Monitoring system for cerebrovascular disease rehabilitation teaching based on AI technology

Abstract

The invention relates to the technical field of monitoring for cerebrovascular disease rehabilitation teaching, and discloses a monitoring system for cerebrovascular disease rehabilitation teaching based on an AI technology. The system comprises an acquisition module, an extraction module, a frame module, an organization module, an analysis module and a rendering module, and sequentially realizes multi-modal teaching data acquisition, knowledge extraction, AI model training, teaching content organization, learner action evaluation and feedback rendering, and constructs an intelligent rehabilitation teaching display interface. According to the application, high-quality acquisition and fusion modeling of multi-modal data in cerebrovascular disease rehabilitation teaching are realized, and an AI teaching system with knowledge-driven and intelligent feedback capabilities is constructed, so that the individuation, systematicness and interactivity levels of rehabilitation action teaching are improved.

Description

Technical Field

[0001] The present application relates to the field of monitoring technology for cerebrovascular disease rehabilitation teaching, and in particular to a monitoring system for cerebrovascular disease rehabilitation teaching based on AI technology. Background Art

[0002] Currently, patients with cerebrovascular diseases rely mainly on face-to-face guidance from rehabilitation therapists during rehabilitation training. In clinical practice, oral demonstration, video teaching, and physical assessment are often used for rehabilitation training teaching and effect evaluation. With the development of technology, some training programs have introduced motion capture, video monitoring, and wearable devices to assist in recording the training process, and use video playback and movement comparison to provide preliminary teaching feedback. However, such methods are still mainly based on linear playback and fixed-point evaluation, and lack intelligent and interactive features. At the data level, most of them only analyze a single modality such as video or angle data, making it difficult to build a multi-dimensional, structured rehabilitation knowledge system; in terms of teaching methods, the teaching content of existing platforms lacks personalized organizational strategies, and cannot provide a dynamic adjustment feedback path for the movement deviations of different patients, affecting rehabilitation efficiency and teaching accuracy.

[0003] Existing rehabilitation training technologies face several key challenges in practical application. First, multimodal data (such as skeletal key points, physiological signals, and instructional text) is not effectively integrated, making it impossible to achieve high-quality teaching data modeling. Second, existing teaching plans are mostly based on template playback, lacking dynamic analysis and personalized feedback mechanisms, making it difficult to evaluate and correct learners' movement quality in real time. Third, knowledge organization is weak, and rehabilitation content lacks graphical management and semantic guidance, resulting in chaotic teaching path planning and one-sided feedback guidance. Overall, it is difficult to meet the level of intelligence and customization required for precise rehabilitation teaching. Summary of the Invention

[0004] This application provides a monitoring system for cerebrovascular disease rehabilitation teaching based on AI technology, which is used to achieve high-quality acquisition and fusion modeling of multimodal data in cerebrovascular disease rehabilitation teaching, and to build an AI teaching system with knowledge-driven and intelligent feedback capabilities, thereby improving the personalization, systematicness and interactivity of rehabilitation movement teaching.

[0005] The present application provides a monitoring system for cerebrovascular disease rehabilitation teaching based on AI technology, and the monitoring system for cerebrovascular disease rehabilitation teaching based on AI technology includes:

[0006] The acquisition module is used to collect and process multimodal data of the rehabilitation expert's demonstration movements to obtain a standardized teaching basic data set including skeletal key point data, video data, and physiological parameter data;

[0007] An extraction module, configured to perform knowledge extraction and association analysis on the standardized teaching basic data set to obtain a rehabilitation teaching knowledge network;

[0008] A framework module is used to input the standardized teaching basic data set and the rehabilitation teaching knowledge network into a multimodal deep learning framework for training to obtain a rehabilitation teaching AI model;

[0009] An organization module, configured to perform multi-modal organization and visualization of teaching content based on the rehabilitation teaching AI model and the rehabilitation teaching knowledge network to obtain a teaching display interface;

[0010] An analysis module is used to evaluate and analyze the learner's movement data based on the rehabilitation teaching AI model to obtain a visual teaching feedback result;

[0011] The rendering module is used to perform visual rendering on the teaching feedback visualization result to obtain a target display interface.

[0012] In the technical solution provided by this application, the acquisition module collects multimodal data of standardized movements of rehabilitation experts, integrates skeletal key point data, video data and physiological parameter data, and solves the problem of relying on single video or movement data in traditional rehabilitation teaching. It significantly enhances the expression ability and multi-dimensional interpretation ability of training data, and provides a complete and structured input basis for subsequent model training. Secondly, the extraction module generates a rehabilitation teaching knowledge network through semantic labeling and structured processing of standardized teaching basic data sets. With the help of the knowledge graph construction method, the implicit association between rehabilitation movements, physiological structures, functional goals and teaching strategies is mapped into an explicit graph structure, so that the teaching content has a semantic basis that is searchable, explainable and inferable. This graphical knowledge structure not only improves the logic of the organization of teaching content, but also provides semantic support for personalized learning path planning. Furthermore, the framework module incorporates a multimodal deep learning model to fuse features from instructional data across different modalities. Specifically, it utilizes a 3D convolutional neural network, a graph convolutional network, a long short-term memory network, and the BERT language model to collaboratively extract spatiotemporal, structural, physiological, and semantic information, ultimately fusing them into a unified instructional representation vector. Using a multi-task learning strategy, this module trains an AI model for rehabilitation instruction capable of motion recognition, decomposition analysis, error detection, and feedback suggestions. A key advantage of this model lies in its improved accuracy in motion analysis, enabling intelligent discrimination and contextual understanding of complex rehabilitation movements. This makes it particularly suitable for rehabilitation scenarios involving movement deviations, compensatory movements, and uneven learning curves. The organization module, based on these AI models and a knowledge graph, constructs a multi-region, multimodal interactive instructional interface. It also incorporates visual attention mechanisms to enhance learners' movement comprehension and visual focus efficiency. The analysis module evaluates and classifies learners' movement data, providing multi-dimensional feedback on accuracy, completeness, stability, and fluency. Graphical representations are used to create intuitive instructional feedback, such as radar charts and trajectory diagrams. Finally, based on the above evaluation results, the rendering module further analyzes user usage behavior and learning paths, identifies common difficulties and personalized needs through data mining, drives the dynamic optimization and update of teaching content and interface, and makes the display interface highly responsive and interactive. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0014] Figure 1This is a schematic diagram of an embodiment of a monitoring system for cerebrovascular disease rehabilitation teaching based on AI technology in an embodiment of the present application;

[0015] Figure 2 This is a schematic block diagram of the structure of a monitoring device for cerebrovascular disease rehabilitation teaching based on AI technology in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] The embodiment of the present application provides a monitoring system for cerebrovascular disease rehabilitation teaching based on AI technology. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0017] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiment of the present application, an embodiment of the monitoring system for cerebrovascular disease rehabilitation teaching based on AI technology includes:

[0018] The acquisition module 101 is used to collect and process multimodal data of the rehabilitation expert's demonstration movements to obtain a standardized teaching basic data set including skeletal key point data, video data, and physiological parameter data;

[0019] Extraction module 102, for performing knowledge extraction and association analysis on the standardized teaching basic data set to obtain a rehabilitation teaching knowledge network;

[0020] The framework module 103 is used to input the standardized teaching basic data set and the rehabilitation teaching knowledge network into the multimodal deep learning framework for training to obtain a rehabilitation teaching AI model;

[0021] The organization module 104 is used to perform multi-modal organization and visualization of teaching content based on the rehabilitation teaching AI model and the rehabilitation teaching knowledge network to obtain a teaching display interface;

[0022] Analysis module 105, used to evaluate and analyze learner movement data based on the rehabilitation teaching AI model to obtain teaching feedback visualization results;

[0023] The rendering module 106 is used to perform visual rendering on the teaching feedback visualization results to obtain a target display interface.

