A machine learning based personalized rehabilitation pathway dynamic planning system

Abstract

The application belongs to the technical field of rehabilitation medicine, and particularly relates to a personalized rehabilitation path dynamic planning system based on machine learning. The application discloses a personalized rehabilitation path dynamic planning system based on machine learning. The system comprises a clinical medical knowledge base, a multi-modal data acquisition terminal, a central cloud processing server, a knowledge-enhanced constraint generator, a dynamic adversarial discriminator and a nonlinear dynamic stability controller. The application converts a medical atlas into a neural network hard constraint, guides path evolution by using an adversarial game mechanism, and realizes real-time fine-tuning or reconstruction of a rehabilitation path based on a phase space model. The application realizes deep integration of medical prior knowledge and a data-driven model, and improves the personalization degree, response sensitivity and global stability of rehabilitation path planning under the premise of ensuring clinical safety.

Description

Technical Field

[0001] This invention belongs to the field of rehabilitation medicine technology, specifically relating to a personalized rehabilitation path dynamic planning system based on machine learning. Background Technology

[0002] With the rapid development of machine learning technology, its application in the field of smart healthcare is becoming increasingly widespread. In precision rehabilitation systems, constructing personalized rehabilitation pathways has become an important trend for improving clinical treatment outcomes and the efficiency of medical resource allocation. Rehabilitation pathway planning not only involves in-depth assessment of patients' functional impairments but also encompasses the development of intervention plans throughout the entire lifecycle, from early intervention to functional reconstruction. By introducing intelligent algorithms, medical systems can process massive amounts of physiological data and provide targeted training programs based on patients' recovery progress. This has clinical value for improving patients' prognosis and quality of life, as well as promoting the standardized development of rehabilitation medicine.

[0003] Data-driven personalized rehabilitation pathway dynamic planning systems are a core component of current rehabilitation robots and telemedicine platforms. They aim to achieve adaptive evolution of rehabilitation strategies by analyzing multi-source, heterogeneous physiological feedback. This type of technology focuses on using time-series prediction models to quantitatively assess rehabilitation effects and to make real-time adjustments to training intensity and movement space based on the patient's performance. Because the rehabilitation process involves complex coupling of physiological parameters and long-term dynamic changes, the system needs to possess model generalization capabilities and a deep integration of clinical knowledge to ensure that interventions achieve optimal recovery efficiency within medical safety boundaries.

[0004] Existing machine learning-based rehabilitation pathway planning schemes exhibit limitations when handling complex clinical scenarios. Traditional models rely excessively on large-scale labeled samples, making them highly susceptible to predictive bias when dealing with rare diseases or non-standard rehabilitation scenarios, thus reducing the medical credibility of the planned pathways. Existing clinical guidelines and expert experience are mostly in the form of external rules, failing to achieve deep coupling between clinical prior knowledge and learning algorithms, resulting in generated pathways that may violate basic anatomical logic and physiological laws. Current technologies often employ linear optimization strategies, which struggle to cope with non-linear fluctuations and plateaus during rehabilitation, failing to strike a balance between maintaining treatment stability and capturing pathological emergencies, leading to system lag or frequent ineffective adjustments. Summary of the Invention

[0005] The purpose of this invention is to provide a personalized rehabilitation pathway dynamic planning system based on machine learning, which can solve the problems of data and clinical knowledge separation, static model-dynamic mismatch, model failure in small sample scenarios, difficulty in digitally embedding clinical knowledge, and imbalance between response sensitivity and stability of rehabilitation pathways in the above-mentioned background technology.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A personalized rehabilitation pathway dynamic programming system based on machine learning includes a clinical medical knowledge base, a multimodal data acquisition terminal, a central cloud processing server, a knowledge-enhanced constraint generator, a dynamic adversarial discriminator, and a nonlinear dynamic stability controller, as follows: The clinical medical knowledge base is used to store a rehabilitation medicine knowledge graph, which includes logical dependencies of rehabilitation stages, contraindications for clinical operations, safety boundaries of physiological indicators, and standardized rehabilitation procedures for different diseases, providing underlying medical logic support for the system.

[0007] The multimodal data acquisition terminal is used to acquire in real time the patient's physiological electrical signals, kinematic posture data, behavioral modal characteristics and patient subjective evaluation information during the rehabilitation training process, and to perform synchronous preprocessing on the acquired multi-source heterogeneous data before transmitting it to the central cloud processing server.

[0008] The central cloud processing server is used to run the rehabilitation pathway planning algorithm architecture, coordinate the data flow and logical calculations between various components, and store the patient's historical rehabilitation records and real-time dynamic data.

[0009] The knowledge-enhanced constraint generator is used to receive patient status data transmitted from the central cloud processing server and medical logic constraints in the clinical medical knowledge base. Through a built-in clinical prior encoder, it transforms medical knowledge into hard-edge constraints of a neural network and generates personalized rehabilitation paths within a specific safe and feasible domain defined by the medical knowledge.

[0010] The dynamic adversarial discriminator is used to evaluate the rehabilitation path output by the knowledge-enhanced constraint generator. It internally constructs a virtual rehabilitation expert model based on expert decision distribution. By inputting the generated rehabilitation path and the patient's historical multidimensional rehabilitation data, it outputs a clinical fit score representing the rationality of the path and forms an adversarial game mechanism with the knowledge-enhanced constraint generator to guide the path generation towards medical effectiveness.

[0011] The nonlinear dynamic stability controller is used to monitor the nonlinear fluctuations in the patient's rehabilitation process in real time. By constructing a high-dimensional phase space model, the patient's rehabilitation process is fitted into a motion trajectory. Based on the degree of deviation of the current state relative to the preset optimal attractor, the decision system performs path fine-tuning or path reconstruction.

[0012] Preferably, the knowledge-enhanced constraint generator integrates a graph neural network architecture. The nodes of the graph neural network correspond to medical entities in the clinical medical knowledge base, and the edges correspond to the physiological coupling relationships between entities. The clinical prior encoder implants attention biases in each layer of the graph neural network, so that when the generator explores the path action space, it automatically avoids action nodes involving contraindications or violating anatomical logic, ensuring that each generated rehabilitation action and its intensity and frequency are within the medically permissible threshold range.

