Stroke rehabilitation training system and method based on deep reinforcement learning

Abstract

The invention relates to the technical field of stroke rehabilitation training systems, in particular to a stroke rehabilitation training system and method based on deep reinforcement learning. The data preprocessing module is in communication connection with the data acquisition module; the feature fusion module is in communication connection with the data preprocessing module; performing spatial-temporal feature fusion based on the preprocessed data; the dynamic modeling module is in communication connection with the feature fusion module; establishing a nonlinear dynamic model based on the fused features; the strategy generation module is in communication connection with the dynamics modeling module and receives the nonlinear dynamics model sent by the dynamics modeling module; according to the nonlinear dynamic model, generating a self-adaptive rehabilitation strategy; the strategy output module is in communication connection with the strategy generation module and receives the self-adaptive rehabilitation strategy sent by the strategy generation module; and a self-adaptive rehabilitation strategy is output and is used for guiding the patient to carry out stroke rehabilitation training, so that the scientificity and effectiveness of the rehabilitation training are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of stroke rehabilitation training systems, in particular to a stroke rehabilitation training system and method based on deep reinforcement learning. Background Art

[0002] Stroke is a neurological disease that seriously threatens human health and often leads to motor dysfunction in patients, severely affecting the quality of life. With the aggravation of population aging, the incidence of stroke shows an upward trend, and the demand for rehabilitation training is becoming increasingly urgent. Traditional stroke rehabilitation training methods mainly rely on manual guidance and fixed training programs, which are difficult to meet the personalized and precise rehabilitation needs of patients.

[0003] In recent years, with the development of artificial intelligence technology, intelligent rehabilitation training systems have gradually become a research hotspot. However, existing intelligent rehabilitation systems still have many limitations. First, most systems only focus on single physiological data, such as electromyogram signals or movement trajectories, and it is difficult to comprehensively grasp the rehabilitation status of patients. Second, existing systems mostly adopt simple machine learning algorithms and lack in-depth understanding and modeling of complex and non-linear human physiological systems. Moreover, the generation of rehabilitation strategies is often static and difficult to dynamically adjust according to the real-time status and progress of patients. Finally, the human-computer interaction interfaces of existing systems are generally not user-friendly enough to stimulate the training enthusiasm of patients, resulting in poor rehabilitation effects.

[0004] These problems seriously restrict the clinical application effect of intelligent rehabilitation systems. Patients often need to experience a long and boring rehabilitation process, and the training effect is less than satisfactory. Medical resources cannot be effectively utilized, and the work burden of medical staff has not been substantially alleviated. Therefore, it is urgent to develop a stroke rehabilitation training system that can provide personalized, adaptive, and efficient rehabilitation training. Summary of the Invention

[0005] The present invention aims to solve the above technical problems and proposes a stroke rehabilitation training system and method based on deep reinforcement learning. Through innovative technologies such as fusion analysis of multi-source physiological data, non-linear dynamics modeling, and deep reinforcement learning strategy generation, the system realizes the formulation and dynamic adjustment of precise and personalized rehabilitation training programs.

[0006] The present invention proposes a stroke rehabilitation training system and method based on deep reinforcement learning, including:

[0007] A data acquisition module, configured to:

[0008] Obtain multi-source physiological data of a patient, where the multi-source physiological data includes surface electromyogram signals, inertial measurement unit data, electroencephalogram signals, and electrocardiogram signals;

[0009] A data preprocessing module, communicatively connected to the data acquisition module, configured to:

[0010] Receive the multi-source physiological data sent by the data acquisition module;

[0011] Based on the multi-source physiological data, perform data preprocessing and feature extraction;

[0012] The feature fusion module, communicatively connected to the data preprocessing module, is used for:

[0013] Receive the preprocessed data sent by the data preprocessing module;

[0014] Based on the preprocessed data, perform spatio-temporal feature fusion;

[0015] The kinetic modeling module, communicatively connected to the feature fusion module, is used for:

[0016] Receive the fused features sent by the feature fusion module;

[0017] Based on the fused features, establish a non-linear kinetic model;

[0018] The strategy generation module, communicatively connected to the kinetic modeling module, is used for:

[0019] Receive the non-linear kinetic model sent by the kinetic modeling module;

[0020] According to the non-linear kinetic model, generate an adaptive rehabilitation strategy;

[0021] The strategy output module, communicatively connected to the strategy generation module, is used for:

[0022] Receive the adaptive rehabilitation strategy sent by the strategy generation module;

[0023] Output the adaptive rehabilitation strategy for guiding the patient to perform stroke rehabilitation training.

[0024] Preferably, the data preprocessing module includes:

[0025] The random matrix processing unit is used to process the multi-source physiological data by using the random matrix theory;

[0026] The topology analysis unit, communicatively connected to the random matrix processing unit, is used to perform topology data analysis on the processed data;

[0027] The feature extraction unit, communicatively connected to the topology analysis unit, is used to obtain the processed data matrix through the topology transformation operator and the Hadamard product operation.

[0028] Preferably, the feature fusion module includes:

[0029] A chaos analysis unit for analyzing the preprocessed data using chaos theory;

[0030] A differential geometry processing unit, communicatively connected to the chaos analysis unit, for performing differential geometry processing on the analyzed data;

[0031] A feature fusion unit, communicatively connected to the differential geometry processing unit, for obtaining a fused feature matrix through calculations using the Laplace - Beltrami operator and Lyapunov exponents.

[0032] Preferably, the dynamic modeling module includes:

[0033] A quantum mechanics modeling unit for preliminarily establishing a dynamic model using quantum mechanics methods;

[0034] A group theory optimization unit, communicatively connected to the quantum mechanics modeling unit, for optimizing the dynamic model using group theory methods;

[0035] A model generation unit, communicatively connected to the group theory optimization unit, for generating a non - linear dynamic model describing the system evolution through variants of equations and group action operators.

[0036] Preferably, the strategy generation module includes:

[0037] A functional analysis unit for performing functional analysis on the non - linear dynamic model;

[0038] A variational method processing unit, communicatively connected to the functional analysis unit, for applying variational methods to process the analysis results;

[0039] A strategy optimization unit, communicatively connected to the variational method processing unit, for obtaining an optimal adaptive rehabilitation strategy by minimizing the energy functional and total variation regularization.

