Scene interactive AI rehabilitation assessment training and health monitoring system

Abstract

The invention discloses a scene interactive AI rehabilitation evaluation training and health monitoring system, and relates to the technical field of rehabilitation medical treatment and artificial intelligence, a semantic perception module is used for collecting and recognizing voice input, facial expressions, action tracks and eye movement paths of a user in a training process, and extracting context parameters; the knowledge-driven training generation module is used for calling a rehabilitation knowledge graph constructed by a graph neural network based on context parameters and individual training history, and generating a multi-path training scheme; training a feedback regulation engine, collecting posture offset, physiological stress and emotion feedback, and dynamically adjusting task difficulty, rhythm and prompt mode based on a dual-channel reinforcement learning model; the prediction module fuses training and monitoring data, and predicts a network identification function degradation risk through degradation driving; the cloud edge fusion platform is used for realizing task quick response and graph strategy iterative updating; according to the invention, the individuation, self-adaption and intelligent prediction capabilities of rehabilitation training are improved, and the rehabilitation effect and the system practicability are obviously optimized.

Description

Technical Field

[0001] The present invention relates to the field of rehabilitation medicine and artificial intelligence technology, and in particular to a scenario-interactive AI rehabilitation assessment training and health monitoring system. Background Art

[0002] During rehabilitation treatment, especially for those with neurological injuries and degenerative diseases in the elderly, traditional rehabilitation methods rely on manual assessment and training guidance from professional therapists, which presents significant limitations: First, assessment results are highly subjective and lack quantitative standards; second, training processes are rigid, lacking personalized and dynamic feedback, making it difficult to adapt to the fluctuations in a patient's condition. Furthermore, while some existing systems have introduced wearable devices to collect physiological signals, and there have been attempts to use virtual reality for sports training, the following technical pain points are common: First, the evaluation system and training system operate separately and lack unified control logic, making it difficult to feed evaluation results back to the training phase in real time, thus forming information silos. Second, the system lacks the ability to embed context into training content. Currently, most training tasks are based on static task libraries and cannot be dynamically matched to user emotions, action deviations, and environmental changes. Third, the health monitoring function is separated from rehabilitation training and only serves as a data collection function. It cannot participate in training strategy generation or actively prompt users to identify behavioral deviations. However, existing technologies lack a system design that integrates semantic scenario recognition, intelligent task generation, modal data fusion analysis, and adaptive training closed-loop control mechanisms. Therefore, building an AI rehabilitation training and assessment system that is highly adaptable, fully data-driven, contextually interactive, and capable of continuous feedback and regulation has become a pressing challenge in this field. Summary of the Invention

[0003] The purpose of the present invention is to provide a scenario-interactive AI rehabilitation assessment training and health monitoring system to address the shortcomings of the background technology.

[0004] To achieve the above objectives, the present invention provides the following technical solutions: a scenario-interactive AI rehabilitation assessment training and health monitoring system, comprising: The semantic perception module is used to analyze the user's voice input, facial expressions, movement patterns, and eye movements during training, and extract contextual parameters that reflect their cognitive and emotional states; The knowledge-driven training generation module, based on the extracted context parameters, combines historical training data and individual difference characteristics, calls on the dynamically evolving cross-modal rehabilitation knowledge graph, performs task adaptation reasoning and multi-path training program construction, and is used to dynamically generate rehabilitation assessment tasks and training paths that are consistent with the current context; The training feedback control engine uses reinforcement learning strategies to dynamically adjust the difficulty level, rhythm parameters, and prompt methods of the current training task in real time based on the user's posture deviation, physiological stress changes, emotional feedback, and task completion rate during training. The prediction module is used to integrate the collected training data and daily health monitoring data into a model, extract the characteristics of high-frequency health events, build a long-term health trend model, and predict the risk level of functional degradation; The cloud-edge fusion platform is equipped with edge computing nodes for low-latency processing and on-site control during training, and cloud servers for long-term data archiving, knowledge graph evolution and cross-user modeling.

[0005] Preferably, the knowledge-driven training generation module includes: The context parameters output by the semantic perception module are jointly encoded with the user's historical training data, evaluation records, and individual physiological feature vectors to generate an individual rehabilitation status index, which serves as the entry key for querying the dynamically evolving knowledge graph. Calling a cross-modal rehabilitation knowledge graph built on a graph neural network. The graph nodes include training action units, evaluation indicators, feedback mechanisms, and their semantic labels. Edges represent task sequence dependencies and adaptation conditions. The information propagation mechanism within the graph is used to retrieve a set of candidate training nodes that meet the current state. Based on the candidate node set and its upstream and downstream dependency paths, training task adaptation reasoning based on edge-weighted scoring is performed to construct a set of multi-path training schemes with high correlation with the current context parameters. The top N schemes are output after sorting by task load gradient, realizing dynamic generation and refined matching of training content.

[0006] Preferably, the multi-path training scheme set construction further includes: Based on the set of candidate task nodes, an improved path guidance generation algorithm is adopted, combined with context parameters and training target constraints. By introducing an adaptive path scoring gating mechanism in the graph neural network, the synergistic effect of different training paths is dynamically evaluated to generate an optimized task path set; wherein, the path scoring gating mechanism constructs a multi-factor control function based on training success rate, user compliance feedback and semantic conflict between paths, dynamically adjusts the path selection priority, and updates the local edge weights of the graph structure.