[0013] Furthermore, when planning a rehabilitation path, the knowledge-enhanced constraint generator defines the order of rehabilitation stages as a topological structure with strong temporal constraints. This topological structure requires that specific functional indicators reach a preset recovery threshold before the training actions of subsequent stages can be initiated, thereby realizing the underlying semantic embedding of clinical knowledge in the generative model.

[0014] Preferably, the dynamic adversarial discriminator is continuously evolved using a deep reinforcement learning algorithm. Its training data includes not only successful rehabilitation cases but also failed cases or cases with substandard results identified through clinical evaluation. By identifying feature patterns in failed cases, the dynamic adversarial discriminator assigns a lower clinical fit score to the potential risk paths output by the knowledge-enhanced constraint generator, forcing the generator to correct its internal weights through backpropagation, thereby enhancing the system's ability to identify and reason about complex clinical scenarios.

[0015] Furthermore, when calculating the clinical fit score, the dynamic adversarial discriminator comprehensively considers the objective indicators fed back by physiological sensors, the professional evaluation recorded by the therapist, and the subjective feelings of the patient's self-evaluation. Through a dynamic weight allocation mechanism, it weights and fuses data from different sources to ensure that the evaluation results can reflect the patient's true recovery status and the degree of fit of the path plan in multiple dimensions.

[0016] Preferably, the nonlinear dynamic stability controller uses Lyapunov stability theory to construct stability discrimination logic, and determines the fluctuation attributes of the current rehabilitation progress by calculating the geometric distance and movement speed between the patient's current rehabilitation state and the target path attractor in phase space in real time.

[0017] Furthermore, when the nonlinear dynamic stability controller detects that the fluctuation of the patient's state is within a preset small oscillation range, it determines that the patient is in a normal physiological fluctuation state. The nonlinear dynamic stability controller sends a fine-tuning instruction to the knowledge-enhanced constraint generator, and outputs progressive path correction suggestions by adjusting the parameters of the generated model to maintain the continuity of the rehabilitation plan.

[0018] Furthermore, when the nonlinear dynamic stability controller detects that the patient's position in phase space deviates from the preset attractor by a distance exceeding a specific threshold and the duration exceeds a predetermined time length, it determines that the system has entered the phase transition critical point, i.e., a plateau or pathological deterioration occurs. The nonlinear dynamic stability controller immediately triggers the path reconstruction mechanism, instructing the knowledge-enhanced constraint generator to re-execute global path planning based on the current abnormal state, and generate a new rehabilitation branch path that bypasses the current obstacle.

[0019] Preferably, the clinical medical knowledge base supports incremental updates. When new rehabilitation technical specifications or medical research results are generated, the central cloud processing server automatically updates the association rules in the rehabilitation medicine knowledge graph through a knowledge extraction algorithm, so that the prior constraint domain of the knowledge-enhanced constraint generator can evolve synchronously with medical progress.

[0020] Furthermore, the multimodal data acquisition terminal has a data quality monitoring function, which can automatically identify sensor detachment, signal interference or abnormal noise. When the data quality is lower than the preset signal-to-noise ratio threshold, the central cloud processing server will switch to a compensation mode based on historical trend prediction and send an early warning signal to the rehabilitation therapist terminal to ensure the data integrity of the dynamic planning process.

[0021] Preferably, the machine learning-based personalized rehabilitation path dynamic planning system further includes an interactive display terminal, which is used to visualize the rehabilitation path scheme generated by the knowledge-enhanced constraint generator and to trace and display the medical knowledge nodes on which each path suggestion is based, providing decision support for rehabilitation therapists and enhancing the interpretability of the system output results.

[0022] Furthermore, the central cloud processing server is connected to the multimodal data acquisition terminal and interactive display terminal through an encrypted communication link to ensure the security of patients' physiological privacy data during transmission and storage, and to ensure real-time response of path planning calculation in scenarios with multiple concurrent user requests through a load balancing mechanism.

[0023] Furthermore, when making path reconstruction decisions, the nonlinear dynamic stability controller uses the patient's historical tolerance information as a key input variable to ensure that the reconstructed path matches the patient's current physiological compensatory capacity in terms of training load, thus avoiding secondary damage caused by excessive path span.

[0024] Preferably, the knowledge-enhanced constraint generator is also equipped with adaptive action space scaling logic, which dynamically adjusts the complexity of selectable actions based on the patient's daily fatigue score. When the fatigue score is higher than a preset safety threshold, the generator automatically locks high-risk action nodes and generates paths only in the set of basic maintenance actions, thus achieving all-weather rehabilitation safety assurance.

[0025] Furthermore, the dynamic adversarial discriminator has a long-term consistency verification function. By comparing the patient's performance data in different rehabilitation cycles, it assesses whether the currently generated path plan is conducive to long-term functional reconstruction, rather than merely pursuing short-term indicator improvement, thus ensuring the scientific nature of the rehabilitation path in the macro dimension.

[0026] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention addresses the problem of missing medical logic in rehabilitation path planning caused by strong data dependence in traditional machine learning models by constructing a knowledge-enhanced constraint generator and deeply integrating rehabilitation medicine knowledge graphs into the underlying logic of graph neural networks. By transforming prior clinical knowledge into hard constraints in the generation process, the system can still generate safe paths that conform to anatomical rules and clinical norms when facing rare diseases or small-sample rehabilitation scenarios, thus improving the system's clinical effectiveness and applicability.

[0027] 2. This invention introduces a dynamic adversarial discriminator and a game-theoretic learning mechanism, continuously discriminating and providing feedback on the generated paths by simulating the decision-making logic of experienced rehabilitation experts. This negative feedback learning loop enables the system to extract deep features from successful and unsuccessful cases, achieving simulation of the clinical reasoning process. The system not only optimizes rehabilitation outcomes but also proactively identifies and avoids potential medical risks, ensuring that the planned paths have higher fit and reliability in complex and ever-changing clinical environments.

