A stroke patient rehabilitation training method and system based on VR and motion recognition

Abstract

This invention relates to the field of rehabilitation medical equipment and intelligent training methods, specifically a rehabilitation training method and system for stroke patients based on VR and motion recognition. The system includes a multimodal data acquisition module, a VR personalized scene generation module, a motion recognition and matching module, a multimodal feedback module, a dynamic safety early warning module, a cloud data management module, and an interactive terminal. The method includes: collecting patient physiological parameters, motion data, and EEG signals; constructing personalized VR interactive scenes based on the patient's condition and rehabilitation stage; identifying deviations between actual motions and target motions in the scene using a multimodal fusion algorithm; strengthening neural connections using visual, auditory, tactile, and olfactory multimodal feedback; dynamically adjusting safety thresholds and providing graded early warnings based on real-time physiological parameters; and enabling collaborative management between doctors, patients, and families through cloud-synchronized data. This invention improves training compliance through immersive VR scenes, and multimodal technology ensures the accuracy of motion recognition and the effectiveness of feedback.

Description

Technical Field

[0001] This invention relates to the field of rehabilitation medical equipment and intelligent training methods, specifically a rehabilitation training method and system for stroke patients based on VR and motion recognition. Background Technology

[0002] Stroke is the leading cause of limb dysfunction worldwide. In my country, over 2 million new stroke patients are diagnosed annually, with 70-80% experiencing varying degrees of hemiplegia, decreased motor coordination, and other sequelae, severely impacting their ability to live independently. Rehabilitation training is a core method for promoting neural remodeling and restoring motor function in stroke patients. Its core goal is to activate the compensatory functions of damaged brain tissue and strengthen neural pathway connections through repetitive, standardized, and personalized movement training. However, current stroke rehabilitation training faces numerous technical bottlenecks, severely restricting rehabilitation effectiveness and safety.

[0003] First, neural remodeling is inefficient due to the monotonous and unengaging nature of traditional training methods. Stroke patients typically undergo long rehabilitation cycles (usually over 6 months), and traditional training often involves repetitive limb movements (such as arm raising and knee flexion), lacking immersion and interactivity. Patient compliance is generally low (only 30%-40%), making it difficult to achieve the "sufficient and standardized" training intensity required for neural remodeling. While existing VR rehabilitation systems introduce virtual scenes, these scenes are highly homogenized and lack personalized adaptation based on the patient's hemiplegic side, muscle strength level, and rehabilitation stage. For example, using the same scene for left-sided and right-sided hemiplegic patients leads to overcompensation on the unaffected side, resulting in poor training effects on the hemiplegic side. Patients with muscle strength below grade 3 experience frustration due to their inability to complete standard range movements, further reducing compliance.

[0004] Secondly, the accuracy of motion recognition is insufficient, and the feedback mechanism is simplistic. Existing systems mostly rely on a single visual sensor (such as a regular camera) for motion capture, which is easily affected by ambient light and limb occlusion. Joint angle recognition errors often exceed 5°, making it impossible to accurately determine the patient's actual exertion state (focusing only on posture and ignoring muscle contraction intensity). Furthermore, feedback methods are mostly limited to single visual cues (such as screen label deviations), lacking multimodal sensory stimulation and failing to effectively strengthen neural connections. Neural remodeling in stroke patients requires coordinated feedback from multiple senses, including vision, hearing, and touch. Single feedback cannot activate sufficient neural pathways, leading to prolonged rehabilitation periods.

[0005] Secondly, the safety protection system is weak and lacks dynamic adaptability. Stroke patients often have cardiovascular dysfunction and muscle imbalances, making them prone to joint injuries and abnormal heart rates during training due to excessive range of motion and load. Existing systems mostly use fixed safety thresholds (such as a uniformly set exercise heart rate ≤130 beats / min), failing to consider individual patient differences (such as age, underlying diseases, and rehabilitation stage) and real-time physiological states (such as muscle fatigue and heart rate fluctuations during training). This results in thresholds that are either too high or too low; thresholds that are too high fail to provide protection, while thresholds that are too low limit training effectiveness. Furthermore, the system lacks real-time monitoring of muscle fatigue and joint load, making it difficult to provide early warnings of potential risks.

[0006] Finally, the lack of a closed-loop data system and insufficient collaborative management capabilities are significant issues. In traditional rehabilitation training, patient training data (proficiency rate of movements, training duration, and deviation records) largely relies on manual recording, resulting in low accuracy and the inability to synchronize data with doctors in real time. Doctors struggle to remotely monitor patient training progress, leading to delays in adjusting rehabilitation plans. Family members also cannot promptly understand the patient's rehabilitation progress, hindering effective supervision and support. While some existing systems possess data storage capabilities, they lack integration with medical institution systems, preventing the achievement of closed-loop management for training, assessment, and adjustment, thus impacting the relevance and timeliness of rehabilitation plans.

[0007] Furthermore, existing technologies do not fully consider the neuroplasticity characteristics of stroke patients. Training program adjustments rely heavily on experience and judgment, failing to optimize based on objective indicators such as electroencephalogram (EEG) signals (e.g., motor cortical potentials, MRCP), resulting in insufficient targeting of neural remodeling. In summary, current stroke rehabilitation training systems suffer from insufficient scenario personalization, low motion recognition accuracy, limited feedback, rigid safety warnings, and a lack of data closure. There is an urgent need for an innovative technological solution that integrates VR immersive experience, high-precision motion recognition, multimodal feedback, dynamic safety protection, and cloud-based collaborative management to overcome existing technological bottlenecks. Summary of the Invention

[0008] To address the following technical problems in existing technologies: 1) Homogeneous VR scenes, lacking personalized adaptation for patients' hemiplegic side, muscle strength level, and rehabilitation stage, resulting in low training compliance and neural remodeling efficiency; 2) Motion recognition relies on a single sensor, lacking accuracy and failing to accurately capture joint angles and muscle exertion states; 3) Limited feedback methods, lacking multi-sensory synergistic stimulation, making it difficult to strengthen neural connections; 4) Fixed safety thresholds, failing to consider individual differences and real-time physiological states, leading to high safety risks; 5) Insufficient data storage and collaborative management capabilities, failing to achieve closed-loop management of training, evaluation, and adjustment; 6) Lack of objective neurophysiological indicators to support rehabilitation program adjustments, resulting in insufficient targeting. This invention provides a rehabilitation training method and system for stroke patients based on VR and motion recognition.

