Spine rehabilitation auxiliary training method, system and device based on deep learning and medium

Abstract

The application relates to a spine rehabilitation auxiliary training method, system, device and medium based on deep learning. The method comprises the following steps: acquiring multi-source heterogeneous data containing spine images, motion postures, electromyographic signals and clinical symptoms and performing feature extraction; performing deep fusion on the extracted multi-dimensional features through a multi-modal feature fusion network to obtain comprehensive features representing rehabilitation states; processing the fused features by using a rehabilitation state evaluation model to obtain structured evaluation data; generating an individualized initial training scheme through a generative adversarial network based on the evaluation data; and dynamically optimizing and adaptively adjusting the initial scheme according to real-time feedback data when a user performs training to generate a final rehabilitation training scheme. The method can effectively overcome the problems of subjectivity in guidance, blind area in supervision, lag in evaluation and rigidity in scheme in the traditional rehabilitation mode, and realizes objective and quantitative evaluation of the whole spine rehabilitation process, generation of individualized schemes and dynamic closed-loop optimization.

Description

Technical Field

[0001] This invention belongs to the field of rehabilitation assistive training technology, and in particular relates to spinal rehabilitation assistive training methods, systems, devices and media based on deep learning. Background Technology

[0002] With the interdisciplinary development of spinal surgery and rehabilitation medicine, patient posture monitoring technology based on wearable sensors and biomechanical analysis has emerged. This technology can objectively record patients' physical activity data to a certain extent. Consequently, current postoperative rehabilitation guidance mainly relies on two methods: one is in rehabilitation institutions, where rehabilitation therapists provide face-to-face, intermittent guidance and simple equipment training based on their experience; the other is for patients to conduct self-training at home by referring to paper or video materials and return to the hospital regularly for evaluation.

[0003] However, traditional methods have significant limitations. First, both the therapist's guidance and the patient's self-perception are subjective, making it difficult to accurately quantify and provide real-time feedback on abstract instructions such as "keep your back straight" and "avoid twisting," easily leading to incorrect training postures, compensatory movements, or even secondary injuries. Second, most of the patient's daily life and training process is in an unsupervised "blind spot," and the spinal load and postural compliance during dynamic activities such as sitting, standing, and walking cannot be continuously monitored and corrected in a timely manner. Third, the assessment of the phased rehabilitation effect is severely delayed, usually relying on outpatient follow-up examinations several weeks or even months later. This discrete assessment cannot dynamically and comprehensively reflect the subtle changes in the rehabilitation process, resulting in untimely and untargeted adjustments to the training plan. Some existing assistive devices have limited functions, often only providing simple postural over-limit reminders, and cannot deeply integrate multi-dimensional physiological and behavioral data for systematic status assessment and personalized rehabilitation decisions.

[0004] Therefore, how to achieve objective quantitative assessment, real-time precise guidance, and dynamic optimization of the entire process of spinal surgery rehabilitation has become a key technical challenge to improve rehabilitation quality and reduce the risk of complications. Summary of the Invention

[0005] Therefore, it is necessary to provide deep learning-based methods, systems, equipment, and media for assistive training in spinal rehabilitation.

[0006] Firstly, this application provides a deep learning-based spinal rehabilitation assisted training method, including:

[0007] S1. Obtain multi-source heterogeneous spinal data of the target user and extract multi-dimensional features from the multi-source heterogeneous data to obtain multi-dimensional feature data; among which, the multi-source heterogeneous data includes spinal imaging data, motion posture data, electromyographic signal data and clinical symptom data;

[0008] S2. Deep feature fusion processing is performed on multi-dimensional feature data through a preset multimodal feature fusion network to obtain fused feature data. The multimodal feature fusion network includes an encoder layer, an attention mechanism layer, and a fusion layer connected in sequence. The fused feature data is used to characterize the comprehensive features of the user's spinal rehabilitation status.

[0009] S3. The fusion feature data is processed by a pre-set rehabilitation status assessment model to classify the status and predict risks, resulting in rehabilitation assessment data. The rehabilitation assessment data includes the current rehabilitation level, degree of functional impairment, and rehabilitation risk index.

[0010] S4. The rehabilitation assessment data is processed by a pre-defined training program generation network to match personalized training programs, resulting in initial training program data. The training program generation network includes a generator and a discriminator.

[0011] S5. Dynamically optimize and adaptively adjust the initial training program data to obtain a rehabilitation training program; the rehabilitation training program includes training movements, execution frequency, intensity gradient arrangement, and phased goal setting.

[0012] In one embodiment, multi-source heterogeneous data of the target user's spine are acquired, and multi-dimensional feature extraction is performed on the multi-source heterogeneous data to obtain multi-dimensional feature data, including:

[0013] S11. Collect spinal imaging data, motion posture data, electromyographic signal data, and clinical symptom text data of the target user;

[0014] S12. Extract morphological features of spinal curvature, intervertebral space height, and vertebral rotation angle from spinal imaging data to obtain imaging feature data;

[0015] S13. Extract the angular velocity and angle time sequence features of trunk flexion, extension, lateral bending and rotation from the motion posture data to obtain posture feature data;

[0016] S14. Extract the average power frequency and median frequency features of the erector spinae and multifidus muscles in a specific frequency band from the electromyographic signal data to obtain electromyographic feature data.

[0017] S15. Extract text feature data from clinical symptom text data;

[0018] S16. Based on image feature data, posture feature data, electromyography feature data, and text feature data, obtain multi-dimensional feature data.

[0019] In one embodiment, deep feature fusion processing is performed on multi-dimensional feature data using a preset multimodal feature fusion network to obtain fused feature data, including:

[0020] S21. Input the image feature data, posture feature data, electromyography feature data and text feature data into the encoder layer, and perform high-dimensional mapping through the corresponding fully connected neural network to obtain a high-dimensional feature vector of the same dimension.

[0021] S22. Input the high-dimensional feature vector into the attention mechanism layer and calculate the cross-modal attention weights;

[0022] S23. Based on cross-modal attention weights, the high-dimensional feature vectors of each modality are weighted and fused in the fusion layer to generate preliminary fused features;

[0023] S24. Perform layer normalization on the preliminary fusion features to obtain fusion feature data that characterizes the comprehensive features of the user's spinal rehabilitation status.

[0024] In one embodiment, the fused feature data is processed by a preset rehabilitation status assessment model to perform status grading and risk prediction, resulting in rehabilitation assessment data, including:

[0025] S31. Input the fused feature data into the first classification network in the rehabilitation status assessment model, and output the current rehabilitation level of the target user;

[0026] S32. Input the fused feature data and the current rehabilitation level into the second regression network in the rehabilitation status assessment model, and output the functional impairment score of the target user.

[0027] S33. Input the fused feature data and functional impairment score into the risk prediction network of the rehabilitation status assessment model, and calculate the rehabilitation risk index using the following formula:

[0028]

[0029] In the formula, As a recovery risk index, To integrate risk-related feature subsets from feature data The risk coefficient obtained through training, This is for the operation of retrieving the maximum value;

[0030] S34. Integrate the current rehabilitation level, functional impairment score, and rehabilitation risk index into structured rehabilitation assessment data.

[0031] In one embodiment, a pre-defined training scheme generation network is used to perform personalized training scheme matching processing on rehabilitation assessment data to obtain initial training scheme data, including:

[0032] S41. Input rehabilitation assessment data into the generator, which is the generative part of the conditional generative adversarial network, and outputs a set of candidate training schemes containing multiple training actions, execution frequency, intensity and target.

[0033] S42. Input each program in the candidate training program set along with the number of rehabilitation assessments into the discriminator. The discriminator outputs a quality score that characterizes the matching degree and rationality of the program.

[0034] S43. Based on the quality score, select the highest-scoring scheme from the candidate training scheme set as the initial training scheme data.

[0035] In one embodiment, the initial training program data is dynamically optimized and adaptively adjusted to obtain a rehabilitation training program, including:

[0036] S51. Decompose the initial training scheme data into a set of parameters that can be adjusted independently. The set of parameters that can be adjusted independently includes the training action sequence, the execution frequency of each action, the intensity gradient, and the setting of phased goals.

[0037] S52. When the target user performs the training action sequence, collect motion posture feedback data and electromyographic feedback data in real time, and input them into the multimodal feature fusion network and rehabilitation status assessment model to generate short-term rehabilitation status change data.

[0038] S53. Based on short-term rehabilitation status change data and rehabilitation risk index, the intensity gradient and execution frequency are dynamically adjusted through optimization algorithms to generate updated training scheme parameters.

[0039] S54. Based on the historical rehabilitation assessment data sequence of the target users, the phased goal setting is adaptively revised to obtain the revised phased goal setting;

[0040] S55. Integrate the updated training program parameters and revised phased goal settings to generate a rehabilitation training program.

[0041] Secondly, this application also provides a deep learning-based spinal rehabilitation assistive training system, comprising:

[0042] The data acquisition and feature module is used to acquire multi-source heterogeneous spinal data of the target user and extract multi-dimensional features from the multi-source heterogeneous data to obtain multi-dimensional feature data; among which, the multi-source heterogeneous data includes spinal imaging data, motion posture data, electromyographic signal data and clinical symptom data;

[0043] The feature data fusion module is used to perform deep feature fusion processing on multi-dimensional feature data through a preset multimodal feature fusion network to obtain fused feature data. The multimodal feature fusion network includes an encoder layer, an attention mechanism layer and a fusion layer connected in sequence. The fused feature data is used to characterize the comprehensive features of the user's spinal rehabilitation status.