[0024] Specifically, the acquisition module 101 assumes the core responsibility of building a multimodal data source. This is achieved by synchronously and accurately collecting various types of data from rehabilitation experts performing standardized rehabilitation movements, and then performing standardized preprocessing on the collected raw data to form a basic teaching data set that is structured, trainable, and reusable. During the acquisition process, the system uses high-frame-rate, high-resolution, multi-angle video equipment to capture the overall movement posture of the rehabilitation expert. At the same time, a three-dimensional skeleton tracking device based on deep vision + inertial measurement unit is used to obtain a three-dimensional spatial coordinate sequence containing 42 skeletal key points. In addition, wearable electromyography sensors and electrocardiography sensors are used to collect muscle activity intensity and cardiac electrophysiological waveforms, and embedded microphones are used to record the expert's oral teaching points. After the acquisition is completed, the data of each modality first undergoes time alignment processing. The image frames, bone data frames and physiological parameter waveforms are aligned to millisecond time resolution through a unified timestamp mechanism. Then, data cleaning is performed, such as removing bone point drift, blurred frames of video images, and artifact interference segments in electrophysiological signals. Then, data normalization processing based on standard templates is used to map each data to a unified scale, format and structure, and data indexes are established according to movement categories. The data are archived into five major categories of rehabilitation movement libraries, including upper limb, lower limb, trunk, balance and daily activity training libraries, to construct a standardized teaching basic data set.

[0025] The extraction module 102 performs knowledge extraction and semantic association operations based on the obtained standardized teaching basic data set. The manual annotation team labels each set of action data with dimensional labels such as action essentials, biomechanical principles, and training goals based on standard action demonstrations, and inputs structured meta-information such as action difficulty, indications, and precautions. Subsequently, the teaching text data is subjected to natural language processing, including word segmentation, entity recognition, keyword extraction, and dependency syntax analysis, to establish a set of terms and a semantic relationship graph between terms. On this basis, a knowledge graph construction strategy is adopted to establish multilateral semantic links between rehabilitation action nodes, anatomical structure nodes, functional goal nodes, common disease nodes, and teaching strategy nodes, and weight normalize the nodes according to their semantic centrality and information entropy distribution, thereby generating a rehabilitation teaching knowledge network with a graph structure. The network is stored in a graph database, so that complex semantic retrieval can be performed later.

[0026] The framework module 103 uses the above-mentioned standardized teaching basic data and rehabilitation teaching knowledge network as training data to input into the multimodal deep learning architecture. First, the skeletal key point data is input into the temporal graph convolutional network (ST-GCN) model to extract the temporal dynamic features of the action; the video frame is input into the three-dimensional convolutional network (3D-CNN) to extract the image features of spatial and temporal fusion; the physiological data is sent to the bidirectional long short-term memory network (Bi-LSTM) to obtain the temporal variation features of the electromyography and electrocardiogram signals; the teaching text is extracted through the BERT model to extract the contextual semantic embedding. The above four types of features are respectively calculated through the attention fusion mechanism to obtain the weighted feature representation and then spliced ​​to generate a unified teaching representation vector, and are input into the graph attention network (GAT) together with the node features in the knowledge graph for joint training to establish the association structure between action-semantics-physiological signals. In order to achieve different teaching objectives, the training process adopts a multi-task learning strategy. By defining three loss functions for action decomposition prediction, error detection and recognition, and teaching advancement recommendation, a rehabilitation teaching AI model that integrates knowledge and multimodal information is constructed. Finally, the functional sub-modules are divided through API encapsulation modules to form a model system with complete teaching guidance and evaluation capabilities.

[0027] The organization module 104 is based on the multi-dimensional teaching vector output by the AI ​​model and the semantic hierarchy of the knowledge network, and logically rearranges the teaching content according to the action category and learning progress. For example, it prioritizes the display of large joint movements, then fine-tunes the movements of small joints, and constructs a teaching process sequence in accordance with the "overall-local-dynamic" model. Afterwards, a multi-screen linkage teaching UI architecture is constructed, which includes a main interface, an auxiliary interface, and an information area. The system calls the encapsulated action decomposition, error demonstration and other modules in the AI ​​model according to the teaching scenario to realize three teaching control modes. Subsequently, the joint model and action trajectory generated by the skeletal key points are converted into a skeleton animation through the three-dimensional rendering engine, and visual highlights, dynamic cursors and trajectory lines are superimposed to realize real-time interactive teaching display.

[0028] The analysis module 105 receives the data of the learner during the training process, including high-frame rate video and motion data collected by sensors, and restores the learner's movements to a skeletal model through a posture estimation algorithm, and sends it to the rehabilitation teaching AI model for matching analysis. The AI ​​model outputs four-dimensional movement quality indicators: accuracy, completeness, stability and fluency. These indicators are converted into visual graphics through data mapping, such as radar charts showing multiple dimensional scores, time series charts showing changes in movement fluency, heat maps for marking areas where movement errors are concentrated, and trajectory comparison charts for comparing the different paths of experts and learners. The model also identifies characteristic deviation points based on anomaly detection methods, such as unstable rhythm, inaccurate posture or compensatory movements. Each type of error is automatically classified and marked, with correction suggestions attached in the graphical interface.

[0029] After receiving the above-mentioned graphic feedback results, the rendering module 106 further performs visual enhancement and interactive processing on them. The system analyzes the user's historical learning path, extracts typical deviation patterns and key difficult points, classifies user group behaviors based on clustering algorithms, and focuses on presenting high-frequency misunderstanding points in the interface; combines user feedback and learning logs to generate personalized recommendation content, and dynamically adjusts the layout and display order of teaching content. For example, for a learner's repeated "insufficient wrist internal rotation" error, the system will highlight the correct trajectory and key points of the action in the teaching display, dynamically scale the wrist area, and add correction suggestion bubbles. Finally, by integrating the interactive control module, a target teaching interface that supports gesture control, voice operation, and automatic teaching navigation is completed.