[0009] The technical solution adopted by the present invention to solve its technical problem is: a rehabilitation training system for stroke patients based on VR and motion recognition, including: a multimodal data acquisition module, used to collect the patient's physiological parameters (heart rate, blood oxygen, surface electromyography signal, muscle temperature), motion data (joint angle, angular velocity, force intensity) and electroencephalogram (EEG) signals; The VR personalized scene generation module is connected to the data acquisition module and has a built-in stroke rehabilitation scene library (including home, game and vocational rehabilitation scenes). Based on the patient's hemiplegic side, muscle strength level, rehabilitation stage and neuroplasticity characteristics, it generates an appropriate immersive VR interactive scene and supports dynamic adjustment of scene difficulty and task complexity. The motion recognition and matching module is connected to the data acquisition module and the VR scene generation module. It uses a multimodal fusion algorithm (visual + inertial + electromyography) to process motion data and compares the deviations (angle, force, speed deviations) between the actual motion and the target motion in the VR scene. The multimodal feedback module is communicatively connected to the action recognition module and the VR scene generation module, respectively. It achieves action deviation correction and neural remodeling enhancement through VR visual prompts, voice guidance, vibration of force feedback devices, and assistance of odor generators. The dynamic safety early warning module is connected to the data acquisition module. Based on the patient's real-time physiological parameters and historical training data, it dynamically generates safety thresholds (heart rate, electromyography, joint load thresholds) and triggers a three-level early warning. The cloud-based data management module communicates with all the modules mentioned above, stores training data, deviation records, early warning information, and rehabilitation assessment reports, and supports integration with medical institution systems and family-side apps; the interactive terminal (VR headset, touch screen) is used to display VR scenes, feedback information, and rehabilitation data, and supports patient input of operation commands.

[0010] Specifically, the multimodal data acquisition module includes a Kinect V3 depth camera (visual sensor), an MPU6050 inertial measurement unit (IMU, worn on key joints of the shoulder / elbow / wrist / hip / knee / ankle), a MYO armband (surface electromyography sensor), a MAX30100 heart rate and blood oxygen sensor, an infrared thermal imaging sensor (muscle temperature), and an EEG cap (for acquiring brain signals).

[0011] Specifically, the VR personalized scene generation module uses the Unity 3D engine to build scenes. The scene library contains three core scene categories: home scenes (clothing, eating, and washing simulations), game scenes (ball hitting and path navigation), and occupational rehabilitation scenes (office operations and handicrafts). Each scene supports 4-6 levels of difficulty adjustment.

[0012] Specifically, the multimodal fusion algorithm of the action recognition and matching module adopts a combination algorithm of Kalman filtering and Bayesian estimation, which integrates the joint coordinates of the visual sensor, the angular velocity of the IMU and the force data of electromyography signals. The action angle recognition accuracy is ≤ ±0.3° and the force intensity recognition accuracy is ≤ ±5%.

[0013] Specifically, the force feedback device of the multimodal feedback module adopts the CyberGlove II flexible force feedback glove, and the odor generator releases corresponding odors according to the scene (such as releasing the aroma of food in a home scene). When the action deviation exceeds 3° or the force deviation exceeds 10%, the multimodal feedback is triggered synchronously.

[0014] Specifically, the three-level warning of the dynamic safety warning module includes: Level 1 warning (parameter reaches 80% of the dynamic threshold, green prompt), Level 2 warning (reaches 90%, yellow pop-up window + vibration), and Level 3 warning (exceeding the limit, red pause + voice alarm). The dynamic threshold is updated based on the rehabilitation progress every training cycle (7 days).

[0015] Specifically, the cloud-based data management module uses a hybrid storage system of blockchain and MySQL, which allows doctors to remotely view data and adjust training plans, while family members receive real-time training progress and warning notifications.

[0016] Specifically, the VR headset of the interactive terminal uses Pico 4 Pro, which supports 4K resolution and 120Hz refresh rate. The touch screen is a 12-inch anti-glare screen with adjustable font size and supports voice interaction.

[0017] A rehabilitation training method for stroke patients based on VR and motion recognition, implemented using the aforementioned rehabilitation training system, includes the following steps: S1: Collect the patient's basic physiological parameters, hemiplegic side information, muscle strength level, electroencephalogram signals and initial movement data through the multimodal data acquisition module; S2: The VR personalized scene generation module generates adapted VR interactive scenes and initial training schemes based on the S1 data; S3: The patient wears a VR headset and sensors, and follows the target movements of the VR scene for training. The data acquisition module collects training data in real time. S4: The action recognition and matching module integrates multimodal data to calculate the deviation between the actual action and the target action; S5: The multimodal feedback module triggers multimodal error correction feedback based on the deviation value to strengthen neural connections; S6: The dynamic safety early warning module monitors physiological parameters in real time, triggers corresponding level warnings, and automatically pauses training if the limit is exceeded; S7: The cloud data management module stores training data, generates rehabilitation assessment reports, and synchronizes them to the doctor's and family's terminals; S8: Doctors adjust the training plan based on the report, the VR scene generation module updates the scene and difficulty, and enters the next round of training.

[0018] Specifically, in step S2, when generating the VR scene, the right visual interaction target is strengthened for patients with left hemiplegia, and the range of motion is simplified (≤60% of the standard range) for patients with muscle strength below grade 3. In step S8, the training program is adjusted every 3-7 days, and is optimized based on the patient's standard rate of motion and changes in motor-related cortical potentials (MRCP) in the electroencephalogram (EEG) signal.