[0040] Preferably, it further includes a strategy optimization module, communicatively connected to the strategy generation module and the strategy output module, for:

[0041] Receiving the adaptive rehabilitation strategy sent by the strategy generation module;

[0042] Optimizing the adaptive rehabilitation strategy using random matrix theory and probability theory;

[0043] Improving the exploration ability and adaptability of the strategy through a random weight matrix and gradient descent method;

[0044] Sending the optimized strategy to the strategy output module.

[0045] Preferably, the data acquisition module includes:

[0046] A surface electromyogram acquisition unit for acquiring the electromyogram signal of a patient;

[0047] An inertial measurement unit for acquiring the joint angle and angular velocity data of a patient;

[0048] An electroencephalogram acquisition unit for acquiring the electroencephalogram signal of a patient;

[0049] An electrocardiogram acquisition unit for acquiring the electrocardiogram signal of a patient;

[0050] A data synchronization unit, communicatively connected to the above-mentioned acquisition units, for time-aligning and unifying the sampling rates of the acquired multi-source physiological data.

[0051] Preferably, it further includes a rehabilitation effect evaluation module, communicatively connected to the strategy output module and the strategy generation module, for:

[0052] Receiving the rehabilitation training data sent by the strategy output module;

[0053] Evaluating the effect of the adaptive rehabilitation strategy based on the rehabilitation training data and preset clinical evaluation indicators;

[0054] Sending an adjustment signal to the strategy generation module according to the evaluation result for adjusting the parameters or structure of the rehabilitation strategy.

[0055] Preferably, it further includes a human-computer interaction module, communicatively connected to the strategy output module, for:

[0056] Receiving the adaptive rehabilitation strategy sent by the strategy output module;

[0057] Converting the adaptive rehabilitation strategy into visual rehabilitation guidance information;

[0058] Displaying the rehabilitation guidance information to the patient through a display device;

[0059] Receiving the feedback information of the patient and sending it to the strategy generation module for further optimizing the rehabilitation strategy.

[0060] A stroke rehabilitation training method based on deep reinforcement learning, including:

[0061] Acquiring the multi-source physiological data of the patient, where the multi-source physiological data includes surface electromyogram signals, inertial measurement unit data, electroencephalogram signals, and electrocardiogram signals;

[0062] Preprocessing and feature extracting the multi-source physiological data;

[0063] Performing spatio-temporal feature fusion based on the preprocessed data;

[0064] Based on the fused features, a non-linear dynamics model is established;

[0065] Based on the non-linear dynamics model, an adaptive rehabilitation strategy is generated;

[0066] The adaptive rehabilitation strategy is outputted to guide the patient in stroke rehabilitation training;

[0067] Among them, the method is implemented using the described system.

[0068] The present invention has the following beneficial effects:

[0069] From a macroscopic perspective, the present invention constructs a closed-loop intelligent rehabilitation ecosystem. This system can not only comprehensively collect the patient's physiological data, but also deeply analyze and model these data, thereby generating the optimal rehabilitation strategy. Such system-level innovation greatly improves the scientificity and effectiveness of rehabilitation training.

[0070] At the system architecture level, the present invention adopts a modular design. The functional modules cooperate closely with each other to form an efficient and scalable technical solution. The cooperation between the data acquisition module and the preprocessing module ensures the high quality of the input data; the coordination between the feature fusion module and the dynamics modeling module provides a solid theoretical basis for strategy generation; the seamless connection between the strategy generation module and the output module ensures the accurate execution of the rehabilitation plan.

[0071] At the algorithm level, the present invention cleverly solves the technical contradictions that are difficult to handle by traditional methods. For example, by introducing quantum mechanics and group theory, the system can simultaneously capture microscopic neuron activities and macroscopic motion patterns, achieving the unity of microscopic accuracy and macroscopic effect. The application of the deep reinforcement learning algorithm solves the problem of real-time optimization of the rehabilitation strategy, enabling the system to continuously adjust the training plan according to the patient's immediate feedback.

[0072] In terms of effects, the present invention demonstrates significant complementarity and synergy. The fusion of multi-source data not only improves the accuracy of the system's assessment of the patient's state, but also provides multi-dimensional information support for the generation of personalized strategies. The combination of deep learning algorithms and traditional rehabilitation theories not only ensures the intelligent level of the system, but also guarantees the scientificity and interpretability of the rehabilitation plan.

[0073] From a microscopic perspective, the present invention has innovative breakthroughs in many technical details. For example, in the data preprocessing stage, the system adopts advanced random matrix theory and topological data analysis methods, greatly improving the efficiency and accuracy of feature extraction. In the dynamics modeling link, the concept of quantum mechanics is introduced, providing a new mathematical tool for describing complex physiological systems.

[0074] Generally speaking, through systematic innovative design and breakthroughs at the algorithm level, the present invention realizes the precision, personalization, and intelligence of stroke rehabilitation training. It can not only significantly improve the rehabilitation effect, shorten the rehabilitation cycle, but also reduce the workload of medical staff and optimize the allocation of medical resources. More importantly, this system is expected to greatly enhance the rehabilitation experience of patients, increase their training enthusiasm, and thus fundamentally improve the rehabilitation effect. This innovative technical solution brings new possibilities to the field of stroke rehabilitation, is expected to play an important role in clinical practice, and bring benefits to more stroke patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 is the overall system architecture diagram of the present invention;

[0076] Figure 2 is the logic block diagram of the data acquisition module of the present invention;

[0077] Figure 3 is the logic block diagram of the data preprocessing module of the present invention;

[0078] Figure 4 is the logic block diagram of the kinetic modeling module of the present invention;

[0079] Figure 5 is the logic block diagram of the strategy generation module of the present invention;

[0080] Figure 6 is the logic block diagram of the human-computer interaction module of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0081] Please refer to the attached Figures 1-6 , the present invention provides a stroke rehabilitation training system and method based on deep reinforcement learning. The system provides personalized and adaptive rehabilitation training strategies for stroke patients through the acquisition, processing, and analysis of multi-source physiological data, combined with deep reinforcement learning technology.