[0030] For example, when teaching "upper limb lifting training", experts record videos, bone data and EMG signals through a standardized platform, and the system collects them synchronously. By aligning the muscle activation pattern in the EMG signal with the time series data of the key points of the skeleton, the training model recognizes the standard electromyographic response and joint trajectory of the movement. After a rehabilitation patient learned this movement, the analysis module identified that the range of motion of his shoulder joint did not reach the standard movement of 90 degrees, and was accompanied by elbow compensation. By comparing the trajectory with the electromyographic intensity map, the system found that the deltoid muscle was not fully activated, and the biceps brachii was used for compensation. The rendering module further highlighted the area in the display interface and provided animation to guide its correction, so that learners can intuitively understand the error location and the direction of improvement, and fully realized the data processing and teaching closed loop from collection, extraction, training, display, analysis to rendering.

[0031] In a specific embodiment, the acquisition module 101 is configured to:

[0032] High-precision motion capture is performed on standardized rehabilitation movements performed by rehabilitation experts to obtain skeletal data containing the three-dimensional coordinates of 42 key points;

[0033] Multi-angle high-definition video recording of standardized rehabilitation movements performed by rehabilitation experts was performed to obtain video material data with a resolution of 4K standard and a frame rate of 60fps;

[0034] Synchronously monitor the physiological state of rehabilitation experts while they perform standardized rehabilitation movements, and obtain physiological parameter data including electromyography and electrocardiography;

[0035] Record the rehabilitation experts' oral teaching points to obtain structured teaching text data containing key technical points of movements, common mistakes and correction methods;

[0036] Perform data cleaning and standardization on skeletal data, video material data, physiological parameter data, and structured teaching text data to obtain standardized data after noise filtering;

[0037] Based on the standardized data, classification and storage index establishment are carried out to obtain a hierarchical rehabilitation movement teaching database, which includes an upper limb rehabilitation movement library, a lower limb rehabilitation movement library, a trunk rehabilitation movement library, a balance training movement library and a daily living activities training library.

[0038] Specifically, the acquisition module 101 is implemented with rehabilitation experts as the data execution subject, and high-precision, full-dimensional data acquisition is performed on the standardized rehabilitation movements performed by them through synchronous multimodal sensing means, and various types of raw data are standardized according to a unified structure, finally forming a trainable teaching data set, and establishing a hierarchical database based on the movement type. Specifically, the acquisition work first deploys an array of multi-angle, synchronously controlled image acquisition devices, using a high-definition camera with 4K resolution and 60fps frame rate to record the entire process of the rehabilitation expert completing the movement from different perspectives, ensuring that the image sequence has continuity and detail expression capabilities in the spatial and temporal dimensions. The camera synchronization signal and all sensors are timestamp-aligned through a unified clock to avoid data mismatch. Skeletal data acquisition uses a motion capture system based on vision and inertial fusion, deploys several IMUs at major joints, obtains their three-dimensional coordinate trajectories within the entire movement cycle, and uses the Kalman filter algorithm to perform noise smoothing to remove error drift caused by instantaneous vibration of the IMU. Each frame of skeletal data contains a spatial position matrix of 42 keypoints, each with three coordinate dimensions, forming a 42×3 matrix sequence. This matrix is ​​arranged in a frame-by-frame sequence to form time-series skeleton data. Physiological parameter data, including electromyography (EMG) and electrocardiography (ECG), are also collected synchronously with the movements using a multi-channel surface electromyography (SEM) recorder and an electrocardiogram (ECG) recorder, respectively. EMG sensors are primarily attached to target muscle groups. For example, during an upper limb lift, electrodes are attached to the anterior deltoid and biceps brachii. The sampling frequency is set to 1000Hz to capture a sequence of muscle activity voltage waveforms. ECG signals are used to determine the degree of physical tension or fatigue during the movement, with a sampling frequency of 500Hz. During data collection, the EMG signal contains high-frequency motion artifacts and power frequency interference. A 50Hz notch filter and bandpass filter are used to cut off and remove noise, preserving the characteristic signal bands of muscle activation. All waveform data are converted into a one-dimensional array sequence in a unified format for subsequent analysis.

[0039] Simultaneously, the acquisition module records the rehabilitation expert's spoken content during the demonstration. Using a far-field microphone array to achieve noise-reduced speech capture, the raw speech is transcribed into text using a speech recognition engine. Combined with natural language processing (NLP) processes, including sentence segmentation, keyword extraction, syntactic analysis, and entity recognition, the oral instructional text is structured into four information fields: key technical points, common error types, descriptions of correction methods, and descriptions of the target training effectiveness. All oral text undergoes manual proofreading to enhance semantic accuracy and is then indexed and mapped to video frame numbers. After the acquisition of the five raw data types—video, skeleton, EMG, ECG, and text—the data cleaning process begins. Image data is first filtered to remove frames with unstable frame rates and motion blur, and distorted frames are removed based on a threshold value from an image clarity evaluation function. Skeletal data uses missing point interpolation and outlier smoothing to remove jumpy frames. Invalid segments are identified and removed from the EMG and ECG spectrograms based on their morphology. The instructional text corrects redundant segments based on keyword distribution and punctuation. All data is then uniformly encoded into a JSON structure, and a data index table is constructed by exercise number, with time start and end stamps and multimodal content identifiers. Different modalities are aligned and calibrated using “action number + timestamp” as the joint key to form a data fusion unit, which is named as a standardized data block.

[0040] After data standardization is complete, the module assigns the movements to a pre-defined rehabilitation movement teaching database based on their category. This database, based on common classification standards in rehabilitation medicine, has five main directories: upper limb rehabilitation movement library, lower limb rehabilitation movement library, trunk rehabilitation movement library, balance training movement library, and daily living activities training library. Each library is further divided into a three-level structure of basic movements, combination movements, and auxiliary training movements, with a tree-like index. Each data block is associated with four label fields in the database: movement name, target muscle group, application scenario, and difficulty level. The data source and version records are also retained for subsequent iterative updates.

[0041] In one embodiment, the extraction module 102 is configured to:

[0042] Expert analysis and annotation of standardized teaching basic data sets were performed to obtain rehabilitation movement teaching knowledge points including movement essentials, biomechanical principles, functional training goals, difficulty levels, and precautions;

[0043] The importance of rehabilitation movement teaching knowledge points was evaluated and weighted, and a hierarchical teaching focus system was obtained.

[0044] Perform natural language processing analysis on the teaching text data in the standardized teaching basic data set to obtain a rehabilitation teaching professional vocabulary database containing core professional terms, their definitions and related relationships;

[0045] Based on the knowledge graph technology, the relationship between the rehabilitation movement teaching knowledge points, the teaching focus system and the rehabilitation teaching professional vocabulary database is mapped to obtain a knowledge network structure containing multiple nodes and connecting edges;

[0046] The nodes in the knowledge network structure are classified and labeled according to rehabilitation action nodes, anatomical structure nodes, functional training target nodes, disease symptom nodes and teaching method nodes to obtain a rehabilitation teaching knowledge network;

[0047] Based on graph database technology, the rehabilitation teaching knowledge network is stored and indexed to obtain a rehabilitation teaching knowledge base that supports complex semantic association retrieval and knowledge reasoning.