[0007] Preferably, the training feedback control engine includes: Collect four types of feedback signals from users during training: posture deviation, physiological stress indicators, real-time emotion classification results, and task completion rate. Construct a state vector through feature normalization and time window aggregation. The state vector is input into a dual-channel deep reinforcement learning network optimized based on a multi-factor attention mechanism to learn the optimal control strategies for the training task difficulty-completion efficiency path and the rhythm-emotional fluctuation response path, and the strategy is updated based on historical execution rewards; According to the network output results, the action complexity, feedback rhythm parameters and prompt interaction mode of the current training task are adaptively adjusted in real time to trigger a fine-grained difficulty climbing or slowing down mechanism, and dynamic incentive feedback is output at the same time.

[0008] Preferably, the reinforcement learning strategy further includes: using a strategy mapping network based on a structurally variable topology to map training feedback signals of different dimensions to the task attribute space, and dynamically reconstructing the strategy response path through graph structure transformation and strategy nested update mechanism; wherein, the strategy mapping network encodes the training state space into a multi-layer plasticity node graph structure, with different task dimensions as target nodes, generates a multi-strategy combination response matrix through cross-layer association learning, and dynamically selects the optimal control path combination according to the user's current state.

[0009] Preferably, the prediction module includes: The rehabilitation movement data and evaluation feedback collected during the training period are synchronized and semantically aligned with the physiological signals and behavioral frequency data from daily health monitoring to construct a spatiotemporal fusion dataset containing time tags, behavioral semantics, and physiological indicators. Using a multi-scale time series-graph convolutional joint model optimized based on a heterogeneous attention mechanism, we extract periodic health fluctuation characteristics and sudden high-frequency event sequences, and perform feature clustering and label attribution on various behavioral-physiological coupling patterns. A long-term health trend map is constructed based on the clustering results, and a prediction result of the functional degradation risk level within a preset period in the future is output through joint modeling of time series trend fitting and degradation sensitivity scoring function.

[0010] Preferably, the long-term health trend map further includes: introducing a degradation-driven variable-scale prediction network, using functional degradation risk sensitivity as a dynamic control factor, and adaptively adjusting the temporal perception window and trend recognition scale of the prediction model; wherein, the long-term health trend map strengthens modeling of high-volatility segments through a convolutional gating mechanism, and implements weak prediction constraints on stable segments.

[0011] In the above technical solution, the technical effects and advantages provided by the present invention are: 1. This invention achieves intelligent management of the entire process, from user status perception and intelligent task allocation to real-time feedback adjustment and long-term health trend prediction, by constructing a closed-loop rehabilitation system that integrates multiple modules, including semantic perception, task generation, training control, trend prediction, and cloud-edge collaboration. Compared with traditional rehabilitation methods that rely on manual assessment and fixed training processes, this invention significantly improves the personalized adaptability of training, real-time response speed, and full-cycle intervention accuracy, effectively addressing the technical bottlenecks of existing rehabilitation systems, such as fixed tasks, delayed feedback, and low user compliance.

[0012] 2. By introducing a dynamic knowledge graph based on graph neural networks, a dual-channel deep reinforcement learning policy control mechanism, and a degradation-driven variable-scale prediction model, the present invention not only achieves structured processing and intelligent response of high-dimensional training feedback data, but also has self-learning and policy evolution capabilities, and can continuously optimize training paths and control strategies as user status changes. At the same time, combined with the efficient data calculation and model synchronization mechanism built on the cloud-edge fusion platform, it ensures low-latency and high-stability operation performance of the system at the edge, and realizes cross-user trend modeling and graph evolution in the cloud, significantly enhancing the system's scalability and deployment practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments described in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0014] Figure 1 This is a mind map of the system modules of the present invention. DETAILED DESCRIPTION

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0016] Example 1, please refer to Figure 1 As shown, the scenario-interactive AI rehabilitation assessment training and health monitoring system described in this embodiment includes: The semantic perception module is used to analyze the user's voice input, facial expressions, movement patterns, and eye movements during training, and extract contextual parameters that reflect their cognitive and emotional states; The knowledge-driven training generation module, based on the extracted context parameters, combines historical training data and individual difference characteristics, calls on the dynamically evolving cross-modal rehabilitation knowledge graph, performs task adaptation reasoning and multi-path training program construction, and is used to dynamically generate rehabilitation assessment tasks and training paths that are consistent with the current context; The training feedback control engine uses reinforcement learning strategies to dynamically adjust the difficulty level, rhythm parameters, and prompt methods of the current training task in real time based on the user's posture deviation, physiological stress changes, emotional feedback, and task completion rate during training. The prediction module is used to integrate the collected training data and daily health monitoring data into a model, extract the characteristics of high-frequency health events, build a long-term health trend model, and predict the risk level of functional degradation; The cloud-edge fusion platform is equipped with edge computing nodes for low-latency processing and on-site control during training, and cloud servers for long-term data archiving, knowledge graph evolution and cross-user modeling.