[0082] First of all, the system of the present invention includes a data acquisition module 1 for obtaining multi-source physiological data of patients. These data include surface electromyogram signals, inertial measurement unit data, electroencephalogram signals, and electrocardiogram signals. The design of the data acquisition module 1 takes into account the special needs of stroke patients and adopts non-invasive sensor technology to ensure the comfort and accuracy of the data acquisition process.

[0083] Communicatively connected to the data acquisition module 1 is the data preprocessing module 2. This module receives multi-source physiological data from the data acquisition module 1 and performs data preprocessing and feature extraction. The main tasks of the data preprocessing module 2 are to remove noise, correct signal deviation, and extract meaningful features. This step is crucial for subsequent analysis as it directly affects the performance and accuracy of the system. Surface electromyogram (EMG) signals evaluate muscle activity and movement control. Inertial measurement units (IMUs) capture joint angles and angular velocities, which help analyze the patient's movement patterns and balance ability. Electroencephalogram (EEG) provides information on the state of the nervous system. Electrocardiogram (ECG) monitors the health of the cardiovascular system. After preliminary acquisition, these data enter the data preprocessing module for cleaning, noise reduction, and feature extraction to ensure the accuracy and reliability of subsequent analysis.

[0084] The present invention uses electrodes attached to the surface of the patient's muscles to collect the electrical signals generated during muscle contraction. Analyzing EMG signals can help evaluate the muscle activity level and identify which muscle groups need more training. For example, during arm rehabilitation, if it is found that the EMG signals of the biceps brachii are weak, targeted exercises can be designed to strengthen the strength of this muscle group. IMU devices are usually installed on parts of the patient's body such as the wrists and ankles to capture changes in joint angles and angular velocities. Through IMU data, the system can monitor the patient's gait pattern in real time. For example, for patients with lower limb stroke, the system can provide personalized gait training programs by analyzing gait data to help restore normal walking ability. The EEG electrode cap is placed on the patient's head to record the electrical signals generated by the activities of brain neurons. EEG signals can be used to evaluate the patient's cognitive function and the state of the nervous system. For example, by analyzing EEG data, the system can detect abnormal activities in specific regions of the brain and adjust the rehabilitation strategy accordingly to promote the recovery of neuroplasticity. The ECG electrodes are attached to the patient's chest to record the electrical activities of the heart. ECG data can help monitor the patient's cardiovascular health. For example, during high-intensity rehabilitation training, the system can adjust the training intensity according to ECG data to ensure the safety of the patient.

[0085] After data preprocessing, the feature fusion module 3 receives the processed data and performs spatio-temporal feature fusion. The purpose of this step is to integrate physiological signals from different sources into a unified feature representation to better capture the patient's overall physiological state.

[0086] The kinetic modeling module 4 then builds a nonlinear kinetic model based on the fused features. This model aims to describe the physiological change laws during the patient's rehabilitation process and provide a theoretical basis for the generation of subsequent rehabilitation strategies.

[0087] The strategy generation module 5 is the core part of this system. It generates adaptive rehabilitation strategies according to the non-linear dynamics model. These strategies are dynamically adjusted according to the patient's real-time state and rehabilitation progress to ensure the effectiveness and safety of training.

[0088] Finally, the strategy output module 6 is responsible for outputting the generated adaptive rehabilitation strategies to guide the patient in stroke rehabilitation training. This may include specific training action guidance, training intensity recommendations, etc.

[0089] In one embodiment of the present invention, the data preprocessing module 2 includes several key sub-units. First is the random matrix processing unit 21, which processes multi-source physiological data using random matrix theory. The application of random matrix theory here is mainly to reduce the dimension of the data while retaining important statistical characteristics.

[0090] For example, the following formula can be used for random projection:

[0091] Y = RX,

[0092] where X is the original data matrix, R is the random projection matrix, and Y is the data matrix after dimension reduction. Suppose there are EMG signals from 5 different muscle groups, and each signal contains 1000 sampling points (i.e., X ∈ R 5×1000 ). To reduce the computational complexity, a randomly generated projection matrix R ∈ R 5×3 can be used to reduce the data dimension to 3 dimensions (i.e., Y ∈ R 3×1000 ). In this way, the system can significantly reduce the amount of calculation while maintaining key features. The data after dimension reduction is easier to process and helps with subsequent feature fusion and dynamics modeling, improving the response speed and accuracy of the system.

[0093] Next, the topological analysis unit 22 performs topological data analysis on the processed data. The purpose of this step is to capture the geometric and topological features of the data, which is very important for understanding the movement patterns and neural activity patterns of stroke patients. Topological data analysis can help discover hidden structures in the data, such as using the persistent homology algorithm:

[0094]

[0095] where H k (X) represents the k-dimensional homology group, is a boundary operator. When analyzing gait data, assume there is a series of IMU data that records the patient's gait patterns at different time periods. Through TDA, key nodes and repetitive patterns in the gait cycle can be identified. For example, by calculating the homology group of the gait cycle, the system can determine the stable and unstable phases of gait, thus designing more targeted gait training programs. This method can discover hidden data structures and help the system better understand and optimize the patient's movement patterns.

[0096] Finally, the feature extraction unit 23 obtains the processed data matrix through a topological transformation operator and Hadamard product operation.

[0097] This step can be expressed as:

[0098]

[0099] where T is the topological transformation operator, ⊙ represents the Hadamard product, matrix, and Z is the final feature matrix.

[0100] Suppose surface electromyography (EMG) and inertial measurement unit (IMU) data of a stroke patient during upper limb rehabilitation training are being processed. The goal is to extract meaningful features from these multi-source physiological signals to develop personalized rehabilitation strategies.

[0101] First, electrodes are used to collect the electrical signals of the patient's upper arm muscle activities, recording the activities of 5 different muscle groups within 10 seconds. Each muscle group has 1000 sampling points, forming a 5×1000 raw data matrix X EMG . IMU devices are installed at the patient's wrist, elbow, and shoulder positions to record the changes in joint angles and angular velocities. Similarly, a 3×1000 raw data matrix X IMU is obtained.