[0019] The beneficial effects of this invention are: 1. Enhance training adherence and solidify the foundation for neural remodeling: The system utilizes a VR personalized scene generation module to construct immersive scenarios tailored to the patient's hemiplegic side, muscle strength level, and rehabilitation stage, avoiding the overcompensation or training frustration caused by the homogeneity of traditional scenarios. Whether simulating home life, fun games, or professional operations, the system enhances the interactivity and enjoyment of training, effectively addressing the issues of long rehabilitation cycles and the monotony of traditional training for stroke patients. This encourages patients to adhere to standardized training, ensuring sufficient and continuous training for neural remodeling. 2. Precise motion recognition and enhanced neural connectivity: Utilizing multimodal motion recognition technology, integrating visual, inertial, and electromyographic data, it can accurately capture details of a patient's joint angles and force intensity, avoiding the limitations of single sensors that are susceptible to environmental interference and have insufficient recognition accuracy. Simultaneously, through multisensory collaborative feedback including visual cues, voice guidance, force feedback vibration, and scene-appropriate odors, it can fully activate neural pathways such as vision, hearing, and touch, strengthening the compensatory function of damaged brain tissue and facilitating more efficient recovery of motor function. 3. Dynamically ensure safety and balance training effectiveness and risk: The dynamic safety early warning module abandons the rigid mode of traditional fixed safety thresholds. It combines the patient's real-time physiological parameters (such as heart rate and muscle temperature) with historical training data to generate personalized dynamic thresholds. Through a three-level early warning mechanism, it promptly alerts potential risks, from mild warnings to suspension of training when the threshold is exceeded. This can effectively avoid safety issues such as joint injury, muscle fatigue, and abnormal heart rate during training. Under the premise of ensuring patient safety, it avoids limiting training effectiveness due to excessively low thresholds, thus achieving a balance between safety and rehabilitation efficiency. 4. Establish a seamless collaborative loop to improve the quality of rehabilitation management: The cloud-based data management module enables real-time storage and sharing of training data, deviation records, and rehabilitation assessment reports. This not only allows doctors to remotely view patients' training progress and adjust rehabilitation plans in a timely manner, avoiding the problems of inaccurate data recording and delayed plan adjustments in traditional manual methods, but also allows family members to understand the patient's rehabilitation progress in real time, providing supervision and support. This forms a complete closed loop of patient training, data synchronization, doctor evaluation, plan optimization, and family collaboration, significantly improving the pertinence and timeliness of rehabilitation management. 5. Optimize the targeting of the program to meet the needs of neurorehabilitation: The system incorporates objective neurophysiological indicators such as EEG signals (e.g., motor cortical potentials, MRCP) into the program adjustment. It no longer relies on single experience judgments, but optimizes the training scenarios, difficulty and tasks based on the actual state of the patient's neural remodeling. This ensures that the rehabilitation program is more in line with the neuroplasticity characteristics of stroke patients, improves the targeting of neural remodeling, avoids the disconnect between the program and the patient's actual rehabilitation needs, and helps patients receive appropriate training support at different stages of rehabilitation. Attached Figure Description

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] Figure 1 This invention provides an architecture diagram of a stroke patient rehabilitation training system based on VR and motion recognition. Figure 2 The flowchart of a rehabilitation training method for stroke patients based on VR and motion recognition provided by the present invention is shown. Detailed Implementation

[0022] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0023] like Figure 1 As shown, the present invention discloses a VR- and motion recognition-based rehabilitation training system for stroke patients. Through multi-module collaboration and multi-technology integration, it constructs a personalized, high-precision, multi-feedback, highly safe, and highly collaborative rehabilitation training system, specifically including the following modules: System composition and functions (1) Multimodal data acquisition module As the data core of the system, it integrates six types of sensors to achieve comprehensive acquisition of physiological parameters, motion data, and neuroelectrophysiological data: Visual sensor: A Kinect V3 depth camera is used to acquire the three-dimensional coordinates of the patient's limbs and the trajectory of joint movements. The sampling frequency is 60fps, the joint angle measurement range is 0-180°, and the initial accuracy is ±0.5°. Inertial Measurement Unit (IMU): The MPU6050 chip is used and worn on six key joints of the patient: shoulder, elbow, wrist, hip, knee and ankle. It collects joint angular velocity (range ±2000° / s) and acceleration (range ±16g) to compensate for the blind spots of the visual sensor. Surface electromyography (EMG) sensor: Using a MYO armband, it collects EMG signals from the hemiplegic limb (sampling rate 200Hz, range 0-10mV) to reflect muscle contraction intensity and fatigue state. Physiological monitoring sensors: MAX30100 heart rate and blood oxygen sensor (heart rate measurement range 30-250 beats / min, blood oxygen accuracy ±1%), infrared thermal imaging sensor (muscle temperature measurement range 32-42℃, accuracy ±0.2℃). EEG acquisition equipment: EEG cap (32 channels, sampling rate 1000Hz), which collects EEG signals during the patient's training process and extracts motor-related cortical potentials (MRCP) to reflect the state of neural remodeling.

[0024] All sensors communicate with the main control unit (Intel Core i7-12700H processor) via USB 3.2 or Bluetooth 5.2, with a data transmission latency of ≤80ms.

[0025] (2) VR Personalized Scene Generation Module An immersive VR scene library was built using the Unity 3D engine, and customized training scenes were generated based on individual patient characteristics. The scenario library is categorized into three core scenarios, each corresponding to different rehabilitation goals: ① Home-based rehabilitation scenarios (dressing, eating, washing, cleaning; goal: to restore self-care ability); ② Game-based rehabilitation scenarios (ball hitting, path navigation, block building; goal: to improve motor coordination and enjoyment); ③ Occupational rehabilitation scenarios (keyboard input, document organization, handicrafts; goal: to restore occupational skills); Personalized adaptation logic: ① Adjust the position of the scene interaction target according to the hemiplegic side (for patients with left hemiplegia, the interaction target is concentrated in the right visual field to guide the left limbs to actively exert force; the opposite is true for patients with right hemiplegia). ② Adjust the range of motion according to the muscle strength level (using the MMT muscle strength grading standard) (Muscle strength level 1-2: 40%-60% of the standard range of motion; Muscle strength level 3-4: 60%-90%; Muscle strength level 5: 100%). ③ Adjust the complexity of the scenario according to the recovery stage (acute phase: single task, static goal; recovery phase: multiple tasks, dynamic goal; sequelae phase: complex task, random goal); Difficulty adjustment mechanism: Each scene supports 4-6 difficulty levels, which can be gradually increased by adjusting the target movement speed (0.5-3m / s), action accuracy requirements (allowable deviation range of 1-5°), and number of tasks (1-5).

[0026] (3) Action recognition and matching module A multimodal fusion algorithm is used to address the problem of insufficient recognition accuracy of a single sensor. Data preprocessing: Denoising the joint coordinates of the visual sensor, the angular velocity of the IMU, and the peak value of the electromyography signal (using median filtering and wavelet transform for denoising). Fusion Algorithms: The Kalman filter algorithm fuses visual and IMU data to optimize the accuracy of joint angle calculation (final accuracy ≤ ±0.3°); the Bayesian estimation algorithm fuses electromyographic signals and angle data to calculate force intensity deviation (accuracy ≤ ±5%). Deviation calculation: Define three types of deviation indicators: joint angle deviation (the difference between the actual angle and the target angle), force intensity deviation (the percentage difference between the actual peak electromyography value and the target peak value), and movement speed deviation (the percentage difference between the actual completion time and the target time). When any deviation exceeds the set threshold (angle 3°, force 10%, speed 20%), the feedback mechanism is triggered.