[0044] The rehabilitation assessment module is used to perform status classification and risk prediction processing on the fused feature data through a preset rehabilitation status assessment model to obtain rehabilitation assessment data; the rehabilitation assessment data includes the current rehabilitation level, degree of functional impairment, and rehabilitation risk index;

[0045] The initial training plan generation module is used to perform personalized training plan matching processing on rehabilitation assessment data through a preset training plan generation network to obtain initial training plan data; the training plan generation network includes a generator and a discriminator.

[0046] The rehabilitation training program generation module is used to dynamically optimize and adaptively adjust the initial training program data to obtain a rehabilitation training program. The rehabilitation training program includes training movements, execution frequency, intensity gradient arrangement, and phased goal setting.

[0047] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect.

[0048] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect.

[0049] The aforementioned deep learning-based spinal rehabilitation assisted training methods, systems, equipment, and media achieve objective quantitative assessment of spinal rehabilitation status through multi-source heterogeneous data acquisition and multi-dimensional feature extraction, combined with multimodal feature fusion and a rehabilitation status assessment model. It utilizes generative adversarial networks to generate personalized initial training plans and dynamically optimizes training parameters and phased goals based on real-time feedback data, forming a precisely tailored rehabilitation training program. This process effectively addresses the problems of traditional rehabilitation guidance being highly subjective, lacking continuous monitoring during training, lagging rehabilitation assessment, and insufficient program specificity. It provides real-time, precise guidance, dynamically matches the rehabilitation process, improves rehabilitation quality, and reduces the risk of secondary injury and complications. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a flowchart illustrating a deep learning-based spinal rehabilitation assisted training method in one embodiment.

[0052] Figure 2 This is a schematic diagram of the structure of a deep learning-based spinal rehabilitation assistive training system in one embodiment. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0054] refer to Figure 1 The document presents a flowchart illustrating the deep learning-based spinal rehabilitation assisted training method provided in this application, which includes the following steps:

[0055] S1. Obtain multi-source heterogeneous spinal data of the target user, and extract multi-dimensional features from the multi-source heterogeneous data to obtain multi-dimensional feature data.

[0056] Optionally, multi-source heterogeneous data include spinal imaging data, motion posture data, electromyographic signal data, and clinical symptom data.

[0057] Optionally, spinal imaging data is acquired using high-precision medical imaging equipment, preferably MRI scanners and computed tomography (CT) scanners. Scans are performed on the cervical, thoracic, lumbar, and sacrococcygeal vertebrae of the user's spine to obtain weighted MRI and CT images. This ensures spatial resolution and clearly presents key pathological features such as vertebral morphology, the degree of intervertebral disc herniation, the extent of spinal canal stenosis, and changes in spinal cord signal. Simultaneously, the image data is stored in a standard format for easy subsequent data processing.

[0058] Optionally, motion posture data is acquired using a multi-sensor fusion acquisition system. This system consists of inertial measurement units deployed at key nodes of the user's head, neck, chest, waist, pelvis, and lower limbs. Each inertial measurement unit integrates a three-axis accelerometer, a three-axis gyroscope, and a three-axis magnetometer, enabling real-time acquisition of acceleration, angular velocity, and magnetic field strength data during various movements such as sitting, standing, walking, bending, and turning. The Euler angles of each body segment are then obtained through a quaternion posture calculation algorithm, achieving accurate characterization of the spinal and related limb motion postures. Furthermore, the system transmits sensor data to an edge computing terminal in real-time via a wireless communication protocol, avoiding data transmission delays.

[0059] Optionally, electromyography (EMG) signal data is acquired through a surface EMG sensor. The sensor electrodes are made of a special material and are deployed on the surface of key muscle groups related to spinal movement, such as the erector spinae, rectus abdominis, quadratus lumborum, and latissimus dorsi. The raw EMG signal is preprocessed by a differential amplifier circuit, a bandpass filter, and a notch filter to remove power frequency interference and baseline drift. Then, the analog signal is converted into a digital signal by an analog-to-digital converter.

[0060] Optionally, clinical symptom data is obtained through a combination of structured electronic questionnaires and face-to-face interviews with rehabilitation therapists. This includes information such as the user's visual analog scale score for pain, functional impairment index, past medical history, type of surgery, and postoperative time. Standardized coding methods are used to convert text and scoring data into quantifiable numerical data.

[0061] Optionally, in the multi-dimensional feature extraction stage, targeted feature extraction algorithms are adopted for different types of multi-source heterogeneous data. For spinal imaging data, a segmentation model is first used to accurately segment key tissues such as vertebrae, intervertebral discs, and spinal cord in MRI and computed tomography images to obtain regions of interest for each tissue. Then, morphological and texture features are extracted. Morphological features include vertebral height, intervertebral disc thickness, spinal canal diameter, area and perimeter of region of interest, etc., which are obtained by extracting tissue edge contours through edge detection algorithms and then using a geometric calculation model. Texture features include parameters such as contrast, correlation, energy, and entropy of the gray-level co-occurrence matrix, which are calculated by setting specific pixel spacing and angle.

[0062] Optionally, for motion attitude data, the attitude angle data collected by each inertial measurement unit is first smoothed by a sliding window, and the mean filtering algorithm is used to remove motion noise. Then, time domain features and frequency domain features are extracted. The time domain features include the mean, standard deviation, maximum value, minimum value, peak factor, kurtosis, etc. of the attitude angles. The frequency domain features are extracted by converting the attitude angle data to the frequency domain through fast Fourier transform, and parameters such as spectral peak value, peak frequency, and spectral energy are extracted.

[0063] Optionally, for electromyography (EMG) signal data, features are extracted based on preprocessing. Time-domain features include integrated EMG value, root mean square value, mean absolute value, waveform length, etc., which are obtained by direct calculation of the processed EMG signal. Frequency-domain features include average power frequency and median frequency, which are obtained through power spectral density analysis.

[0064] Optionally, for clinical symptom data, feature encoding and normalization are employed. Categorical data is converted into numerical vectors using one-hot encoding, and continuous data is mapped to a specified interval using min-max normalization, ensuring dimensional consistency across different dimensions of clinical data. Finally, the features extracted from each type of data are summarized to obtain multi-dimensional feature data with unified dimensions and standardized format.

[0065] S2. Deep feature fusion processing is performed on multi-dimensional feature data through a preset multimodal feature fusion network to obtain fused feature data.

[0066] Optionally, the multimodal feature fusion network includes an encoder layer, an attention mechanism layer, and a fusion layer connected in sequence, and the fused feature data is used to characterize the comprehensive features of the user's spinal rehabilitation status.

[0067] Optionally, the preset multimodal feature fusion network adopts an end-to-end deep learning architecture, and the structure and function of each layer are designed as follows: The encoder layer is composed of parallel convolutional neural network branches and long short-term memory network branches, which are used to perform targeted feature encoding on different types of multidimensional feature data to achieve dimensionality reduction of high-dimensional features and enhancement of key information.

[0068] Optionally, a convolutional neural network branch is used to process spatially structured data such as spinal image features. This branch contains multiple convolutional layers, pooling layers, and fully connected layers. The convolutional layers are used to extract different levels of spatial features from the image features. Each convolutional layer is followed by a pooling layer to reduce feature dimensionality, retain key features, and prevent overfitting. Finally, the convolutional and pooled features are integrated through a fully connected layer to output the image encoded features.

[0069] Optionally, a Long Short-Term Memory (LSTM) branch is used to process temporal data such as motion posture features and electromyography (EMG) signal features. This branch contains multiple LTM layers and fully connected layers, and uses dropout layers to prevent overfitting. The input data are time-series aligned sequences of motion posture features and EMG signal features. The LTM branch extracts key features in the temporal dimension by modeling the long-term dependencies of the temporal data. Finally, the features output by the LTM network are integrated through fully connected layers to output temporal encoded features.

[0070] Optionally, for structured numerical features such as clinical symptom features, a fully connected encoder is used for encoding. This encoder contains multiple fully connected layers and converts clinical symptom features into structured coded features through multiple linear transformations and nonlinear activations.

[0071] Optionally, the attention mechanism layer employs a multi-head self-attention mechanism to adaptively assign weights to the three types of encoded features output by the encoder layer, highlighting features that significantly impact spinal rehabilitation status assessment and suppressing interference from redundant features. Specifically, the three types of encoded features are first concatenated to obtain a concatenated feature vector, which is then input into the multi-head self-attention module. This module contains multiple parallel self-attention heads, implemented through linear transformations of the query, key, and value of the concatenated feature vector. These linear transformations are performed using multiple different fully connected layers, thereby calculating the attention weight for each feature dimension. The attention weights are calculated using the scaled dot product attention formula, as follows:

[0072]

[0073] In the above formula, This represents the query vector, used to characterize the feature information that needs to be focused on at the moment; This represents the key vector, used for matching calculations with the query vector to measure the degree of association between features; This represents a value vector, used to perform a weighted summation of feature information based on the calculated attention weights; This indicates that matrix multiplication is performed on the transpose of the query vector and the key vector to obtain the similarity matrix between features; This represents the square root of the attention head dimension, used to scale the similarity matrix and avoid excessively large values ​​due to high dimensionality. The dimension of the attention head is represented; the softmax function is used to convert the similarity matrix into an attention weight matrix, so that the sum of the weights is 1, thereby achieving weight allocation for different feature dimensions; finally, the attention-weighted feature vector is obtained by matrix multiplication of the attention weight matrix and the value vector.