[0017] The following is a detailed description of the semantic perception module in the present invention in combination with the implementation methods. This module is one of the core subsystems of the situational interactive AI rehabilitation assessment training and health monitoring system. It is mainly responsible for realizing multimodal information perception, contextual semantic understanding and cognitive state modeling of rehabilitation users during the training process, and providing key input parameters for subsequent training task generation and feedback control.

[0018] This step integrates multiple sensors and perception models to collect multi-source behavioral data such as voice input, facial muscle changes, body movement trajectories, and eye movement paths during rehabilitation training. To improve collection accuracy and time consistency, the present invention adopts the following specific technical means: The micro-expression capture network uses a lightweight CNN model based on convolutional modeling of local facial regions. It is deployed on the front end of a high-definition camera and performs real-time frame analysis on high-frequency changing areas such as the user's eye corners, mouth corners, and eyebrows to extract information on facial micro-muscle changes. The speech spectrum emotion recognition model is built on a multi-layer bidirectional LSTM architecture. Its input is a Mel spectrum graph, and its output is a continuous probability distribution of emotional tendency. It can model the temporal dependency of speech intonation. The gait-eye movement joint perception algorithm simultaneously dispatches an infrared motion recognition camera and an eye tracking module, using the user's stride change rate, center of gravity shift, and gaze stability as feature vectors to form a behavior-vision coupling indicator stream. All collected data are uniformly timestamped by this module to improve the synchronization alignment quality of cross-modal data in subsequent modeling steps and effectively eliminate the feature offset problem caused by asynchronous sampling.

[0019] After completing synchronous acquisition and time alignment, the present invention introduces a Transformer recognition model based on multimodal attention fusion to encode the above multi-source data to generate a stable, high-dimensional context representation vector. The model has the following structure and features: The core structure design based on Transformer allows global dependency modeling of each modal information under the self-attention mechanism; A cross-modal attention fusion mechanism is introduced. By configuring an independent attention weight matrix for each modal channel (semantic channel, emotional channel, behavioral channel, etc.), the model can automatically learn the relative importance of each modality in different scenarios. In particular, the weight coefficients of the semantic channel and the emotional channel are updated in real time based on historical recognition accuracy, implementing an adaptive dynamic adjustment mechanism for channel weights. That is, when a modality performs better in recognition of the current user sample, its weight in the vector fusion stage will increase, enhancing the ability to express information. The final output is the fused context representation vector, which encodes the user's behavioral intention, emotional state and their mutually coupled dynamic patterns in the current training scenario in the feature space.

[0020] The context representation vector is fed as input into the emotion-cognition dual-channel discriminant network. This network uses a parallel structure to identify the user's emotional state and cognitive load level respectively. The key technical path is as follows: The emotion recognition channel uses an emotional state classifier that combines an improved GRU with a contextual attention mechanism, capable of modeling emotional change sequences and detecting sudden emotional anomalies. The cognitive load identification channel builds a cognitive response intensity scoring mechanism based on the cross-coding of residual attention network and training interaction history features to identify user attention concentration, slow response trends, and signs of mental fatigue. The outputs of the two channels are integrated through a fusion layer to generate structured context parameters, including but not limited to the current semantic context label, sentiment score vector, cognitive load level, and credibility confidence coefficient; The structured context parameters will be passed as key inputs to the knowledge-driven training generation module and the training feedback control engine to guide their subsequent task path generation and real-time adjustment of training content.

[0021] By introducing a multimodal synchronous acquisition mechanism and a Transformer fusion network structure, the semantic perception module of the present invention can accurately perceive the user's true intention and psychological state in a dynamic training environment, achieving a situational understanding accuracy far exceeding that of traditional single-modal input systems. The use of a weight adaptive adjustment mechanism and a dual-channel discrimination structure design can enhance the system's adaptability to different modal signals in different individuals or situations, significantly improving the robustness and generalization ability of contextual judgment. This embodiment specifically relates to the implementation method and functional flow of the knowledge-driven training generation module in the present invention. This module is the core intelligent unit of the situational interactive AI rehabilitation assessment training and health monitoring system. Its main function is to generate personalized and situational training assessment paths based on the user's current status, and to collaborate with the semantic perception module and the training feedback control engine to construct an adaptive, closed-loop rehabilitation training logic.

[0022] This step integrates the structured context parameters output by the semantic perception module with the user's individual training data accumulated over a long period of time in the system to construct a comprehensive representation of their current physiological state, psychological state, and behavioral ability. Specifically, it includes: Extract historical training data, task completion status, evaluation results, training preference information, etc. from the user profile database; Synchronously call the user's current physiological monitoring data (such as heart rate variability, movement balance, myoelectric intensity distribution, etc.); The above data are combined with contextual parameters (such as cognitive load level, emotion type, and interaction semantic label) for feature splicing and unified coding; An individual recovery status index is generated through a deep feature extraction network (such as MLP or a variant BERT structure). This index serves as the primary key for querying the knowledge graph to locate the training task group that matches the current status.