[0027] (4) Multimodal feedback module Based on the theory of neural remodeling, neural connections are strengthened through multi-sensory collaborative feedback: Visual feedback: The target action is highlighted in the VR scene, and the deviation is marked with a red arrow to indicate the direction of correction (e.g., "The elbow joint needs to bend 5° more"). Auditory feedback: Using real human voice (120 words / minute, adjustable volume), the type of deviation and correction method are clearly identified, and encouraging voice (such as "Your movements are more standard now, keep it up!") is played simultaneously. Tactile feedback: The CyberGlove II force feedback glove has 18 built-in miniature linear motors that correspond to the finger and wrist joints. The motors vibrate at the deviation points (the vibration intensity is positively correlated with the magnitude of the deviation). Olfactory feedback: Scene-adaptive odor generators (such as releasing the aroma of food in a home dining scene or the fragrance of flowers in a garden scene) enhance the immersion in the scene and activate the olfactory-motor neural connection; Feedback synchronization mechanism: Visual, auditory and tactile feedback are triggered simultaneously, and olfactory feedback is continuously present in the scene to ensure multi-sensory synergistic stimulation.

[0028] (5) Dynamic safety early warning module Safety thresholds are dynamically generated based on individual patient data to achieve end-to-end security protection: Dynamic threshold generation: Based on the patient's baseline physiological parameters (resting heart rate, maximum muscle strength, joint range of motion), dynamic adjustments are made in conjunction with the rehabilitation stage and training intensity. ① Heart rate threshold = resting heart rate + (220 - age - resting heart rate) × training intensity coefficient (0.4 in the acute phase, 0.6 in the recovery phase, and 0.7 in the sequelae phase); ②Electromyography threshold = peak electromyography value at the beginning of training × 0.8 (to avoid excessive muscle fatigue); ③ Joint load threshold = safe load calculated based on the musculoskeletal finite element model (combined with patient bone density data); Level 3 early warning mechanism: ① Level 1 Warning (Parameters reach 80% of the threshold): A green warning box is displayed in the corner of the VR scene, with a voice prompt "Maintain the current rhythm and breathe steadily"; ② Level 2 warning (up to 90%): A yellow pop-up window covers 1 / 3 of the scene area, the corresponding part of the glove vibrates continuously, and a voice prompt says "You are approaching the safety limit, slow down appropriately"; ③ Level 3 warning (over the limit): A red pop-up window covers the entire screen, the system automatically pauses the VR scene demonstration, the gloves vibrate strongly, and a voice alarm says "Training has been paused, please take a break" while the warning data is recorded. Threshold update: The dynamic threshold is updated every 7 days (1 training cycle) based on the patient's rehabilitation progress (percentage of correct movement, degree of muscle strength improvement) to ensure a balance between protection and training effects.

[0029] (6) Cloud Data Management Module Establish a closed-loop management system that fosters collaboration among patients, doctors, and their families: Data storage: A hybrid storage architecture of blockchain and MySQL is adopted. Blockchain ensures that the data is tamper-proof (storing basic patient information, diagnosis results, and rehabilitation plans), while MySQL stores training data (movement deviation records, early warning records, and training duration). Local backup (128GB SSD) and cloud backup provide dual protection. Data analysis: Historical data is processed using machine learning algorithms (random forest) to generate rehabilitation assessment reports, including core indicators: changes in the standard rate of movement, the extent of muscle strength improvement, the neural remodeling index (calculated based on changes in the peak value of EEG signals MRCP), and the incidence of safety events; Collaboration features: ① Doctor's side: The web-based system allows real-time viewing of patient training data and assessment reports, and remote adjustment of training plans (modification of scenarios, difficulty, and action parameters), with a response time of ≤5 minutes; ② Family member's side: The APP receives training progress push notifications (daily / weekly reports), level 2 and above warning notifications, and supports voice calls with patients (through the built-in microphone of the VR headset).

[0030] (7) Interactive terminal Adapted to the usage habits of stroke patients, improving ease of operation: VR headset: Pico 4 Pro, 4K resolution, 120Hz refresh rate, 105° field of view to reduce eye strain; built-in speaker and microphone to support voice interaction; Touchscreen: 12-inch anti-glare capacitive screen, font supports 4 levels of magnification (standard, large, extra large, extremely large), interface adopts large icon design (icon size ≥ 2cm × 2cm), core functions (start training, pause, view report) can be triggered with one click; Voice interaction: Supports 10 core commands ("Start training", "Pause", "Adjust difficulty", "Contact doctor"), with a recognition rate of ≥98% and is compatible with accented Mandarin.

[0031] like Figure 2 As shown, the stroke patient rehabilitation training method based on VR and motion recognition described in this invention is implemented based on the above-mentioned system and includes the following steps: S1: Patient registration and data collection. Patients input basic information (age, gender, stroke onset time, hemiplegic side, underlying diseases) through an interactive terminal, and doctors enter muscle strength grade (MMT classification), bone density, and diagnosis results; patients wear multimodal sensors and complete 3 sets of basic movements (arm raising, knee flexion, grasping), and initial movement data, resting heart rate, blood oxygen, electromyography signals, and electroencephalography signals are collected.

[0032] S2: Personalized Scene and Solution Generation. The VR personalized scene generation module is based on the S1 data, selects suitable scenes (such as left hemiplegia, muscle strength grade 3, patients in the recovery period, select "home eating scene" plus "ball hitting game scene"), and sets the initial difficulty (level 3), training duration (30 minutes / time), and training frequency (1 time / day).

[0033] S3: Immersive training begins. The patient wears a VR headset and sensors, enters a VR scene, and follows the target movements for training; a virtual coach is set up in the scene to demonstrate the standard movements and provide real-time encouragement (such as "Good, hold this position for 3 seconds").

[0034] S4: Real-time motion recognition and deviation calculation. The data acquisition module collects 60 sets of data per second, and the motion recognition and matching module calculates the angle, force, and speed deviations through a multimodal fusion algorithm.

[0035] S5: Multimodal feedback error correction. When the deviation exceeds the threshold, the multimodal feedback module simultaneously triggers visual, auditory, and tactile feedback to guide the patient to adjust the movement; if the standard rate of the movement is ≥95% for 3 consecutive times, the virtual coach provides encouraging feedback (such as "Great job, you have mastered this movement!").

[0036] S6: Dynamic Safety Monitoring and Early Warning. The dynamic safety early warning module monitors heart rate, electromyography, and joint load parameters in real time and triggers corresponding level warnings. If a level three warning is triggered, the system automatically pauses training and pushes rest suggestions (such as "It is recommended to rest for 10 minutes and replenish water").

[0037] S7: Data Storage and Report Generation. After training, the cloud data management module stores the training data, generates a daily rehabilitation report, and synchronizes it to the patient, doctor, and family terminals.