[0074] Optionally, the output features of multiple attention heads are then concatenated and integrated through a fully connected layer to obtain attention-enhanced features. The fusion layer employs a combination of cross-attention fusion and adaptive feature adjustment to achieve deep fusion and dimensional unification of the attention-enhanced features. First, the attention-enhanced features are input into the cross-attention module, which introduces learnable fusion parameters to further interact and fuse encoded features from different sources, thereby mining complementary information between features of different modalities. Subsequently, the output features of the cross-attention module are linearly transformed through a fully connected layer, using GELU as the activation function to enhance the non-linear expressive power of the features. Finally, the output features of the fully connected layer are normalized through a batch normalization layer to eliminate dimensional differences between different feature dimensions, resulting in fused feature data.

[0075] Optionally, the training process of the multimodal feature fusion network adopts a supervised training method. The training dataset contains multi-source heterogeneous data, fused feature data and corresponding rehabilitation status labels. The loss function adopts the cross-entropy loss function, and the optimizer adopts the Adam optimizer. Through iterative training, the network converges to the preset loss threshold to ensure that the network has good feature fusion performance.

[0076] S3. The fusion feature data is processed by a preset rehabilitation status assessment model to perform status classification and risk prediction, thereby obtaining rehabilitation assessment data.

[0077] Optionally, rehabilitation assessment data may include the current rehabilitation level, degree of functional impairment, and rehabilitation risk index.

[0078] Optionally, the preset rehabilitation status assessment model adopts a deep learning-based multi-task learning architecture, which can simultaneously complete three tasks: rehabilitation status classification, functional impairment assessment, and rehabilitation risk index prediction. The input of the model is fused feature data, and the output is the corresponding rehabilitation assessment data.

[0079] Optionally, the architecture of the rehabilitation status assessment model includes a shared feature extraction layer, three task-specific output layers, and a loss fusion module. The shared feature extraction layer is used to further deepen and abstract the fused feature data, providing a unified high-level feature representation for the three tasks. This layer contains multiple fully connected layers and dropout layers. The dropout layers are used to prevent the model from overfitting. After processing by this layer, the fused feature data is converted into high-level shared features.

[0080] Optionally, a dedicated output layer for the rehabilitation status grading task maps high-level shared features to corresponding rehabilitation level labels. These rehabilitation levels are divided into multiple levels according to clinical guidelines for spinal rehabilitation, with different levels corresponding to different degrees of functional impairment and training recommendations. This output layer employs a fully connected layer containing the corresponding number of neurons, using the Softmax activation function. By calculating the probability distribution for each rehabilitation level, the level with the highest probability is used as the predicted result for the current rehabilitation level.

[0081] Optionally, a dedicated output layer for assessing the degree of spinal dysfunction is used to quantitatively evaluate the user's spinal dysfunction. It employs a continuous value prediction method, with the output range corresponding to the quantification interval of the dysfunction degree; higher values ​​indicate more severe dysfunction. This output layer uses a fully connected layer containing a single neuron, with the sigmoid activation function. By mapping high-level shared features to a specified interval and then scaling them, a quantitative score of the dysfunction degree is obtained.

[0082] Optionally, a dedicated output layer for the rehabilitation risk index prediction task is used to predict the probability of secondary injury, complications, and other risks to the user during subsequent rehabilitation. The output range corresponds to the quantification interval of the risk probability, with higher values ​​indicating higher rehabilitation risk. This output layer uses a fully connected layer containing a single neuron and employs the sigmoid activation function. By modeling risk-related information from high-level shared features, the quantitative prediction result of the rehabilitation risk index is obtained.

[0083] The loss fusion module integrates the loss functions of the three tasks to achieve multi-task joint training of the model. Specifically, the cross-entropy loss function is used for the rehabilitation status grading task to measure the difference between the predicted rehabilitation level probability distribution and the true level label; the mean squared error loss function is used for both the functional impairment assessment task and the rehabilitation risk index prediction task to measure the error between the predicted continuous values ​​and the true values. The loss functions of the three tasks are fused into a total loss function through weighted summation, with the weight coefficients determined experimentally based on the importance of each task.

[0084] The model training process employs the Adam optimizer, iteratively training the total loss function until it converges to a preset threshold. After training, the fused feature data is input into the rehabilitation status assessment model. The model, through a shared feature extraction layer and task-specific output layers, outputs the current rehabilitation level, functional impairment score, and rehabilitation risk index; these three components together constitute the rehabilitation assessment data. To ensure the reliability of the assessment results, validation is performed after model output. The results are compared with those from human assessments by clinical rehabilitation therapists. If the error exceeds a preset threshold, the model parameters are fine-tuned until the assessment results meet clinical accuracy requirements.

[0085] S4. The rehabilitation assessment data is processed by a pre-set training program generation network to match personalized training programs, resulting in initial training program data.

[0086] Optionally, the training scheme generation network includes a generator and a discriminator.

[0087] Optionally, the preset training scheme generation network adopts a generative adversarial network architecture. Through adversarial training between the generator and the discriminator, the generator can learn the mapping relationship between rehabilitation assessment data and the optimal training scheme, thereby generating a training scheme that meets the user's personalized needs.

[0088] The generator is used to convert rehabilitation assessment data into structured initial training scheme data, while the discriminator is used to distinguish the training scheme data generated by the generator from the actual optimal training scheme data. Through the adversarial game between the two, the quality and rationality of the training scheme generated by the generator are improved.

[0089] Optionally, the generator adopts an encoder-decoder architecture. The input is rehabilitation assessment data, including the current rehabilitation level, functional impairment score, and rehabilitation risk index. First, the rehabilitation assessment data is feature-encoded and converted into a high-dimensional feature vector. The decoder then maps the high-dimensional feature vector into structured initial training scheme data.

[0090] Specifically, the encoder part of the generator contains multiple fully connected layers. Through multiple linear transformations and nonlinear activations, it transforms the rehabilitation assessment data into a high-dimensional feature vector. To enhance the expressive power of the features, a batch normalization layer and a dropout layer are set after each fully connected layer of the encoder to prevent overfitting. The decoder part of the generator adopts a structure combining deconvolutional neural networks and fully connected layers. First, the high-dimensional feature vector is expanded through multiple deconvolutional layers. Then, the output of the deconvolutional layers is transformed into a feature vector of a specified dimension through a global average pooling layer. Finally, multiple fully connected layers perform linear transformations on the feature vector, and finally output the initial training scheme data vector. This vector contains the total number of training scheme parameters, including the training action type encoding, the execution frequency of each action, the intensity level, the training duration, and the stage target parameters.

[0091] Optionally, to ensure that the generated training protocol data conforms to clinical rehabilitation standards, the activation function is set to Tanh after the output layer of the decoder, the output data is mapped to a specified interval, and then converted into values ​​that conform to actual clinical significance through inverse normalization processing.

[0092] Optionally, the discriminator adopts an architecture that combines convolutional neural networks and fully connected layers. The input is a concatenated vector of rehabilitation assessment data and training scheme data, and the output is the true probability corresponding to the concatenated vector, that is, to determine whether the input training scheme data is the true optimal scheme or a scheme generated by the generator.

[0093] Specifically, the discriminator consists of multiple convolutional layers, pooling layers, and fully connected layers. The convolutional layers are used to extract features from different levels of the concatenated vector, and the activation function is LeakyReLU. Each convolutional layer is followed by a pooling layer to reduce the feature dimensionality. Then, the features after convolution and pooling are integrated through a fully connected layer, and the activation function is LeakyReLU. Finally, a fully connected layer containing a single neuron outputs the true probability, and the activation function is Sigmoid. The output range corresponds to the probability interval. An output value close to 1 indicates that it is judged as a real solution, and a value close to 0 indicates that it is judged as a generated solution.

[0094] Optionally, the training process of the training scheme generation network adopts an adversarial training approach. The training dataset includes a large amount of rehabilitation assessment data and corresponding real optimal training scheme data. The real optimal training scheme data is formulated by senior rehabilitation therapists based on clinical experience and rehabilitation guidelines. During training, the generator and discriminator are trained alternately: First, the generator parameters are fixed, and the discriminator is trained. By calculating the classification loss of the discriminator on real scheme data and generated scheme data, the cross-entropy loss function is used, and the discriminator parameters are updated using the gradient descent algorithm, enabling the discriminator to accurately distinguish between real schemes and generated schemes. Subsequently, the discriminator parameters are fixed, and the generator is trained. By calculating the adversarial loss when the generated scheme data is judged as a real scheme by the discriminator, the cross-entropy loss function is used. Combined with the reconstruction loss between generated scheme data and real scheme data, the mean squared error loss function is used, and the generator parameters are updated using the gradient descent algorithm, so that the scheme generated by the generator can be as close as possible to the real optimal scheme. The total loss function is a weighted sum of the adversarial loss and the reconstruction loss. The weight coefficients are determined experimentally to balance the adversarial training effect and reconstruction accuracy.