[0028] 3. This invention innovatively applies nonlinear dynamics theory to the dynamic adjustment of rehabilitation pathways through a nonlinear dynamic stability controller. By accurately modeling the patient's rehabilitation trajectory in phase space, the system can scientifically distinguish between normal physiological fluctuations and pathological state changes. This mechanism achieves a dialectical unity of stable following and flexible reconstruction, avoiding frequent path changes caused by oversensitivity while ensuring timely response when entering a plateau or experiencing deterioration, thus improving rehabilitation efficiency and alleviating patient anxiety.

[0029] 4. This invention constructs a highly interpretable human-computer collaboration architecture. Because the generation process of the rehabilitation pathway is controlled by entities and relationships in the clinical medical knowledge base, every instruction output by the system can be traced back to specific medical evidence. This provides rehabilitation therapists with clear and transparent auxiliary decision-making information, enhances the trust of both doctors and patients in the intelligent system, and achieves a complementary advantage between professional clinical experience and high-performance computing capabilities, providing a solid technical guarantee for realizing precise and personalized intelligent rehabilitation.

[0030] 5. The system architecture of this invention possesses excellent scalability and adaptability. Through modular design, the system can integrate multi-dimensional physiological feedback data and dynamically update the knowledge base according to the latest clinical guidelines. This continuous evolution capability ensures that the rehabilitation pathway planning system can keep pace with the forefront of medical development, achieving an optimal balance between different rehabilitation stages and diverse individual needs, thus possessing clinical application value and social benefits. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the overall technical solution architecture according to the present invention; Figure 2 This is a schematic diagram of data flow according to the present invention; Figure 3 This is a schematic diagram illustrating the principle framework of the knowledge-enhanced constraint generator in this invention, which generates paths within the medical safety feasible domain through a clinical prior encoder. Figure 4 The following is a logical flowchart of the adversarial game mechanism formed by the dynamic adversarial discriminator of the present invention using a virtual rehabilitation expert model and a generator; Figure 5 This is a flowchart illustrating the logical flow of the nonlinear dynamic stability controller based on a high-dimensional phase space model for fluctuation monitoring and decision feedback according to the present invention. Figure 6 This is a logical flowchart illustrating the process of implementing medical knowledge node traceability display and path scheme visualization in the interactive display terminal of the present invention. Detailed Implementation

[0032] Example 1: Please refer to the appendix Figure 1 To be continued Figure 6 A personalized rehabilitation pathway dynamic planning system based on machine learning includes a clinical medical knowledge base, a multimodal data acquisition terminal, a central cloud processing server, a knowledge-enhanced constraint generator, a dynamic adversarial discriminator, and a nonlinear dynamic stability controller.

[0033] The clinical medical knowledge base is configured to store and manage a highly structured rehabilitation medicine knowledge graph. This knowledge graph is not a simple collection of entries, but a knowledge network with deep semantic relationships constructed through a resource description framework. The rehabilitation medicine knowledge graph contains logical dependencies between rehabilitation stages, defining the sequential logic between different rehabilitation stages and the functional criteria that must be met to enter the next stage. The clinical medical knowledge base stores clinical contraindications, used to define the range of actions or intensity thresholds that are absolutely prohibited during rehabilitation training. The base also includes safety boundaries for physiological indicators, such as heart rate fluctuation range, lower limit of blood oxygen saturation, and extreme values ​​of blood pressure fluctuation, as well as standardized rehabilitation procedures for different diseases (such as hemiplegia after stroke, spinal cord injury, and post-fracture surgery). The clinical medical knowledge base supports incremental updates. When new rehabilitation technical specifications or medical research results are generated, the system uses a built-in knowledge extraction engine and natural language processing technology to automatically extract new entities and relationships from medical literature, achieving self-evolution of the knowledge base and providing underlying medical logic support and safety barriers for the entire planning system.

[0034] The multimodal data acquisition terminal consists of a hardware sensor array distributed across key parts of the patient's body. It is configured to acquire multi-source heterogeneous data from the patient in real-time and at high frequency during rehabilitation training. The acquisition dimensions include, but are not limited to: physiological electrical signals (such as electromyography and electrocardiogram signals), kinematic posture data (joint angles, angular velocities, and linear accelerations acquired through inertial measurement units and optical sensors), behavioral modal characteristics (such as gait cycles, movement smoothness, and synergistic movement indices), and subjective patient evaluation information acquired through an interactive interface (such as pain perception and fatigue level recorded by visual analog scales). The multimodal data acquisition terminal integrates a data preprocessing unit for filtering and denoising the raw signals, baseline drift correction, and multi-channel synchronization. In terms of transmission logic, the terminal has a data quality monitoring function, capable of calculating the signal-to-noise ratio in real time. When sensor detachment, severe signal interference, or abnormal noise exceeding a preset ratio is detected, the terminal automatically identifies the abnormal state and notifies the central server. If the data quality is lower than the preset signal-to-noise ratio threshold, the system will activate the historical trend prediction compensation mode based on long short-term memory networks, using the patient's previous health data to fill in the missing values, ensuring data continuity, and issuing a visual warning to the therapist's terminal.

[0035] The central cloud processing server, serving as the computing and data hub of the entire system, is configured to run complex rehabilitation pathway planning algorithms. This server is built using a high-performance computing cluster with a containerized computing environment deployed internally, responsible for coordinating data flow between the clinical medical knowledge base, generator, discriminator, and controller. Its storage unit maintains the patient's full-cycle historical rehabilitation records, including historical training records, functional assessment scale scores, imaging examination results, and real-time dynamic vital signs. The central cloud processing server connects to the multimodal data acquisition terminal and subsequent interactive display terminals via encrypted communication links (such as using a secure transport layer protocol). To handle high-concurrency rehabilitation requests, the server implements a dynamic load balancing mechanism, scheduling GPU processing units in real-time for deep learning model inference and training based on the urgency of the computing tasks and the availability of computing resources, ensuring real-time response in pathway planning decisions.