[0102] First, these raw data are preprocessed, including steps such as denoising and normalization, to obtain the dimensionality-reduced data matrices Y EMG ∈R 5×500 and Y IMU ∈R 3×500 .

[0103] Next, the preprocessed data is further processed using the topological transformation operator T. Assume the persistent homology algorithm is used to extract topological features. Specifically, the homology group can be calculated within each time window to capture the geometric and topological structures in the data. For example, for Y EMG , its 0-dimensional and 1-dimensional homology groups can be calculated to obtain a new feature matrix T(Y EMG )∈R 5×500 . Similarly, for YIMU , its homology group can also be calculated to obtain

[0104] Now, these two feature matrices need to be combined. To this end, a pre-defined Hadamard matrix H ∈ R 8× 5 00 is introduced. This matrix can contain some prior knowledge or weight coefficients to emphasize the importance of certain specific features.

[0105] Suppose and are combined into an 8×500 matrix Then, perform the Hadamard product operation:

[0106]

[0107] where H ∈ R 8×500 , and the resulting matrix Z ∈ R 8 ×500.

[0108] In this specific example, assume the following feature matrix Z is obtained:

[0109]

[0110] Based on these eigenvalues, the system can identify which muscle groups perform well during training and which need more attention. For example, if it is found that the eigenvalue of muscle D is always high while the eigenvalue of muscle E is low, the system can recommend increasing the strength training for muscle E.

[0111] Through topological transformation and Hadamard product operation, the system can extract more representative and discriminative features, thus better reflecting the patient's physiological state. By combining multiple physiological signals, the system can more comprehensively understand the patient's rehabilitation needs, formulate more accurate and personalized rehabilitation strategies. The extracted features can help the system predict potential problems during the rehabilitation process in advance and adjust the training plan in a timely manner to ensure the safety and effectiveness of the rehabilitation process.

[0112] The feature extraction unit 23 extracts meaningful features from multi-source physiological signals through the topological transformation operator and Hadamard product operation, significantly enhancing the understanding of the patient's overall physiological state. This method not only improves the personalization level of the rehabilitation strategy but also can predict and address potential problems in advance, thus significantly improving the rehabilitation effect. Specifically, by fusing multiple signals such as EMG and IMU, the system can discover hidden patterns and relationships, formulate more accurate rehabilitation plans, and ensure the safety and effectiveness of the training process.

[0113] In another embodiment of the present invention, the feature fusion module 3 includes a chaos analysis unit 31, a differential geometry processing unit 32, and a feature fusion unit 33. The chaos analysis unit 31 analyzes the preprocessed data using chaos theory. In the context of stroke rehabilitation, chaos analysis can help understand the complex dynamic behavior of the patient's nervous system. For example, the maximum Lyapunov exponent can be calculated:

[0114]

[0115] where λ is the maximum Lyapunov exponent, and δx(t) represents the distance between two initially close orbits in the phase space.

[0116] The differential geometry processing unit 32 performs differential geometry processing on the analyzed data. The purpose of this step is to capture the intrinsic geometric structure of the data, which is crucial for understanding the changes in the patient's movement patterns.

[0117] For example, the Riemannian metric can be used to describe the data manifold:

[0118] ds 2 =g ij dx i dx j ,

[0119] where g ij are the components of the metric tensor.

[0120] The feature fusion unit 33 obtains the fused feature matrix through the calculation of the Laplace - Beltrami operator and the Lyapunov exponent. The Laplace - Beltrami operator can help capture the global and local structures of the data:

[0121]

[0122] where Δ g is the Laplace - Beltrami operator, g is the Riemannian metric, and f is a function defined on the manifold.

[0123] Through these processing steps, the system of the present invention can extract rich feature information from multi - source physiological data, providing a solid basis for the subsequent generation of rehabilitation strategies. This multi - modal and multi - scale data processing method can comprehensively capture the physiological state of the patient, thereby realizing more accurate and personalized rehabilitation training.

[0124] By fusing multiple physiological signals, the system can understand the physiological state of the patient more comprehensively, rather than just the information provided by a single signal. This helps to discover hidden patterns and relationships, thereby formulating more effective rehabilitation strategies.

[0125] Suppose a stroke patient is undergoing upper limb rehabilitation training. By fusing data from EMG, IMU, and EEG, the system can simultaneously monitor muscle activity, joint angle changes, and the activity of brain neurons. For example, if during a specific movement, the EMG signal shows weak muscle activity but the EEG signal shows abnormally active brain activity, this may indicate that the patient's brain is attempting to compensate for muscle weakness. Based on this information, the system can adjust the training plan, increase strength training for that muscle group, and incorporate neurofeedback training to optimize the coordination between the brain and muscles.

[0126] Through multimodal data fusion, the system can identify individual differences and formulate personalized rehabilitation strategies accordingly. This approach is more accurate and effective than relying solely on a single type of physiological signal.

[0127] Suppose a stroke patient is undergoing gait training. The system detects significant asymmetry during walking through IMU data and abnormal activity in certain brain regions through EEG data analysis. Through the analysis of the feature fusion unit 33, the system can generate a feature matrix that includes gait asymmetry and brain activity patterns. Based on this feature matrix, the system can design a training plan specifically targeting the patient's gait asymmetry and brain function recovery, including using robotic assistive devices for gait correction training and combining cognitive training to promote brain function recovery.

[0128] By fusing and analyzing multi-source data, the system can predict potential problems during the rehabilitation process in advance and adjust the rehabilitation strategy in a timely manner, thereby improving the predictability and controllability of the rehabilitation effect.

[0129] Suppose a stroke patient is undergoing rehabilitation training. The system continuously monitors the patient's electrocardiogram (ECG) and electromyogram (EMG) and discovers that the patient has an overloaded heart during high-intensity training. Through the analysis of the feature fusion unit 33, the system can identify the relationship between this heart load and the specific training intensity. Based on this information, the system can recommend reducing the training intensity or adjusting the training frequency to ensure the patient's safety and avoid over-fatigue. In addition, the system can further optimize the training plan based on the patient's real-time feedback (such as self-reported fatigue) to ensure the smooth progress of the rehabilitation process.