[0048] Specifically, the extraction module 102 relies on a structured and standardized data foundation. Its primary task is to perform semantic-level knowledge mining and structural organization on the standardized teaching basic data set generated by the aforementioned acquisition module, ultimately constructing a rehabilitation teaching knowledge network with semantic retrieval capabilities. In actual implementation, the extraction module first retrieves the identification information of each data unit and, relying on pre-set index fields in the standardized data, synchronously loads the corresponding video, skeletal movements, electromyographic signals, and electrocardiographic data with structured text. Experts are then organized to analyze and annotate each training movement in sections. During operation, the video frames and skeletal animation synchronization trajectories are viewed through the teaching visual interface. The muscle group activation sequence and intensity waveform are analyzed in combination with the electromyographic signals to determine the quality of movement completion and key links. The analysis process breaks each movement into several key sub-segments, each of which is accompanied by technical essentials, biomechanical principles, functional goal description, movement difficulty level, and safety precautions. These contents are converted into structured tags, each containing a field name, field value, and time period index, and the tags are bound to the original movement data.

[0049] After completing the knowledge tagging, the system evaluates the importance of all the generated knowledge point tags. The evaluation method is based on weighted scoring based on the frequency of the tag, the degree of coupling between the tag and the high-frequency error type, and the strength of the association between the tag and the target rehabilitation part. Taking the coupling degree calculation as an example, the system counts the frequency of co-occurrence of the error type "insufficient shoulder abduction" with multiple action tags to form a label-error co-occurrence matrix, and uses the TF-IDF idea to normalize the weight of each tag, thereby forming a hierarchical knowledge structure based on the semantics of rehabilitation teaching, in which high-weight tags are promoted to teaching points. For example, the biomechanical parameter of "the maximum angle of shoulder abduction is controlled between 85° and 95°" is rated as a high-weight item and included in the key prompt tag library.

[0050] The extraction module uses natural language processing techniques to semantically decompose the specialized knowledge contained in the instructional text. The process first involves sentence segmentation and lexical analysis of the transcribed text. A named entity recognition model is used to identify entity terms such as rehabilitation exercise names, anatomical structures, common conditions, and equipment names. Dependency parsing then identifies the relationship structure between these entities. For example, in the sentence "The biceps brachii muscle plays the primary contractile role in elbow flexion," the entities are "biceps brachii" and "elbow flexion," and the semantic relationship is "role-dominant," forming the triple "biceps brachii, dominant, elbow flexion." All triples are organized into a vocabulary, known as the Rehabilitation Teaching Professional Terminology Glossary. Each term is accompanied by a definition, synonyms, antonyms, and frequency of occurrence. During the knowledge graph construction phase, the extraction module uses the aforementioned knowledge point labels, key label system, and professional vocabulary triples as graph construction elements, defining node and edge types. Node types include rehabilitation exercise nodes, anatomical structure nodes, functional training target nodes, disease symptom nodes, and teaching method nodes. Edge types include semantic links such as "belongs to," "leads to," "guides," "restricts," and "substitutes." For example, the node "active shoulder abduction 90°" connects to the node "supraspinatus" through the edge "main activated muscle group" and also connects to "periarthritis of shoulder" through the edge "common indications." The graph is constructed using an RDF triple structure, and all nodes and edges are uniformly numbered and incorporated into the graph database.

[0051] The generated graph structure is then subjected to node classification, with different node types labeled to construct a rehabilitation teaching knowledge network that supports semantic query and reasoning. To support efficient retrieval and updates, the module stores this network in a graph database that supports attribute graph modeling, such as Neo4j. Index fields include node name, node type, associated action ID, and knowledge label level, enabling multi-hop semantic path retrieval and node relationship reasoning using a graph query language.

[0052] In a specific embodiment, the framework module 103 is used to:

[0053] Preprocess the different modal data in the standardized teaching basic data set to obtain standardized training data including action video data, skeleton key point data, physiological parameter data and teaching text data;

[0054] A multimodal feature extraction network structure was constructed for the normalized training data, resulting in a feature extraction module that includes a 3D convolutional neural network for processing video data, a temporal graph convolutional network for processing skeletal data, a long short-term memory network for processing physiological parameter data, and a BERT model for processing text data.

[0055] The multimodal features extracted by the feature extraction module are fused with the attention mechanism to obtain a unified teaching representation vector;

[0056] The teaching representation vector and the node features in the rehabilitation teaching knowledge network are jointly trained to obtain a deep learning model that integrates knowledge network information.

[0057] The deep learning model integrating knowledge network information is trained with multi-task tasks of action decomposition teaching, error detection teaching, and advanced guidance teaching tasks to obtain a trained multifunctional teaching model;

[0058] The multifunctional teaching model is modularly packaged and interface defined to obtain a rehabilitation teaching AI model with the functions of movement quality assessment, error movement detection and teaching, personalized teaching content generation and rehabilitation advanced path planning.

[0059] Specifically, the implementation of framework module 103 revolves around converting multimodal data from the acquisition and extraction modules into trainable deep learning input with knowledge fusion capabilities. First, the four types of raw data obtained during the acquisition phase—action video data, skeletal keypoint data, physiological parameter data, and instructional text data—are preprocessed. The core objectives of this preprocessing are to remove redundancy, correct anomalies, and unify the format and timeline to meet the requirements of subsequent model synchronization training. When processing video data, the original 4K 60fps video sequence is frame-extracted into a single-frame image sequence. Redundant frames without significant motion changes are removed using an inter-frame differencing method. Subsequently, the resolution and compression format are unified to form a tensor sequence that meets the input requirements of a three-dimensional convolutional neural network. In the skeletal data, each frame contains 42 keypoints. The 3D coordinates of each keypoint are sequentially arranged into a 42×3 array. After forming a time series, local time series segments are constructed using a sliding window method to enhance the model's ability to capture temporal context. The electromyogram (EMG) and electrocardiogram (ECG) waveforms in the physiological parameter data were uniformly sampled at a frequency of 1000 Hz. Bandpass filtering was used to retain their effective bands. Time window segmentation and normalization were then performed, aligning the time series length of each sample with the skeletal data. The teaching text data was encoded using the BERT model's word embedding converter, which first performed word segmentation, stop word removal, and spelling correction. The data was then fed into the BERT tokenizer to generate a fixed-length embedding representation, ensuring uniform dimensionality and semantic expression across all modal data.

[0060] The feature extraction network construction phase then begins. Video data is fed into a three-dimensional convolutional neural network (3D-CNN). This network structure uses convolution kernels to extract image variation features in both temporal and spatial dimensions, capturing subtle dynamic changes in movement, such as the time series variation of elbow flexion and extension angles. Skeletal keypoint data is fed into a temporal graph convolutional network (ST-GCN). ST-GCN treats skeletal points as nodes in a graph structure, extracting skeletal structural relationships through spatial convolution and the dynamic evolution of movement through temporal convolution, such as the change in joint angles from the starting position to the finished position. Physiological data is fed into a bidirectional long short-term memory network (Bi-LSTM), which uses its bidirectional gating mechanism to perform forward and backward modeling of the temporal waveform, extracting temporal features such as muscle contraction cycles and heart rate trends. Text data, encoded by the BERT model, is directly fed into a fully connected layer to extract deep semantic embeddings representing instructional intent, movement goals, and operational prompts.

