rehabilitation stages. The trajectory features store typical movement patterns that indicate an impending fall. The real-time posture trajectory is compared with the pre-stored high-risk trajectory templates for similarity. A match is determined when the similarity exceeds 0.8.

[0113] When the system detects that the posture keypoint sequence matches high-risk action features or that the pain index exceeds a preset safety threshold, it immediately interrupts the execution of the current training plan and pauses the decision output of the Dyna-Q reinforcement learning model. The pain index in the subjective feedback data adopts a visual simulation rating method, with a value range of 0 to 10, where 0 represents no pain and 10 represents severe pain. In this embodiment, the preset safety threshold is set to 7.

[0114] Generate an alert message and send it to the therapist's terminal, mark it as a triggered safety event, and record it in the safety log;

[0115] The system receives intervention instructions from the therapist's terminal, adjusts the subsequent training plan according to the instructions, and writes the adjusted training plan parameters as safety constraints into the rehabilitation knowledge graph, while simultaneously updating the personalized weights of relevant entity instances on the corresponding rehabilitation path. When the therapist returns an intervention instruction through the terminal, the system parses the instruction into quantifiable safety constraints. The safety constraints are then written into the rehabilitation knowledge graph in the form of "taboo relationships".

[0116] Upon receiving the training resumption instruction, the decision output of the Dyna-Q reinforcement learning model is restarted to continue the rehabilitation training.

[0117] In this embodiment, triggering a security event and recording it in the security log specifically includes:

[0118] When a security event is triggered by the medical safety boundary rule base, the system automatically collects the continuous posture key point sequence before the trigger time, the decision trajectory of the Dyna-Q reinforcement learning model during that time period, and the state prediction results of the corresponding environmental model unit, and constructs a causal chain data of the security event containing decision input, execution process and triggering cause; in this embodiment, the time range before the trigger time is set to 1 minute.

[0119] The causal chain data of safety incidents is stored as a structured event record and a visual playback interface is provided. The visual playback interface allows therapists to annotate key problem nodes in the event process.

[0120] The system acquires key problem nodes marked by therapists, extracts new high-risk action feature templates from the corresponding posture key point sequences, and adds these templates to the high-risk action library of the medical safety boundary rule base. When a key problem node marked by a therapist belongs to the "action error" type, the system automatically performs the following operations: extracts posture sequence segments within 0.5 seconds before and after the marked timestamp, calculates the mean, variance, and rate of change of each joint angle within the segment; extracts the joint trajectory features of the segment, including displacement direction, peak acceleration, and motion smoothness; combines these features into a new high-risk action feature template and compares its similarity with existing templates in the high-risk action library; if the similarity with existing templates is less than 0.6, it is confirmed as a new risk pattern, the new template is added to the high-risk action library, and associated with the marked surgical type and rehabilitation stage.

[0121] Based on the causal chain data of the safety incident and the therapist's annotation results, the prediction bias of the environmental model unit for high-risk actions is calculated, and the prediction parameters of the environmental model unit are corrected using the prediction bias. The correction mechanism adopts a sample-weighted update: the complete transfer sample of the safety incident is added to the retrieval library of the environmental model unit and assigned a base weight that is twice that of ordinary samples.

[0122] Example 1: To verify the feasibility of this invention in practice, it was applied to a group of patients with different types of rehabilitation training. This experiment mainly targeted patients after joint replacement surgery, fracture rehabilitation patients, and patients with other common diseases, aiming to solve some problems existing in traditional rehabilitation methods, such as shortage of therapist resources, low patient training compliance, poor rehabilitation effects or secondary injuries caused by improper movements, etc.

[0123] In traditional rehabilitation, therapists typically adjust training plans manually based on the patient's medical history and recovery progress, and monitor the patient's movements in real time during training. However, this method has some significant drawbacks. Due to the limited number of therapists, patients cannot receive sufficient attention, especially in home or community rehabilitation centers where therapists cannot constantly monitor the patient's progress. The patient's rehabilitation progress and training intensity are difficult to adjust in a timely manner based on actual performance, and often the training plan fails to be dynamically optimized based on the patient's real-time feedback, leading to unsatisfactory training results.