[0038] S8: Iterative optimization of the program. Based on the rehabilitation report and the neural remodeling index of EEG signals, the doctor adjusts the training program every 3-7 days (e.g., if the standard movement rate is ≥90% for 2 consecutive weeks, increase the difficulty by 1 level; if the neural remodeling index improves slowly, change the scene or increase the training time); the VR scene generation module updates the scene and parameters, and enters the next round of training.

[0039] Example Example 1: Basic Multimodal VR Rehabilitation Training System (Applicable to Home Scenarios) 1. System Composition This embodiment is a basic system suitable for stroke patients undergoing home rehabilitation during the recovery period (3 months after onset). The core configuration is as follows: Multimodal data acquisition module: Kinect V3 depth camera, 6 MPU6050IMUs, MYO armband (worn on the left arm by patients with left hemiplegia), MAX30100 heart rate and blood oxygen sensor, infrared thermal imaging sensor (no EEG cap, reducing costs); the main control unit is an Intel Core i5-12400H processor, and data transmission is via USB 3.2 wired connection.

[0040] VR Personalized Scene Generation Module: Home rehabilitation scene (3 sub-scenes: dressing, eating, and washing) built with Unity 3D engine, supporting 4 levels of difficulty adjustment; Personalized adaptation logic: For patients with left hemiplegia, clothing, tableware, and toiletries in the scene are all placed in the patient's right field of vision, and the range of motion is set to 70% of the standard range according to muscle strength level 3.

[0041] Action recognition and matching module: Kalman filter + Bayesian estimation fusion algorithm, angle recognition accuracy ±0.3°, force intensity recognition accuracy ±5%, deviation threshold set at angle 3°, force 10%, and speed 20%.

[0042] Multimodal feedback module: CyberGlove II force feedback glove (left side), VR scene visual cues, voice feedback, and odor generator (releasing food aroma in eating scenes and mint aroma in washing scenes).

[0043] Dynamic safety early warning module: The dynamic threshold is based on the patient's resting heart rate (72 beats / min), age 65 years, and recovery period training intensity coefficient of 0.6. The calculated heart rate threshold is 72 + (220-65-72) × 0.6 = 119 beats / min; the electromyography threshold is the initial maximum peak electromyography value (0.8mV) × 0.8 = 0.64mV; the joint load threshold is 15N (calculated based on bone density T-value - 1.0).

[0044] Cloud-based data management module: MySQL cloud storage (Alibaba Cloud server), patient-side APP, family-side APP, and doctor-side web interface, supporting training data viewing and treatment plan adjustment.

[0045] Interactive terminal: Pico 4 ProVR headset, 12-inch anti-glare touch screen, voice interaction supports core command recognition.

[0046] 2. Work Process Taking a home-based feeding scenario training for a patient with left hemiplegia, muscle strength grade 3, in the recovery period (65 years old, resting heart rate 72 beats / min, 4 months after onset): S1: Data Collection. The patient inputs information via a touchscreen, and the doctor enters the muscle strength grade 3, left hemiplegia, and recovery period diagnosis; the patient wears a sensor to complete arm raising and grasping movements, and initial movement data (maximum range of motion of the left shoulder joint 90°, peak electromyography 0.8mV) and resting heart rate 72 beats / min are collected.

[0047] S2: Scene Generation. The system generates a "home eating scene" (goal: use your left hand to pick up a spoon from the table on the right, scoop up food and put it in your mouth), with an initial difficulty level of 3, a movement range of 70% (standard range 120°, actual requirement 84°), and a training time of 30 minutes.

[0048] S3: Training begins. The patient puts on a VR headset and gloves and enters the scene: the spoon on the virtual dining table is located on the right side (30cm away from the patient's left hand). The virtual coach demonstrates the action of "raising the arm - grasping - scooping - putting into the mouth", with voice prompts "slowly raise the left arm, feel the shoulder joint exert force, and keep the speed even".

[0049] S4: Motion recognition. The patient begins training. Kinect V3 acquires the left shoulder joint angle (actually 80°), IMU acquires the angular velocity (10° / s), and MYO armband acquires the peak electromyography (0.7mV). The fusion algorithm calculates an angle deviation of 4° (exceeding the threshold of 3°) and a force deviation of 12.5% ​​(exceeding the threshold of 10%).

[0050] S5: Multimodal feedback. A red arrow marks the position of the left shoulder joint in the VR scene, prompting "Raise it another 4°"; a voice prompt says "The left arm is not raised enough, the force is slightly weak, please adjust"; the force feedback glove's shoulder motor vibrates at a medium intensity; the odor generator releases the aroma of food to enhance immersion.

[0051] S6: Safety Monitoring. After 15 minutes of training, the patient's heart rate rose to 115 beats / min (reaching 96.6% of the threshold), triggering a level 2 warning: a yellow pop-up window appeared, the glove vibrated continuously, and a voice prompt said "You are approaching the heart rate threshold, slow down your movements"; after the patient adjusted the speed, the heart rate dropped to 110 beats / min, and the warning was lifted.

[0052] S7: Data and Reports. At the end of the training, the system recorded a 78% accuracy rate of movement, a peak heart rate of 115 beats per minute, and no level 3 warnings. A report was generated in the cloud, showing that "the range of motion of the left shoulder joint has improved by 5° compared to the initial value, and the accuracy rate of movement needs to be further improved," and was synchronized to the family member's and doctor's devices.

[0053] S8: Solution Optimization. After reviewing the report, the doctor remotely adjusts the difficulty to level 3 (maintains it), adds a "grip strength training" sub-task, updates the scene parameters, and it takes effect in the next training session.

[0054] 3. Application Effects In this embodiment, the basic system was applied to 25 stroke patients recovering at home (13 with left hemiplegia and 12 with right hemiplegia, muscle strength grade 2-4). After 8 weeks of continuous use, the test results showed: Rehabilitation outcomes: The average FMA motor function score of patients' limbs improved by 18.6 points (initial average 35.2 points, final average 53.8 points), of which 18 patients had an FMA score improvement of ≥15 points, accounting for 72%; the standardization rate of movements improved from an initial average of 62% to a final average of 89%; and the neural remodeling-related indicators (motor unit recruitment rate calculated based on electromyography signals) improved by 45%.

[0055] Safety: Only one Level 2 warning (heart rate close to the threshold) occurred within 8 weeks, with no Level 3 warnings or safety events (joint injuries, muscle strains, etc.). The safety event rate was 4%, far lower than the 18% of traditional home training.

[0056] Adherence: Patients trained an average of 6.8 times per week (target 7 times), with a training adherence rate of 97.1%, significantly higher than the 42% of traditional training; 23 patients (92%) reported that "the scenarios were interesting, the operation was simple, and the feedback was timely".