[0061] To ensure that the features of these different modalities possess a unified semantic expression, the framework module fuses their outputs using an attention mechanism. This fusion process first sets initial attention weight parameters for each modality, calculates the gradient backpropagation impact of each modality on the action classification accuracy in the training set, and adjusts the fusion weights based on this impact to form a final unified teaching representation vector. This vector incorporates visual features from the image, structural features from the skeleton, physiological biological features, and semantic features from the text, resulting in a high-dimensional, multi-source, context-sensitive action understanding vector.

[0062] After completing the construction of the teaching representation vector, the framework module jointly trains it with the node features in the rehabilitation teaching knowledge network. The node features are provided by the knowledge graph constructed in the extraction module, including the standard essentials vector of the action node, the anatomical attribute vector of the target muscle group node, etc., and are propagated through a multi-hop graph structure through a graph neural network (such as GAT), so that the teaching representation vector can absorb the upstream and downstream semantic constraints from the knowledge network during the training process. For example, when training a "shoulder abduction" action model, the representation vector not only contains the visual trajectory of the action itself, but also integrates the force pattern features from the supraspinatus muscle node and the common error prompts from the teaching misunderstanding node, thereby forming a semantically enhanced deep model.

[0063] This deep model is optimized and trained using a multi-task learning mechanism. The multiple tasks are action decomposition teaching, error detection teaching, and advanced guidance teaching. Each task has an independent loss function and is optimized simultaneously during joint training. The action decomposition teaching task uses sequence labeling to mark key nodes on the action timeline, such as the start point, peak point, and end point. The error detection task is a binary classification task that determines the degree of match between the current action and the standard action. The advanced guidance task uses a contrastive learning framework to calculate the cosine similarity between the representation vector and the advanced target vector to determine whether the advanced standard is met.

[0064] The multifunctional teaching model, after training, is modularized and packaged during the export phase. Each functional module, such as "movement quality assessment," "error movement detection and feedback," "personalized teaching content generation," and "rehabilitation path planning," is encapsulated as a callable API. The package structure includes input data standards, a call flow chart, and output format definitions, facilitating dynamic invocation by subsequent teaching content organization modules.

[0065] In one embodiment, the organization module 104 is configured to:

[0066] The teaching contents are logically sorted according to the knowledge structure in the rehabilitation teaching knowledge network, and a teaching content sequence is obtained that is organized from basic to advanced, from overall to partial, and from static to dynamic.

[0067] Construct a multi-screen linkage interface layout architecture based on the teaching content sequence, and obtain a multi-area teaching interface structure including a main display area, an auxiliary display area, and an information display area;

[0068] The teaching function modules in the rehabilitation teaching AI model are called and configured to obtain a teaching display control system that supports standard teaching mode, comparative teaching mode, and decomposition teaching mode;

[0069] The video data and skeleton key point data in the standardized teaching basic data set are visualized and rendered according to the teaching display control system to obtain dynamic teaching display elements including a three-dimensional skeleton model and motion trajectory curves;

[0070] Based on the visual attention guidance technology, the dynamic teaching display elements are marked with key points to obtain a visual guidance layer including dynamic cursor, area highlight and trajectory tracking line;

[0071] The multi-region teaching interface structure, dynamic teaching display elements and visual guidance layer are integrated and assembled to obtain a teaching display interface that supports multi-perspective observation and layered display.

[0072] Specifically, the implementation of the organization module 104 relies on the rehabilitation teaching knowledge network constructed by the extraction module. This network includes multiple teaching-related nodes and their relationship edges, such as action nodes, anatomical structure nodes, functional training target nodes, and teaching method nodes. Before organizing the teaching content, the node relationship graph in the knowledge network is first read. Based on the semantic level, connection density, and task label of the nodes, a directed graph sorting method based on topological sorting is used to establish a hierarchical sequence of teaching content. This sorting uses priority rules based on the order of "from basic to advanced," "from overall to local," and "from static to dynamic." For example, if the node "Shoulder abduction basic training" is connected to the nodes "Supraspinatus activation guidance" and "Central control coordination improvement" in the knowledge graph, and the latter is a high-level training target node, the sorting algorithm determines that "Shoulder abduction basic training" should be displayed first. The sorting rules use weighted scoring based on the node in-degree and edge label attributes. The scoring results are used to establish a teaching content sequence, forming a training path automatically arranged according to logical and semantic relationships. After the teaching content sequence is generated, the organization module calls the interface builder module to establish a multi-screen interactive teaching display structure. The structure consists of a main display area, an auxiliary display area, and an information display area. The main display area is used to play rehabilitation expert demonstration videos or rendered skeletal animations, the auxiliary display area is used to display the electromyogram, electrocardiogram, and timeline of key movement nodes of the current movement, and the information display area is used to load teaching key points, movement structure diagrams, and error type prompts. Each module is bound to a specific content ID in the front-end UI, and content is dynamically loaded in a data-driven manner. When the user switches teaching units, there is no need to refresh the entire interface; only the content in the corresponding area is updated.

[0073] In order to make the display content interactive, the organization module further calls the teaching function module in the rehabilitation teaching AI model. The function modules include "standard teaching mode", "comparative teaching mode" and "action decomposition teaching mode". In the standard teaching mode, the display content only contains standard demonstrations and relevant key texts; in the comparative teaching mode, the skeletal trajectory of the learner's motion data is superimposed on the standard trajectory, and the trajectory offset is represented by the difference; in the action decomposition teaching mode, the organization module automatically plays the key frame animation and pauses at the key node according to the action key frame segmentation points returned by the AI ​​model, highlighting the teaching points through edge prompts and magnification effects. The control system automatically adjusts the teaching mode in the background based on the path weight between the user behavior data and the knowledge graph nodes to improve teaching efficiency.

[0074] In the actual rendering of presentation content, the organization module imports preprocessed video frames and skeletal keypoint data into a 3D visualization engine to generate a 3D skeletal model animation. The skeletal model uses 42 keypoints as driving nodes and employs a skeletal constraint drive mechanism to establish angular relationships between bone segments. Keypoint trajectory interpolation generates a coherent motion flow. The motion trajectory curve generates a smooth path line by reading the coordinate changes of the keypoints in each frame. This curve is plotted in 3D space to illustrate the range of motion and joint changes. The trajectory line is color-coded based on velocity or acceleration to convey the stability and rhythm of the movement.

[0075] To enhance user focus on key movement structures, the organization module introduces a visual attention guidance mechanism. It first analyzes which key points or bone segments in the current frame are the key points of instruction. It then overlays a dynamic cursor on the rendering layer for real-time tracking, or highlights target muscle groups. The visual guidance layer contains three types of layers: a dynamic cursor track layer, a regional highlight layer, and a track tracking line layer. These three types of information are overlaid on the 3D animation with increasing transparency to ensure visual clarity and focus. The organization module combines and binds the aforementioned multi-region interface structure, dynamic teaching display elements, and the visual guidance layer. Each module is bidirectionally bound to the corresponding data ID, creating a teaching display interface that supports multi-perspective viewing and layered display. The perspective switching function rotates based on the skeletal coordinate system, updating the viewing angle in real time through the 3D coordinate transformation matrix. The layered display function uses layer control to allow users to view the skeleton layer, muscle activation layer, or attention guidance layer separately, meeting different teaching needs.