[0130] The feature fusion unit 33 generates a unified feature representation by fusing multiple physiological signals, which can significantly enhance the understanding of the patient's overall physiological state. This method not only improves the personalization level of the rehabilitation strategy but also can predict and address potential problems in advance, thus significantly improving the rehabilitation effect. Specifically, by fusing multiple signals such as EMG, IMU, and EEG, the system can discover hidden patterns and relationships, formulate more precise rehabilitation plans, and ensure the safety and effectiveness of the training process.

[0131] Preferably, each module and unit in the system of the present invention are connected by a high-speed data bus to ensure the real-time and reliability of data transmission. At the same time, the system is also equipped with a high-performance computing unit to support the operation of complex data processing and deep reinforcement learning algorithms.

[0132] The method and system of the present invention have multiple advantages in practical applications. First, it can adapt to the individual differences of different patients and provide personalized rehabilitation programs. Second, through real-time data analysis and strategy adjustment, it can respond in a timely manner to the changes during the patient's rehabilitation process and maximize the rehabilitation effect. Moreover, the application of deep reinforcement learning enables the system to continuously learn and optimize, and its performance will continue to improve as the usage time increases.

[0133] Generally speaking, the present invention provides an innovative stroke rehabilitation training solution, which combines advanced data analysis techniques, deep learning algorithms with traditional rehabilitation theories, providing new possibilities for the rehabilitation of stroke patients. Through continuous data collection, analysis, and strategy optimization, this system is expected to significantly improve the effect of stroke rehabilitation, shorten the rehabilitation period, and ultimately improve the patient's quality of life.

[0134] In the system of the present invention, the dynamics modeling module 4 is a key component, which is responsible for establishing a non-linear dynamics model describing the patient's rehabilitation process. Preferably, this module includes a quantum mechanics modeling unit 41, a group theory optimization unit 42, and a model generation unit 43. This structural design aims to make full use of modern physics and mathematical theories to more accurately describe the complex physiological changes during the rehabilitation process of stroke patients.

[0135] The quantum mechanics modeling unit 41 initially establishes a dynamics model using quantum mechanics methods. In this process, the patient's physiological state can be regarded as a quantum system, and its evolution can be described by a variant of the equation:

[0136]

[0137] where, |ψ(t)) represents the quantum state of the system, is the Hamilton operator, is the reduced Planck constant.

[0138] Suppose we are studying the ion channel dynamics between neurons. Through a quantum mechanics model, we can simulate the opening and closing behavior of ion channels and predict the optimal stimulation time and intensity. For example, by analyzing the dynamics of a specific neuron population, the system can suggest applying electrical stimulation at a specific time point to maximize the restoration of neuroplasticity. Suppose we are studying the ion channel dynamics between neurons. Through a quantum mechanics model, we can simulate the opening and closing behavior of ion channels and predict the optimal stimulation time and intensity. For example, by analyzing the dynamics of a specific neuron population, the system can suggest applying electrical stimulation at a specific time point to maximize the restoration of neuroplasticity.

[0139] The advantage of this quantum mechanics modeling method is that it can capture microscopic processes that are difficult to describe by traditional classical models, such as the quantum effects of neurons. The group theory optimization unit 42 then uses group theory methods to optimize the dynamic model. The application of group theory here is mainly to describe the symmetry and invariance of the system, which is crucial for understanding the restoration process of the movement patterns of stroke patients. For example, Lie groups can be used to describe continuous symmetries:

[0140] g(t) = exp(tX),

[0141] where g(t) is an element of the Lie group, X is an element of the Lie algebra, and t is a parameter. Through group theory optimization, the system of the present invention can better capture the invariant features and symmetry structures during the rehabilitation process.

[0142] In upper limb rehabilitation training, suppose we need to analyze the continuous movements of the patient's arm. Through group theory optimization, the system can identify and utilize the consistent patterns during the rehabilitation process. For example, by analyzing the symmetry of arm rotation, the system can design more effective exercise programs, such as performing repetitive movement training within a specific angular range to promote faster recovery.

[0143] Through symmetry analysis, the system can identify and utilize the consistent patterns during the rehabilitation process and design more effective training programs

[0144] Finally, the model generation unit 43 generates a non - linear dynamic model describing the system evolution through variants of equations and group action operators. This step can be expressed as:

[0145]

[0146] where Ψ is the wave function of the system and G is the group action operator. This model can comprehensively consider quantum effects and system symmetries, providing a solid theoretical basis for the generation of subsequent rehabilitation strategies.

[0147] Another important component of the present invention is the strategy generation module 5. This module is responsible for generating an adaptive rehabilitation strategy based on the non-linear dynamics model. Preferably, the strategy generation module 5 includes a functional analysis unit 51, a variational method processing unit 52, and a strategy optimization unit 53.

[0148] The functional analysis unit 51 performs a functional analysis on the non-linear dynamics model. The purpose of this step is to transform complex dynamics problems into functional optimization problems. For example, an energy functional can be defined:

[0149] The functional analysis unit 51 performs a functional analysis on the non-linear dynamics model. The purpose of this step is to transform complex dynamics problems into functional optimization problems. For example, an energy functional can be defined:

[0150]

[0151] where Ω is the configuration space of the system, and V(x) is the potential energy function.

[0152] The variational method processing unit 52 then applies the variational method to process the analysis results. The application of the variational method here is mainly to find the extreme value of the energy functional, which corresponds to the optimal state of the system. For example, the Euler-Lagrange equation can be used

[0153]

[0154] where L is the Lagrangian function.

[0155] Suppose it is necessary to develop a personalized rehabilitation plan for a specific patient. By defining the energy functional E[ψ], the optimal training intensity and frequency can be found. For example, for a patient with a shoulder injury, the system can adjust the parameters in the energy functional to find the best training plan suitable for the current state, such as training three times a week for 30 minutes each time. This method allows the system to find the optimal rehabilitation strategy and maximize the patient's rehabilitation effect.

[0156] The strategy optimization unit 53 obtains the optimal adaptive rehabilitation strategy by minimizing the energy functional and the total variation regularization. This step can be expressed as:

[0157]

[0158] where is the optimal rehabilitation strategy, is the energy functional, is the total variation regularization term, and λ is the regularization parameter.

