A stroke patient hand function rehabilitation training system based on a hybrid brain-computer interface

Abstract

A hand function rehabilitation training system for stroke patients based on a hybrid brain-computer interface is disclosed. The system comprises a multimodal signal synchronous acquisition platform that simultaneously acquires EEG, surface electromyography (EMG), and inertial signals through a hardware and software-integrated synchronous triggering method. A motion intention recognition module utilizes an ensemble learning algorithm combined with a multimodal probability fusion model to recognize motion intentions. An adaptive feedback adjustment module analyzes the patient's motor function in real time and dynamically adjusts training parameters. The rehabilitation training module supports selective, individual, or synchronous execution of virtual interactive tasks and the driving of external rehabilitation equipment. This invention significantly enhances the relevance, engagement, and active participation of rehabilitation training through a hardware-software fusion synchronization strategy and multimodal information collaborative decoding, combined with adaptive dynamic feedback, gamified interactive training, and multi-device expansion capabilities. It enables proactive, personalized, and multi-scenario adapted efficient hand function rehabilitation training after stroke, effectively promoting the recovery of hand motor function in patients.

Description

Technical Field

[0001] This invention belongs to the field of rehabilitation engineering technology and relates to a hand function rehabilitation training system for stroke patients based on a hybrid brain-computer interface. Background Technology

[0002] Stroke, as a common neurological disease, often leads to limb motor dysfunction, with hand dysfunction being particularly prevalent, severely impacting patients' daily living abilities and quality of life. Traditional rehabilitation training methods, such as occupational therapy and physical therapy, rely primarily on the therapist's experience and the patient's passive participation, resulting in drawbacks such as poor initiative, lack of enjoyment, difficulty in objectively quantifying and evaluating training effectiveness, and difficulty in dynamically adjusting training programs based on the patient's real-time abilities.

[0003] With the development of brain-computer interface (BCI) technology, its application in rehabilitation medicine has provided new ideas for the functional recovery of stroke patients. However, single-modal signals have inherent limitations in decoding motor intentions. Relying solely on electroencephalogram (EEG) signals results in a low signal-to-noise ratio and susceptibility to EEG artifacts; classification performance is also low when the classification dimension is high. Using only surface electromyography (sEMG) signals is difficult for patients with severe functional impairments due to weak muscle activity, making signal acquisition challenging and hindering accurate capture of motor intentions. Inertial measurement units (IMUs) can reflect motor behavior but cannot detect motor intentions in advance. Integrating EEG, sEMG, and IMU through hybrid BCI technology can overcome the inherent limitations of single signals, creating a synergistic effect of complementary advantages and offsetting defects. However, in terms of intent recognition algorithms, existing algorithms mostly adopt the direct concatenation of multi-source features. For example, patent CN202111286589 discloses a method for recognizing motion imagery intent based on multimodal signals, which directly concatenates the extracted multimodal features into a high-dimensional vector and then inputs it into a single classifier for recognition. This direct concatenation method not only easily leads to the "curse of dimensionality" but also ignores the huge differences in the dependence of stroke patients on each modality signal due to different degrees of damage.

[0004] Currently, the lack of standardized interfaces in multi-source signal acquisition devices complicates device connections and leads to poor compatibility during signal acquisition, increasing the difficulty and cost of system construction. More importantly, significant signal deviations exist between different signals. This is mainly due to asynchronous sampling clocks of different acquisition units, differences in hardware circuit characteristics, and the influence of electromagnetic interference and varying transmission path lengths during signal transmission. Existing synchronization solutions mostly employ single hardware or software triggering methods, which are insufficient to cope with complex and variable signal transmission environments. While hardware triggering offers high synchronization accuracy, it is limited by hardware connections. When device interfaces are mismatched or numerous, pulse signal distribution and transmission are prone to attenuation and distortion, resulting in decreased synchronization effectiveness. Software triggering relies on network protocols for timestamp marking, which can solve interface compatibility issues to some extent, but in situations of network congestion or unstable transmission delays, the accuracy of timestamps is difficult to guarantee, and nanosecond-level time synchronization accuracy is even more challenging to achieve. This signal deviation causes multimodal signals to be inaccurately aligned on the time axis, leading to errors in subsequent feature extraction and motion intent recognition, severely impacting the performance of the rehabilitation training system.

[0005] Furthermore, adaptive feedback adjustment and gamification of rehabilitation training are key to improving rehabilitation outcomes. Existing rehabilitation systems often rely on preset parameters for feedback adjustment, failing to dynamically adjust auxiliary parameters based on the patient's real-time motor function, resulting in insufficient personalization. Rehabilitation training games frequently employ fixed difficulty settings, lacking a dynamic matching mechanism with the patient's motor abilities, making it difficult to balance the challenge and feasibility of training, and easily leading to decreased patient motivation. Simultaneously, game performance evaluation indicators are singular, failing to comprehensively reflect the patient's accuracy in recognizing motor intentions, response speed, and motor coordination, thus affecting the scientific validity of difficulty adjustment.

[0006] In summary, current hand function rehabilitation training for stroke patients suffers from several problems, including low initiative, low rehabilitation efficiency, lack of personalization, lack of functional feedback, poor interactivity, inaccurate perception of motor intentions, and insufficient training engagement. Therefore, developing a rehabilitation training system that can accurately perceive patients' motor intentions, provide personalized feedback, and offers high interactivity and engagement is of significant practical importance for improving the effectiveness of hand function rehabilitation for stroke patients. To address these issues, there is an urgent need to develop a hand function rehabilitation training system for stroke patients based on a hybrid brain-computer interface, aiming to solve many problems existing in current technologies and provide a more effective solution for hand function rehabilitation in stroke patients. Summary of the Invention

[0007] To overcome the aforementioned deficiencies in the existing technology, this invention provides a hand function rehabilitation training system for stroke patients based on a hybrid brain-computer interface. By organically integrating multimodal signal collaborative acquisition, personalized intention recognition, dynamic feedback adjustment, and adaptive rehabilitation training, it achieves precise rehabilitation training and is suitable for active, precise, and personalized rehabilitation training for patients with hand motor dysfunction after stroke.