[0076] In one embodiment, the analysis module 105 is configured to:

[0077] Conduct multimodal data collection on the learner's rehabilitation training process to obtain learner motion data including high-definition video data and motion parameter data collected by wearable devices;

[0078] The learner's movement data is input into the rehabilitation teaching AI model for movement evaluation calculation, and evaluation index data including four dimensions of movement accuracy, completeness, stability and fluency are obtained;

[0079] Perform data visualization conversion on the evaluation indicator data to obtain a graphical feedback interface including radar charts, heat maps, trajectory comparison charts, and time series charts;

[0080] Based on the graphical feedback interface, movement deviation detection and analysis are carried out to obtain graphical correction guidance including deviation angle annotation, correct movement trajectory guide lines and key posture correction auxiliary lines;

[0081] The common errors in the learner's movement data are classified and identified, and three types of error point marks are obtained, including compensatory movement errors, force errors and rhythm errors.

[0082] The graphical feedback interface, graphical correction guidance and error point marking are integrated and assembled to obtain a teaching feedback visualization result that supports multi-level display.

[0083] Specifically, the analysis module 105 performs structured collection, model reasoning, feature analysis, graphical processing and error classification on the multimodal data generated by the learner during the rehabilitation training, and finally outputs a visual teaching feedback interface. In the data collection link, the camera equipment installed in the teaching space first obtains high-definition video data. The video records the complete movement process of the trainee at a frequency of 60fps, supplemented by the skeleton capture device to record the skeleton key point data in real time, and the wearable motion monitoring device (such as IMU sensor) obtains the motion parameters such as limb angle, acceleration and angular velocity. All data are synchronized using a unified timestamp to constitute the trainee's motion data, where the skeleton data is arranged as a three-dimensional coordinate sequence of 42 key points by frame, and the motion parameters are arranged in time sequence as an angle-acceleration-time series matrix, which constitutes the analysis input source. After the acquisition is completed, the motion data is sent as input to the rehabilitation teaching AI model completed by the above training. The model has the ability to evaluate the quality of movement and recognize errors. Its model structure consists of a multimodal feature extraction network and a fusion discriminant layer. The spatial structure and temporal evolution characteristics of the skeleton key point data are extracted by the time-series graph convolutional network, and the rhythmicity and speed change trend of the motion parameter data are extracted by the bidirectional long short-term memory network. The video frame can be used to assist in image comparison analysis, which can be disabled in this embodiment. In the model output, each action instance will be mapped to an action representation vector and the distance in the feature space will be calculated with the standard action representation vector. The dimensions of the distance vector are defined as accuracy, completeness, stability and fluency respectively. The four index values ​​are obtained by jointly optimizing the multi-task loss function in the model. For example, the accuracy index calculates the posture similarity between the standard and the learner's key posture through the skeleton key frame matching algorithm, the completeness is calculated by comparing the action time coverage with the target amplitude, the stability evaluates the consistency of the action through the acceleration variation coefficient of the key point, and the fluency depends on the comprehensive score of the first-order derivative change trend of the timing parameter curve.

[0084] After obtaining the four-dimensional indicator vector, the analysis module performs a graphical conversion process. Accuracy, completeness, stability, and fluency are used as data inputs for the four axes of the radar chart to form an overall score overview. The range of motion of each key point in the skeletal data is mapped to a heat map color intensity value, which is rendered in color on the three-dimensional skeleton model. For example, the joint area with excessive movement will be marked in red, and the static or insufficiently active area will be blue; the trajectory comparison chart directly draws the spatial difference line between the standard trajectory and the learner's trajectory. The path with a larger difference is represented by a bold or dotted line in the figure. The timing chart uses the time axis as the horizontal axis and key parameters such as the shoulder lifting angle and elbow flexion and extension speed as the vertical axis to depict the trend of movement changes.

[0085] Based on the feedback graphics, the module detects and analyzes movement deviations by performing frame-level scanning of trajectory differences and angle errors, marking nodes that exceed the error threshold as deviation points. Each deviation point is associated with three elements: a deviation angle annotation, a guide line for the correct movement trajectory, and an auxiliary line for correcting key postures. These three pieces of information are superimposed on the skeletal animation to guide learners in making corrections. Visual information is not only presented through graphics, but also accompanied by text prompts, such as "The elbow is too high and needs to be controlled below 90°" or "The shoulder is not stable, pay attention to the coordinated force of the muscle groups."

[0086] Common error types are identified based on the classifier results from previous AI model training. The model calculates cosine similarity between the input action representation vector and predefined error type vectors, and outputs error labels such as compensatory movement error, force error, and rhythm error. Each error type is labeled with a corresponding time period and specific joint position, forming an error point marker, which is visually identified by an icon of different shapes and colors on the skeleton model.

[0087] The above-mentioned graphical feedback elements are unified and integrated into a teaching feedback visualization result interface. The interface structure is a multi-layer stacked layout, including a skeletal animation layer, a thermal rendering layer, a graphic guidance layer, a text prompt layer and an error mark layer. Each layer is bound to the data ID to ensure content consistency. The display interface supports layered viewing and interactive operations. For example, clicking on the error mark will pop up a text explanation box, and sliding the timeline can view the posture and indicator value of any frame.

[0088] In one embodiment, the rendering module 106 is configured to:

[0089] Data analysis and evaluation of the teaching feedback visualization results were performed to obtain an evaluation index system covering four dimensions: learning efficiency, knowledge mastery, skill improvement, and teaching satisfaction.

[0090] Based on the evaluation index system, a multi-dimensional analysis of the teaching feedback visualization results was conducted to obtain three optimization directions: content layer, technology layer, and interaction layer;

[0091] Conduct data mining on learners' usage habits and learning paths in the teaching feedback visualization results to obtain common learning difficulties and personalized learning needs data;

[0092] Input common learning difficulties and personalized learning needs data into the optimization algorithm to optimize the teaching content organization and obtain an updated rehabilitation teaching knowledge base;

[0093] Based on the updated rehabilitation teaching knowledge base, the teaching display interface is optimized in terms of layout and visual design, resulting in an optimized teaching interface structure.

[0094] The optimized teaching interface structure is integrated with the dynamic visual elements in the rehabilitation teaching AI model to obtain a target display interface that supports multiple perspectives and interactive operations.

[0095] Specifically, the rendering module 106 first calls the teaching feedback visualization results output by the analysis module 105 and reanalyzes its structured content to establish a data-driven evaluation index system, which includes four dimensions: learning efficiency, knowledge mastery, skill improvement, and teaching satisfaction. The input data for the analysis and processing consists of multiple structures, including the four-dimensional score vector in the radar chart data, the key point activation distribution matrix in the heat map, the spatial error vector in the trajectory comparison map, and the error mark label in the text prompt log. Learning efficiency is evaluated by calculating the time required for learners to complete a standard action, the number of repetitions, and the average error point density; knowledge mastery is weighted and scored based on the number of times the teaching text keywords are triggered and the learner's interactive behavior; skill improvement is calculated by comparing the performance trend of the same action in different time periods and calculating the Euclidean distance between learners' action vectors in the model feature space; and teaching satisfaction is comprehensively scored based on the user's feedback behavior in the interactive interface, such as dwell time, operation frequency, and the sentiment analysis results of subjective evaluation statements.