[0124] To address these issues, this invention employs an intelligent rehabilitation system based on pose estimation, reinforcement learning, and knowledge graphs. In the experiment, patients undergo rehabilitation training via a smart terminal device. The device captures the patient's training movements in real time using a camera and transmits the image data to a PoseNet model for pose estimation. The PoseNet model accurately extracts the patient's joint coordinates and generates a sequence of pose key points. Simultaneously, the system inputs the patient's medical record information into a rehabilitation knowledge graph and initializes a personalized rehabilitation training plan using the graph's paths.

[0125] During training, patients' subjective feedback is recorded in real time via terminal devices and input into the Dyna-Q reinforcement learning model. The Dyna-Q model combines movement standardization scores with patient feedback data to evaluate the effectiveness of the training plan in real time and dynamically adjust training intensity, frequency, and content. All training data and adjustments are synchronously updated in the rehabilitation knowledge graph, ensuring the system optimizes subsequent training plans based on the patient's individualized needs.

[0126] The system also includes a built-in medical safety boundary rule library for real-time detection of anomalies during patient training. When the system detects a patient reporting severe pain, the training automatically pauses and triggers a manual intervention process. The therapist can then adjust the instructions based on the patient's specific situation and apply them to subsequent training plans to ensure patient safety.

[0127] To verify the effectiveness of this invention, we divided 100 patients into two groups: one group used traditional rehabilitation methods, and the other group used the intelligent rehabilitation system of this invention. The experimental results showed that using the intelligent rehabilitation system of this invention not only significantly improved the patients' rehabilitation outcomes but also effectively reduced the workload of therapists and improved patient compliance and satisfaction.

[0128] To comprehensively compare the differences in rehabilitation effects between this invention and traditional methods, we conducted statistical analysis on the rehabilitation data of the two groups of patients. The table below shows the main indicators of patients in rehabilitation training, including rehabilitation time, patient satisfaction, therapist workload, training compliance, movement standardization score, pain score, and fatigue score.

[0129] Table 1: Comparison of the effects of rehabilitation methods

[0130] index This invention (intelligent rehabilitation system) Traditional method (fully therapist supervised) Recovery time (weeks) 8 10 Patient satisfaction (0-10) 9 7 Therapist workload (hours / week) 12 20 Training compliance (%) 82% 75% Performance score (0-1) 0.89 0.83 Pain score (0-10) 2 4 Fatigue rating (0-10) 3 5 Training intensity (standardized score) 8 6 Training frequency (times / week) 5 4

[0131] By comparing and analyzing the data of the two groups of patients during the rehabilitation process, it can be seen that the present invention has advantages over traditional rehabilitation methods.

[0132] In terms of rehabilitation time, approximately two weeks were saved. The intelligent system, by adjusting the training plan in real time, can optimize the intensity and frequency of training based on the patient's progress, thus improving rehabilitation efficiency. Patient satisfaction has improved. The average satisfaction score of patients using the intelligent rehabilitation system is 9 points. Personalized training programs and real-time feedback mechanisms make patients feel more involved in the rehabilitation process, increasing their satisfaction. Regarding therapist workload, the use of the intelligent system has significantly reduced the workload of therapists. The system can automatically adjust and provide feedback, reducing the need for therapist intervention. In terms of training adherence, the adherence rate of the intelligent rehabilitation system is 82%, which is an improvement compared to 75% for traditional methods. Personalized, real-time training feedback has improved patients' training enthusiasm and persistence. The motor standardization score has also improved. The intelligent system helps patients maintain more standardized movements, with an average standardization score of 0.89, compared to 0.83 for traditional methods. This indicates that the intelligent system can better guide patients to standardize movements and reduce the risks caused by non-standard movements. In terms of pain and fatigue scores, it shows that the system effectively reduces discomfort by dynamically adjusting the training intensity.

[0133] In summary, this invention demonstrates advantages in improving rehabilitation efficiency, enhancing patient satisfaction, reducing therapist workload, improving compliance, and enhancing training effectiveness through an intelligent rehabilitation management system.