[0008] To achieve the above objectives, the present invention employs the following technical solution: A hand function rehabilitation training system for stroke patients based on a hybrid brain-computer interface includes a multimodal signal synchronous acquisition platform, a motor intention recognition module, an adaptive feedback adjustment module, and a rehabilitation training module. The aforementioned multimodal signal synchronous acquisition platform is the information input source for the entire system, used to synchronously acquire multimodal signals, including the patient's electroencephalogram (EEG) signals, surface electromyography (EMG) signals, and inertial signals, providing data support for subsequent functional assessment of motor intention recognition and adaptive feedback regulation. The motion intent recognition module preprocesses the multimodal signal, extracts and normalizes features, generates multimodal feature vectors, and uses an ensemble learning algorithm to recognize motion intent from the multimodal feature vectors. The adaptive feedback adjustment module normalizes and filters the electromyographic features extracted by the motion intention recognition module, then uses a support vector machine to assess motor function, and finally adjusts the adjustment based on the assessed motor function level and the comprehensive adjustment coefficient in the rehabilitation training module. K The training parameters are adaptively adjusted and then fed into the rehabilitation training module. The rehabilitation training module includes a virtual interaction unit and an external device expansion unit. The virtual interaction unit matches the difficulty of a single interactive gesture game task according to the training parameters, and simultaneously converts the movement intention results and adaptive auxiliary parameters into physical control commands, which are then sent to the external device expansion unit to trigger the external rehabilitation device to perform corresponding physical auxiliary actions. After a single rehabilitation training session, a comprehensive adjustment coefficient is calculated. K And the comprehensive adjustment coefficient K Return to the adaptive feedback adjustment module.

[0009] The multimodal signal synchronization acquisition platform includes a synchronization mode decision unit, a multimodal synchronization control unit, a synchronization triggering unit, and a signal acquisition unit; The synchronization mode decision unit dynamically selects the hardware or software trigger path based on the signal transmission environment parameters to ensure the synchronization and stability of signal acquisition. The multi-mode synchronization control unit dynamically switches between hardware and software triggering modes based on the triggering path determined by the synchronization mode decision unit to adapt to different environments and signal acquisition requirements. The aforementioned synchronization triggering unit includes a hardware synchronization triggering channel and a software synchronization triggering channel. The hardware synchronization triggering channel generates pulse signals and drives the sampling clock phase-locking of the EEG, EMG and inertial acquisition devices through an electrical isolation circuit to ensure clock synchronization of each signal acquisition. The software synchronization triggering channel is connected to a logic clock synchronization protocol stack to assign nanosecond-level timestamps to each acquisition device to achieve time synchronization. The signal acquisition unit includes EEG, surface electromyography and inertial motion signal acquisition devices, and performs signal acquisition according to the instructions of the synchronous triggering unit.

[0010] The hardware synchronization triggering channel of the synchronization triggering unit includes a signal generator unit, a standard signal conditioning unit, an electrical isolation unit, a pulse distribution unit, and a synchronization interface unit. The signal generator unit generates square wave pulse signals with a frequency of 1kHz to 20kHz. The standard signal conditioning unit eliminates signal jitter through a Schmitt trigger and can generate different voltages to be compatible with different devices. The electrical isolation unit uses optocouplers to block electrical interference between medical devices, ensuring patient safety. The pulse distribution unit distributes the synchronization signal indiscriminately to a corresponding number of devices through an analog switch. The synchronization interface unit includes three interface types: DIN, BNC, and DB25, to be compatible with different devices.

[0011] The software synchronization triggering channel of the synchronization triggering unit includes a protocol timing unit, a timestamp marking unit, and a dynamic delay compensation unit. The protocol timing unit implements the Precise Time Protocol (PTP) based on the IEEE 1588 standard protocol stack. The timestamp marking unit adds a nanosecond-level timestamp to the trigger message at the data link layer. The dynamic delay compensation unit predicts and corrects network transmission jitter through a Kalman filter.

[0012] The motion intent recognition module, specifically the motion intent recognition steps, are as follows: Step 1: Bandpass filtering is performed on the acquired EEG and EMG signals, and mean filtering is performed on the acquired inertial signals; Step 2: Based on the filtered electromyographic signals, complete the detection of the starting point of movement; Step 3: Based on the detected starting point of motion, extract a certain length of multimodal signal from the starting point, extract features from the active segment signals respectively, and construct EEG signal feature sets, EMG signal feature sets and inertial signal feature sets in groups; Step 4: Perform maximum and minimum value normalization on all feature sets; Step 5: Repeat steps 1 to 4 to collect different types of motion patterns, collect feature vector sample sets of different motion patterns, and then construct a parallel ensemble learning algorithm to build a motion intent recognition model; Step 6: Use the trained motion intent recognition model to identify the multi-source signals generated by subsequent motion patterns in real time; In step 3, feature extraction is performed on the signals, and feature sets for electroencephalogram (EEG), electromyography (EMG), and inertial signals are constructed in groups. The results are as follows: The EEG signal feature set includes: amplitude of motor-related cortical potentials (MRCP); slope of MRCP; event-related desynchronization (ERD); event-related synchronization (ERS); wavelet coefficients; wavelet energy; and embedding entropy.

[0013] The electromyographic signal feature set includes: mean absolute amplitude (MAV); variance (VAR); Wilson amplitude (WAMP); wavelength (WL); number of zero crossings (ZC); number of slope sign changes (SSC); third-order spectral moments (SM3); and sixth-order linear autoregressive coefficients (AR6).

[0014] The inertial signal feature set includes: mean (MEAN); standard deviation (STD); maximum value (MAX); minimum value (MIN); extreme values; interquartile range (IQR); and signal amplitude area (SMA).

[0015] The method for constructing the motion intent recognition model using the ensemble learning algorithm in step 5 is as follows: ① Train an AdaBoost classification model based on EEG, EMG, and inertial feature sets respectively, and output a probability vector. P 1 、P 2 、 P 3; ② Assign weight coefficients to the output probability vector l 1 、l 2 、l 3. Generate a weighted probability vector l 1 P 1 、l 2 P 2 、l 3 P 3; ③ The weighted vector [ l 1 P 1, l 2 P 2, l 3 P 3] Perform splicing and fusion to obtain the fused feature vector; ④ The fused feature vector is processed by the Linear Discriminant Analysis (LDA) model in machine learning to output the motion intent category label.

[0016] The weighting coefficient is l 1. l 2. l 3 represents the model hyperparameters. The specific setting method is as follows: ① The weighting coefficients satisfy the normalization constraint: l 1+ l 2 + λ³ = 1, and l 1. l 2. l 3≥0; ② Set the initial l 1. l 2. l 3. Optimization range and step size: Calculate the classification accuracy of the model through grid search; ③ Determine the parameters of the motion intention recognition model l 1. l 2. l The optimal value of 3 is determined based on the training set data to complete the training of the motion intention recognition model.