[0096] After constructing a four-dimensional evaluation index system, the rendering module uses a multi-dimensional factor analysis algorithm to perform attribution cluster analysis on the feedback data, thereby identifying optimization directions for teaching feedback content at the content, technical, and interaction levels. Optimization at the content level is determined based on the difference between knowledge mastery and skill improvement. For example, if an action involves multiple teaching points but is only partially triggered by users, and skill improvement is slow, it indicates that the content complexity needs to be adjusted. Optimization at the technical level is evaluated based on the degree of deviation between learning efficiency and accuracy. For example, if accuracy is still low despite high repetition, it indicates that the action comparison reference or guidance mechanism provided by the model is not clear enough. The interaction level is determined based on the relationship between satisfaction and interaction behavior density. For example, if perspective and function switching is frequent but feedback satisfaction is low, it indicates that the interaction design needs to be optimized. These evaluation results are then fed into the behavior analysis module to mine and process user usage habits and learning paths. The learning path consists of each learner's action sequence and interaction behavior log, structured as a time series vector, including the starting action, click nodes, repeatedly practiced action numbers, switching mode records, etc. A sequence pattern mining algorithm (such as PrefixSpan) is used to extract high-frequency path patterns and mark them as common learning paths; action nodes with high training failure rates in all paths are compared and extracted as common learning difficulties; in addition, cluster analysis of the types of errors that repeatedly occur in specific actions of specific users, feedback click behaviors and learning time are used to output personalized learning needs labels, such as "difficulty in rhythm control" and "key content requires illustrations".

[0097] Data on common learning difficulties and individual learning needs is then fed into a teaching content organization optimization algorithm, which updates the rehabilitation teaching knowledge base based on two key principles: content reordering and model adaptation. Content reordering involves reordering or splitting nodes in the original teaching path that are prone to deviation into smaller sub-action units. Model adaptation selects appropriate teaching method modules based on learner behavior tags, such as prioritizing graphical instruction for "visual learners" and interactive operation modules for "operational learners." The updated results are then written back to the rehabilitation teaching knowledge base in the form of a graph structure.

[0098] After the teaching knowledge base is updated, the rendering module regenerates the layout structure of the teaching interface. By evaluating data-driven interface layout optimization and visual style adjustments, the main display area rearranges the order of teaching module calls based on user frequency, the auxiliary area automatically recommends frequently used functions based on behavioral data, and the information display area adds user-specific prompts and common error focus areas. In the visual design, the color coding scheme adjusts the main color according to the high-frequency error areas of the heat map to improve attention focus, and the dynamic element refresh strategy is optimized to behavior-triggered refresh to reduce information interference. All optimized teaching interface structures are integrated into the current rehabilitation teaching AI model call chain to generate a target display interface. This interface supports multiple perspective observation functions, such as viewing the skeletal model movement process from different angles, and supports interactive operations, such as gesture switching mode, voice calling teaching points, and clicking on highlighted areas to enter decomposition teaching. Each visual element in the target display interface is bound to a specific data source ID and knowledge node ID to ensure that all dynamic feedback, visual prompts and interactive behaviors are traceable and can be fed back into subsequent training tasks.

[0099] For example, during a "trunk rotation control training" task, the system identified that most users performed insufficient trunk rotation between frames 5 and 9, and learners frequently clicked on the "Notes" text area for this movement, resulting in low feedback satisfaction scores. The rendering module, based on analysis, identified dual issues at the content and interaction levels. Combining this with the learning path, it determined that this node was a difficult movement for many users. The optimization algorithm then split the movement into two segments for instruction: the first segment focused solely on lumbar rotation, while the second segment guided coordinated back movements. Dynamic icon prompts for this segment were added to the interface, along with voice instruction synchronization, and visual guide lines were bolded. The resulting target display interface displayed a movement control chart and shoulder blade posture correction prompts to the right of the standard animation. Clicking on keyframe nodes allowed users to switch perspectives and view changes in spinal curves. Throughout the entire process, the data flow, from analysis results to evaluation modeling, content optimization, and interface updates, was structured and tracked. The rendering module achieved a deep reconstruction of feedback information and iterative optimization of the user teaching experience.

[0100] above Figure 1 From the perspective of modular functional entities, the monitoring system for cerebrovascular disease rehabilitation teaching based on AI technology in the embodiment of the present invention is described in detail. The following describes in detail the monitoring equipment for cerebrovascular disease rehabilitation teaching based on AI technology in the embodiment of the present invention from the perspective of hardware processing.

[0101] Figure 2: This is a structural diagram of a monitoring device for cerebrovascular disease rehabilitation teaching based on AI technology provided by an embodiment of the present invention. The monitoring device 200 for cerebrovascular disease rehabilitation teaching based on AI technology may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 210 (for example, one or more processors) and a memory 320, and one or more storage media 230 (for example, one or more mass storage device terminals) storing application programs 233 or data 232. Among them, the memory 220 and the storage medium 230 can be temporary storage or permanent storage. The program stored in the storage medium 230 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations in the monitoring device 200 for cerebrovascular disease rehabilitation teaching based on AI technology. Furthermore, the processor 210 can be configured to communicate with the storage medium 230, and execute a series of instruction operations in the storage medium 230 on the monitoring device 200 for cerebrovascular disease rehabilitation teaching based on AI technology to implement the steps of the above-mentioned monitoring system for cerebrovascular disease rehabilitation teaching based on AI technology.

[0102] The AI-based cerebrovascular disease rehabilitation teaching monitoring device 200 may further include one or more power supplies 240, one or more wired or wireless network interfaces 250, one or more input and output interfaces 260, and / or one or more operating systems 231, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. It will be understood by those skilled in the art that Figure 2 The structure of the monitoring device for cerebrovascular disease rehabilitation teaching based on AI technology shown does not constitute a limitation on the monitoring device for cerebrovascular disease rehabilitation teaching based on AI technology provided by the present invention, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0103] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A monitoring system for cerebrovascular disease rehabilitation teaching based on AI technology, characterized by: The AI-based cerebrovascular disease rehabilitation teaching monitoring system includes: The acquisition module is used to collect and process multimodal data of the rehabilitation expert's demonstration movements to obtain a standardized teaching basic data set including skeletal key point data, video data, and physiological parameter data; An extraction module, configured to perform knowledge extraction and association analysis on the standardized teaching basic data set to obtain a rehabilitation teaching knowledge network; A framework module is used to input the standardized teaching basic data set and the rehabilitation teaching knowledge network into a multimodal deep learning framework for training to obtain a rehabilitation teaching AI model; An organization module, configured to perform multi-modal organization and visualization of teaching content based on the rehabilitation teaching AI model and the rehabilitation teaching knowledge network to obtain a teaching display interface; An analysis module is used to evaluate and analyze the learner's movement data based on the rehabilitation teaching AI model to obtain a visual teaching feedback result; The rendering module is used to perform visual rendering on the teaching feedback visualization result to obtain a target display interface.