[0023] This module relies on a cross-modal rehabilitation knowledge graph built on a graph neural network (GNN) structure to model the semantics, logic, and dependency structure of multi-dimensional training tasks: The nodes in the knowledge graph include: training action units (such as lifting and pressing, gait stability control, etc.), evaluation indicators (such as ROM, movement consistency), interactive feedback mechanisms (such as graphic prompts, voice guidance), and each node is annotated with task labels, intensity levels, adaptation contexts, and user success rates. The weight of the edge between nodes represents the logical order, dependency, conflict risk, etc. between tasks, while also considering the degree of contextual matching; Based on the message passing mechanism of the graph neural network, the system retrieves the most relevant training task node set in the graph according to the current rehabilitation status index to form a candidate training node set; This retrieval mechanism is different from the traditional static task calling scheme. It has the ability to dynamically update the graph structure, adaptively reorder, and evaluate connectivity based on task dependency semantics.

[0024] The system further performs the task adaptation reasoning process based on the above candidate node set: Each candidate node is scored based on the "edge-weighted scoring mechanism". Scoring factors include: the matching degree between the node and the current context parameters, historical completion efficiency, user emotional acceptance, etc. Based on the scoring results, the training path combinations with relative advantages are identified, the task sequence planning algorithm is executed, and multiple training path branches are constructed; In each path, nodes are arranged in order, and paths can intersect. The system output sorts the paths according to the task load intensity gradient, target diversity, and time accessibility, forming a set of top N multi-path training solutions for selection. This collection, as the core content of the individual training plan, will be passed into the training feedback regulation engine for dynamic operation control.

[0025] To further enhance the intelligence and coordination of training paths, this module designs an improved path guidance generation algorithm and introduces a path scoring gating mechanism. It has the following key technical features: Using the GNN topology, we introduce a learnable gating variable into each candidate task path and adjust the weight of the selection probability of each task node in the path. The gating mechanism integrates multiple dimensions of path regulation factors, including: ① training success rate (based on historical data regression modeling); ② user compliance feedback (such as frequent interruptions and mood swings); ③ semantic conflict between paths (used to avoid content duplication or load redundancy); Construct a multi-factor control function, combine the above weights to dynamically score and re-prioritize the candidate path set, and realize adaptive dynamic adjustment of path selection priority; At the same time, if certain paths show poor feedback or increased task execution risk in recent use, the system will automatically update the local edge weights of the knowledge graph structure through the edge weight reduction mechanism to achieve fine-tuning and dynamic evolution of the graph.

[0026] This module uses a state index-driven graph retrieval mechanism to effectively achieve precise correspondence between semantic perception results and training tasks, breaking the technical barrier of separating training content and state recognition in traditional rehabilitation systems; the introduced knowledge graph structure and message passing mechanism based on graph neural networks can make task recommendation results more robust and personalized in terms of multi-task logic and context adaptation; through the path scoring gating mechanism and improved path generation algorithm, it realizes the coordinated regulation and individual dynamic adaptation of training task paths for the first time, which not only improves the execution success rate and user acceptance of task combinations, but also automatically adjusts the graph structure and has self-learning and self-optimization capabilities.

[0027] The training feedback control engine provided in an embodiment of the present invention, as one of the core control units of the scenario-interactive AI rehabilitation assessment training and health monitoring system, is mainly responsible for responding to the user's multi-dimensional feedback status during the training process in real time, dynamically adjusting the training task parameters based on the reinforcement learning strategy, and realizing refined and adaptive rehabilitation process control to improve training efficiency and user compliance.

[0028] During rehabilitation training, the system collects multiple types of feedback signals from the user in each training unit in real time, including: Posture offset: The deviation between the user's actual movement and the target movement measured by the motion capture device; Physiological stress indicators: such as heart rate variability (HRV), galvanic skin response (EDA), respiratory rhythm, and other indicators that reflect the user's physiological stress level; Real-time emotion classification results: Output from the speech spectrum emotion recognition model and facial expression recognition algorithm to construct emotional state labels; Task completion rate: Generates a quantitative score based on the evaluation criteria set by the system, such as the number of times the action standard is completed, the completion time and the error ratio.

[0029] The above multi-dimensional feedback signals are normalized and aggregated in a time window to form a stable, inputtable state vector, which reflects the user's overall physiological, psychological, and executive state in the current training stage.

[0030] To achieve optimal control of training parameters, this module designs a dual-channel deep reinforcement learning network (Dual-DQN-Attn) based on a multi-factor attention mechanism optimization, including the following core contents: The first channel (task difficulty-completion efficiency channel): used to model the strengthening relationship between changes in training task complexity and user completion efficiency; The second channel (rhythm-emotional fluctuation channel): is used to study the influence path between prompt frequency, feedback rhythm and user emotional fluctuation; Both channels are based on the DQN structure and introduce an attention mechanism to adaptively focus on the most sensitive factors (such as sudden changes in stress indicators or task completion breakpoints) in the state-action space. The strategy update mechanism is optimized based on the historical return function, introducing a trade-off mechanism between short-term rewards and long-term compliance benefits to ensure that the strategy has the ability to take into account both real-time and scalability.

[0031] Through the trained network model, the system can quickly output control suggestions under any state input, and guide the specific adjustment strategies of the action complexity, feedback rhythm and prompt method of the current training task.