[0017] The adaptive feedback adjustment module includes a motor function assessment unit and a parameter adaptive adjustment unit. The motor function assessment unit constructs a motor function assessment model based on the electromyographic features extracted by the motor intention recognition module to assess the patient's motor function status. The steps are as follows: Step 1: Perform maximum and minimum value normalization on the electromyographic features extracted by the motion intention recognition module; Step 2: Use the Fisher Score algorithm to filter the normalized features and determine the optimal combination of electromyographic features; Step 3: Construct an evaluation model based on support vector machines, using the Brunnstrom scale's hand function impairment level classification as the model's supervisory signal; Step 4: Complete the construction of the hand function evaluation model through model parameter optimization and SVM model training; Step 5: Use the trained hand function assessment model to evaluate the electromyographic features extracted by the subsequent motion intention recognition module and output the motion function level.

[0018] In step 3, the Brunnstrom scale for hand dysfunction is divided into stages I, II, III, IV, V, and VI. The severity of hand, upper limb, and lower limb dysfunction is classified into four levels, with stage 1 being the most severe and stage 4 being close to normal. Specifically, Brunnstrom stage I is classified as stage 1, Brunnstrom stages II and III as stage 2, Brunnstrom stages IV and V as stage 3, and Brunnstrom stage VI as stage 4. The classification level serves as the mentor supervision signal for the classification model.

[0019] In step 4, the SVM model is trained, and its kernel function uses the radial basis function (REF). The hyperparameters C and g are selected through grid search to find the combination of C and g with the highest classification accuracy as the optimal value. The parameter adaptive adjustment unit adjusts the parameters based on the motor function level and the comprehensive adjustment coefficient in the rehabilitation training module. K The specific steps for adaptively adjusting training parameters are as follows: Step 1: Before the patient begins the first rehabilitation training task, set the initial training parameters that match the patient's Brunnstrom stage as the basic difficulty of the training. Step 2: After the single-round rehabilitation training task is completed, the motor function level output by the motor function assessment unit and the comprehensive adjustment coefficient returned by the rehabilitation training module are used as the basis for the assessment. K Adjust the training parameters and save them as training parameters for the next game; Step 3: During the daily rehabilitation training, Step 2 is executed repeatedly; after all training tasks for the day are completed, the system extracts and saves the last generated training parameters and marks them as the patient's daily limit training parameters. Step 4: At the start of the next rehabilitation training cycle or the following day, the system retrieves the extreme training parameters for that day and uses these extreme training parameters as the starting point of the new cycle, and then re-enters the dynamic loop of Step 2 and Step 3.

[0020] The motor function level in step 2 corresponds to the training parameter of the interval time between different types of tasks ( T interval The higher the patient's motor function level, the corresponding T interval The shorter; the mapping relationship is configured as follows: when the motor function assessment is level 1, T interval Set to 8 seconds; when the motor function assessment is level 2, T interval Set to 6 seconds; when the motor function assessment is level 3, T interval Set to 4 seconds; when the motor function assessment is level 4, T interval Set to 2 seconds; The comprehensive adjustment coefficient in step 2 K The adjusted training parameters are the time for a single-class task ( T task When patients perform their first rehabilitation training task, the default time is for a single category of task; after the single-game task training is completed, the time is adjusted according to the aforementioned comprehensive adjustment coefficient. K The adaptive dynamic parameter tuning logic for the value is specifically as follows: If K If the score is ≥0.8, the patient's current challenge level is deemed insufficient, and the next training session will begin. T task Reduce by 0.5s; if 0.5 < K <0.8, indicating the patient's current performance is a good match; maintain the current level.T task ;like K If the score is ≤0.5, the patient's current challenge level is deemed too high, and the next training session will begin. T task Increase by 0.5s.

[0021] The rehabilitation training module is equipped with a virtual interaction unit and an external device expansion unit, which can support the selection of virtual interactive tasks or external rehabilitation devices. The virtual interaction unit provides interactive gesture game tasks, adjusts the game difficulty based on training parameters fed back by the parameter adaptive adjustment unit, and generates matching audiovisual feedback in real time. Combining real-time recognition of the patient's movement intentions with game state analysis, it calculates the comprehensive adjustment coefficient of a single game of rehabilitation game training. K The virtual interaction unit comprises an interaction configuration subunit, a game execution subunit, a performance evaluation subunit, and a difficulty determination subunit. The interaction configuration subunit adjusts the initial game difficulty level based on the training parameters fed back by the parameter adaptive adjustment unit. The game execution subunit drives the training game process based on real-time motion intent recognition results and simultaneously generates the game interface and corresponding audio signals. The performance evaluation subunit calculates the task score after each game. S c Average recognition time T r Average actual time of action T a and delay offset coefficient D c The calculation formula is: In the formula, n To accurately identify the number of actions, N The total number of actions. S c The larger the value, the higher the task score and the higher the recognition accuracy; in the formula... Indicates the first i The actual recognition time for each action Indicates the first i The time from receiving the instruction to actually starting the movement. D c To determine the degree of deviation in the average motion intent at the current difficulty level, D c A larger value indicates a greater degree of deviation and a longer motion intent recognition time; Difficulty Decision Subunit: Calculates the comprehensive adjustment coefficient. K and will K The calculation formula is returned to the adaptive feedback adjustment module as follows: KA quantitative indicator reflecting the match between current athletic ability and the difficulty level set by the system. K A higher value indicates a higher completion rate for the current difficulty setting; The external device expansion unit converts the motion intention and adaptive auxiliary parameters into physical control commands, driving the external rehabilitation device to perform corresponding physical auxiliary actions.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a multimodal signal synchronous acquisition platform. By combining hardware synchronous triggering mechanism with software synchronous triggering mechanism, it achieves high-precision time synchronous acquisition of EEG signals, surface electromyography signals and inertial motion signals. It effectively solves the problem of time alignment difficulties between different acquisition devices, thereby providing a complete, accurate and time-consistent data foundation for subsequent multimodal signal fusion and motion intention recognition, and improving the overall stability and reliability of the system recognition.

[0023] 2. The motion intention recognition module of this invention achieves accurate recognition of the patient's motion intention by extracting multimodal features from electroencephalogram (EEG), electromyogram (EMG), and inertial signals, and then fusing them using a multi-model learning method. Compared to single-signal recognition methods, this invention can fully utilize the complementary information between different signals, effectively improving the accuracy and anti-interference ability of motion intention recognition, and providing more reliable control commands for the rehabilitation training process.