[0057] Example 2: Cloud-based closed-loop VR rehabilitation training system (institutional + home collaboration) 1. System Composition This embodiment, based on embodiment 1, enhances cloud collaboration and remote doctor guidance functions, and is suitable for rehabilitation hospitals and home-based collaborative rehabilitation scenarios. The core configuration optimizations are as follows: Multimodal data acquisition module: A new EEG cap (32 channels) is added to acquire EEG signals and extract MRCP peak values. Other sensor configurations are the same as in Example 1.

[0058] Cloud data management module: upgraded to a hybrid storage system of blockchain and MySQL, connecting to the HIS systems of three rehabilitation hospitals (using the HL7 medical data exchange standard); the doctor-side system adds a "plan template library" (categorized by stroke type and rehabilitation stage), supporting one-click generation of personalized plans.

[0059] Family App: Added "Real-time Monitoring" function (with patient authorization, family members can view real-time VR scene footage) and "Training Reminder" (daily training time can be set, and the app will push notifications).

[0060] The configuration of other modules is the same as in Example 1.

[0061] 2. Work Process Taking the "occupational rehabilitation scenario (keyboard input)" training of a 58-year-old patient with right hemiplegia, muscle strength grade 4, in the recovery period (3 months after onset, home rehabilitation after discharge from the rehabilitation hospital): S1: Data Collection. Data collection is completed at the rehabilitation hospital when the patient is discharged. The doctor enters the muscle strength grade 4, right hemiplegia, and recovery period. The EEG cap collects the MRCP peak value (2.5μV) at rest. After discharge, the home system synchronizes the hospital data, and there is no need to collect it again.

[0062] S2: Scene Generation. The system generates a "professional keyboard input scene" based on hospital data (goal: use the right hand to type virtual keyboard letters in a specified order). The initial difficulty level is 4, the movement range is 90% (standard range is 100°), the training time is 40 minutes / session, and the frequency is once a day.

[0063] S3: Training Start. The patient starts training at home. In the VR scene, the virtual keyboard is located on the right side (adapted to right hemiplegia). The virtual coach demonstrates the "wrist extension - finger tapping - wrist relaxation" movement. The data acquisition module collects movement data and EEG signals (MRCP peak 3.0μV) in real time.

[0064] S4: Motion Recognition and Feedback. During the patient's first training session, the finger tapping angle deviated by 2° (not exceeding the threshold), but the force intensity deviated by 8% (not exceeding the threshold). The system provided positive feedback: "The movement is standard, the force is stable, keep it up"; the letters in the VR scene turned green, indicating that it was correct.

[0065] S5: Safety Warning. After 25 minutes of training, the infrared thermal imaging sensor detected a muscle temperature of 38.5℃ in the patient's right forearm (exceeding the fatigue threshold of 38℃), triggering a Level 1 warning: a green message "Muscles are fatigued, switch to relaxation exercises"; the system automatically switches to wrist relaxation mode, and training is paused for 5 minutes.

[0066] S6: Cloud Collaboration. Doctors viewed patient training data in real time through the hospital's HIS system and found that the MRCP peak value increased from 2.5μV to 3.2μV (significant neural remodeling effect), but the finger tapping speed was too slow (0.8 times / second, target 1.2 times / second). The doctor remotely adjusted the plan: increased the scenario difficulty to level 5, increased the finger tapping speed requirement (1.0 times / second), and sent voice guidance: "Please increase the tapping speed appropriately and pay attention to maintaining stable force."

[0067] S7: Family Collaboration. When a patient triggers a Level 1 alert during training, the family member's app will send a notification: "Patient's muscles are fatigued, relaxation exercises have been switched." The family member can check the patient's status in real time and encourage them via voice call: "Hang in there, relax and continue."

[0068] S8: Program Iteration. During the 8-week training period, the doctor adjusted the program 3 times based on the weekly rehabilitation reports (increasing the difficulty from level 4 to level 6, and expanding the scenario from "keyboard input" to "file organization"); the system automatically optimized the movement parameters and extended the demonstration time of complex movements based on the changes in the peak value of the EEG signal MRCP (4.8μV in the late stage).

[0069] 3. Application Effects In this embodiment, the cloud-based closed-loop system was applied to 20 stroke recovery patients (10 receiving hospital + home-based collaborative rehabilitation, and 10 receiving purely home-based rehabilitation), 8 rehabilitation doctors, and 20 family members. After 8 weeks of continuous use, the results showed: Rehabilitation outcomes: The average FMA score of patients in the collaborative rehabilitation group improved by 22.3 points (36.5 points initially, 58.8 points at the end), while the score of patients in the pure home-based group improved by 17.8 points; the neural remodeling index (MRCP peak change) of the collaborative group improved by 62%, while that of the pure home-based group improved by 43%; the standardization rate of movements in the collaborative group reached 93% at the end, while that of the pure home-based group was 86%.

[0070] Doctor efficiency: The average response time for remote guidance is 3.2 minutes, and treatment plans can be adjusted without on-site visits; each doctor can manage 15-20 patients at the same time, increasing work efficiency by 60% (in the traditional model, each doctor can only manage 5-8 patients).

[0071] Family and patient satisfaction: Family awareness of patient recovery increased from 25% in the traditional model to 98%; patient satisfaction with "remote doctor guidance" reached 95%, and the recognition of "family supervision" reached 90%; training compliance in the collaborative rehabilitation group was 98%, and in the purely home-based group it was 92%.

[0072] Example 3: AI Adaptive VR Rehabilitation Training System (Neural Remodeling Enhancement) 1. System Composition This embodiment, based on embodiment 2, integrates AI reinforcement learning and EEG signal feedback optimization functions, and is suitable for patients in the sequelae stage of stroke (more than 6 months after onset). The core configuration optimization is as follows: Multimodal data acquisition module: Retains the EEG cap, and adds a muscle fatigue monitoring algorithm (based on the fusion calculation of the root mean square value (RMS) of electromyography signal and infrared thermal imaging temperature).

[0073] VR Personalized Scene Generation Module: Integrates reinforcement learning algorithm (DQN deep Q network) to analyze patient movement standard rate, muscle fatigue, and MRCP peak value of EEG signal in real time, and dynamically adjust scene tasks, difficulty, and rest intervals.

[0074] Action recognition and matching module: Added AI action intent prediction function, which predicts the patient's next action based on historical action data and optimizes the timing of feedback in advance (feedback delay reduced from 80ms to 50ms).

[0075] Multimodal Feedback Module: Added EEG feedback function. When the peak value of EEG signal MRCP increases (enhanced neural activation), an animation prompt of "accelerated neural remodeling" appears in the VR scene to strengthen positive stimulation.