[0032] Based on the output of the dual-channel network, the training feedback control engine performs the following dynamic adjustment operations: Fine-tune the action complexity parameters in real time. If the system detects a rapid drop in task completion rate under high load, it will automatically switch to a simplified task path. Adjust the feedback rhythm parameters to encrypt or slow down the prompt content; Dynamically switch prompt interaction modes, such as switching from graphic prompts to voice guidance or multimodal overlay; Synchronously output adjustable motivational feedback content (such as encouraging voice, point-enhancing animation, etc.) to enhance the user's emotional stability and training enthusiasm.

[0033] To further improve the adaptability of the policy model to user status changes, this module introduces a Dynamic Strategy Mapping Network (DSMN) based on a structurally variable topology. Its core design includes: The training state space is encoded into a graph structure consisting of multiple layers of plastic nodes, where nodes represent state feature combinations and edges represent state transition possibilities. Each training control dimension (movement complexity, rhythm frequency, prompt method) is defined as an independent target strategy node, which receives the strategy conduction input of the current state diagram; A cross-layer association learning mechanism is introduced to build a feedback dependency graph between different state scales, and a nested policy update algorithm is used to dynamically generate response paths. Construct a multi-strategy combination response matrix. The elements in the matrix represent the adaptation score of each strategy combination in the current state. The system dynamically selects the control path combination based on the highest score and performs strategy reconstruction. At the same time, the strategy execution results are reversely evaluated with the final task completion rate and user physiological response, and fed back to the DSM network for topology fine-tuning, which has the ability of self-evolution and strategy memory.

[0034] By collecting multi-dimensional feedback signals and using state vectors as input to the reinforcement learning network, the system has the ability to deeply perceive the user's physiological, emotional, and behavioral states, far exceeding the feedback strategy driven only by completion rate in traditional training systems; the introduced dual-channel DQN structure can simultaneously consider the interactive dependencies between efficiency-difficulty and rhythm-emotion, achieving a more fine-grained multi-objective control effect; a structurally variable strategy mapping network is used to construct a dynamic mapping mechanism between feedback status and control response, which does not rely on fixed control strategy templates and has high flexibility and response stability in nonlinear and complex scenarios.

[0035] The prediction module provided in this embodiment is used to integrate modeling of a user's rehabilitation training period and daily health status, identifying health fluctuations and predicting individual functional degradation risk levels. By integrating multi-source behavioral data, physiological signals, and assessment feedback, this module uses a heterogeneous modeling mechanism to construct a long-term health trend map. Furthermore, it introduces a degradation-driven multi-scale prediction model to achieve high-precision, dynamic perception of degradation risk prediction and intervention support.

[0036] This step aims to perform temporal alignment and semantic normalization on the multimodal data distributed in the training process and daily monitoring. Specifically, it includes: Rehabilitation training data: Collect rehabilitation movement data generated by users in various training tasks, and provide real-time evaluation feedback (such as movement completion, error rate, and reaction time); Daily health monitoring data: including physiological indicators such as heart rate, blood pressure, body temperature, and behavioral frequency data such as number of steps and duration of sitting recorded by the user's wearable device; The data is time-aligned using a unified time tag, and a semantic annotation mechanism driven by behavioral events is introduced to logically divide the data into "event units"; Finally, a spatiotemporal fusion dataset with a three-dimensional structure of time labels, behavioral semantics, and physiological indicators is constructed, providing a unified input representation for subsequent modeling.

[0037] In order to effectively identify behavioral-physiological coupling features that have a strong influence on degradation trends from fused data, this module introduces a multi-scale temporal-graph convolutional joint model (MS-TGCN) based on the optimization of heterogeneous attention mechanism: Time series modeling channel: Using a multi-scale LSTM and variable perception window structure, it extracts the user's health status fluctuation trajectory at different time granularities and distinguishes periodic behavior patterns from sudden abnormal events; Graph convolutional modeling channel: A heterogeneous graph structure is constructed based on the coupling relationship between behavioral and physiological variables. Nodes represent variable features, and edges represent association strength. The nonlinear dependencies between variables are modeled through the GAT mechanism. The model introduces a cross-channel attention mechanism to dynamically adjust the participation of each feature at different time periods to form a highly recognizable coupled feature vector; The extracted features are classified into behavioral patterns through the K-morphological clustering algorithm and assigned labels in the graph structure, forming the structural basis for subsequent trend graph modeling.

[0038] After feature extraction and pattern attribution, the system constructs a long-term health trend map based on the clustering results. The map includes the following design elements: Node type: represents different health event categories (such as stable activity, severe fluctuation, mild degradation, potential risk, etc.); Edge relationships: represent temporal logic, variable evolution trends, and behavioral inducing paths. Edge weights are determined by event co-occurrence frequency and causal relationship modeling algorithms. For data series of different time periods, perform time series trend fitting based on weighted regression to identify the trend direction of health indicators over time; A degradation sensitivity scoring function is introduced to integrate factors such as the completion rate decline rate, emotional fluctuation intensity, and physiological parameter fluctuation amplitude to form a risk score for event nodes, which is used to mark key risk nodes in the trend path.