[0159] The system of the present invention further includes a strategy optimization module 6, which is communicatively connected to the strategy generation module 5 and the strategy output module 7. The main function of the strategy optimization module 6 is to further optimize the generated adaptive rehabilitation strategy to improve its exploration ability and adaptability.

[0160] Preferably, the strategy optimization module 6 optimizes the adaptive rehabilitation strategy using random matrix theory and probability theory. For example, the stochastic gradient descent method can be used:

[0161]

[0162] where, is the current strategy, η is the learning rate, is the loss function. By introducing randomness, the system of the present invention can better explore the strategy space and avoid falling into local optimal solutions. In an embodiment of the present invention, the data acquisition module 1 includes a plurality of dedicated acquisition units. The surface electromyography acquisition unit 11 is used to acquire the myoelectric signals of the patient, which is crucial for evaluating muscle activity and movement control. The inertial measurement unit 12 is used to acquire the joint angle and angular velocity data of the patient, which helps to analyze the patient's movement pattern and balance ability. The electroencephalogram acquisition unit 13 and the electrocardiogram acquisition unit 14 are respectively used to acquire the electroencephalogram signals and electrocardiogram signals of the patient, and these signals can provide important information about the state of the patient's nervous system and cardiovascular system.

[0163] In particular, the system of the present invention further includes a data synchronization unit 15, which is communicatively connected to the above-mentioned acquisition units. The main task of the data synchronization unit 15 is to perform time alignment and sampling rate unification on the acquired multi-source physiological data. This step is crucial for subsequent data analysis and feature fusion because different physiological signals may have different sampling rates and time delays.

[0164] Preferably, the data synchronization unit 15 uses the Dynamic Time Warping (DTW) algorithm for time alignment. The objective function of the DTW algorithm can be expressed as:

[0165]

[0166] where, X and Y are two time series to be aligned, w is the alignment path, and d is the distance function.

[0167] Through this carefully designed data acquisition and synchronization process, the system of the present invention can obtain high-quality, synchronized multimodal physiological data, laying a solid foundation for subsequent analysis and strategy generation. The integration of this multi-source data not only improves the accuracy of the system's assessment of the patient's state but also provides rich information support for the formulation of personalized rehabilitation strategies. The system of the present invention also includes a rehabilitation effect evaluation module 8, which is communicatively connected to the strategy output module 7 and the strategy generation module 5. The main function of the rehabilitation effect evaluation module 8 is to evaluate the effect of the adaptive rehabilitation strategy and provide feedback based on the evaluation results to further optimize the rehabilitation strategy.

[0168] Preferably, the rehabilitation effect evaluation module 8 receives the rehabilitation training data from the strategy output module 7 and conducts comprehensive analysis in combination with preset clinical evaluation indicators. These clinical evaluation indicators may include, but are not limited to, the Fugl-Meyer score, the Barthel index, and the modified Rankin scale, etc. The rehabilitation effect evaluation module 8 uses machine learning algorithms, especially deep learning models, to analyze the patient's rehabilitation progress.

[0169] For example, in an embodiment of the present invention, the rehabilitation effect evaluation module 8 uses a long short-term memory network (LSTM) to predict the patient's rehabilitation trajectory. The core formula of the LSTM model can be expressed as:

[0170] f t =σ(W f ·[h t-1 ,x t +b f ),

[0171] i t =σ(W i ·[h t-1 ,x t +b i ),

[0172]

[0173] o t =σ(W o ·[h t-1 ,x t +b o ),

[0174] h t =o t *tanh(C t ),

[0175] where f t 、i t and o t are the forget gate, input gate, and output gate respectively, and C tis the cell state, h t is the hidden state, and W and B are the weight and bias parameters.

[0176] Based on the prediction results of the LSTM model, the rehabilitation effect evaluation module 8 can calculate the deviation between the actual rehabilitation progress and the expected progress. If the deviation exceeds a preset threshold, such as the 95% confidence interval, the rehabilitation effect evaluation module 8 will send an adjustment signal to the policy generation module 5. This dynamic evaluation and feedback mechanism ensures that the rehabilitation strategy can respond in a timely manner to the individual needs and progress of the patient.

[0177] The system of the present invention further includes a human-machine interaction module 9, which is communicatively connected to the policy output module 7. The design of the human-machine interaction module 9 aims to improve the usability of the system and the compliance of the patient, which is crucial for the long-term effect of stroke rehabilitation training.

[0178] In a preferred embodiment of the present invention, the human-machine interaction module 9 includes a visualization unit 91 and a feedback processing unit 92. The visualization unit 91 is responsible for converting the adaptive rehabilitation strategy sent by the policy output module 7 into intuitive and easy-to-understand visual information. For example, 3D animations can be used to show the correct movement postures, or progress bars can be used to display the completion status of the rehabilitation training.

[0179] The visualization unit 91 employs advanced computer graphics technologies, such as real-time rendering and physical simulation, to ensure the authenticity and smoothness of the visual effects. The rendering process can be represented by the following simplified rendering equation:

[0180] L o (x, ω o , λ, t) = L e (x, ω o , λ, t) + ∫ Ω f r (x, ω i , ω o , λ, t)L i (x, ω i , λ, t)(ω i ·n)dω i ,

[0181] where L 0 is the outgoing radiance, L e is the self-emitting radiance, f r is the bidirectional reflectance distribution function (BRDF), and L iis the incident radiance, λ is the wavelength, and t is the time. The feedback processing unit 92 is responsible for receiving the patient's feedback information and converting it into a form that the system can understand. Such feedback may include voice commands, gesture recognition, or touch screen input, etc. The feedback processing unit 92 uses natural language processing (NLP) and computer vision technologies to parse the patient's feedback. For example, for voice feedback, a recurrent neural network (RNN) can be used for speech recognition:

[0182] h t = tanh(W hx x t + W hh h t-1 + b h ),

[0183] where h t is the hidden state at time t, x t is the input, W is the weight matrix, and b is the bias term.