[0095] Optionally, the training process uses the Adam optimizer to iteratively train the total loss function to converge to a preset threshold. After training, rehabilitation assessment data is input into the generator, which, through encoder-decoder processing, outputs initial training scheme data. This data includes key parameters such as training action type, execution frequency of each action, intensity gradient parameters, training duration, and phased goals.

[0096] S5. Dynamically optimize and adaptively adjust the initial training program data to obtain a rehabilitation training program.

[0097] Optionally, the rehabilitation training program includes training movements, execution frequency, intensity gradient arrangement, and phased goal setting.

[0098] Optionally, dynamic optimization and adaptive adjustment mainly include three core links: individual difference adaptation adjustment, real-time training status feedback optimization, and clinical standard verification and adjustment. Each link is closely connected to form a closed-loop optimization mechanism.

[0099] Optionally, the individual difference adaptation adjustment stage mainly targets individual characteristics such as the user's age, gender, height, weight, type of surgery, and postoperative time, making targeted adjustments to the initial training program data. First, an individual difference influencing factor model is constructed. This model is trained based on a large amount of clinical rehabilitation data and can quantify the degree of influence of different individual characteristics on the training program parameters.

[0100] Specifically, the user's individual characteristic parameters are input into the individual difference influencing factor model. The model outputs adjustment coefficients for each training scheme parameter, including training intensity adjustment coefficients, training frequency adjustment coefficients, and movement difficulty adjustment coefficients. Subsequently, the corresponding parameters in the initial training scheme data are corrected based on these adjustment coefficients. For example, if the initial training intensity level is a certain level, and the user has specific individual characteristics, the individual difference influencing factor model outputs a corresponding intensity adjustment coefficient; the adjusted training intensity level is then obtained by multiplying the two. If the initial training movements include movements of a certain difficulty level, and the user's postoperative time is short with a low movement difficulty adjustment coefficient, this movement is replaced with a substitute movement of the corresponding difficulty level, and the key parameters of the movement are adjusted. Simultaneously, based on the user's exercise capacity limits obtained through prior collection and analysis of movement posture data, the range parameters of the training movements are restricted to ensure that the training movements are performed within the user's safe range of motion.

[0101] The real-time training status feedback optimization process dynamically adjusts training program parameters by collecting real-time data from the user during the execution of the initial training plan, achieving closed-loop optimization of the training program. Specifically, real-time data acquisition employs a multi-source data acquisition system to collect real-time data on the user's movement posture, electromyography signals, and the user's real-time subjective feelings during training. This data, including pain and fatigue levels, is collected through the interactive module on the training device.

[0102] Subsequently, a real-time training status assessment model was constructed. The input to this model consisted of real-time collected motion posture data, electromyographic signal data, and subjective feeling data. The motion posture data included the standardization score and completion rate of the training movements, the electromyographic signal data included the degree of muscle activation and the presence of abnormal electromyographic signals, and the subjective feeling data included the pain score. The model output was the training status assessment result, which included scores in three dimensions: training intensity suitability, movement standardization, and physical endurance. The score range for each dimension corresponded to the quantitative range of suitability, standardization, or endurance. The higher the score, the better the suitability, standardization, or endurance.

[0103] Dynamic adjustment rules are formulated based on the training status assessment results. If the training intensity suitability score is lower than the specified threshold, it indicates that the intensity is too high, so the training intensity is reduced by one level, and the duration of each training session is reduced. If the training intensity suitability score is higher than the specified threshold, it indicates that the intensity is too low, so the training intensity is increased by the corresponding level, or the number of repetitions of the training movements is increased. If the movement standardization score is lower than the specified threshold, the current training movement is paused, and a standard movement demonstration is shown to the user through prompts, and a breakdown training session for that movement is added. If the physical tolerance score is lower than the specified threshold, such as a pain score exceeding a specified value, training is stopped immediately, and the training plan is adjusted to rest and relaxation training, including spinal stretching, muscle relaxation, etc., and the user's condition is restored before reassessment and adjustment.

[0104] In addition, by using a sliding window to cumulatively analyze the real-time training status evaluation results, if the scores of each dimension in multiple consecutive training sessions are higher than the specified threshold, it indicates that the current training scheme is well-suited and the training difficulty can be gradually increased according to the preset intensity gradient; if the score of a certain dimension in multiple consecutive training sessions is lower than the specified threshold, it is necessary to re-examine the rationality of the initial training scheme and make significant adjustments based on individual differences.

[0105] The clinical standard verification and adjustment process aims to ensure that the optimized training program conforms to clinical guidelines and medical standards for spinal rehabilitation, avoiding unreasonable training content. Specifically, a clinical standard verification database is constructed, which includes authoritative clinical guidelines in the field of spinal rehabilitation, preventive measures for common complications, and contraindicated training movements for different disease types.

[0106] The training program data, after being adapted to individual differences and optimized with real-time training status feedback, is compared and verified with information in the clinical standard verification library: if the training program contains contraindicated training movements, the movement is immediately deleted and replaced with a clinically recommended alternative movement; if the rate of increase in training intensity exceeds the upper limit of the clinical standard, the rate of increase in intensity is adjusted to conform to the standard range; if the phased goals are set beyond reasonable clinical expectations, the phased goals are reset according to clinical rehabilitation patterns.

[0107] Simultaneously, a manual review mechanism by senior rehabilitation therapists is introduced. Training programs that pass initial screening undergo final review by these therapists. If potential safety risks or areas for improvement are identified, adjustments are made based on clinical experience. After dynamic optimization and adaptive adjustments through these three stages, the final rehabilitation training program is obtained. This program clearly includes core elements such as training movements, execution frequency, intensity gradient arrangement, and phased goal setting, providing users with personalized, safe, and effective guidance for spinal rehabilitation training. The training movements include specific movement names, movement specifications, and demonstration points; the execution frequency includes the number of daily training sessions, the interval between each session, and the number of training days per week; the intensity gradient arrangement includes the initial intensity, the periodic intensity increase, and the intensity ceiling; and the phased goal setting includes short-term, medium-term, and long-term goals, corresponding to the expected functional recovery within different periods.

[0108] The aforementioned deep learning-based spinal rehabilitation assisted training method achieves objective quantitative assessment of spinal rehabilitation status through multi-source heterogeneous data acquisition and multi-dimensional feature extraction, combined with multimodal feature fusion and a rehabilitation status assessment model. It utilizes generative adversarial networks to generate personalized initial training plans and dynamically optimizes training parameters and phased goals based on real-time feedback data and historical trends. This effectively addresses the problems of traditional rehabilitation guidance being highly subjective, lacking continuous supervision during training, lagging rehabilitation assessment, and insufficient targeted plans. The entire process is supported by objective data and real-time dynamic adjustments, enabling precise training guidance, timely correction of improper postures, reduction of secondary injury risk, and improved rehabilitation quality. This provides strong support for the precise and personalized management of the entire spinal rehabilitation process.

[0109] In an optional embodiment, multi-source heterogeneous data of the target user's spine are acquired, and multi-dimensional feature extraction is performed on the multi-source heterogeneous data to obtain multi-dimensional feature data, including:

[0110] S11. Collect spinal imaging data, motion posture data, electromyographic signal data, and clinical symptom text data of the target user.

[0111] Optionally, spinal imaging data is acquired using medical imaging equipment, scanning the core area of ​​the user's spinal lesion and adjacent segments to obtain imaging data including spinal bony structures and soft tissues. The image data is uniformly converted to a standard medical imaging format and stored on a medical data server. Motion posture data is acquired using a wearable inertial measurement unit (IMU) array. This array consists of multiple IMU sensors fixed to specific segments of the user's head, cervical spine, thoracic spine, lumbar spine, pelvis, and lower limbs. It simultaneously acquires triaxial acceleration, triaxial angular velocity, and triaxial magnetic field strength data. The real-time acquired data is transmitted to an edge computing terminal for temporary storage and preprocessing via a wireless communication protocol. Electromyography (EMG) signal data is acquired using surface electromyography (SEM) sensors. These sensors employ dedicated biocompatible electrodes, selecting muscle groups around the spine closely related to the rehabilitation state as monitoring sites. The data is connected to the data acquisition system via wired transmission. During acquisition, power frequency interference and motion artifacts are simultaneously eliminated to ensure signal integrity. Clinical symptom text data is collected collaboratively in two ways: first, by calling the hospital's electronic medical record system interface to obtain structured text information such as the user's age, gender, type of surgery, duration of illness, and past medical history; second, by collecting data through standardized clinical questionnaires to obtain unstructured clinical symptom descriptions such as the user's pain experience and degree of activity limitation. Finally, the two types of text data are integrated into a unified clinical symptom text dataset.

[0112] S12. Extract morphological features of spinal curvature, intervertebral disc height, and vertebral rotation angle from spinal imaging data to obtain imaging feature data.