[0036] The knowledge-enhanced constraint generator is the core module designed to generate personalized rehabilitation plans. Instead of blindly learning data distribution, it receives patient status data from a central cloud processing server and retrieves corresponding medical logical constraints from a clinical medical knowledge base. Internally, it deeply integrates a graph neural network architecture, where nodes correspond to medical entities in the knowledge graph (such as specific muscle groups, specific movements, and specific rehabilitation indicators), and edges correspond to physiological coupling or causal logical relationships between entities. The knowledge-enhanced constraint generator incorporates a clinical prior encoder, which transforms qualitative rules from medical knowledge into quantitative hard-edge constraints and node attention biases in the neural network. In the search space of the generated paths, the encoder sets the weights of movement paths that violate anatomical logic or involve contraindications to a minimum value (approaching 0), exploring paths within a specific safe and feasible domain defined by medical knowledge. This mechanism ensures that every generated rehabilitation movement and its intensity and frequency parameters are logically medically interpretable and safe.

[0037] The dynamic adversarial discriminator is configured to deeply review and score the initial paths output by the knowledge-enhanced constraint generator. Internally, this discriminator constructs a virtual rehabilitation expert model based on expert decision distribution. Unlike traditional discriminators, it not only distinguishes between genuine and fake data but also focuses on the clinical rationality of the paths. The discriminator's input receives the path scheme output by the generator and the patient's multidimensional historical rehabilitation data (including objective sensor data, therapist's professional notes, and the patient's psychological assessment results). The discriminator outputs a value between 0 and 1, representing the clinical fit score. The higher the score, the more the path aligns with the clinical decision-making habits of experienced rehabilitation experts. A game-theoretic learning loop is constructed between the dynamic adversarial discriminator and the knowledge-enhanced constraint generator: the generator aims to produce paths that achieve high scores, while the discriminator enhances its ability to identify seemingly reasonable but actually ineffective paths through reinforcement learning by introducing a large number of negative cases (such as cases of poor rehabilitation outcomes or secondary injuries). This adversarial process continuously adjusts the generator's internal weights through a backpropagation algorithm, gradually enabling it to master complex clinical reasoning logic.

[0038] The nonlinear dynamic stability controller is configured as the dynamic monitoring and closed-loop feedback center of the entire system, mainly used to solve the nonlinear fluctuation problem in the rehabilitation process. This controller utilizes nonlinear dynamics theory to model the patient's current rehabilitation process as a point moving in a high-dimensional phase space, while the ideal rehabilitation goal is defined as an attractive optimal attractor trajectory. The controller receives multimodal data streams in real time and calculates the geometric distance of the current state point relative to the preset attractor trajectory and its velocity in phase space. Based on Lyapunov stability theory, the controller constructs a hierarchical discrimination logic: when the calculated distance and velocity are within a preset small-amplitude oscillation range, the patient is determined to be in a normal physiological fluctuation state, and the controller sends fine-tuning parameters to the generator for progressive path optimization to maintain training continuity; however, when the distance exceeds a specific threshold and the duration exceeds a predetermined length, the system is determined to have entered a phase transition critical point, i.e., a rehabilitation problem, plateau, or pathological deterioration occurs. The controller immediately forces a path reconstruction mechanism, requiring the generator to jump out of the current action set and perform global path branch planning based on the latest abnormal state.

[0039] Furthermore, the knowledge-enhanced constraint generator defines the sequence of rehabilitation stages as a topological structure with strong temporal constraints when constructing the path plan. Under this structure, functional indicators (such as joint range of motion and muscle strength level) must reach a preset recovery threshold before the system allows the activation of training action nodes for the corresponding subsequent stages. For example, in lower limb rehabilitation planning, if the patient's sitting balance ability does not meet the preset numerical requirements, the generator's action space will automatically lock the action nodes related to walking training. This underlying semantic embedding mechanism eliminates the risk of overtraining across stages from the physical architecture level. The generator is also equipped with adaptive action space scaling logic, which dynamically adjusts the complexity and load level of selectable actions based on the patient's daily fatigue score. When the fatigue score is higher than a preset safety threshold, the generator automatically enters a protection mode, selecting only low-intensity rehabilitation tasks in the basic maintenance action set.

[0040] Furthermore, the dynamic adversarial discriminator employs a dynamic weighting mechanism when calculating the clinical fit score. Throughout different rehabilitation cycles, the discriminator automatically switches the weight given to data from different sources. For example, in the early stages of rehabilitation, the discriminator assigns higher weight to objective indicators from physiological sensors (such as heart rate balance) to ensure safety; in the later stages of rehabilitation, it assigns higher weight to therapist evaluations and patient self-assessments to assess the patient's functional integration level and psychological adaptability. This discriminator also features a long-term consistency verification function, comparing the patient's historical performance across multiple rehabilitation cycles to assess whether the currently generated path is conducive to achieving long-term functional reconstruction goals, thus preventing the system from falling into the trap of pursuing short-term indicator improvements at the expense of long-term recovery potential.

[0041] Furthermore, when executing path reconstruction decisions, the nonlinear dynamic stability controller uses the patient's historical tolerance information (such as the upper limit of exercise intensity during past pain episodes) as a key input variable. When generating new branch paths, the controller mandates that the training load growth slope of the new path be lower than the average growth rate of the patient's historical tolerance, ensuring that the reconstructed path can both overcome the current rehabilitation plateau and avoid motor injury due to excessive path span. In the phase space model, this control strategy manifests as a curvature constraint on the motion trajectory of state points, ensuring its smooth evolution towards new attractors.