[0184] The feedback processing unit 92 is responsible for receiving the patient's feedback information and converting it into a form that the system can understand. Such feedback may include voice commands, gesture recognition, or touch screen input, etc. The feedback processing unit 92 uses natural language processing (NLP) and computer vision technologies to parse the patient's feedback. For example, for voice feedback, a recurrent neural network (RNN) can be used for speech recognition:

[0185] h t = tanh(W hx x t + W hh h t-1 + b h ),

[0186] where h t is the hidden state at time t, x t is the input, W is the weight matrix, and b is the bias term.

[0187] The design of the human-computer interaction module 9 not only improves the usability of the system, but also provides the patient with a richer feedback channel, which helps to enhance the patient's participation and the pertinence of rehabilitation training.

[0188] Finally, the present invention also provides a stroke rehabilitation training method based on deep reinforcement learning. This method is closely related to the above system and includes steps such as data acquisition, preprocessing, feature fusion, dynamics modeling, policy generation, and output.

[0189] In this method, multi-source physiological data of the patient are first obtained, including surface electromyogram signals, inertial measurement unit data, electroencephalogram signals, and electrocardiogram signals. These data are collected by high-precision sensors and are subject to preliminary filtering and noise reduction processing.

[0190] Subsequently, the obtained multi-source physiological data are preprocessed and feature extracted. This step involves operations such as data cleaning, normalization, and dimensionality reduction. For example, principal component analysis (PCA) can be used for dimensionality reduction:

[0191] Y = XW,

[0192] where X is the original data matrix, W is the feature vector matrix, and Y is the data matrix after dimensionality reduction. Next, spatio-temporal feature fusion is performed based on the preprocessed data. This step aims to capture the correlation and temporal characteristics between different physiological signals. Tensor decomposition methods such as Tucker decomposition can be used:

[0193]

[0194] where, is the original data tensor, is the core tensor, and A, B, and C are factor matrices.

[0195] Based on the fused features, a non-linear dynamics model is established. This step utilizes the quantum mechanics and group theory methods mentioned above to capture the complex dynamic behavior of the system.

[0196] Finally, based on the non-linear dynamics model, an adaptive rehabilitation strategy is generated, and its output is used to guide the patient in stroke rehabilitation training. The strategy generation process uses deep reinforcement learning algorithms such as proximal policy optimization (PPO):

[0197]

[0198] where, r t (θ) is the probability ratio, is the advantage function estimate, and ∈ is the clipping parameter.

[0199] Through this method, the present invention can provide personalized and adaptive rehabilitation training programs for stroke patients, effectively improving the rehabilitation effect and efficiency. The innovation of this method lies in its integration of multidisciplinary cutting-edge technologies such as quantum computing, group theory, and deep learning, bringing new ideas and solutions to the field of stroke rehabilitation.

[0200] To verify the superiority of the stroke rehabilitation training system and its method based on deep reinforcement learning of the present invention, a series of experiments were conducted to compare the effects of the embodiments of the present invention with those of two comparative examples. These experiments aimed to simulate real stroke rehabilitation training scenarios and evaluate the performance of the system in improving the rehabilitation effect and efficiency of patients.

[0201] Embodiment 1 adopted the complete system of the present invention, including core technologies such as multi-source data acquisition, deep reinforcement learning strategy generation, and adaptive rehabilitation programs. Comparative example 1 adopted a traditional fixed rehabilitation program without personalized adjustment. Comparative example 2 used a simple machine learning algorithm to generate rehabilitation programs but lacked the adaptive ability of deep reinforcement learning.

[0202] Sixty hemiplegic patients after stroke were selected to participate in a 12-week rehabilitation training, with 20 people in each group. The main evaluation indicators included the Fugl-Meyer upper limb motor function score (FMA-UE), the ability of activities of daily living (ADL) score, rehabilitation training compliance, and patient satisfaction. These indicators comprehensively reflected the rehabilitation effect, the improvement of the patient's quality of life, and the practicality of the system.

[0203] The FMA-UE score used the standard Fugl-Meyer assessment scale with a full score of 66 points. The ADL score used the modified Barthel index (MBI) with a full score of 100 points. The rehabilitation training compliance was measured by the proportion of patients completing the specified training tasks, with a full score of 100%. The patient satisfaction used a Likert 5-point scale, with 1 point being the lowest and 5 points being the highest.

[0204] The experimental results are shown in the following table:

[0205] Evaluation index Example 1 Comparative example 1 Comparative example 2 Improvement in FMA-UE score 18.5±2.3 10.2±1.8 13.7±2.1 Improvement in ADL score 22.3±2.7 12.8±2.2 16.5±2.4 Compliance with rehabilitation training 92％±3％ 75％±5％ 83％±4％ Patient satisfaction 4.6±0.3 3.2±0.5 3.8±0.4

[0206] It can be clearly seen from the experimental results that Embodiment 1 of the present invention was significantly superior to the two comparative examples in all evaluation indicators. The FMA-UE score of Embodiment 1 increased by 18.5 points, which was 81.4% higher than that of Comparative Example 1 and 35.0% higher than that of Comparative Example 2. This indicates that the system of the present invention can more effectively improve the upper limb motor function of patients, which benefits from the system's adaptive rehabilitation strategy and precise movement guidance.

[0207] In terms of the increase in ADL score, Embodiment 1 also performed excellently, being 74.2% and 35.2% higher than those of Comparative Example 1 and Comparative Example 2 respectively. This means that the present invention not only improved the motor function of patients but also significantly improved their ability to take care of themselves in daily life, which has a direct positive impact on the quality of life of patients.

[0208] The compliance of rehabilitation training is an important indicator to measure the practicality of the system and the acceptance of patients. The compliance of Example 1 is as high as 92%, far higher than 75% of Comparative Example 1 and 83% of Comparative Example 2. This result shows that the system of the present invention can better stimulate the training enthusiasm of patients, probably due to its personalized training plan and intuitive human-computer interaction interface.

[0209] In terms of patient satisfaction, Example 1 also leads by a large margin, reaching 4.6 points (out of 5 points). This reflects the high recognition of the patients for this system, probably due to its remarkable rehabilitation effect and good user experience.

[0210] These experimental results fully prove the superiority of the present invention in the field of stroke rehabilitation training. Through the deep reinforcement learning algorithm, the system can dynamically adjust the rehabilitation strategy according to the real-time state and progress of the patient, achieving true personalization and precision rehabilitation. The fusion and analysis of multi-source data enable the system to comprehensively grasp the physiological state of the patient, so as to formulate a more scientific and effective training plan.