[0024] 3. This invention incorporates an adaptive feedback adjustment module, which assesses the patient's muscle function during training and dynamically adjusts the auxiliary parameters of rehabilitation training based on the assessment results. This allows the training process to be intelligently adjusted according to the patient's actual motor ability, thereby achieving a personalized rehabilitation training plan and improving training efficiency and rehabilitation effectiveness.

[0025] 4. The rehabilitation training module of this invention supports two methods: rehabilitation training games and external rehabilitation equipment. The rehabilitation training games incorporate gamified interactive design, enabling patients to complete rehabilitation training tasks in an interactive training environment through virtual interactive tasks and real-time audiovisual feedback mechanisms. Simultaneously, the system can adaptively adjust the task difficulty and training parameters based on the patient's training performance, ensuring the training is both challenging and appropriately difficult, thereby increasing the patient's enjoyment and active participation, enhancing training compliance, and promoting the recovery of hand function in stroke patients.

[0026] 5. The rehabilitation training module of the present invention connects to rehabilitation equipment (such as robotic gloves, electrical stimulation devices or other rehabilitation aids) by setting an external device expansion unit. It can convert the results of motion intention recognition and training parameters into control signals to drive the external rehabilitation equipment, thereby realizing a training mode in which virtual training and physical rehabilitation equipment work together, further improving the effect of rehabilitation training and expanding the application scope of the system.

[0027] In summary, this invention significantly enhances the relevance, engagement, and active participation of rehabilitation training by organically integrating multimodal signal collaborative acquisition, hardware and software fusion synchronization strategies, multimodal information collaborative decoding, personalized intent recognition, dynamic feedback adjustment, and adaptive rehabilitation training. It enables proactive, personalized, and multi-scenario-adaptive efficient rehabilitation training for hand function after stroke, and is suitable for proactive, precise, and personalized rehabilitation training for patients with hand motor dysfunction after stroke. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the overall design scheme of the system of the present invention.

[0029] Figure 2 This is a structural block diagram of the multimodal signal synchronous acquisition platform described in this invention.

[0030] Figure 3 This is a block diagram of the hardware synchronous triggering channel structure described in this invention.

[0031] Figure 4 This is a block diagram of the software synchronization triggering channel structure described in this invention.

[0032] Figure 5 This is a flowchart of the motion intent recognition process described in this invention.

[0033] Figure 6 This is a diagram of the multimodal probability fusion model based on ensemble learning described in this invention.

[0034] Figure 7 Flowchart for assessing motor function levels based on electromyography.

[0035] Figure 8 This is a schematic diagram of the timing parameters of the rehabilitation training task in this invention. Detailed Implementation

[0036] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and examples.

[0037] See Figure 1This embodiment discloses a hand function rehabilitation training system for stroke patients based on a hybrid brain-computer interface, including a multimodal signal synchronous acquisition platform, a motor intention recognition module, an adaptive feedback adjustment module, and a rehabilitation training module; the modules work together to form a complete rehabilitation training closed loop.

[0038] See Figure 2 The multimodal signal synchronous acquisition platform includes a synchronous mode decision unit, a multimodal synchronous control unit, a synchronous triggering unit, and a signal acquisition unit. This platform is used to synchronously acquire patients' electroencephalogram (EEG), surface electromyography (EMG), and inertial signals—multimodal signals—and serves as the information input source for the entire system, providing comprehensive and accurate data support for subsequent motor intention recognition and adaptive feedback adjustment motor function assessment.

[0039] The synchronization mode decision unit selects the trigger path by real-time monitoring of signal transmission environment parameters. In this embodiment, the synchronization mode decision unit uses the bit error rate as the core criterion and the signal-to-noise ratio as an auxiliary verification criterion; when the bit error rate is below 10... -6 When the signal-to-noise ratio is ≥30dB, select hardware trigger mode; when the bit error rate is below 10... -6 However, when the signal-to-noise ratio is <30dB, a signal filtering stage is added before execution in hardware trigger mode; when the bit error rate is higher than 10... -6 Regardless of the signal-to-noise ratio, the software trigger mode is selected, and delay compensation is enhanced by using a Kalman filter.

[0040] The multi-mode synchronization control unit dynamically switches the enabling state of the hardware trigger channel and the software trigger channel according to the trigger path instruction determined by the synchronization mode decision unit: when the hardware synchronization path is selected, the control unit outputs a hardware trigger enable signal and closes the software trigger channel; when the software synchronization path is selected, it outputs a software trigger enable signal and closes the hardware trigger channel to avoid dual-mode conflict.

[0041] In this embodiment, the multi-mode synchronous control unit uses a 32-bit microcontroller based on the ARM Cortex-M4 core as the core control chip. This chip integrates high-speed GPIO, timers, Ethernet controllers, and other peripherals, and can directly interface with hardware trigger interfaces and network communication modules, meeting the real-time and compatibility requirements of mode switching. In specific implementations, microcontrollers with equivalent performance, such as the GigaDevice GD32F407 series or STMicroelectronics STM32F4 series, can be used. This invention is not limited to a specific chip model.

[0042] The aforementioned synchronization triggering unit includes a hardware synchronization triggering channel and a software synchronization triggering channel. The two channels are mutually exclusive and adapted to different acquisition scenarios to achieve high-precision time synchronization of multi-modal signals.

[0043] See Figure 3In this embodiment, the hardware synchronization triggering channel of the synchronization triggering unit is composed of a signal generator unit, a standard signal conditioning unit, an electrical isolation unit, a pulse distribution unit, and a synchronization interface unit connected in sequence. Signal generator unit: generates square wave synchronization pulse signals with a frequency of 1kHz~20kHz to provide a unified clock reference for the acquisition equipment; Standard signal conditioning unit: Built-in Schmitt trigger to shape input pulses, effectively eliminating signal jitter and glitches, and ensuring steep pulse edges; it is also equipped with a DIP switch to flexibly select the output +5V or +3.3V level, compatible with the interface level requirements of mainstream EEG, EMG and inertial acquisition equipment on the market; Electrical isolation unit: High-speed optocouplers are used to achieve electrical isolation with an isolation voltage of not less than 3000V, blocking grounding interference and leakage current between medical devices, meeting medical electrical safety standards, and ensuring patient safety. Pulse distribution unit: Employs a multi-channel analog switch to distribute the conditioned synchronization pulse to multiple acquisition devices without attenuation or delay, thereby achieving synchronous triggering of multi-channel signals; Synchronization interface unit: integrates three standard interfaces: DIN, BNC, and DB25, which are adapted to the hardware trigger interfaces of different types of acquisition devices to improve system compatibility.