[0039] To further improve the system's responsiveness in early identification of degradation risks, this module introduces a degradation-sensitive scalable prediction network (DSS-PNet). The specific structure and process are as follows: The system uses the risk sensitivity of functional degradation as a dynamic control factor to drive the model to automatically select the optimal time window and feature perception scale; The DSS-PNet model introduces a convolutional gating mechanism to enhance the prediction weights of data channels marked as high-volatility segments in the trend map, while implementing weak prediction constraints on low-volatility and stable segments, thereby achieving adaptive optimization between resource allocation and recognition accuracy. The model outputs the degradation risk level results within the preset future period, which are divided into multiple levels of labels (such as no risk, mild degradation, accelerated degradation, acute degradation, etc.), and also provides the trend chart of key variables corresponding to each level; This mechanism effectively solves the problems of low accuracy and delayed response of traditional static time window prediction models when processing highly variable rehabilitation data, and can provide early warnings and trigger intervention recommendations for health risks.

[0040] This module achieves unified perception and behavioral pattern recognition among multi-source health data by constructing a spatiotemporal fusion dataset and adopting heterogeneous modeling and graph structure attribution technology, surpassing the limitations of existing systems that only model single-dimensional indicator trends. By utilizing a multi-scale time series-graph convolutional model, it can simultaneously identify and label periodic behaviors and sudden high-frequency events, improving the system's robustness to unstable states. The introduction of long-term health trend maps and degradation sensitivity scoring mechanisms enables the system to have the ability to perform structured modeling and quantitative risk analysis of long-term changes in individual health trends. In particular, the degradation-driven variable-scale prediction network (DSS-PNet) combined with the convolutional gating mechanism significantly improves the prediction resolution of degradation segments, achieving early identification, detailed response and stable update of risk states, and is highly intelligent and practical.

[0041] The cloud-edge fusion platform proposed in this paper is built on a heterogeneous distributed computing architecture. It aims to leverage the low-latency characteristics of edge computing and the strong modeling capabilities of cloud computing to achieve multi-level collaborative processing of training task data processing, control decision-making, and knowledge graph evolution. Based on task type and latency sensitivity, the platform uses hierarchical modeling and data flow distribution mechanisms to create a highly flexible and intelligent task scheduling and model evolution architecture.

[0042] The system first partitions and categorizes various tasks and builds an edge-cloud collaborative task scheduling strategy based on the following three core indicators: Task type: divided into real-time interactive tasks (such as semantic recognition, training feedback), periodic modeling tasks (such as trend modeling, strategy update) and batch optimization tasks (such as graph structure adjustment); Delay sensitivity: Tasks that need to be completed within <200ms, such as modality fusion perception, training rhythm control, and individual task prompts, are judged to be highly delay-sensitive; Data load intensity: If the data processing volume or model inference overhead involved in a single task exceeds the computing power threshold of the edge device, it will be processed by the cloud.

[0043] Based on the above indicators, the system builds an edge priority scheduling rule library, which includes: Edge computing nodes are assigned to handle real-time control tasks, modal fusion perception, and training feedback adjustment logic; Cloud servers are used for cross-cycle modeling and model optimization of non-real-time tasks to ensure optimal overall response efficiency and resource utilization.

[0044] This module uses a cross-layer nested collaborative model framework, which decomposes the complete AI model into a lightweight sub-model (edge) + full mother model (cloud) architecture, and executes the following process: Lightweight sub-models are deployed on edge nodes, such as streamlined semantic perception models and training feedback discrimination models with reduced channels, focusing on short-term and rapid response. During the operation of the sub-model, intermediate data such as user training behavior summary, stress response characteristics, and predicted probability results are continuously generated; All intermediate results are uploaded to the cloud periodically (e.g., every 10 minutes) using a compressed transmission protocol; After receiving the summary information uploaded by the edge, the cloud-based master model performs model weight fine-tuning, policy function reconstruction, and generates an individualized model parameter update package that is fed back to the edge node; The cloud-edge model parameter synchronization follows a dual-track mechanism of "periodic batch + event triggering": on the one hand, the basic model is updated every 24 hours; on the other hand, when the edge detects a training failure signal (such as a sudden drop in success rate or increased emotional fluctuations), a temporary rapid model revision is immediately triggered.

[0045] The cloud server also undertakes cross-user trend modeling and rehabilitation knowledge graph evolution management tasks, including: Conduct trend fusion modeling on multi-user data streams to identify the commonalities of group rehabilitation paths and the heterogeneity of individual training preferences; Build a dynamic graph evolution engine to restructure the training task nodes and strategy paths in the original graph, including: Node importance assessment: Dynamic scoring based on frequency, completion rate, and context relevance; Edge weight redistribution: construct a new weight distribution based on path usage, risk level, and feedback response; Node merging and pruning strategy: automatically remove low-frequency / redundant task nodes and merge semantically redundant task branches; The evolved knowledge graph will serve as the input basis for new task generation and reasoning strategies, and will be used to drive the generation of a new round of personalized training paths.