[0113] Optionally, the spinal imaging data is first preprocessed. Standard format image data is read from a professional data reading library, and grayscale value normalization is performed to eliminate imaging differences between different devices. Gaussian filtering is used for noise reduction to improve image quality. Image segmentation algorithms are used to automatically segment the spinal bony structures and accurately locate key areas such as vertebrae and intervertebral spaces. Spinal curvature extraction employs a curve fitting algorithm. Multiple feature points are selected along the segmented spinal centerline, and a least-squares method is used to fit a curve to the spinal centerline. The spinal curvature feature value is calculated based on the curvature of the fitted curve. Intervertebral space height extraction is achieved through distance calculation. The upper and lower endplates of adjacent vertebrae are located in the segmented image, multiple measurement points are selected to calculate the vertical distance between the endplates, and the average value is taken as the height feature value of the intervertebral space. Vertebral rotation angle extraction uses contour matching and angle calculation methods. The edge contour features of the vertebra are extracted and matched with the standard vertebral contour. The rotation angle of the actual vertebral contour relative to the standard contour is calculated as the vertebral rotation angle feature value. After standardizing the extracted spinal curvature, intervertebral space height, and vertebral rotation angle features, they are integrated to form image feature data.

[0114] S13. Extract the angular velocity and angle time sequence features of trunk flexion, extension, lateral bending and rotation from the motion posture data to obtain posture feature data.

[0115] Optionally, the raw motion attitude data is first preprocessed by using Kalman filtering to eliminate sensor noise. Kalman filtering includes two core stages: prediction and update. The state equation for the prediction stage is... The observation equation is The Kalman gain formula for the update phase is: The state update formula is: The covariance update formula is: The parameters are explained below: for The system state vector at time t, which includes three dimensions: position, velocity, and acceleration; This is the state transition matrix, used to describe the transition of the system state from... Time's up The relationship of time transition; for Time-based control input; This is process noise; for The observed value at time; This is the observation matrix, used to map the system state to the observation space; To observe noise; for The covariance matrix of the time-state prediction; Kalman gain is used to balance the confidence of predicted and observed values. To observe the noise covariance matrix; For based on state at any moment Predicted value at any time; After merging observations The estimated optimal state at time t; for The covariance matrix of the time-optimal state estimate; The matrix is ​​the identity matrix. By fusing preprocessed acceleration and angular velocity data through complementary filtering, the Euler angles of the user's torso in each motion direction are calculated, including pitch, roll, and yaw angles, corresponding to torso flexion / extension, lateral bending, and rotational movements, respectively. Based on the calculated Euler angle time-series data, angular and angular velocity time-series features for each motion direction are extracted: angular time-series features include the maximum and minimum angles, the range of angle changes, and the rate of angle change during the motion; angular velocity time-series features include the maximum and minimum angular velocities, the rate of change of angular velocity, and the time of peak angular velocity occurrence. After feature alignment and standardization, the extracted time-series features are integrated to form attitude feature data.

[0116] S14. Extract the average power frequency and median frequency features of the erector spinae and multifidus muscles in a specific frequency band from the electromyographic signal data to obtain electromyographic feature data.

[0117] Optionally, the electromyography (EMG) signal data is first preprocessed, using notch filtering to eliminate power frequency interference, and wavelet transform to extract the effective signal components. The core formula of wavelet transform is: In the formula These are wavelet coefficients. This is a scaling parameter used to control the scaling of the wavelet. These are translation parameters used to control the translation of the wavelet. This is a time-domain function of the electromyography signal. For time variables, This is the mother wavelet function. After preprocessing, for the electromyographic signals corresponding to the erector spinae and multifidus muscles, a Fast Fourier Transform (FFT) is used to convert the time-domain signals to the frequency domain. The core formula of the FFT is: In the formula This is the frequency domain output value. For frequency point number, To change the number of points, The first time-domain input signal One sampling point, The unit is the imaginary number. Within the frequency domain, specific frequency bands closely related to muscle activity are selected, and the average power frequency and median frequency within these bands are calculated: the average power frequency is the weighted average of the power spectral density within the band, with the weight being the power at each frequency point; the median frequency is the frequency point where the power spectral area within the band is divided into two equal parts. The average power frequency and median frequency of the erector spinae and multifidus muscles in specific frequency bands are extracted separately, and these feature values ​​are integrated to form electromyographic feature data.

[0118] S15. Extract text feature data from clinical symptom text data.

[0119] Optionally, the clinical symptom text data is first preprocessed: unstructured text is segmented and stop words are removed to convert the text into a standardized word sequence; structured text is regularized to unify data format and representation. A word embedding model is used to convert the preprocessed text word sequence into low-dimensional dense word vectors. The word embedding model is pre-trained on a large amount of clinical text corpus and can effectively capture the semantic information of medical terms. For a single clinical symptom text, the average pooling method is used to aggregate all its word vectors to obtain the global vector representation of the text; for a text set composed of multiple clinical symptom texts, the individual text vectors are concatenated and dimensionality compressed to obtain a text feature vector with a unified dimension. At the same time, the classification information in the clinical symptom text is processed using one-hot encoding, and the continuous text information is standardized. The processed features are fused with the text vector features to finally form the text feature data.

[0120] S16. Based on image feature data, posture feature data, electromyography feature data, and text feature data, obtain multi-dimensional feature data.

[0121] Optionally, the image feature data, posture feature data, electromyography (EMG) feature data, and text feature data are first standardized to map the values ​​of each feature dimension to the same data distribution range, eliminating dimensional differences between different feature dimensions. The four types of feature data are then integrated into a high-dimensional feature vector using feature concatenation, with the concatenation order being image features, posture features, EMG features, and text features, ensuring structural consistency of the feature vector. Redundant features are removed from the concatenated high-dimensional feature vector, and features with variance thresholding are selected to retain those with variance greater than a preset threshold, retaining key features that can effectively distinguish different rehabilitation states. The resulting high-dimensional feature vector is the multi-dimensional feature data, which comprehensively integrates information from spinal structure, motor function, muscle activity, and clinical symptoms, providing comprehensive feature support for subsequent rehabilitation status assessment.

[0122] In an optional embodiment, deep feature fusion processing is performed on multi-dimensional feature data using a preset multimodal feature fusion network to obtain fused feature data, including:

[0123] S21. Input the image feature data, posture feature data, electromyography feature data and text feature data into the encoder layer, and perform high-dimensional mapping through the corresponding fully connected neural network to obtain a high-dimensional feature vector of the same dimension.

[0124] Optionally, the encoder layer consists of four parallel fully connected neural network encoders, corresponding to image feature data, pose feature data, electromyography (EMG) feature data, and text feature data, respectively. Each encoder maintains a consistent network structure, consisting of multiple fully connected layers. Each fully connected layer is followed by a batch normalization layer and an activation function. The ReLU activation function is chosen to effectively introduce non-linear transformation and enhance feature representation capabilities. For each type of input modal feature data, it is input into the corresponding fully connected neural network encoder. Through the linear transformation of each fully connected layer and the non-linear transformation of the activation function, modal features of different dimensions and distributions are mapped to a high-dimensional feature space of the same dimension. Through this high-dimensional mapping process, image features, pose features, EMG features, and text features are all converted into high-dimensional feature vectors of a unified dimension, providing a foundation for subsequent cross-modal attention calculation and feature fusion.

[0125] S22. Input the high-dimensional feature vector into the attention mechanism layer and calculate the cross-modal attention weights.

[0126] Optionally, the attention mechanism layer employs a cross-modal multi-head attention mechanism, capable of simultaneously capturing multiple association patterns between different modalities. First, the four types of high-dimensional feature vectors corresponding to images, pose, electromyography, and text are mapped into query vector matrices, key vector matrices, and value vector matrices, respectively. This mapping process is implemented through a fully connected layer, and the dimension of the mapping matrix is ​​determined based on the dimension of the high-dimensional feature vectors. The multi-head attention mechanism includes multiple parallel attention heads, each independently calculating cross-modal attention weights. For each attention head, the dot product of the query vector matrix and the transpose of the key vector matrix is ​​calculated to obtain a similarity matrix between different modalities. The similarity matrix is ​​normalized using a Softmax function to obtain the attention weight matrix between each modality. The attention weight matrices calculated by each attention head are concatenated, and a linear transformation layer converts the concatenated weight matrix into a cross-modal attention weight vector of uniform dimension. This vector contains the importance information of each modality feature relative to the rehabilitation status assessment task.

[0127] S23. Based on cross-modal attention weights, the high-dimensional feature vectors of each modality are weighted and fused in the fusion layer to generate preliminary fused features.

[0128] Optionally, the fusion layer employs a weighted summation strategy, multiplying the high-dimensional feature vectors corresponding to image feature data, posture feature data, electromyography feature data, and text feature data by the corresponding weight values ​​in the cross-modal attention weight vector to obtain the weighted feature vector for each modality. The magnitude of the weight value directly reflects the contribution of the corresponding modality feature to the rehabilitation status assessment; the larger the weight value, the greater the influence of that modality feature in the fusion process. The four weighted modal feature vectors are then summed element-wise to obtain the preliminary fused feature vector. This vector integrates the key information from each modality and can initially reflect the comprehensive characteristics of the user's spinal rehabilitation status. During the weighted fusion process, to avoid the excessive dominance of a single modality feature, a weight adjustment mechanism is set to ensure that each modality feature can participate reasonably in the fusion, thereby improving the robustness of the fused features.

[0129] S24. Perform layer normalization on the preliminary fusion features to obtain fusion feature data that characterizes the comprehensive features of the user's spinal rehabilitation status.