[0044] See Figure 4 In this embodiment, the software synchronization triggering channel of the synchronization triggering unit consists of a protocol timing unit, a timestamp marking unit, and a dynamic delay compensation unit. Protocol timing unit: The master-slave clock system is built based on the IEEE 1588 Precision Time Protocol (PTP). The protocol stack is implemented using hardware timestamp chips such as the DP83640 Ethernet PHY chip. The synchronization accuracy can reach ±10ns, which meets the high precision requirements of multi-device network synchronization. Timestamp marking unit: At the data link layer, a nanosecond-level timestamp is added to each frame of synchronization trigger message to accurately record the time of data generation and provide a benchmark for subsequent delay compensation; Dynamic delay compensation unit: It adopts a Kalman filter, whose state equation and observation equation are adaptively adjusted according to the historical data of network transmission. It can effectively predict and correct network transmission jitter, keep the time synchronization error within 50ns, and ensure the stability and accuracy of software synchronization.

[0045] The signal acquisition unit includes an EEG acquisition device, a surface electromyography (EMG) acquisition device, and an inertial signal acquisition device, used to acquire EEG signals, EMG signals, and inertial signals. In this embodiment, the electrodes of the EEG acquisition device are placed at the C3, C4, C2, FC3, and FC4 positions corresponding to the motor cortex of the brain, used to acquire signals related to hand movement execution, with a sampling rate of 1000Hz. The EMG acquisition device is attached to the extensor digitorum, flexor digitorum superficialis, abductor pollicis brevis, and flexor carpi ulnaris, with a sampling rate of 2000Hz. The inertial signal acquisition devices are all fixed to the extensor digitorum, flexor digitorum superficialis, abductor pollicis brevis, flexor carpi ulnaris, and the dorsal surface of the wrist, with a sampling rate of 200Hz.

[0046] The motion intent recognition module, deployed on the host computer, is responsible for processing and analyzing the signals acquired by the multimodal signal synchronous acquisition platform to accurately identify the patient's motion intent. The motion intent recognition process is as follows: Figure 5 As shown, the steps are as follows: Step 1: The EEG signal was filtered using a 4th-order Butterworth bandpass filter with a filtering range of 0.5-50Hz; the EMG signal was filtered using a 4th-order Butterworth bandpass filter with a filtering range of 20-200Hz; and the inertial signal was filtered using a 5-point smoothing filter.

[0047] Step 2: Based on the filtered electromyographic (EMG) signal, a threshold method is used to detect the motor initiation point. First, the root mean square (RMS) value of the EMG signal in the resting state is calculated, and three times this value is used as the threshold. When the RMS value of the EMG signal exceeds the threshold for five consecutive sampling points, it is determined as the motor initiation point.

[0048] Step 3: Based on the detected motion initiation point, extract the multi-mode signal 300ms after the initiation point as the active segment signal. Perform feature extraction on the active segment signal separately, and construct EEG signal feature sets, EMG signal feature sets, and inertial signal feature sets in groups: The EEG signal feature set includes: amplitude of motor-related cortical potentials (MRCP); slope of MRCP; event-related desynchronization (ERD); event-related synchronization (ERS); wavelet coefficients; wavelet energy; and embedding entropy.

[0049] The electromyographic signal feature set includes: mean absolute amplitude (MAV); variance (VAR); Wilson amplitude (WAMP); wavelength (WL); number of zero crossings (ZC); number of slope sign changes (SSC); third-order spectral moments (SM3); and sixth-order linear autoregressive coefficients (AR6).

[0050] The inertial signal feature set includes: mean (MEAN); standard deviation (STD); maximum value (MAX); minimum value (MIN); extreme values; interquartile range (IQR); and signal amplitude area (SMA).

[0051] Step 4: Perform maximum and minimum value normalization on all feature sets. The normalization function is as follows: In the formula: x The original values ​​of the features, z The value after feature normalization. max( x ) represents the maximum value among the feature vectors in the training samples, min( x ) represents the minimum value among the feature vectors in the training samples.

[0052] Step 5: Repeat steps 1 to 4 to collect different types of motion patterns, collect feature vector sample sets of different motion patterns, and then build a multimodal probabilistic fusion motion intention recognition model based on the ensemble learning algorithm; Step 6: Use the trained motion intent recognition model to identify the multi-source signals generated by subsequent motion patterns in real time.

[0053] In step 5 above, the process of constructing a multimodal probabilistic fusion motion intent recognition model using the ensemble learning algorithm is as follows: Figure 6 As shown, the specific implementation method is as follows: ① Train an AdaBoost classification model based on EEG, EMG, and inertial feature sets respectively, and output a probability vector. P 1 、P 2 、 P 3; ② Assign weight coefficients to the output probability vector l 1 、l 2 、l 3. Generate a weighted probability vector l 1 P 1 、l 2 P 2 、l 3 P 3. The weighting coefficients satisfy the normalization constraint: l 1 +λ 2 +λ 3=1, and l 1 、l 2 、l 3≥0; ③ The weighted vector [ l 1 P 1, l 2 P 2, l 3 P 3] Perform splicing and fusion to form a combined feature vector; splice the weighted probability vectors into a fused feature vector; ④ The fused feature vector is processed by a linear discriminant analysis (LDA) model to output a motion intent category label; ⑤ In this embodiment, the initial value is set. l 1 = 0.3 l 2 = 0.3 l A parameter search grid is constructed with a step size of 0.1 and a value of 3=0.4. The grid search algorithm is used to traverse all valid parameter combinations, and the classification accuracy is used as the evaluation index. ⑥ Determine the parameters of the motion intent recognition model l 1. l 2. l The optimal value of 3 is determined based on the training set data to complete the training of the motion intention recognition model.

[0054] The adaptive feedback adjustment module can adjust rehabilitation training parameters according to the patient's motor function status, improving the targeting and effectiveness of training. This module includes a motor function assessment unit and a parameter adaptive adjustment unit.