[0042] Furthermore, the machine learning-based personalized rehabilitation pathway dynamic planning system also includes an interactive display terminal. This terminal is configured as a visual front-end device connected to the central cloud server, such as a tablet or large-screen monitor. The interactive display terminal not only displays the generated rehabilitation pathway steps, but more importantly, it has a medical knowledge traceability display function. When a rehabilitation therapist clicks on a training suggestion, the terminal extracts relevant medical knowledge nodes from the clinical medical knowledge base, graphically displaying the anatomical basis, physiological principles, and corresponding clinical guidelines behind the suggestion. This transparent design enhances the system's interpretability, enabling rehabilitation therapists to understand the algorithm's logic and, when necessary, manually fine-tune the generated constraint boundaries through the interactive interface, achieving highly efficient human-machine collaborative decision-making.

[0043] Example 2: As an alternative and supplement to Example 1, this example describes a personalized rehabilitation path dynamic planning system based on edge computing and distributed storage architecture. This architecture is particularly suitable for large-scale remote home rehabilitation scenarios, which can reduce the system's dependence on the central cloud network and improve the local processing capability of privacy data.

[0044] In this embodiment, the system also includes a clinical medical knowledge base, a multimodal data acquisition terminal, a central cloud processing server, a knowledge-enhanced constraint generator, a dynamic adversarial discriminator, and a nonlinear dynamic stability controller, but it has been specifically optimized in terms of hardware deployment and logical collaboration mode.

[0045] In this embodiment, the clinical medical knowledge base adopts a distributed deployment model. Its core global knowledge base is stored in the central cloud processing server, while sub-knowledge bases for specific diseases are cached in edge computing gateways deployed on the user side. These sub-knowledge bases contain commonly used medical rules, safety thresholds, and corresponding rehabilitation path templates highly relevant to the specific patient's condition. This distributed storage structure allows the system to maintain basic security monitoring and path generation functions based on the locally cached sub-knowledge bases even in the event of network interruptions or high latency.

[0046] In this embodiment, the multimodal data acquisition terminal integrates an edge processing unit. This unit is configured to perform preliminary feature extraction at the beginning of data acquisition, such as extracting feature vectors from the raw electromyographic signals using wavelet transform algorithms, instead of uploading massive amounts of raw sampled data. This not only reduces the burden on the communication link, but more importantly, by performing the feature transformation of sensitive physiological data locally, privacy desensitization is achieved, ensuring that the patient's original biological information does not leave the local terminal environment.

[0047] In this embodiment, the knowledge-enhanced constraint generator is divided into a cloud-based global optimization module and an edge-based real-time fine-tuning module. The cloud-based global optimization module periodically analyzes the patient's long-term rehabilitation trend, utilizes its powerful computing resources to train and update the parameters of a complex graph neural network, and distributes the updated model weights to the edge gateway. The edge-based real-time fine-tuning module then dynamically corrects the rehabilitation path with low latency within a relatively small parameter space based on the patient's daily physiological fluctuations, climate changes, and the patient's immediate feedback. This cloud-edge collaborative generation mechanism ensures both the long-term scientific validity of the planning scheme and meets the need for immediate response in a home-based rehabilitation environment.

[0048] In this embodiment, the dynamic adversarial discriminator incorporates a federated learning mechanism. Due to the highly heterogeneous and privacy-sensitive nature of patient data distribution in home rehabilitation scenarios, the discriminator's virtual rehabilitation expert model iterates using a federated learning protocol. Each edge computing gateway performs gradient descent on the discriminator model locally using local data, uploading only the encrypted model gradients to the central cloud processing server for aggregation, and then feeding back the aggregated global discriminator parameters to each terminal. This approach allows the system to continuously enhance its clinical reasoning and risk identification capabilities by leveraging experience from diverse cases worldwide, while protecting patient privacy.

[0049] In this embodiment, the nonlinear dynamic stability controller is specifically enhanced to address the uncertainties of the home environment. The controller is configured to incorporate home physical environment parameters (such as floor friction coefficient, ambient light level, and temperature) as additional external perturbation variables into the high-dimensional phase space model. When the controller detects that environmental perturbations may cause the patient's state to deviate from the optimal attractor, it automatically shrinks the action space of the knowledge-enhanced constraint generator. For example, in the early morning when lighting is insufficient, the controller automatically reduces the difficulty gradient of balance training movements to compensate for the risk of falls caused by environmental factors.

[0050] Furthermore, in this embodiment, the interactive display terminal is integrated into the patient's mobile smart device or smart TV system. This terminal uses augmented reality technology to project the generated rehabilitation path movements onto the patient's real-time motion images. The interactive display terminal is configured to interact in real-time with a nonlinear dynamic stability controller. When the controller determines that the patient's movements have deviated, the terminal provides correction guidance through the augmented reality layer and displays a real-time compliance score of the current movement given by a dynamic adversarial discriminator, guiding the patient to complete high-quality rehabilitation training in an unattended home environment.

[0051] Furthermore, in this embodiment, the central cloud processing server also acts as a knowledge scheduler. Based on the rehabilitation effects reported from each edge device, it analyzes the contribution rate of different medical knowledge constraints to the actual rehabilitation progress. If a certain medical constraint shows a low degree of fit in multiple similar cases, the server will trigger the self-checking logic of the clinical medical knowledge base, prompting medical experts to review and correct the timeliness or scope of application of that knowledge node, thus improving the accuracy of the system's underlying logic in a closed loop.

[0052] Example 3: As a further detailed implementation of Examples 1 and 2, this example describes a personalized rehabilitation path dynamic planning system that focuses on multi-agent collaborative control, aiming to solve the conflict resolution and multi-objective optimization problem of rehabilitation path planning in complex comorbid scenarios (such as patients suffering from cardiovascular disease and orthopedic movement disorders at the same time).

[0053] In this embodiment, the knowledge-enhanced constraint generator is constructed as a multi-agent collaborative architecture, including a first path generation submodule (focusing on musculoskeletal system reconstruction) and a second path generation submodule (focusing on cardiopulmonary function maintenance). These two submodules are each connected to a corresponding professional knowledge branch in the clinical medical knowledge base and generate path suggestions for their respective domains based on the same set of multimodal data streams.