[0130] Optionally, layer normalization can be used to improve the stability of the fused features and accelerate the training and convergence speed of subsequent models. The core formula for layer normalization is: In the formula To initially fuse the elements in the feature vector, For the normalized elements, To initially fuse the mean of the feature vectors, To initially fuse the variance of the feature vectors, This is a local minimum value, used to avoid the denominator being zero. For scaling parameters, For translation parameters, and All parameters are learnable and adaptively adjusted through model training. Layer normalization standardizes the mean and variance of the initial fused feature vectors, adjusting the numerical distribution of the feature vectors to a standard normal distribution and eliminating distributional differences in fused features between different samples. The feature vectors after layer normalization are the fused feature data, which can accurately represent the comprehensive characteristics of the user's spinal rehabilitation status, providing high-quality input features for subsequent rehabilitation status assessment models.

[0131] In an optional embodiment, the fused feature data is processed by a preset rehabilitation status assessment model to perform status grading and risk prediction, resulting in rehabilitation assessment data, including:

[0132] S31. Input the fused feature data into the first classification network in the rehabilitation status assessment model, and output the current rehabilitation level of the target user.

[0133] Optionally, the first classification network adopts a multi-layer fully connected network structure. The input is fused feature data, and the number of network layers is determined according to the dimension of the fused features and the number of rehabilitation levels. Each fully connected layer is followed by a batch normalization layer and a ReLU activation function to improve the network's non-linear expressive power and training stability. The network's output layer uses the Softmax activation function, the formula of which is: In the formula The input value is used to map the network output to the [0,1] interval, obtaining the probability distribution of each rehabilitation level. Rehabilitation levels are divided into multiple levels according to clinical rehabilitation guidelines, with different levels corresponding to different degrees of functional impairment and rehabilitation needs. The first classification network is trained by minimizing the cross-entropy loss function, the formula for which is: In the formula The one-hot encoded value of the real label. The model predicts the first Class probability, This represents the total number of rehabilitation level categories. During the inference phase, the fused feature data is input into the trained first classification network to obtain the probability distribution of each rehabilitation level. The level with the highest probability is selected as the target user's current rehabilitation level.

[0134] S32. Input the fused feature data and the current rehabilitation level into the second regression network in the rehabilitation status assessment model, and output the functional impairment score of the target user.

[0135] Optionally, the second regression network adopts a multi-layer fully connected network structure. The input is an encoded vector of fused feature data and the current rehabilitation level. The current rehabilitation level is converted into a vector form through one-hot encoding and concatenated with the fused feature data as the network input. The hidden layer structure of the network is similar to that of the first classification network, containing multiple fully connected layers, batch normalization layers, and a ReLU activation function. The output layer uses a linear activation function, directly outputting the predicted value of the functional impairment score. The functional impairment score quantifies the degree of damage to the user's spinal function; a higher score indicates a more severe functional impairment. The second regression network is trained by minimizing the mean squared error loss function, the formula of which is: In the formula A true score of the degree of functional impairment. To predict the score, The sample size is [number]. During the inference phase, the fused feature data and the encoded vector of the current rehabilitation level are input into the trained second regression network, which outputs a score of the target user's functional impairment level.

[0136] S33. Input the fused feature data and functional impairment score into the risk prediction network of the rehabilitation status assessment model, and calculate the rehabilitation risk index using the following formula:

[0137]

[0138] In the formula, As a recovery risk index, To integrate risk-related feature subsets from feature data The risk coefficient obtained through training, This is for retrieving the maximum value.

[0139] In the above formula for calculating the rehabilitation risk index, It is a rehabilitation risk index used to assess the risk of secondary injury or stagnation of rehabilitation progress within a specific future period. The value ranges from 0 to 1, where 0 indicates no risk and 1 indicates extremely high risk. The risk-related feature subset is obtained by filtering the fused feature data through a feature selection algorithm, and contains key features that can reflect rehabilitation risk. For bias terms, and The risk coefficients obtained through training are used to adjust the weights of the impact of the functional impairment score and the maximum value of the risk-related feature subset on the rehabilitation risk index. The maximum value operation is used to extract the largest feature value from the risk-related feature subset. This value highlights the feature information most likely to lead to rehabilitation risk. The core function of the risk prediction network is to determine the maximum feature value through training. , and Three risk coefficients are used. During training, the mean squared error between the true and predicted values ​​of the rehabilitation risk index is used as the loss function, and the risk coefficients are optimized using a gradient descent algorithm. In the inference phase, the fused feature data and the functional impairment score are input into the risk prediction network. First, a subset of risk-related features from the fused feature data is selected. Calculate its maximum value Substituting these values ​​into the above formula yields the rehabilitation risk index.

[0140] S34. Integrate the current rehabilitation level, functional impairment score, and rehabilitation risk index into structured rehabilitation assessment data.

[0141] Optionally, a structured format for the rehabilitation assessment data is first defined, including fields for rehabilitation level, functional impairment score, and rehabilitation risk index. Each field corresponds to a specific data type and value range. The current rehabilitation level, functional impairment score, and rehabilitation risk index are then filled into their respective fields, along with a data generation timestamp and data source identifier to ensure the traceability of the assessment data. The entered data is then validated to check if the data type and value range meet the preset requirements. If any data anomalies are found, the process is repeated at the previous level. The resulting structured rehabilitation assessment data clearly and comprehensively reflects the target user's spinal rehabilitation status, providing accurate decision-making basis for the subsequent generation of personalized training programs.

[0142] In an optional embodiment, a preset training scheme generation network is used to perform personalized training scheme matching processing on rehabilitation assessment data to obtain initial training scheme data, including:

[0143] S41. Input rehabilitation assessment data into the generator, which is the generative part of the conditional generative adversarial network, and outputs a set of candidate training schemes containing multiple training actions, execution frequencies, intensities and goals.

[0144] Optionally, the generator employs a multi-layer fully connected network structure, belonging to the generative part of a conditional generative adversarial network. Its input is a structured vector of rehabilitation assessment data, which is transformed into a fixed-dimensional conditional vector through an embedding layer. Simultaneously, a random noise vector is input to enhance the diversity of the generation scheme. The random noise vector follows a normal distribution, and the probability density function of the normal distribution is... In the formula The mean, Standard deviation The variables are random. The generator's hidden layers contain multiple fully connected layers, each using the LeakyReLU activation function, and a batch normalization layer is added to prevent pattern collapse, improving the diversity and stability of the generated schemes. The generator's output layer uses a linear activation function, outputting parameter vectors for multiple candidate training schemes. Each parameter vector corresponds to a candidate training scheme, including training action encoding, execution frequency parameters, intensity parameters, and stage target parameters. The training action encoding corresponds to specific training actions in a preset rehabilitation action library, the execution frequency parameter corresponds to the training time arrangement, the intensity parameter corresponds to the difficulty and load of the training action, and the stage target parameter corresponds to the target value for different rehabilitation stages. The decoding module converts the output parameter vectors into specific candidate training schemes, forming a candidate training scheme set containing multiple candidate schemes.

[0145] S42. Input each program in the candidate training program set along with the number of rehabilitation assessments into the discriminator. The discriminator outputs a quality score that characterizes the matching degree and rationality of the program.

[0146] Optionally, the discriminator also employs a multi-layer fully connected network structure, belonging to the discriminant part of a conditional generative adversarial network. Its input consists of the parameter vector of the candidate training scheme and the conditional vector of the rehabilitation assessment data, which are concatenated as the discriminator's input vector. The discriminator's hidden layers use the LeakyReLU activation function, with a Dropout layer added to enhance generalization ability and reduce overfitting. The output layer uses the Sigmoid activation function, outputting a probability value between 0 and 1. This probability value is the quality score, representing the degree of matching between the input candidate training scheme and the rehabilitation assessment data, as well as the scheme's rationality. The closer the quality score is to 1, the better the candidate training scheme matches the user's rehabilitation status and the higher its rationality; the closer it is to 0, the less suitable the candidate training scheme is or the more rational it is. The discriminator is co-optimized with the generator through adversarial training. During training, the binary cross-entropy loss function is used as the optimization objective to distinguish between real high-quality training schemes and candidate training schemes generated by the generator, continuously improving its ability to discriminate the quality of training schemes. During the inference phase, each program in the candidate training program set is input along with rehabilitation assessment data into the discriminator after training is completed, and a quality score for each candidate program is obtained.

[0147] S43. Based on the quality score, select the highest-scoring scheme from the candidate training scheme set as the initial training scheme data.

[0148] Optionally, the quality scores of all candidate training programs in the candidate training program set are first sorted in descending order, and the candidate program with the highest score is selected as the initial training program. If multiple candidate programs have the same quality score and are all the highest scores, the similarity between these programs and the user's historical rehabilitation training programs is further calculated, and the program with the highest similarity is selected as the initial training program to ensure the continuity and adaptability of the program to the user. The selected initial training program is then checked for completeness, verifying whether it contains core elements such as training movements, execution frequency, intensity, and goals. If any elements are missing, a new set of candidate programs is generated. The final determined initial training program is the initial training program data, which is generated based on the user's rehabilitation assessment data and has strong personalization and targeting.

[0149] In an optional embodiment, the initial training program data is dynamically optimized and adaptively adjusted to obtain a rehabilitation training program, including:

[0150] S51. Decompose the initial training scheme data into a set of parameters that can be adjusted independently.