[0055] In this embodiment, the motor function assessment unit constructs an assessment model of the patient's hand motor function based on the electromyographic features output by the motor intention recognition component, thereby achieving automatic grading of the motor function status of stroke patients. The flowchart of the motor function assessment model is as follows: Figure 7 As shown, the specific implementation steps are as follows: Step 1: Perform maximum and minimum value normalization on the electromyographic features MAV, VAR, WAMP, WL, ZC, SSC, SM3, and AR6 extracted by the motion intention recognition module; Step 2: The Fisher Score algorithm is used to score all normalized electromyographic features. A higher feature score indicates a greater contribution to the classification task. In this embodiment, the top 50% of features are selected based on the scoring results to form the optimal feature combination, achieving feature dimensionality reduction. Step 3: Construct a motor function assessment model based on Support Vector Machine (SVM), using the Brunnstrom scale hand dysfunction level as the classification supervision signal. In this embodiment, the Brunnstrom stages are graded into four levels, with level 1 being the most severe and level 4 being close to normal. Brunnstrom stage I is classified as level 1, Brunnstrom stages II and III as level 2, Brunnstrom stages IV and V as level 3, and Brunnstrom stage VI as level 4. The classification levels serve as the supervisory signals for the classification model. Step 4: The kernel function of the SVM model uses a radial basis function (RBF), whose performance depends on the hyperparameter penalty factor C and the kernel function parameter g. In this embodiment, C and g are set from 2... -16 to 2 16The parameters are divided into exponentially equal parts, and the parameter space is traversed by grid search. The classification accuracy is used as the evaluation index. The combination of C and g that achieves the highest classification accuracy is selected as the optimal parameter value to complete the training of the SVM model. Step 5: After training is completed, the evaluation model can input new electromyographic feature data in real time and directly output the patient's motor function level through the motor function evaluation model.

[0056] The parameter adaptive adjustment unit in this embodiment adjusts the parameters based on the motor function level output by the motor function assessment section and the comprehensive adjustment coefficient fed back by the rehabilitation training section. K The parameters of rehabilitation training tasks are dynamically adjusted to achieve personalized and progressive control of rehabilitation training. The specific implementation steps are as follows: Step 1: Before the patient first enters the rehabilitation training task, the initial training parameters are automatically matched according to their Brunnstrom stage as the basic difficulty of the training, ensuring that the initial training difficulty is appropriate for the patient's functional level. Step 2: After the single-round rehabilitation training task is completed, the parameter adaptive adjustment unit adjusts the parameters based on the motor function level output by the motor function assessment section and the comprehensive adjustment coefficient returned by the rehabilitation training module. K Adjust the training parameters and save them as the rehabilitation training parameters for the next game; Rehabilitation training parameters include two categories: interval time between different types of tasks ( T interval ) and time for individual category tasks ( T task ), the timing parameters of the rehabilitation training task are as follows Figure 8 As shown; In this embodiment, based on the level of motor function, T interval Adaptive adjustments are made; when the motor function assessment is Level 1. T interval Set to 8 seconds; when the motor function assessment is level 2, T interval Set to 6 seconds; when the motor function assessment is level 3, T interval Set to 4 seconds; when the motor function assessment is level 4, T interval Set to 2 seconds; In this embodiment, T task Multiple discrete time increments from 1 second to 5 seconds, with a step interval of 0.5 seconds between adjacent increments, are set as the default setting for patients' first rehabilitation training. T task It is 3s; according to the comprehensive adjustment coefficient K ,right Ttask Dynamic adjustment is performed; if K If the score is ≥0.8, the patient's current challenge level is deemed insufficient, and the next training session will begin. T task Reduce by 0.5s; if 0.5 < K <0.8, indicating the patient's current performance is a good match; maintain the current level. T task ;like K If the score is ≤0.5, the patient's current challenge level is deemed too high, and the next training session will begin. T task Increase by 0.5s; Step 3: Execute multiple training sessions within a single day, updating the training parameters after each session. When all training sessions for the day are completed, the parameter adaptive adjustment unit saves the training parameters generated in the last session as the maximum training parameters for that day. Step 4: At the start of the next rehabilitation training cycle or the following day, the system automatically loads the daily limit training parameters as the initial training parameters for the new cycle and continues to execute the dynamic parameter tuning loop.

[0057] The rehabilitation training module is equipped with a virtual interaction unit and an external device expansion unit, which can support the selection of virtual interactive tasks or external rehabilitation devices.

[0058] The virtual interaction unit comprises four sub-units: interaction configuration, game execution, performance evaluation, and difficulty determination. The interactive configuration subunit adaptively adjusts the training parameters of the unit based on the parameters to set the initial game difficulty level; In this embodiment, the game execution subunit uses the Unity3D game engine to develop the game "Rebuilding the Hand." The game scene is a virtual rehabilitation center, where patients control a virtual hand to perform gesture tasks such as grasping, moving, and placing based on recognized motor intentions. When a patient completes a task, a cheerful prompt sound is played; when a task fails, a failure prompt sound is played.

[0059] The performance evaluation subunit automatically calculates the task score after each game. S c Average recognition time T r Average actual time of action T a and delay offset coefficient D c The calculation method is as follows: In the formula, n To accurately identify the number of actions, N The total number of actions. S cA larger value indicates a higher task score and higher recognition accuracy. In the formula... Indicates the first i The actual recognition time for each action Indicates the first i The time from receiving the instruction to actually starting the movement. D c To determine the degree of deviation in the average motion intent at the current difficulty level, D c A larger value indicates a greater degree of offset and a longer motion intent recognition time. The difficulty judgment subunit is calculated based on the above. S c , D c Calculate the comprehensive adjustment coefficient K and will K Return to the adaptive feedback adjustment module. K The calculation formula is as follows: K A quantitative indicator reflecting the match between current athletic ability and the difficulty level set by the system. K A higher value indicates a higher level of completion for the current difficulty setting.

[0060] In this embodiment, the external device expansion unit converts the motion intention recognition results and training parameters into physical control commands to drive the external rehabilitation device: ① Rehabilitation robot gloves: assist patients in making fists, extending fingers, and opposing thumbs with palms; ② Functional electrical stimulator: Electrically stimulates the target muscles of the hand to enhance muscle activation; ③ Wrist and finger rehabilitation robotic hand: Provides active assistance for wrist flexion and extension.

[0061] This invention significantly improves the reliability and anti-interference ability of motion intention recognition by integrating hardware and software synchronization strategies and multimodal information collaborative decoding; combined with adaptive dynamic feedback, gamified interactive training and multi-device expansion capabilities, it greatly enhances the pertinence, fun and active participation of rehabilitation training, and can realize active, personalized and multi-scenario adapted efficient rehabilitation training of hand function after stroke, effectively promoting the recovery of patients' hand motor function.