2. The AI-based cerebrovascular disease rehabilitation teaching monitoring system according to claim 1 is characterized in that: The acquisition module is used to: High-precision motion capture is performed on standardized rehabilitation movements performed by rehabilitation experts to obtain skeletal data containing the three-dimensional coordinates of 42 key points; Recording the standardized rehabilitation movements performed by the rehabilitation expert in high-definition video from multiple angles to obtain video material data with a 4K resolution and a frame rate of 60 fps; Synchronously monitoring the physiological state of the rehabilitation expert during the performance of standardized rehabilitation movements to obtain physiological parameter data including electromyogram and electrocardiogram; Recording the rehabilitation expert's oral teaching points to obtain structured teaching text data containing key technical points of movements, common errors, and correction methods; Performing data cleaning and standardization processing on the skeleton data, the video material data, the physiological parameter data, and the structured teaching text data to obtain standardized data after noise filtering; Based on the standardized data, classification and storage index establishment are performed to obtain a hierarchical rehabilitation movement teaching database, which includes an upper limb rehabilitation movement library, a lower limb rehabilitation movement library, a trunk rehabilitation movement library, a balance training movement library and a daily living activities training library.

3. The AI-based cerebrovascular disease rehabilitation teaching monitoring system according to claim 1 is characterized in that: The extraction module is used to: Performing expert analysis and annotation processing on the standardized teaching basic data set to obtain rehabilitation movement teaching knowledge points including movement essentials, biomechanical principles, functional training goals, difficulty levels and precautions; Performing importance evaluation and weight distribution on the rehabilitation movement teaching knowledge points to obtain a teaching focus system with a hierarchical structure; Performing natural language processing analysis on the teaching text data in the standardized teaching basic data set to obtain a rehabilitation teaching professional vocabulary library containing core professional terms and their definitions and associations; Based on the knowledge graph technology, the relationship between the rehabilitation movement teaching knowledge points, the teaching focus system and the rehabilitation teaching professional vocabulary library is mapped to obtain a knowledge network structure including multiple nodes and connecting edges; Classifying and labeling the nodes in the knowledge network structure according to rehabilitation action nodes, anatomical structure nodes, functional training target nodes, disease symptom nodes, and teaching method nodes to obtain the rehabilitation teaching knowledge network; The rehabilitation teaching knowledge network is stored and indexed based on graph database technology to obtain a rehabilitation teaching knowledge base that supports complex semantic association retrieval and knowledge reasoning.

4. The AI-based cerebrovascular disease rehabilitation teaching monitoring system according to claim 1 is characterized in that: The framework module is used to: Preprocessing the different modal data in the standardized teaching basic data set to obtain standardized training data including action video data, skeleton key point data, physiological parameter data and teaching text data; Constructing a multimodal feature extraction network structure for the normalized training data to obtain a feature extraction module including a three-dimensional convolutional neural network for processing video data, a temporal graph convolutional network for processing skeletal data, a long short-term memory network for processing physiological parameter data, and a BERT model for processing text data; Performing attention mechanism fusion processing on the multimodal features extracted by the feature extraction module to obtain a unified teaching representation vector; Jointly training the teaching representation vector and the node features in the rehabilitation teaching knowledge network to obtain a deep learning model that integrates knowledge network information; Performing multi-task training of action decomposition teaching, error detection teaching, and advanced guidance teaching tasks on the deep learning model integrating the knowledge network information to obtain a trained multifunctional teaching model; The multifunctional teaching model is modularly packaged and interface defined to obtain the rehabilitation teaching AI model with the functions of movement quality assessment, error movement detection and teaching, personalized teaching content generation and rehabilitation advanced path planning.

5. The AI-based cerebrovascular disease rehabilitation teaching monitoring system according to claim 1 is characterized in that: The organization module is used to: Logically sorting the teaching contents according to the knowledge structure in the rehabilitation teaching knowledge network to obtain a teaching content sequence organized from basic to advanced, from overall to partial, and from static to dynamic; Constructing a multi-screen linkage interface layout architecture based on the teaching content sequence to obtain a multi-area teaching interface structure including a main display area, an auxiliary display area, and an information display area; Calling and configuring the teaching function module in the rehabilitation teaching AI model to obtain a teaching display control system that supports standard teaching mode, comparative teaching mode, and decomposition teaching mode; Visually rendering the video data and skeleton key point data in the standardized teaching basic data set according to the teaching display control system to obtain a dynamic teaching display element including a three-dimensional skeleton model and a motion trajectory curve; Based on the visual attention guidance technology, the dynamic teaching display elements are marked with key points to obtain a visual guidance layer including a dynamic cursor, area highlight and trajectory tracking line; The multi-region teaching interface structure, the dynamic teaching display elements and the visual guidance layer are integrated and assembled to obtain the teaching display interface that supports multi-view observation and layered display.

6. The AI-based cerebrovascular disease rehabilitation teaching monitoring system according to claim 1 is characterized in that: The analysis module is used to: Conduct multimodal data collection on the learner's rehabilitation training process to obtain learner motion data including high-definition video data and motion parameter data collected by wearable devices; Inputting the learner's movement data into the rehabilitation teaching AI model for movement evaluation calculation to obtain evaluation index data including four dimensions: movement accuracy, completeness, stability and fluency; Performing data visualization conversion processing on the evaluation index data to obtain a graphical feedback interface including a radar chart, a heat map, a trajectory comparison chart, and a time series chart; Based on the graphical feedback interface, movement deviation detection and analysis are performed to obtain graphical correction guidance including deviation angle annotation, correct movement trajectory guide lines and key posture correction auxiliary lines; Classifying and identifying common errors in the learner's movement data to obtain three types of highlighted error point marks, including compensatory movement errors, force errors, and rhythm errors; The graphical feedback interface, the graphical correction guide and the error point mark are integrated and assembled to obtain the teaching feedback visualization result supporting multi-level display.

7. The AI-based cerebrovascular disease rehabilitation teaching monitoring system according to claim 1 is characterized in that: The rendering module is used to: Performing data analysis and evaluation on the teaching feedback visualization results to obtain an evaluation index system including four dimensions: learning efficiency, knowledge mastery, skill improvement, and teaching satisfaction; Perform a multi-dimensional analysis of the teaching feedback visualization results based on the evaluation index system to obtain three optimization directions: content layer, technology layer, and interaction layer; Performing data mining on the learners' usage habits and learning paths in the teaching feedback visualization results to obtain common learning difficulties and personalized learning needs data; Inputting the common learning difficulties and personalized learning needs data into an optimization algorithm to optimize the teaching content organization and obtain an updated rehabilitation teaching knowledge base; Based on the updated rehabilitation teaching knowledge base, the interface layout and visual design of the teaching display interface are optimized to obtain an optimized teaching interface structure; The optimized teaching interface structure is integrated with the dynamic visual elements in the rehabilitation teaching AI model to obtain the target display interface that supports observation from multiple perspectives and interactive operations.

Images