[0054] In this embodiment, the central cloud processing server runs a conflict resolution algorithm layer. When a conflict arises between the high-intensity weight-bearing training proposed by the first path generation submodule and the heart rate protection limit proposed by the second path generation submodule (e.g., the exercise intensity required for orthopedic rehabilitation causes the heart rate to exceed the safety threshold of cardiology), the conflict resolution algorithm layer calls the stability analysis data provided by the nonlinear dynamic stability controller. The conflict resolution layer is configured to prioritize vital sign safety, and by dynamically adjusting the attention bias, it forcibly reduces the intensity of the conflicting movements until the cardiopulmonary indicators return to the preset stability domain within the phase space.

[0055] In this embodiment, the dynamic adversarial discriminator evolves into a multi-dimensional evaluation matrix. It not only assesses the clinical fit of the pathway but also introduces a rehabilitation cost function discriminative dimension to calculate the physiological and psychological energy consumed in achieving specific rehabilitation goals. The discriminator is configured to penalize pathways that, while effective, result in low patient psychological tolerance. This mechanism forces the generator to seek a Pareto optimal solution between the often conflicting goals of rehabilitation speed and patient comfort when searching for pathways.

[0056] In this embodiment, the nonlinear dynamic stability controller incorporates multi-scale time window analysis. It is configured to simultaneously monitor stability at both the microscale (seconds, for fall prevention and acute sign monitoring) and the macroscale (days, for rehabilitation trend monitoring). At the microscale, when the controller detects chaotic signs in the patient's movement trajectory via the Lyapunov index, it immediately cuts off the power output to prevent accidents. At the macroscale, it analyzes the long dwell time of the state point near the phase space attractor to determine whether the patient has entered the rehabilitation pseudo-plateau phase, and based on this, the instruction generator introduces novel rehabilitation variables with appropriate stimulation to activate neural plasticity.

[0057] Furthermore, in this embodiment, the multimodal data acquisition terminal adds a computer vision-based micro-expression analysis unit. This unit is configured to capture the facial micro-expression features of the patient when performing specific path actions, extracting feature components representing pain, anxiety, or a sense of accomplishment. These emotional features are input as nonlinear features into the dynamic adversarial discriminator. The discriminator analyzes the correlation between emotional features and physiological electrical signals to determine whether the current rehabilitation intensity exceeds the patient's psychological defense boundary. For example, if the patient's electromyographic signals show that they still have energy but their micro-expression shows extreme fear, the discriminator will lower the clinical fit score of that path, prompting the generator to reconstruct a path branch with higher psychological safety.

[0058] Furthermore, in this embodiment, the clinical medical knowledge base includes a dedicated drug-motor interaction subgraph. This subgraph defines the potential impact of various clinical medications the patient is taking on motor performance and physiological responses (e.g., some medications may cause low resting heart rate or exercise-induced hypotension). The knowledge-enhanced constraint generator incorporates drug factors as underlying correction parameters when planning paths. For example, when it detects that the patient has taken a specific diuretic on a given day, the generator automatically locks training modules involving a high risk of dehydration. This interdisciplinary, in-depth constraint further enhances the system's professional depth and safety in handling complex clinical conditions.

[0059] Furthermore, in this embodiment, the interactive display terminal supports scheme simulation. Before formally executing the generated rehabilitation path, the therapist can start simulation prediction on the terminal. The system uses stored historical files and current path parameters to pre-simulate the patient's possible functional evolution trajectory over a future period in virtual space. This evolution trajectory is calculated by the central cloud server based on a nonlinear dynamics model. If the simulation results show that the state point has a high probability of deviating from the optimal attractor, the therapist can manually intervene through the interactive terminal to modify specific hard constraint weights in the generator, achieving path pre-intervention based on prediction.

[0060] The system in the above embodiments, through precise logical interlocking and data interaction between modules, not only solves the black-box problem commonly encountered in the application of machine learning in the field of rehabilitation, but also, through the introduction of knowledge graphs and dynamic systems, enables rehabilitation path planning to possess the rigor of a senior therapist and the flexibility to cope with complex and ever-changing physiological states.

[0061] The personalized rehabilitation pathway dynamic planning system based on machine learning provided in this invention has the core advantage of robust and evolvable architecture. By combining rigid medical knowledge with a flexible data-driven model, the system can maintain diagnostic and planning accuracy even in small sample environments. This interdisciplinary technological integration provides a new paradigm for precision rehabilitation in the field of smart healthcare, not only shortening the functional recovery period for patients but also demonstrating significant engineering practice value and social benefits in reducing medical costs and alleviating the shortage of rehabilitation resources.

[0062] In practical deployment, the central cloud processing server can also interface with the hospital's electronic medical record system through a standardized interface. By quickly extracting the patient's past medical history, allergy history, and surgical records, the system can automatically match the closest initial rehabilitation attractor track from the clinical medical knowledge base when the patient first connects, shortening the system's initial cold start time. The system architecture supports cross-device communication protocols, enabling real-time command interaction with various types of rehabilitation exoskeletons, gait reduction robots, and functional electrical stimulation devices. The path commands output by the nonlinear dynamic stability controller can be converted into torque compensation parameters or electrical stimulation pulse parameters for the underlying hardware devices, achieving the integration of planning decisions and physical execution.

[0063] To address the common psychological plateau in rehabilitation, the system described in this invention utilizes a game-theoretic mechanism of a dynamic adversarial discriminator to proactively design motivating pathways that are moderately challenging yet remain within a medically safe range. These pathways, precisely guided by a nonlinear dynamic stability controller, provide patients with sufficient functional feedback stimulation without causing physiological damage. This biofeedback-based pathway evolution logic not only focuses on the recovery of limb function but also considers the maintenance of the patient's psychological state during long-term rehabilitation, embodying a patient-centered modern rehabilitation philosophy.