Claims

1. A hand function rehabilitation training system for stroke patients based on a hybrid brain-computer interface, characterized in that, It includes a multimodal signal synchronous acquisition platform, a motion intention recognition module, an adaptive feedback adjustment module, and a rehabilitation training module; The aforementioned multimodal signal synchronous acquisition platform is the information input source for the entire system, used to synchronously acquire multimodal signals, including the patient's electroencephalogram (EEG) signals, surface electromyography (EMG) signals, and inertial signals, providing data support for subsequent functional assessment of motor intention recognition and adaptive feedback regulation. The motion intent recognition module preprocesses the multimodal signal, extracts and normalizes features, generates multimodal feature vectors, and uses an ensemble learning algorithm to recognize motion intent from the multimodal feature vectors. The adaptive feedback adjustment module normalizes and filters the electromyographic features extracted by the motion intention recognition module, then uses a support vector machine to assess motor function, and finally adjusts the adjustment based on the assessed motor function level and the comprehensive adjustment coefficient in the rehabilitation training module. K The training parameters are adaptively adjusted and then fed into the rehabilitation training module. The rehabilitation training module includes a virtual interaction unit and an external device expansion unit. The virtual interaction unit matches the difficulty of a single interactive gesture game task according to the training parameters, and simultaneously converts the movement intention results and adaptive auxiliary parameters into physical control commands, which are then sent to the external device expansion unit to trigger the external rehabilitation device to perform corresponding physical auxiliary actions. After a single rehabilitation training session, a comprehensive adjustment coefficient is calculated. K And the comprehensive adjustment coefficient K Return to the adaptive feedback adjustment module.

2. The hand function rehabilitation training system for stroke patients based on a hybrid brain-computer interface according to claim 1, characterized in that, The multimodal signal synchronization acquisition platform includes a synchronization mode decision unit, a multimodal synchronization control unit, a synchronization triggering unit, and a signal acquisition unit; The synchronization mode decision unit dynamically selects the hardware or software trigger path based on the signal transmission environment parameters to ensure the synchronization and stability of signal acquisition. The multi-mode synchronization control unit dynamically switches between hardware and software triggering modes based on the triggering path determined by the synchronization mode decision unit to adapt to different environments and signal acquisition requirements. The aforementioned synchronization triggering unit includes a hardware synchronization triggering channel and a software synchronization triggering channel. The hardware synchronization triggering channel generates pulse signals and drives the sampling clock phase-locking of the EEG, EMG and inertial acquisition devices through an electrical isolation circuit to ensure clock synchronization of each signal acquisition. The software synchronization triggering channel is connected to a logic clock synchronization protocol stack to assign nanosecond-level timestamps to each acquisition device to achieve time synchronization. The signal acquisition unit includes EEG, surface electromyography and inertial motion signal acquisition devices, and performs signal acquisition according to the instructions of the synchronous triggering unit.

3. The hand function rehabilitation training system for stroke patients based on a hybrid brain-computer interface according to claim 2, characterized in that, The hardware synchronization triggering channel of the synchronization triggering unit includes a signal generator unit, a standard signal conditioning unit, an electrical isolation unit, a pulse distribution unit, and a synchronization interface unit; the signal generator unit generates square wave pulse signals with a frequency of 1kHz to 20kHz; the standard signal conditioning unit eliminates signal jitter through a Schmitt trigger and can generate different voltages to be compatible with different devices; The electrical isolation unit uses optocoupler isolation to block electrical interference between medical devices, ensuring patient safety; the pulse distribution unit distributes synchronization signals indiscriminately to a corresponding number of devices through analog switches; the synchronization interface unit includes three interface types: DIN, BNC, and DB25, to be compatible with different devices.

4. The hand function rehabilitation training system for stroke patients based on a hybrid brain-computer interface according to claim 2, characterized in that, The software synchronization triggering channel of the synchronization triggering unit includes a protocol timing unit, a timestamp marking unit, and a dynamic delay compensation unit. The protocol timing unit implements the Precise Time Protocol (PTP) based on the IEEE 1588 standard protocol stack. The timestamp marking unit adds a nanosecond-level timestamp to the trigger message at the data link layer. The dynamic delay compensation unit predicts and corrects network transmission jitter through a Kalman filter.

5. A hand function rehabilitation training system for stroke patients based on a hybrid brain-computer interface according to claim 1, characterized in that, The motion intent recognition module, specifically the motion intent recognition steps, are as follows: Step 1: Bandpass filtering is performed on the acquired EEG and EMG signals, and mean filtering is performed on the acquired inertial signals; Step 2: Based on the filtered electromyographic signals, complete the detection of the starting point of movement; Step 3: Based on the detected starting point of motion, extract a certain length of multimodal signal from the starting point, extract features from the active segment signals respectively, and construct EEG signal feature sets, EMG signal feature sets and inertial signal feature sets in groups; Step 4: Perform maximum and minimum value normalization on all feature sets; Step 5: Repeat steps 1 to 4 to collect different types of motion patterns, collect feature vector sample sets of different motion patterns, and then use parallel ensemble learning algorithms to build a motion intention recognition model; Step 6: Use the trained motion intent recognition model to identify the multi-source signals generated by subsequent motion patterns in real time.

6. A hand function rehabilitation training system for stroke patients based on a hybrid brain-computer interface according to claim 5, characterized in that, In step 3, feature extraction is performed on the signals, and feature sets for electroencephalogram (EEG), electromyography (EMG), and inertial signals are constructed in groups. The results are as follows: The EEG signal feature set includes: amplitude of motor-related cortical potentials; slope of MRCP; event-related desynchronization; event-related synchronization; wavelet coefficients; wavelet energy; embedding entropy; The electromyographic signal feature set includes: mean absolute amplitude; variance; Wilson amplitude; wavelength; number of zero crossings; number of slope sign changes; 3rd order spectral moments; and 6th order linear autoregressive coefficients (AR6). The feature set of inertial signals includes: mean; standard deviation; maximum value; minimum value; extreme value; interquartile range; and signal amplitude area.

7. A hand function rehabilitation training system for stroke patients based on a hybrid brain-computer interface according to claim 5, characterized in that, The method for constructing the motion intent recognition model using the ensemble learning algorithm in step 5 is as follows: ① Train an AdaBoost classification model based on EEG, EMG, and inertial feature sets respectively, and output a probability vector. P 1 、P 2 、P 3; ② Assign weight coefficients to the output probability vector λ 1 , λ 2 , λ 3. Generate a weighted probability vector λ 1 P 1 , λ 2 P 2 , λ 3 P 3; ③ The weighted vector [ λ 1 P 1, λ 2 P 2, λ 3 P 3] Perform splicing and fusion to obtain the fused feature vector; ④ The fused feature vector is processed by a linear discriminant analysis model in machine learning to output a motion intent category label; The weighting coefficient λ 1. λ 2. λ 3 represents the model hyperparameters. The specific setting method is as follows: ① The weighting coefficients satisfy the normalization constraint: λ 1+ λ 2+ λ 3=1, and λ 1. λ 2. λ 3≥0; ② Set the initial λ 1. λ 2. λ 3. Optimization range and step size: Calculate the classification accuracy of the model through grid search; ③ Determine the parameters of the motion intention recognition model λ 1. λ 2. λ The optimal value of 3 is determined based on the training set data to complete the training of the motion intention recognition model.