[0151] Optionally, the independently adjustable parameter sets include the training movement sequence, the execution frequency of each movement, the intensity gradient, and the phased goal setting. The training movement sequence parameter set contains all training movements in the plan and their execution order; the execution frequency parameter set for each movement contains frequency information such as the daily execution count, duration of each execution, and number of days per week for each training movement; the intensity gradient parameter set contains the intensity variation pattern of each training movement, such as the gradient increase or decrease coefficients for movement amplitude and load intensity; and the phased goal setting parameter set contains the time division for different rehabilitation phases and the corresponding rehabilitation goal indicators for each phase. Each parameter set is parameterized, converting the parameters into quantifiable and calculable numerical forms, clarifying the value range and adjustment step size of each parameter, and providing a clear operational basis for subsequent dynamic adjustments.

[0152] S52. When the target user performs the training action sequence, collect motion posture feedback data and electromyographic feedback data in real time, and input them into the multimodal feature fusion network and rehabilitation status assessment model to generate short-term rehabilitation status change data.

[0153] Optionally, during the user's execution of the training sequence, wearable inertial measurement unit arrays and surface electromyography (EMG) sensors are deployed to collect motion posture feedback data and EMG feedback data in real time. The acquisition frequency is matched to the execution rhythm of the training movements to ensure complete capture of the dynamic data of each training movement. Real-time features are extracted from the collected motion posture feedback data and EMG feedback data using feature extraction methods to obtain real-time multi-dimensional feature data. This real-time multi-dimensional feature data is input into a multimodal feature fusion network for feature fusion to obtain real-time fused feature data. The real-time fused feature data is then input into a rehabilitation status assessment model to obtain real-time rehabilitation level, real-time functional impairment score, and real-time rehabilitation risk index. By comparing the real-time rehabilitation status indicators with those at the time of the initial training plan, the change in indicators and the rate of change are calculated to generate short-term rehabilitation status change data. This data reflects the trend of the user's rehabilitation status changes during short-term training.

[0154] S53. Based on short-term rehabilitation status change data and rehabilitation risk index, the intensity gradient and execution frequency are dynamically adjusted through optimization algorithms to generate updated training scheme parameters.

[0155] Optionally, the gradient descent optimization algorithm is selected as the core optimization algorithm, with short-term rehabilitation status change data and rehabilitation risk index as optimization objectives: When the short-term rehabilitation status change data shows that the user's rehabilitation progress is good and the rehabilitation risk index is lower than the preset threshold, the intensity gradient parameter is appropriately increased through the optimization algorithm to increase the intensity of training actions, and the execution frequency can be appropriately increased to accelerate the rehabilitation process; when the short-term rehabilitation status change data shows that the user's rehabilitation progress is slow or the rehabilitation risk index is higher than the preset threshold, the intensity gradient parameter is decreased through the optimization algorithm to reduce the intensity of training actions and the execution frequency, avoiding overtraining that could lead to secondary injury. The objective function of the optimization algorithm is to maximize the rehabilitation progress rate while minimizing the rehabilitation risk, and the expression of the objective function is: In the formula For the rate of recovery progress, These are the weighting coefficients. This is the rehabilitation risk index. An optimization algorithm is used to solve the objective function to obtain the adjustment amounts for the intensity gradient and execution frequency. Based on these adjustments, the corresponding parameter set is adjusted to generate updated training scheme parameters.

[0156] S54. Based on the historical rehabilitation assessment data sequence of the target users, the phased goal setting is adaptively revised to obtain the revised phased goal setting.

[0157] Optionally, historical rehabilitation assessment data of the target users is first collected to form a historical rehabilitation assessment data sequence. This sequence includes indicators such as the user's rehabilitation level, functional impairment score, and rehabilitation risk index at different time points. Time series analysis is used to analyze the trend of the historical rehabilitation assessment data sequence, fitting trend curves of rehabilitation indicators over time to predict the trend of rehabilitation indicators in the future. Combining the predicted trends of rehabilitation indicators with clinical rehabilitation target standards, the phased goals in the initial training program are revised: if the predicted rate of rehabilitation progress is higher than the initial target expectation, the phased goals are appropriately increased; if the predicted rate of rehabilitation progress is lower than the initial target expectation, the phased goals are appropriately decreased, ensuring that the revised phased goals are both challenging and achievable. Simultaneously, referring to the target requirements of different rehabilitation stages in clinical rehabilitation guidelines, the time division of the phased goals is optimized to make the goal setting more consistent with clinical rehabilitation patterns, ultimately resulting in the revised phased goal setting.

[0158] S55. Integrate the updated training program parameters and revised phased goal settings to generate a rehabilitation training program.

[0159] Optionally, the updated training program parameters (including adjusted training movement sequences, execution frequencies, and intensity gradients) are integrated with the revised stage-specific goal settings, and the corresponding fields are filled in according to the preset program format. The integrated program undergoes a logical consistency check, verifying whether the training movement sequences match the intensity gradients and execution frequencies, and whether the stage-specific goal settings are consistent with the training program parameters. If logical conflicts exist, the corresponding steps are returned for readjustment. The validated programs are standardized, using clear and easy-to-understand language to specify the operational procedures for each training movement, the detailed arrangement of the execution frequency, the stage-by-stage change table of the intensity gradient, and the specific target values ​​for each stage, generating a standardized rehabilitation training program document. This rehabilitation training program fully considers the user's real-time rehabilitation status changes and historical rehabilitation trends, possessing strong personalization and adaptability, and can effectively guide users in spinal rehabilitation training.

[0160] The aforementioned deep learning-based spinal rehabilitation assisted training method integrates multi-source heterogeneous data such as spinal imaging, movement posture, electromyographic signals, and clinical symptoms. Through automated feature extraction, fusion, and decision-making via a deep learning model, it achieves a comprehensive and objective quantitative assessment of the spinal rehabilitation status. Based on this assessment, a highly personalized initial training plan can be generated, and the plan's intensity and objectives can be dynamically optimized and adaptively adjusted based on real-time feedback data during execution. This effectively overcomes the inherent shortcomings of traditional rehabilitation, such as subjective guidance, blind spots in supervision, delayed assessment, and rigid plans. It provides patients with scientifically accurate and continuously adaptable rehabilitation training guidance that spans both in-hospital and home settings, helping to improve the safety and effectiveness of rehabilitation training and reduce the risk of secondary injury.

[0161] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0162] Based on the same inventive concept, this application also provides a deep learning-based spinal rehabilitation assistive training system for implementing the aforementioned deep learning-based spinal rehabilitation assistive training method. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more deep learning-based spinal rehabilitation assistive training system embodiments provided below can be found in the limitations of the deep learning-based spinal rehabilitation assistive training method described above, and will not be repeated here.

[0163] In one exemplary embodiment, such as Figure 2 As shown, a schematic diagram of the structure of a deep learning-based spinal rehabilitation assistive training system 10 is provided, including:

[0164] The data acquisition and feature module 11 is used to acquire multi-source heterogeneous spinal data of the target user and extract multi-dimensional features from the multi-source heterogeneous data to obtain multi-dimensional feature data; among which, the multi-source heterogeneous data includes spinal imaging data, motion posture data, electromyographic signal data and clinical symptom data;

[0165] The feature data fusion module 12 is used to perform deep feature fusion processing on multi-dimensional feature data through a preset multimodal feature fusion network to obtain fused feature data. The multimodal feature fusion network includes an encoder layer, an attention mechanism layer and a fusion layer connected in sequence. The fused feature data is used to characterize the comprehensive features of the user's spinal rehabilitation status.

[0166] The rehabilitation assessment module 13 is used to perform status classification and risk prediction processing on the fused feature data through a preset rehabilitation status assessment model to obtain rehabilitation assessment data; wherein, the rehabilitation assessment data includes the current rehabilitation level, degree of functional impairment and rehabilitation risk index;

[0167] The initial training plan generation module 14 is used to perform personalized training plan matching processing on rehabilitation assessment data through a preset training plan generation network to obtain initial training plan data; wherein the training plan generation network includes a generator and a discriminator;

[0168] The rehabilitation training program generation module 15 is used to dynamically optimize and adaptively adjust the initial training program data to obtain a rehabilitation training program. The rehabilitation training program includes training movements, execution frequency, intensity gradient arrangement, and phased goal setting.

[0169] Furthermore, the data acquisition and feature module 11 is also used for:

[0170] S11. Collect spinal imaging data, motion posture data, electromyographic signal data, and clinical symptom text data of the target user;

[0171] S12. Extract morphological features of spinal curvature, intervertebral space height, and vertebral rotation angle from spinal imaging data to obtain imaging feature data;

[0172] S13. Extract the angular velocity and angle time sequence features of trunk flexion, extension, lateral bending and rotation from the motion posture data to obtain posture feature data;

[0173] S14. Extract the average power frequency and median frequency features of the erector spinae and multifidus muscles in a specific frequency band from the electromyographic signal data to obtain electromyographic feature data.

[0174] S15. Extract text feature data from clinical symptom text data;

[0175] S16. Based on image feature data, posture feature data, electromyography feature data, and text feature data, obtain multi-dimensional feature data.