[0064] The system's security mechanism permeates every stage of data flow. From signal-to-noise ratio monitoring at the multimodal data acquisition terminal, to encrypted transmission on the central cloud processing server, to the medical prior hard constraints of the knowledge-enhanced constraint generator, and the Lyapunov discrimination of the nonlinear dynamic stability controller, multi-layered risk avoidance logic collectively constructs an all-weather, multi-dimensional rehabilitation safety net. Even in the event of occasional prediction failures in the machine learning model, the underlying clinical medical knowledge base acts as the final physical circuit breaker, forcibly locking rehabilitation actions within a safe area. This balance between determinism and randomness is a characteristic of this invention's system compared to traditional purely data-driven models.

[0065] In summary, this invention, through deep collaboration of multiple modules, not only achieves automated planning of rehabilitation pathways but also makes substantial technological advancements in the scientific rigor, safety, personalization, and interpretability of dynamic planning. By combining motion stability analysis in high-dimensional phase space with logical game theory in generative adversarial networks, the system successfully captures and addresses the extremely complex nonlinear physiological process of human rehabilitation. Whether in centralized applications in professional rehabilitation institutions or decentralized applications in the homes of patients, this system provides stable, efficient, and professional rehabilitation decision support, demonstrating broad industrial application prospects.

[0066] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. In all respects, the embodiments should be considered exemplary and not restrictive, and the scope of the invention is defined by the appended claims rather than the foregoing description; therefore, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A personalized rehabilitation pathway dynamic planning system based on machine learning, characterized in that, include: A clinical medicine knowledge base is used to store a rehabilitation medicine knowledge graph that includes logical dependencies in the rehabilitation stage, contraindications for clinical procedures, safety boundaries of physiological indicators, and standardized rehabilitation processes. A multimodal data acquisition terminal is used to acquire multi-source heterogeneous data of patients during rehabilitation training in real time, and transmit the multi-source heterogeneous data after synchronization preprocessing. The central cloud processing server is connected to the multimodal data acquisition terminal and is used to run the rehabilitation path planning algorithm architecture, coordinate the data flow within the system, and store the patient's historical rehabilitation records and real-time dynamic data. A knowledge-enhanced constraint generator is connected to the central cloud processing server and the clinical medical knowledge base, respectively. It is used to transform medical knowledge into hard-edge constraints of neural networks through a built-in clinical prior encoder, and generate personalized rehabilitation path schemes within the safe and feasible domain defined by medical knowledge. A dynamic adversarial discriminator is used to construct a virtual rehabilitation expert model based on expert decision distribution. By inputting the generated rehabilitation path and the patient's historical multidimensional rehabilitation data, it outputs a clinical fit score representing the rationality of the path and forms an adversarial game mechanism with the knowledge-enhanced constraint generator. The nonlinear dynamic stability controller is used to fit the patient's rehabilitation process into a motion trajectory by constructing a high-dimensional phase space model. Based on the degree of deviation of the current rehabilitation state from the preset optimal attractor, it decides to execute path fine-tuning instructions or path reconstruction instructions.

2. The personalized rehabilitation pathway dynamic planning system based on machine learning according to claim 1, characterized in that: The knowledge-enhanced constraint generator integrates a graph neural network architecture. The nodes in the graph neural network architecture correspond to medical entities in the clinical medical knowledge base, and the edges in the graph neural network architecture correspond to physiological coupling relationships or causal logic relationships between medical entities in the clinical medical knowledge base. The clinical prior encoder is used to implant attention biases in each layer of the graph neural network architecture, so that when the knowledge-enhanced constraint generator explores the path action space, it sets the weights of action paths involving contraindications or violating anatomical logic to a minimum value close to 0, ensuring that each generated rehabilitation action and its corresponding intensity and frequency parameters are within the medically permissible threshold range. The knowledge-enhanced constraint generator also defines the order of rehabilitation stages as a topological structure with strong temporal constraints. The topological structure with strong temporal constraints requires that the functional indicators can only be activated after reaching the preset recovery threshold, so as to realize the underlying semantic embedding of clinical knowledge in the generative model.

3. The personalized rehabilitation path dynamic planning system based on machine learning according to claim 2, characterized in that: The dynamic adversarial discriminator is configured to evolve using a deep reinforcement learning algorithm, and its training data includes successful rehabilitation cases, failed cases, and cases with substandard results as identified through clinical evaluation. The dynamic adversarial discriminator identifies the characteristic patterns in the failed cases, assigns a low clinical fit score to the potentially risky paths output by the knowledge-enhanced constraint generator, and corrects the internal weights of the knowledge-enhanced constraint generator through a backpropagation algorithm. The dynamic adversarial discriminator is equipped with a dynamic weight allocation mechanism when calculating the clinical fit score, which comprehensively considers the objective indicators fed back by physiological sensors, the professional evaluation recorded by the therapist, and the subjective feelings of the patient's self-evaluation. In the early stages of rehabilitation, the dynamic adversarial discriminator assigns high weights to the objective indicators fed back by the physiological sensors; In the mid-to-late stages of rehabilitation, the dynamic adversarial discriminator increases the weighting of professional evaluation and subjective feelings, and performs long-term consistency checks by comparing the patient's performance data at different rehabilitation cycles.

4. The machine learning-based personalized rehabilitation pathway dynamic planning system according to claim 3, characterized in that: The nonlinear dynamic stability controller uses Lyapunov stability theory to construct stability discrimination logic. By calculating the geometric distance and movement speed between the patient's current rehabilitation state and the target path attractor in phase space in real time, it judges the fluctuation attributes of the current rehabilitation progress. When the nonlinear dynamic stability controller detects that the fluctuation of the patient's state is within a preset small oscillation range, it determines that the patient is in a normal physiological fluctuation state and sends a fine-tuning instruction to the knowledge-enhanced constraint generator, which outputs a progressive path correction suggestion by adjusting the parameters of the generated model. When the nonlinear dynamic stability controller detects that the patient's position in phase space deviates from the geometric distance of the preset attractor by more than a certain threshold and the duration exceeds a predetermined time length, it determines that the system has entered the phase transition critical point and immediately triggers the path reconstruction mechanism, instructing the knowledge-enhanced constraint generator to re-execute global path planning according to the current abnormal state, and generate a new rehabilitation branch path that bypasses the current rehabilitation obstacle.