[0134] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A rehabilitation status assessment method based on pattern recognition, characterized in that, Includes the following steps: Acquire patient rehabilitation training images and medical record information; Input rehabilitation training images into the PoseNet pose estimation model to extract joint coordinates and generate a sequence of pose key points. Input medical record information and posture key point sequences into the rehabilitation knowledge graph, and initialize the rehabilitation training plan based on the path relationships in the rehabilitation knowledge graph; The subjective feedback data of patients during the training process is obtained. The posture key point sequence and subjective feedback data are input into the Dyna-Q reinforcement learning model. The Dyna-Q reinforcement learning model evaluates the effect of the current training plan based on the action standardization score and subjective feedback, and dynamically adjusts the training intensity and frequency, and outputs the adjusted training plan parameters. The paths in the rehabilitation knowledge graph are updated based on the adjusted training plan parameters, and the training plan is iteratively optimized based on the daily execution feedback data using the incremental learning mechanism of the Dyna-Q reinforcement learning model. During each pose estimation, dynamic pruning is performed on the PoseNet pose estimation model based on the complexity of the training task, selecting the network layers to be activated. A medical safety boundary rule base is constructed to detect the sequence of key posture points and subjective feedback data in real time during rehabilitation training. When a patient reports severe pain, the training is forcibly paused and a manual intervention process is triggered. The manual correction instruction is received, and the manually corrected training plan is applied to subsequent rehabilitation training until the rehabilitation assessment indicators reach the preset standards.

2. The rehabilitation status assessment method based on pattern recognition according to claim 1, characterized in that, The acquisition of patient rehabilitation training images and medical record information specifically includes: The system terminal obtains the patient's medical record information and configures initial training parameters for the patient based on the medical record information. The initial training parameters include training actions, training intensity and training frequency, and generate a personalized initial rehabilitation plan. Patients use mobile terminal camera devices to capture real-time training video streams and extract rehabilitation training images from the real-time training video streams. Rehabilitation training images are image frames continuously captured by a mobile terminal camera device when patients perform training movements under the guidance of a personalized initial rehabilitation plan.

3. The rehabilitation status assessment method based on pattern recognition according to claim 2, characterized in that, The generated attitude key point sequence specifically includes: Input the rehabilitation training images into the PoseNet pose estimation model to obtain the initial pose key point sequence and the confidence level of each joint coordinate; Obtain information about the surgical site or the site of functional impairment from the medical record to identify the joints to focus on; Joints with a confidence level below a preset threshold in the initial pose keypoint sequence are identified as joints to be completed. When the joint to be completed is a key focus joint, the coordinates of the visible joints adjacent to the joint to be completed are obtained. Based on the personalized bone length ratio and joint range of motion constraints in the medical record information, and combined with the historical coordinates of the joint to be completed in the preceding image frame, the completion coordinates are calculated using an interpolation algorithm based on human kinematic constraints. Replace the corresponding joint coordinates in the initial pose keypoint sequence with the completed coordinates to generate a complete pose keypoint sequence that conforms to the individual anatomical characteristics of the patient.

4. The rehabilitation status assessment method based on pattern recognition according to claim 3, characterized in that, The rehabilitation knowledge graph specifically includes: Obtain knowledge sources in rehabilitation medicine, extract entity types and relation types from these sources, and construct an ontology model for the rehabilitation domain. Collect historical rehabilitation case data, extract entity instances and inter-entity relationship instances from the historical rehabilitation case data, map the entity instances and relationship instances to the rehabilitation domain ontology model, and generate a basic rehabilitation knowledge graph; The individual characteristic parameters contained in the medical record information are obtained, the personalized weight of each entity instance is calculated based on the individual characteristic parameters, and the personalized weight is added as an attribute to the corresponding entity instance in the basic rehabilitation knowledge graph to generate a rehabilitation knowledge graph containing personalized weights. Obtain the patient's medical record information, input the medical record information into a rehabilitation knowledge graph containing personalized weights, retrieve rehabilitation paths that match the medical record information, and extract recommended training parameters from the retrieval results; The recommended training parameters are presented to the therapist's terminal, and the therapist's confirmation or adjustment instructions for the recommended training parameters are received. A personalized initial rehabilitation plan is generated based on the confirmation or adjustment instructions.