[0076] Dynamic safety early warning module: The AI ​​algorithm predicts potential risks (such as a continuous increase in muscle fatigue and heart rate changes in patients in real time) and triggers an early warning 2 minutes in advance.

[0077] 2. Work Process Taking a 62-year-old patient with left hemiplegia, muscle strength grade 3, and in the sequelae stage (8 months after onset, FMA score of 42, slow rate of neural remodeling) as an example, the following training was conducted using a "game-based rehabilitation scenario (ball hitting)": S1: Data Collection. The patient wears a full set of sensors to collect basic data: resting heart rate 68 bpm, left shoulder joint range of motion 85°, peak EMG value 0.7mV, peak EEG MRCP value 2.2μV; the muscle fatigue threshold is set as EMG RMS ≥ 0.5mV and temperature ≥ 38.2℃.

[0078] S2: Scene Generation. The system generates a "ball hitting scene" (a virtual basketball is thrown from the right, and the patient hits the basketball into the basket with their left hand). The initial difficulty level is 4. The AI ​​algorithm predicts "low neural remodeling efficiency" based on the patient data and sets the initial training strategy as: low intensity and long rest interval (5 minutes of training, 2 minutes of rest).

[0079] S3: Training Start. The patient enters the scene, and a virtual basketball is thrown at a speed of 0.8m / s; the data acquisition module collects motion data in real time (left shoulder joint angle 88°, deviation 2°, not exceeding the threshold), EEG MRCP peak value 2.4μV (9% increase from the initial value), and muscle temperature 37.5℃.

[0080] S4: AI Adaptive Adjustment. The AI ​​algorithm analyzes the movement standard rate (85%), MRCP peak increase, and muscle fatigue prevention, and dynamically adjusts: basketball speed is increased to 1.0m / s (increased difficulty), training interval is extended to 6 minutes (2 minutes of rest); at the same time, EEG feedback is triggered, and the VR scene displays the animation "Enhanced neural activation, keep going!"

[0081] S5: Risk Prediction and Early Warning. After 18 minutes of training, the AI ​​algorithm predicts "muscle fatigue will occur in 5 minutes" based on electromyography (RMS) (0.45mV) and temperature (38.0℃), triggering a Level 1 warning 2 minutes in advance: "Muscles are about to become fatigued, prepare to switch to relaxation exercises"; 2 minutes later, the system automatically switches to a shoulder relaxation scenario to avoid fatigue exceeding the limit.

[0082] S6: Multimodal feedback optimization. If a patient's striking motion deviates by 4°, the AI ​​motion intent prediction module anticipates that "the patient may not be able to adjust in time," and the feedback timing is 30ms ahead. Force feedback glove vibration and voice prompt are triggered simultaneously, allowing the patient to correct the error in time, reducing the deviation to 1°.

[0083] S7: Cloud and Solution Optimization. After training, the system recorded a 91% accuracy rate in movements, a peak MRCP value of 2.8μV (an improvement of 27%), and muscle fatigue within limits. The doctor received a report, and the AI ​​algorithm generated a solution suggestion: "The patient's neural remodeling efficiency has improved, so the training intensity can be increased. We recommend adding a 'block building scenario'." The doctor adopted the suggestion and updated the solution remotely.

[0084] S8: Long-term adaptive training. During the 8-week training period, the AI ​​algorithm dynamically adjusted the training strategy based on the patient's performance: the first 2 weeks focused on low-intensity adaptation, the middle 4 weeks gradually increased the difficulty, and the last 2 weeks strengthened the training of complex movements; finally, the difficulty of the scenario increased from level 4 to level 6, and the training time was extended from 30 minutes to 45 minutes.

[0085] 3. Application Effects In this embodiment, the AI ​​adaptive system was applied to 15 patients in the sequelae of stroke (6-12 months after onset, FMA score 38-45). After 8 weeks of continuous use, the results showed: Rehabilitation outcomes: Patients' FMA scores improved by an average of 15.7 points (41.3 points initially, 57.0 points at the end), significantly higher than the 8-10 points achieved with traditional post-sequelae training; the neural remodeling index (MRCP peak change) improved by an average of 78%, with 12 patients (80%) showing a neural remodeling rate that more than doubled compared to the first 6 months; the standardization rate of movements reached 95% at the end, and the number of training interruptions due to muscle fatigue decreased to 0.5 times / week.

[0086] Safety: The AI ​​risk prediction accuracy rate reached 89%, with no Level 3 warnings or safety incidents within 8 weeks, and the occurrence rate of Level 2 warnings was only 3 times per person, all of which were prevented from escalating through early intervention.

[0087] Adaptability: 14 out of 15 patients (93%) reported that "the system can adjust the difficulty according to their own condition, and it is neither too easy nor too difficult"; the training compliance rate reached 99%, and no patients gave up training due to discomfort or poor results.

[0088] Comparison Example To verify the inventiveness and superiority of the present invention, three control groups were set up to compare with Examples 1-3. The test subjects were 60 stroke patients (20 in each group, matched for age, gender, and severity of illness). They underwent continuous training for 8 weeks. The test indicators included the improvement of FMA score, the rate of standard movement, the incidence of safety events, and training compliance.

[0089] Comparison with Example 1: Single visual recognition plus general VR scene system System configuration: Only Kinect V3 visual sensor is used, VR scene is general scene (no hemiplegic side, muscle strength adaptation), feedback method is only visual prompts, safety threshold is fixed (heart rate ≤130 beats / min), no cloud collaboration function.

[0090] Test results: The FMA score improved by an average of 9.2 points, the accuracy rate of movement was 65% at the end of the test, the incidence of safety incidents was 15% (3 cases of muscle strain), and the training compliance rate was 68%.

[0091] Comparative conclusion: The improvement in FMA of Embodiment 1 of the present invention (multimodal recognition + personalized scene + multimodal feedback) is 2.02 times that of Control Example 1, the action standardization rate is increased by 36.9%, the security incident rate is reduced by 73.3%, and the compliance rate is increased by 42.8%, highlighting the creative advantages of multimodal fusion and personalized adaptation.

[0092] Compare with Example 2: Multimodal recognition with fixed security thresholds and a cloud-free closed-loop system System configuration: The sensor is the same as in Example 2 (including multimodal acquisition), but the safety threshold is fixed (heart rate ≤120 beats / min), and there is no cloud-based linkage function with doctors / family members. The adjustment of the plan depends on the patient's own recording and feedback.

[0093] Test results: FMA score improved by an average of 14.5 points, the standard movement rate reached 80% at the end of the test, the safety incident rate was 8% (1 case of joint discomfort), and the training compliance rate was 75%.