5. The machine learning-based personalized rehabilitation pathway dynamic planning system according to claim 4, characterized in that: When the nonlinear dynamic stability controller makes path reconstruction decisions, it uses the patient's historical tolerance information as a key input variable, which includes the upper limit of exercise intensity when pain occurred in the past. When generating new branch paths, the nonlinear dynamic stability controller forces the training load growth slope of the new path to be lower than the average growth rate of the patient's historical tolerance. By constraining the curvature of the motion trajectory of the state point, it ensures that the reconstructed path scheme maintains a smooth trajectory during the evolution towards the new attractor, and avoids motion damage caused by excessive path span. Meanwhile, the knowledge-enhanced constraint generator is also equipped with adaptive action space scaling logic, which dynamically adjusts the complexity of selectable actions based on the patient's daily fatigue score. When the fatigue score is higher than a preset safety threshold, it automatically locks high-risk action nodes and generates paths only in the set of basic maintenance actions.

6. The personalized rehabilitation pathway dynamic planning system based on machine learning according to claim 5, characterized in that: The multimodal data acquisition terminal consists of a hardware sensor array distributed on key parts of the patient's body, used to acquire multi-source heterogeneous data including electromyography signals, electrocardiogram signals, joint angles, angular velocity of movement, linear acceleration, gait cycle, smoothness of movement and synergistic movement index. The multimodal data acquisition terminal integrates a data preprocessing unit, which is used to perform filtering and noise reduction, baseline drift correction and multi-channel synchronization processing, and to calculate the signal-to-noise ratio of the signal in real time to perform data quality monitoring functions. When the signal-to-noise ratio is detected to be lower than a preset threshold, the central cloud processing server initiates a historical trend prediction and compensation mode based on long short-term memory networks, fills in missing values ​​using the patient's historical health data, and sends an early warning signal to external terminals.

7. The machine learning-based personalized rehabilitation pathway dynamic planning system according to claim 6, characterized in that: The central cloud processing server is built with a high-performance computing cluster and deployed in a containerized computing environment, and is connected to the multimodal data acquisition terminal through an encrypted communication link; The central cloud processing server has a dynamic load balancing mechanism that is used to schedule the graphics processing unit in real time to perform inference and training of deep learning models based on the urgency of the computing tasks and the usage of computing resources. The clinical medical knowledge base supports incremental updates. The central cloud processing server uses a built-in knowledge extraction engine and natural language processing technology to automatically extract new entities and relationships from medical literature and update the association rules in the rehabilitation medicine knowledge graph, so that the prior constraint domain of the knowledge-enhanced constraint generator evolves synchronously with the results of medical research. The central cloud processing server is also connected to the hospital's electronic medical record system. By extracting the patient's past medical history and surgical records, the initial rehabilitation attractor track is matched by the clinical medical knowledge base.

8. The machine learning-based personalized rehabilitation pathway dynamic planning system according to claim 7, characterized in that: The system also includes an interactive display terminal, which is configured as a visualization front-end device connected to the central cloud processing server, used to visualize the rehabilitation path scheme generated by the knowledge-enhanced constraint generator. The interactive display terminal has a medical knowledge traceability display function. When it receives a query command for training suggestions, the interactive display terminal extracts and graphically displays the associated medical knowledge nodes from the clinical medical knowledge base. The medical knowledge nodes include anatomical basis, physiological principles and corresponding clinical guidelines. The interactive display terminal also supports scheme simulation function. Using the stored historical archives and current path parameters, the central cloud processing server simulates the patient's functional evolution trajectory in the future cycle in the virtual space based on the nonlinear dynamic model, providing decision support for manual intervention.

9. The machine learning-based personalized rehabilitation pathway dynamic planning system according to claim 8, characterized in that: The knowledge-enhanced constraint generator is constructed as a multi-agent collaborative architecture, including a first path generation submodule focusing on musculoskeletal system reconstruction and a second path generation submodule focusing on cardiopulmonary function maintenance. The central cloud processing server runs a conflict resolution algorithm layer. When the training intensity proposed by the first path generation submodule causes the patient's heart rate to exceed the safety threshold set by the second path generation submodule, the conflict resolution algorithm layer calls the stability analysis data provided by the nonlinear dynamic stability controller. With the safety of vital signs as the first priority, it forcibly reduces the intensity of the conflicting part of the action by dynamically adjusting the attention bias until the cardiopulmonary index returns to the preset stable domain in the phase space. The multimodal data acquisition terminal also includes a computer vision-based micro-expression analysis unit, which is used to capture the emotional feature components when the patient performs actions, and input the emotional feature components into the dynamic adversarial discriminator to help determine whether the intensity of rehabilitation exceeds the patient's psychological defense boundary.

10. The machine learning-based personalized rehabilitation pathway dynamic planning system according to claim 9, characterized in that: The clinical medical knowledge base adopts a distributed deployment mode, with its core global knowledge base stored in the central cloud processing server, and sub-knowledge bases for specific diseases cached in the edge computing gateway deployed on the user side. The knowledge-enhanced constraint generator is divided into a cloud-based global optimization module and an edge-based real-time fine-tuning module. The cloud-based global optimization module periodically analyzes the patient's long-term rehabilitation trend and distributes model weights. The edge-based real-time fine-tuning module performs low-latency dynamic corrections within a small parameter space based on the patient's immediate feedback. The dynamic adversarial discriminator introduces a federated learning mechanism, whereby each edge computing gateway uses local data to iterate the model locally and uploads the encrypted model gradient to the central cloud processing server for aggregation, thereby realizing the evolution of expert models in a privacy-desensitized environment. The nonlinear dynamic stability controller also introduces external environmental parameters as disturbance variables into the high-dimensional phase space model, and automatically shrinks the action space of the knowledge-enhanced constraint generator according to the probability of fall risk caused by environmental disturbances.

Images