5. The rehabilitation status assessment method based on pattern recognition according to claim 4, characterized in that, The Dyna-Q reinforcement learning model specifically includes: The posture key point sequence, subjective feedback data, and training parameters corresponding to the current rehabilitation path in the rehabilitation knowledge graph are jointly constructed into a state vector; An environmental model unit is constructed to predict the state changes of patients after implementing the adjusted training plan, and simulated experience samples are generated based on the prediction results. The real training experience samples and simulated experience samples are stored together in the experience memory set. Q-value updates are performed based on the experience memory set, and the value function is iteratively updated using an incremental learning approach. Newly acquired daily execution feedback data covers the value estimates corresponding to historically low-relevance samples during the update process. Upon receiving a manual correction instruction, the manually corrected training plan parameters are converted into external reward signals and input into the value update unit to perform reinforcement correction on the Q value of the corresponding state-action pair. After completing the value function update, the output training plan adjustment parameters are synchronously written into the rehabilitation knowledge graph to update the entity weights of the corresponding rehabilitation path.

6. The rehabilitation status assessment method based on pattern recognition according to claim 5, characterized in that, The dynamic pruning of the PoseNet pose estimation model specifically includes: Obtain the image complexity features of the current rehabilitation training images. The image complexity features include image resolution, number of human targets, and degree of joint occlusion. Based on the image complexity features, the current pose estimation task is mapped to one of several preset computational accuracy levels; Obtain the average confidence level of the action prescriptive scores output by the Dyna-Q reinforcement learning model for historical tasks of the same accuracy level. If the average confidence level is lower than the preset accuracy threshold and the system resource utilization rate does not exceed the load limit, the current task will be automatically upgraded to a higher accuracy level and the entire network layer will be activated to participate in the calculation. If the average confidence level is higher than the preset accuracy threshold or the system resource utilization rate exceeds the load limit, the current task will be automatically downgraded by one level of calculation accuracy and some network layers will be skipped. Based on the final determined target accuracy level, the corresponding set of network layers is activated from the PoseNet pose estimation model, and joint coordinates are extracted to generate a sequence of pose key points.

7. The rehabilitation status assessment method based on pattern recognition according to claim 6, characterized in that, The medical safety boundary rule base specifically includes: Real-time acquisition of pain index from posture key point sequences and subjective feedback data; The sequence of posture key points is compared with a pre-stored high-risk motion library, which stores joint angle thresholds and trajectory features. When the system detects that the posture key point sequence matches high-risk action features or that the pain index exceeds a preset safety threshold, it immediately interrupts the execution of the current training plan and pauses the decision output of the Dyna-Q reinforcement learning model. Generate an alert message and send it to the therapist's terminal, mark it as a triggered safety event, and record it in the safety log; The system receives intervention instructions from the therapist's terminal, adjusts the subsequent training plan according to the intervention instructions, and writes the adjusted training plan parameters as safety constraints into the rehabilitation knowledge graph, while simultaneously updating the personalized weights of relevant entity instances on the corresponding rehabilitation path. Upon receiving the training resumption instruction, the decision output of the Dyna-Q reinforcement learning model is restarted to continue the rehabilitation training.

8. The rehabilitation status assessment method based on pattern recognition according to claim 7, characterized in that, The specific steps for triggering a security event and recording it in the security log include: When a security event is triggered by the medical safety boundary rule base, the system automatically collects the continuous attitude key point sequence before the trigger time, the decision trajectory of the Dyna-Q reinforcement learning model during that period, and the state prediction results of the environmental model unit to construct the causal chain data of the security event. The causal chain data of safety incidents is stored as a structured event record and a visual playback interface is provided. The visual playback interface allows therapists to annotate key problem nodes in the event process. Obtain the key problem nodes marked by the therapist, extract new high-risk action feature templates from the posture key point sequence corresponding to the key problem nodes, and add the new high-risk action feature templates to the high-risk action library of the medical safety boundary rule base. Based on the causal chain data of safety events and the annotation results of therapists, the prediction bias of the environmental model unit for high-risk actions is calculated, and the prediction parameters of the environmental model unit are corrected using the prediction bias.

Images