[0094] Comparative conclusion: Embodiment 2 of the present invention (dynamic security threshold + cloud closed loop) has a 53.8% higher FMA improvement, a 16.25% higher action standard rate, a 50% lower security incident rate, and a 30.7% higher compliance rate, demonstrating the technical advantages of dynamic early warning and cloud collaboration.

[0095] Compare with Example 3: Multimodal recognition + personalized scene + AI-free adaptive system System configuration: The sensors and scenarios are the same as in Example 3, but there is no AI reinforcement learning and EEG feedback function. The training program is adjusted according to a fixed cycle (7 days) and does not take into account real-time muscle fatigue and neural state.

[0096] Test results: FMA scores improved by an average of 10.3 points, the accuracy rate of movement was 86% at the end of the test, the incident rate was 6%, and the training compliance rate was 85%.

[0097] Comparative conclusions: Embodiment 3 of the present invention (AI adaptation plus EEG feedback) showed a 52.4% higher improvement in FMA, a 10.5% higher rate of action standardization, an 83.3% lower rate of safety incidents, and a 16.5% higher compliance rate, verifying the inventive value of AI adaptation and neural remodeling-guided feedback.

[0098] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of protection claimed by the present invention. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A rehabilitation training system for stroke patients based on VR and motion recognition. Its features are, Includes: a multimodal data acquisition module, used to collect patients' physiological parameters, movement data, and electroencephalogram (EEG) signals; The VR personalized scene generation module is connected to the data acquisition module and has a built-in stroke rehabilitation scene library. Based on the patient's hemiplegic side, muscle strength level, rehabilitation stage and neuroplasticity characteristics, it generates an appropriate immersive VR interactive scene and supports dynamic adjustment of scene difficulty and task complexity. The action recognition and matching module is communicatively connected to the data acquisition module and the VR scene generation module. It uses a multimodal fusion algorithm to process action data and compares the deviation between the actual action and the target action in the VR scene. The multimodal feedback module is communicatively connected to the action recognition module and the VR scene generation module, respectively. It achieves action deviation correction and neural remodeling enhancement through VR visual prompts, voice guidance, vibration of force feedback devices, and assistance of odor generators. The dynamic safety early warning module is connected in communication with the data acquisition module. Based on the patient's real-time physiological parameters and historical training data, it dynamically generates safety thresholds and triggers a three-level early warning. The cloud-based data management module communicates with all the modules mentioned above, stores training data, deviation records, early warning information, and rehabilitation assessment reports, and supports integration with medical institution systems and family-side apps; the interactive terminal is used to display VR scenes, feedback information, and rehabilitation data, and supports patient input of operation commands.

2. The stroke patient rehabilitation training system based on VR and motion recognition according to claim 1, characterized in that: The multimodal data acquisition module includes a Kinect V3 depth camera, an MPU6050 inertial measurement unit, a MYO armband, a MAX30100 heart rate and blood oxygen sensor, an infrared thermal imaging sensor, and an EEG cap.

3. The stroke patient rehabilitation training system based on VR and motion recognition according to claim 1, characterized in that: The VR personalized scene generation module uses the Unity 3D engine to build scenes. The scene library contains three core scene categories: home scenes, game scenes, and vocational rehabilitation scenes. Each scene supports 4-6 levels of difficulty adjustment.

4. The stroke patient rehabilitation training system based on VR and motion recognition according to claim 1, characterized in that: The multimodal fusion algorithm of the motion recognition and matching module adopts a combination of Kalman filtering and Bayesian estimation algorithm, which integrates the joint coordinates of the visual sensor, the angular velocity of the IMU and the force data of electromyography signals. The motion angle recognition accuracy is ≤ ±0.3° and the force intensity recognition accuracy is ≤ ±5%.

5. A stroke patient rehabilitation training system based on VR and motion recognition according to claim 1, characterized in that: The force feedback device of the multimodal feedback module adopts the CyberGlove II flexible force feedback glove, and the odor generator releases the corresponding odor according to the scene. When the action deviation exceeds 3° or the force deviation exceeds 10%, the multimodal feedback is triggered synchronously.

6. The stroke patient rehabilitation training system based on VR and motion recognition according to claim 1, characterized in that: The dynamic safety early warning module has three levels of warnings: Level 1 warning, which displays a green prompt when the parameter reaches 80% of the dynamic threshold; Level 2 warning, which displays a yellow pop-up and vibrates when the parameter reaches 90%; and Level 3 warning, which displays a red pause and voice alarm when the parameter exceeds the limit. The dynamic threshold is updated based on the rehabilitation progress each training cycle.

7. A stroke patient rehabilitation training system based on VR and motion recognition according to claim 1, characterized in that: The cloud-based data management module uses a hybrid storage system of blockchain and MySQL, enabling doctors to remotely view data and adjust training plans, while family members receive real-time training progress and alerts.

8. A rehabilitation training system for stroke patients based on VR and motion recognition according to claim 1, characterized in that: The VR headset of the interactive terminal uses Pico 4 Pro, which supports 4K resolution and 120Hz refresh rate. The touch screen is a 12-inch anti-glare screen with adjustable font size and supports voice interaction.

9. A rehabilitation training method for stroke patients based on VR and motion recognition, wherein the method is implemented using the rehabilitation training system according to any one of claims 1-8, characterized in that, Includes the following steps: S1: Collect the patient's basic physiological parameters, hemiplegic side information, muscle strength level, electroencephalogram signals and initial movement data through the multimodal data acquisition module; S2: The VR personalized scene generation module generates adapted VR interactive scenes and initial training schemes based on the S1 data; S3: The patient wears a VR headset and sensors, and follows the target movements of the VR scene for training. The data acquisition module collects training data in real time. S4: The action recognition and matching module integrates multimodal data to calculate the deviation between the actual action and the target action; S5: The multimodal feedback module triggers multimodal error correction feedback based on the deviation value, strengthening neural connections; S6: The dynamic safety early warning module monitors physiological parameters in real time, triggers corresponding level warnings, and automatically suspends training if the limits are exceeded. S7: The cloud data management module stores training data, generates rehabilitation assessment reports, and synchronizes them to the doctor's and family's terminals; S8: Doctors adjust the training plan based on the report, the VR scene generation module updates the scene and difficulty, and enters the next round of training.

10. A method for rehabilitation training of stroke patients based on VR and motion recognition according to claim 9, characterized in that: In step S2, when generating the VR scene, the interactive target of the right visual field is strengthened for patients with left hemiplegia, and the range of motion is simplified for patients with muscle strength below grade 3. In step S8, the training program is adjusted every 3-7 days, and is optimized based on the patient's standard rate of movement and changes in motor-related cortical potentials in the EEG signal.

Images