[0211] It is particularly worth mentioning that the present invention performs outstandingly in improving patient compliance. This is particularly important for stroke rehabilitation, because rehabilitation is often a long and difficult process, and maintaining a high level of training enthusiasm is crucial for the rehabilitation effect. Through the vivid visualization interface and timely progress feedback, this system has successfully improved the participation and persistence of patients.

[0212] In summary, the stroke rehabilitation training system and method based on deep reinforcement learning of the present invention show significant advantages in improving the motor function of patients, enhancing the quality of life, and increasing training compliance. This method combining advanced algorithms and user-friendly design brings new possibilities to the field of stroke rehabilitation, is expected to play an important role in clinical practice, and bring benefits to more stroke patients.

[0213] It should be noted that the above are only the preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A stroke rehabilitation training system based on deep reinforcement learning, characterized in that: include: Data acquisition module for: Acquiring multi-source physiological data of the patient, wherein the multi-source physiological data includes surface electromyography signals, inertial measurement unit data, electroencephalogram signals, and electrocardiogram signals; A data preprocessing module is connected to the data acquisition module for: Receiving multi-source physiological data sent by the data acquisition module; Based on the multi-source physiological data, perform data preprocessing and feature extraction; The feature fusion module is connected to the data preprocessing module for: Receiving the preprocessed data sent by the data preprocessing module; Based on the preprocessed data, performing spatiotemporal feature fusion; The dynamic modeling module is connected to the feature fusion module for: Receiving the fused features sent by the feature fusion module; Based on the fused features, a nonlinear dynamics model is established; A strategy generation module is in communication with the kinetic modeling module and is used to: Receiving the nonlinear dynamic model sent by the dynamic modeling module; generating an adaptive rehabilitation strategy according to the nonlinear dynamic model; A strategy output module is connected to the strategy generation module for: receiving the adaptive rehabilitation strategy sent by the strategy generation module; The adaptive rehabilitation strategy is outputted to guide the patient to perform stroke rehabilitation training.

2. The system according to claim 1, characterized in that The data preprocessing module comprises: A random matrix processing unit, used for processing the multi-source physiological data using random matrix theory; A topology analysis unit, connected to the random matrix processing unit for performing topology data analysis on the processed data; The feature extraction unit is connected to the topology analysis unit for obtaining a processed data matrix through a topology transformation operator and a Hadamard product operation.

3. The system according to claim 1, characterized in that The feature fusion module includes: A chaos analysis unit, used for analyzing the pre-processed data using chaos theory; A differential geometry processing unit, which is in communication connection with the chaos analysis unit and is used for performing differential geometry processing on the analyzed data; The feature fusion unit is connected to the differential geometry processing unit for obtaining a fused feature matrix through Laplace-Be ltrami operator and Lyapunov index calculation.

4. The system according to claim 1, characterized in that The kinetic modeling module includes: The quantum mechanics modeling unit is used to preliminarily establish a dynamics model using quantum mechanics methods; A group theory optimization unit, in communication with the quantum mechanics modeling unit, for optimizing the dynamics model using a group theory method; A model generation unit is connected to the group theory optimization unit for Variants of the equations and group action operators generate nonlinear dynamic models that describe the evolution of the system.

5. The system according to claim 1, characterized in that The strategy generation module includes: A functional analysis unit, used for performing functional analysis on the nonlinear dynamics model; a variational method processing unit, which is in communication with the functional analysis unit and is used to apply the variational method to process the analysis results; The strategy optimization unit is in communication with the variational method processing unit and is used to obtain an optimal adaptive rehabilitation strategy by minimizing energy functional and total variation regularization.

6. The system according to claim 1, characterized in that It also includes a strategy optimization module, which is in communication with the strategy generation module and the strategy output module and is used to: receiving the adaptive rehabilitation strategy sent by the strategy generation module; Utilizing random matrix theory and probability theory to optimize the adaptive rehabilitation strategy; Improve the exploration and adaptability of the strategy through random weight matrices and gradient descent methods; The optimized strategy is sent to the strategy output module.

7. The system according to claim 1, characterized in that The data acquisition module comprises: Surface electromyography acquisition unit, used to obtain electromyography signals of patients; An inertial measurement unit, used to obtain the patient's joint angle and angular velocity data; An EEG acquisition unit, used to obtain the patient's EEG signal; ECG acquisition unit, used to obtain ECG signals of patients; The data synchronization unit is connected to the above-mentioned acquisition units for time alignment and sampling rate unification of the acquired multi-source physiological data.

8. The system according to claim 1, characterized in that It also includes a rehabilitation effect evaluation module, which is in communication with the strategy output module and the strategy generation module and is used to: Receiving the rehabilitation training data sent by the strategy output module; Based on the rehabilitation training data and preset clinical evaluation indicators, evaluating the effect of the adaptive rehabilitation strategy; According to the evaluation result, an adjustment signal is sent to the strategy generation module to adjust the parameters or structure of the rehabilitation strategy.

9. The system according to claim 1, characterized in that It also includes a human-computer interaction module, which is in communication with the strategy output module and is used to: receiving the adaptive rehabilitation strategy sent by the strategy output module; converting the adaptive rehabilitation strategy into visual rehabilitation guidance information; Display rehabilitation guidance information to patients through display devices; Receive patient feedback and send it to the strategy generation module for further optimizing the rehabilitation strategy.

10. A stroke rehabilitation training method based on deep reinforcement learning, characterized in that: include: Acquiring multi-source physiological data of the patient, wherein the multi-source physiological data includes surface electromyography signals, inertial measurement unit data, electroencephalogram signals, and electrocardiogram signals; Preprocessing and feature extraction of the multi-source physiological data; Based on the preprocessed data, perform spatiotemporal feature fusion; According to the fused features, a nonlinear dynamics model is established; generating an adaptive rehabilitation strategy based on the nonlinear dynamic model; Outputting the adaptive rehabilitation strategy to guide the patient to perform stroke rehabilitation training; Wherein, the method is implemented using the system described in any one of claims 1-9.

Images