8. A hand function rehabilitation training system for stroke patients based on a hybrid brain-computer interface according to claim 1, characterized in that, The adaptive feedback adjustment module includes a motor function assessment unit and a parameter adaptive adjustment unit. The motor function assessment unit constructs a motor function assessment model based on the electromyographic features extracted by the motor intention recognition module to assess the patient's motor function status. The steps are as follows: Step 1: Perform maximum and minimum value normalization on the electromyographic features extracted by the motion intention recognition module; Step 2: Use the Fisher Score algorithm to filter the normalized features and determine the optimal combination of electromyographic features; Step 3: Construct an evaluation model based on support vector machines, using the Brunnstrom scale's hand function impairment level classification as the model's supervisory signal; Step 4: Complete the construction of the hand function assessment model through model parameter optimization and model training; Step 5: Use the trained hand function assessment model to evaluate the electromyographic features extracted by the subsequent motion intention recognition module and output the motion function level. In step 3, the Brunnstrom scale for hand dysfunction is divided into stages I, II, III, IV, V, and VI. The severity of hand, upper limb, and lower limb dysfunction is classified into four levels, with stage 1 being the most severe and stage 4 being close to normal. Specifically, Brunnstrom stage I is classified as stage 1, Brunnstrom stages II and III as stage 2, Brunnstrom stages IV and V as stage 3, and Brunnstrom stage VI as stage 4. The classification level serves as the mentor supervision signal for the classification model. In step 4, the model training uses a radial basis function as its kernel function, and the hyperparameters C and g are selected through grid search to find the optimal combination of C and g with the highest classification accuracy.

9. A hand function rehabilitation training system for stroke patients based on a hybrid brain-computer interface according to claim 8, characterized in that, The parameter adaptive adjustment unit adjusts the parameters based on the motor function level and the comprehensive adjustment coefficient in the rehabilitation training module. K The specific steps for adaptively adjusting training parameters are as follows: Step 1: Before the patient begins the first rehabilitation training task, set the initial training parameters that match the patient's Brunnstrom stage as the basic difficulty of the training. Step 2: After the single-round rehabilitation training task is completed, the motor function level output by the motor function assessment unit and the comprehensive adjustment coefficient returned by the rehabilitation training module are used as the basis for the assessment. K Adjust the training parameters and save them as training parameters for the next game; Step 3: During the daily rehabilitation training, Step 2 is executed repeatedly; after all training tasks for the day are completed, the system extracts and saves the last generated training parameters and marks them as the patient's daily limit training parameters. Step 4: At the start of the next rehabilitation training cycle or the following day, the system retrieves the extreme training parameters for that day and uses these extreme training parameters as the starting point of the new cycle, and then re-enters the dynamic loop of Step 2 and Step 3. The motor function level in step 2 corresponds to the training parameter of the interval time between different types of tasks ( T interval The higher the patient's motor function level, the corresponding T interval The shorter; the mapping relationship is configured as follows: when the motor function assessment is level 1, T interval Set to 8 seconds; when the motor function assessment is level 2, T interval Set to 6 seconds; when the motor function assessment is level 3, T interval Set to 4 seconds; when the motor function assessment is level 4, T interval Set to 2 seconds; The comprehensive adjustment coefficient in step 2 K The adjusted training parameters are the time for a single-class task ( T task When patients perform their first rehabilitation training task, the default time is for a single category of task; after the single-game task training is completed, the time is adjusted according to the aforementioned comprehensive adjustment coefficient. K The adaptive dynamic parameter tuning logic for the value is specifically as follows: If K If the score is ≥0.8, the patient's current challenge level is deemed insufficient, and the next training session will begin. T task Reduce by 0.5s; if 0.5 < K <0.8, indicating the patient's current performance is a good match; maintain the current level. T task ;like K If the score is ≤0.5, the patient's current challenge level is deemed too high, and the next training session will begin. T task Add 0.5s.

10. A hand function rehabilitation training system for stroke patients based on a hybrid brain-computer interface according to claim 1, characterized in that, The rehabilitation training module is equipped with a virtual interaction unit and an external device expansion unit, which can support the selection of virtual interactive tasks or external rehabilitation devices. The virtual interaction unit provides interactive gesture game tasks, adjusts the game difficulty based on training parameters fed back by the parameter adaptive adjustment unit, and generates matching audiovisual feedback in real time. Combining real-time recognition of the patient's movement intentions with game state analysis, it calculates the comprehensive adjustment coefficient of a single game of rehabilitation game training. K The virtual interaction unit includes an interaction configuration subunit, a game execution subunit, a performance evaluation subunit, and a difficulty judgment subunit. The interaction configuration subunit sets the initial game difficulty level based on the training parameters fed back by the parameter adaptive adjustment unit. The game execution subunit drives the training of the game process based on real-time motion intent recognition results, and simultaneously generates the game interface and corresponding audio signals; The performance evaluation subunit calculates the task score after each game. S c Average recognition time T r Average actual time of action T a and delay offset coefficient D c The calculation formula is: In the formula, n To accurately identify the number of actions, N The total number of actions. S c The larger the value, the higher the task score and the higher the recognition accuracy; in the formula... Indicates the first i The actual recognition time for each action Indicates the first i The time from receiving the instruction to actually starting the movement. D c To determine the degree of deviation in the average motion intent at the current difficulty level, D c A larger value indicates a greater degree of deviation and a longer motion intent recognition time; Difficulty Decision Subunit: Calculates the comprehensive adjustment coefficient. K and will K The calculation formula is returned to the adaptive feedback adjustment module as follows: K A quantitative indicator reflecting the match between current athletic ability and the difficulty level set by the system. K The higher the value, the higher the completion rate of the current difficulty setting; The external device expansion unit converts the motion intention and adaptive auxiliary parameters into physical control commands, driving the external rehabilitation device to perform corresponding physical auxiliary actions.

Images