[0046] The evolved graph structure and related model update results will be transmitted back to the edge via an incremental synchronization mechanism: The system only synchronizes structural changes (such as new nodes and changes in edge weights) to avoid the communication burden caused by full-graph synchronization; Deploy a graph receiving and parsing module at the edge to perform structure insertion and conflict resolution on the received data to ensure local consistency of the graph structure; The synchronously updated graph can immediately drive the next round of training task allocation, allowing personalized evolution results to be quickly implemented, and forming a closed-loop "data-model-knowledge-response" full-link self-evolution mechanism.

[0047] The cloud-edge fusion platform of the present invention opens up the response link between real-time perception, feedback control and strategy evolution by constructing a multi-factor scheduling strategy, and realizes the coordinated unification of edge rapid processing and cloud-side deep modeling; the system effectively takes into account model performance and computing resource limitations through nested model structure layered deployment and parameter synchronization mechanism, avoids the problem of edge response lag, and maintains long-term learning ability; the introduced dynamic graph evolution mechanism has structural autonomy and semantic differentiation capabilities, can actively adapt to changes in user behavior patterns and output refined training path update logic, which is different from the traditional fixed task recommendation mechanism; the incremental synchronization mechanism and edge structure analysis mechanism ensure that the model and graph iteration results can be efficiently deployed, constructing a "low latency-strong adaptability-fast iteration" intelligent computing architecture with practical value in rehabilitation systems.

[0048] Example 2: In order to verify the effectiveness of the scenario-interactive AI rehabilitation assessment training and health monitoring system proposed in the present invention in practical applications, a prototype system was constructed and deployed in the neurorehabilitation department of a tertiary rehabilitation hospital. A system evaluation test was conducted on 20 stroke patients. The test period was 4 consecutive weeks, with 3 training sessions per week and each training session lasting 30 minutes.

[0049] Experimental environment and data source: Hardware composition: Use a set of edge computing terminals (including GPU modules), high-definition cameras, voice pickup modules, eye trackers and multi-channel physiological data acquisition belts; Software system: Deploys the modules of the present invention, including the semantic perception module, knowledge-driven training generation module, training feedback control engine, prediction module and cloud-edge fusion platform; Test subjects: 20 patients with upper limb movement disorders and mood swing disorders after stroke, who had basic cognitive abilities.

[0050] Semantic Perception Module Test Results: 30 randomly selected semantic task commands (e.g., "raise your hand to throw a ball," "stretch your fist forward") were used to integrate voice commands, facial expressions, and eye movement data. The system calculates the recognition accuracy by comparing the manually recorded results: The average context recognition accuracy reached 91.3%; The average latency of semantic state–action deviation detection was <450ms; The AUC for distinguishing high and low cognitive load states reached 0.87.

[0051] Dynamic generation of training tasks: The system generates a set of five training paths based on the initial context and automatically selects the optimal task chain after review by the doctor; The user adaptation completion rate increased from 61.2% in the first week to 86.5% in the fourth week; Compared with the fixed task group (control group), the completion rate of the personalized task path increased by 22.7%, and the user compliance score (Likert five-level) increased by 1.4 points.

[0052] Feedback control and reinforcement strategy adjustment performance: collecting feedback such as posture error and HRV (heart rate variability) during each training session to control the training difficulty; During the first two training sessions, the user's HRV fluctuated significantly. After the system automatically reduced the pace and difficulty, the HRV returned to the baseline ±8% from the third training session. The average dynamic policy response delay is 650ms, which is better than the traditional static parameter adjustment method (average 3 seconds + manual judgment); The difficulty increase strategy was triggered an average of 2.4 times per week, showing a positive growth trend, reflecting that the system reasonably guides users to gradually increase their training load based on their status.

[0053] Health trend prediction and degradation warning capabilities: using MS-TGCN and DSS-PNet to jointly model training behavior and daily heart rate / activity frequency; Early identification of functional degradation signals (e.g., continuous decrease in training completion rate >15%, low emotional response) in 3 patients with cognitive fluctuations; The average prediction accuracy in the first five days was 88.2%. After the early warning and intervention recommendation mechanism was triggered, the training effects of the two patients were stabilized and further deterioration was avoided.

[0054] Overall system deployment performance: Average edge processing time for tasks: 320ms; Model incremental update cycle: synchronize once every morning; Graph reconstruction efficiency: single node change insertion time <100ms; The network stability and delay tolerance capability were verified through the hospital's internal QoS test.

[0055] Control group and comparison results: 10 patients in the control group were assigned to a traditional manual rehabilitation method (fixed tasks + periodic assessment + no emotional feedback). The main evaluation indicators were compared as follows: This embodiment clearly demonstrates that by integrating modules such as semantic perception, multi-path training generation, enhanced feedback control, and degradation trend prediction, the present invention can achieve the following: Significantly improve the personalized fit of rehabilitation training; effectively regulate the user's psychological and physiological load during training; identify risk trends in real time and intervene proactively; and achieve superior training efficiency and compliance improvement compared to traditional methods. This invention offers significant advantages in practicality, responsiveness, and sustainable intervention capabilities, making it highly valuable for application and promotion.

[0056] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. A scenario-interactive AI rehabilitation assessment training and health monitoring system, characterized by: include: The semantic perception module is used to analyze the user's voice input, facial expressions, movement patterns, and eye movements during training, and extract contextual parameters that reflect their cognitive and emotional states; The knowledge-driven training generation module, based on the extracted context parameters, combines historical training data and individual difference characteristics, calls on the dynamically evolving cross-modal rehabilitation knowledge graph, performs task adaptation reasoning and multi-path training program construction, and is used to dynamically generate rehabilitation assessment tasks and training paths that are consistent with the current context; The training feedback control engine uses reinforcement learning strategies to dynamically adjust the difficulty level, rhythm parameters, and prompt methods of the current training task in real time based on the user's posture deviation, physiological stress changes, emotional feedback, and task completion rate during training. The prediction module is used to integrate the collected training data and daily health monitoring data into a model, extract the characteristics of high-frequency health events, build a long-term health trend model, and predict the risk level of functional degradation; The cloud-edge fusion platform is equipped with edge computing nodes for low-latency processing and on-site control during training, and cloud servers for long-term data archiving, knowledge graph evolution and cross-user modeling.

2. The scenario-interactive AI rehabilitation assessment training and health monitoring system according to claim 1, characterized in that: The knowledge-driven training generation module includes: The context parameters output by the semantic perception module are jointly encoded with the user's historical training data, evaluation records, and individual physiological feature vectors to generate an individual rehabilitation status index, which serves as the entry key for querying the dynamically evolving knowledge graph. Calling a cross-modal rehabilitation knowledge graph built on a graph neural network. The graph nodes include training action units, evaluation indicators, feedback mechanisms, and their semantic labels. Edges represent task sequence dependencies and adaptation conditions. The information propagation mechanism within the graph is used to retrieve a set of candidate training nodes that meet the current state. Based on the candidate node set and its upstream and downstream dependency paths, training task adaptation reasoning based on edge-weighted scoring is performed to construct a set of multi-path training schemes with high correlation with the current context parameters. The top N schemes are output after sorting by task load gradient, realizing dynamic generation and refined matching of training content.

3. The scenario-interactive AI rehabilitation assessment training and health monitoring system according to claim 2, characterized in that: The multi-path training scheme set construction further includes: Based on the set of candidate task nodes, an improved path guidance generation algorithm is adopted, combined with context parameters and training target constraints. By introducing an adaptive path scoring gating mechanism in the graph neural network, the synergistic effect of different training paths is dynamically evaluated to generate an optimized task path set; wherein, the path scoring gating mechanism constructs a multi-factor control function based on training success rate, user compliance feedback and semantic conflict between paths, dynamically adjusts the path selection priority, and updates the local edge weights of the graph structure.

4. The scenario-interactive AI rehabilitation assessment training and health monitoring system according to claim 1, characterized in that: The training feedback control engine includes: Collect four types of feedback signals from users during training: posture deviation, physiological stress indicators, real-time emotion classification results, and task completion rate. Construct a state vector through feature normalization and time window aggregation. The state vector is input into a dual-channel deep reinforcement learning network optimized based on a multi-factor attention mechanism to learn the optimal control strategies for the training task difficulty-completion efficiency path and the rhythm-emotional fluctuation response path, and the strategy is updated based on historical execution rewards; According to the network output results, the action complexity, feedback rhythm parameters and prompt interaction mode of the current training task are adaptively adjusted in real time to trigger a fine-grained difficulty climbing or slowing down mechanism, and dynamic incentive feedback is output at the same time.

5. The scenario-interactive AI rehabilitation assessment training and health monitoring system according to claim 4, characterized in that: The reinforcement learning strategy further includes: using a strategy mapping network based on a structurally variable topology to map training feedback signals of different dimensions to the task attribute space, and dynamically reconstructing the strategy response path through graph structure transformation and strategy nesting update mechanism; wherein, the strategy mapping network encodes the training state space into a multi-layer plasticity node graph structure, with different task dimensions as target nodes, generates a multi-strategy combination response matrix through cross-layer association learning, and dynamically selects the optimal control path combination based on the user's current state.

6. The scenario-interactive AI rehabilitation assessment training and health monitoring system according to claim 1, characterized in that: The prediction module includes: The rehabilitation movement data and evaluation feedback collected during the training period are synchronized and semantically aligned with the physiological signals and behavioral frequency data from daily health monitoring to construct a spatiotemporal fusion dataset containing time tags, behavioral semantics, and physiological indicators. Using a multi-scale time series-graph convolutional joint model optimized based on a heterogeneous attention mechanism, we extract periodic health fluctuation characteristics and sudden high-frequency event sequences, and perform feature clustering and label attribution on various behavioral-physiological coupling patterns. A long-term health trend map is constructed based on the clustering results, and a prediction result of the functional degradation risk level within a preset period in the future is output through joint modeling of time series trend fitting and degradation sensitivity scoring function.

7. The scenario-interactive AI rehabilitation assessment training and health monitoring system according to claim 6, characterized in that: The long-term health trend map further includes: introducing a degradation-driven variable-scale prediction network, using functional degradation risk sensitivity as a dynamic control factor, and adaptively adjusting the temporal perception window and trend recognition scale of the prediction model; wherein, the long-term health trend map uses a convolutional gating mechanism to strengthen modeling of high-volatility segments and implement weak prediction constraints on stable segments.

Images