[0176] Furthermore, the feature data fusion module 12 is also used for:

[0177] S21. Input the image feature data, posture feature data, electromyography feature data and text feature data into the encoder layer, and perform high-dimensional mapping through the corresponding fully connected neural network to obtain a high-dimensional feature vector of the same dimension.

[0178] S22. Input the high-dimensional feature vector into the attention mechanism layer and calculate the cross-modal attention weights;

[0179] S23. Based on cross-modal attention weights, the high-dimensional feature vectors of each modality are weighted and fused in the fusion layer to generate preliminary fused features;

[0180] S24. Perform layer normalization on the preliminary fusion features to obtain fusion feature data that characterizes the comprehensive features of the user's spinal rehabilitation status.

[0181] Furthermore, rehabilitation assessment module 13 is also used for:

[0182] S31. Input the fused feature data into the first classification network in the rehabilitation status assessment model, and output the current rehabilitation level of the target user;

[0183] S32. Input the fused feature data and the current rehabilitation level into the second regression network in the rehabilitation status assessment model, and output the functional impairment score of the target user.

[0184] S33. Input the fused feature data and functional impairment score into the risk prediction network of the rehabilitation status assessment model, and calculate the rehabilitation risk index using the following formula:

[0185]

[0186] In the formula, As a recovery risk index, To integrate risk-related feature subsets from feature data The risk coefficient obtained through training, This is for the operation of retrieving the maximum value;

[0187] S34. Integrate the current rehabilitation level, functional impairment score, and rehabilitation risk index into structured rehabilitation assessment data.

[0188] Furthermore, the initial training scheme generation module 14 is also used for:

[0189] S41. Input rehabilitation assessment data into the generator, which is the generative part of the conditional generative adversarial network, and outputs a set of candidate training schemes containing multiple training actions, execution frequency, intensity and target.

[0190] S42. Input each program in the candidate training program set along with the number of rehabilitation assessments into the discriminator. The discriminator outputs a quality score that characterizes the matching degree and rationality of the program.

[0191] S43. Based on the quality score, select the highest-scoring scheme from the candidate training scheme set as the initial training scheme data.

[0192] Furthermore, the rehabilitation training program generation module 15 is also used for:

[0193] S51. Decompose the initial training scheme data into a set of parameters that can be adjusted independently. The set of parameters that can be adjusted independently includes the training action sequence, the execution frequency of each action, the intensity gradient, and the setting of phased goals.

[0194] S52. When the target user performs the training action sequence, collect motion posture feedback data and electromyographic feedback data in real time, and input them into the multimodal feature fusion network and rehabilitation status assessment model to generate short-term rehabilitation status change data.

[0195] S53. Based on short-term rehabilitation status change data and rehabilitation risk index, the intensity gradient and execution frequency are dynamically adjusted through optimization algorithms to generate updated training scheme parameters.

[0196] S54. Based on the historical rehabilitation assessment data sequence of the target users, the phased goal setting is adaptively revised to obtain the revised phased goal setting;

[0197] S55. Integrate the updated training program parameters and revised phased goal settings to generate a rehabilitation training program.

[0198] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the deep learning-based spinal rehabilitation assisted training method as described above.

[0199] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0200] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0201] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A deep learning-based spinal rehabilitation assisted training method, characterized in that, The method includes: S1. Obtain multi-source heterogeneous spinal data of the target user, and extract multi-dimensional features from the multi-source heterogeneous data to obtain multi-dimensional feature data; wherein, the multi-source heterogeneous data includes spinal imaging data, motion posture data, electromyographic signal data and clinical symptom data; S2. Deep feature fusion processing is performed on multi-dimensional feature data through a preset multimodal feature fusion network to obtain fused feature data; wherein, the multimodal feature fusion network includes an encoder layer, an attention mechanism layer and a fusion layer connected in sequence, and the fused feature data is used to characterize the comprehensive features of the user's spinal rehabilitation status; S3. The fused feature data is processed by a preset rehabilitation status assessment model to perform status classification and risk prediction to obtain rehabilitation assessment data; wherein, the rehabilitation assessment data includes the current rehabilitation level, degree of functional impairment and rehabilitation risk index; S4. The rehabilitation assessment data is processed by a preset training scheme generation network to match personalized training schemes, thereby obtaining initial training scheme data; wherein the training scheme generation network includes a generator and a discriminator. S5. The initial training plan data is dynamically optimized and adaptively adjusted to obtain a rehabilitation training plan; the rehabilitation training plan includes training movements, execution frequency, intensity gradient arrangement and phased goal setting.

2. The method according to claim 1, characterized in that, The process involves acquiring multi-source heterogeneous spinal data of the target user and extracting multi-dimensional features from the multi-source heterogeneous data to obtain multi-dimensional feature data, including: S11. Collect spinal imaging data, motion posture data, electromyographic signal data, and clinical symptom text data of the target user; S12. Extract morphological features of spinal curvature, intervertebral space height, and vertebral rotation angle from the spinal imaging data to obtain imaging feature data; S13. Extract the angular velocity and angle time sequence features of the trunk flexion, extension, lateral bending and rotation from the motion posture data to obtain posture feature data; S14. Extract the average power frequency and median frequency features of the erector spinae and multifidus muscles in a specific frequency band from the electromyographic signal data to obtain electromyographic feature data. S15. Extract text feature data from the clinical symptom text data; S16. Based on the image feature data, posture feature data, electromyographic feature data, and the text feature data, the multi-dimensional feature data is obtained.

3. The method according to claim 2, characterized in that, The process involves performing deep feature fusion processing on multi-dimensional feature data using a pre-defined multimodal feature fusion network to obtain fused feature data, including: S21. Input the image feature data, posture feature data, electromyography feature data and text feature data into the encoder layer, and perform high-dimensional mapping through the corresponding fully connected neural network to obtain a high-dimensional feature vector of the same dimension. S22. Input the high-dimensional feature vector into the attention mechanism layer and calculate the cross-modal attention weights; S23. Based on the cross-modal attention weights, the high-dimensional feature vectors of each modality are weighted and fused in the fusion layer to generate preliminary fused features; S24. Perform layer normalization on the preliminary fusion features to obtain the fusion feature data used to characterize the comprehensive features of the user's spinal rehabilitation status.

4. The method according to claim 3, characterized in that, The rehabilitation assessment data is obtained by performing state classification and risk prediction processing on the fused feature data through a preset rehabilitation status assessment model, including: S31. Input the fused feature data into the first classification network of the rehabilitation status assessment model, and output the current rehabilitation level of the target user; S32. Input the fused feature data and the current rehabilitation level into the second regression network in the rehabilitation status assessment model, and output the functional impairment score of the target user. S33. Input the fused feature data and the functional impairment score into the risk prediction network of the rehabilitation status assessment model, and calculate the rehabilitation risk index using the following formula: In the formula, As a recovery risk index, To integrate risk-related feature subsets from feature data The risk coefficient obtained through training, This is for retrieving the maximum value. S34. Integrate the current rehabilitation level, the functional impairment score, and the rehabilitation risk index into structured rehabilitation assessment data.

5. The method according to claim 4, characterized in that, The process of generating personalized training plans by a pre-set training plan generation network to match the rehabilitation assessment data with individualized training plans to obtain initial training plan data includes: S41. Input the rehabilitation assessment data into the generator, which is the generation part in the conditional generative adversarial network, and outputs a set of candidate training schemes containing multiple training actions, execution frequencies, intensities and goals. S42. Input each of the candidate training schemes in the set of candidate training schemes and the number of rehabilitation assessments into the discriminator. The discriminator outputs a quality score to characterize the matching degree and rationality of the schemes. S43. Based on the quality score, select the scheme with the highest score from the set of candidate training schemes as the initial training scheme data.

6. The method according to claim 5, characterized in that, The process of dynamically optimizing and adaptively adjusting the initial training program data to obtain a rehabilitation training program includes: S51. The initial training scheme data is decomposed into a set of independently adjustable parameters, which includes the training action sequence, the execution frequency of each action, the intensity gradient, and the stage target setting. S52. When the target user performs the training action sequence, real-time motion posture feedback data and electromyographic feedback data are collected and input into the multimodal feature fusion network and the rehabilitation status assessment model to generate short-term rehabilitation status change data. S53. Based on the short-term rehabilitation status change data and the rehabilitation risk index, dynamically adjust the intensity gradient and execution frequency through an optimization algorithm to generate updated training scheme parameters. S54. Based on the historical rehabilitation assessment data sequence of the target user, the phased goal setting is adaptively revised to obtain the revised phased goal setting; S55. Integrate the updated training program parameters and the revised phased goal settings to generate the rehabilitation training program.

7. A deep learning-based spinal rehabilitation assistive training system, characterized in that, The system includes: The data acquisition and feature module is used to acquire multi-source heterogeneous spinal data of the target user and extract multi-dimensional features from the multi-source heterogeneous data to obtain multi-dimensional feature data; wherein, the multi-source heterogeneous data includes spinal imaging data, motion posture data, electromyographic signal data and clinical symptom data; The feature data fusion module is used to perform deep feature fusion processing on multi-dimensional feature data through a preset multimodal feature fusion network to obtain fused feature data; wherein, the multimodal feature fusion network includes an encoder layer, an attention mechanism layer and a fusion layer connected in sequence, and the fused feature data is used to characterize the comprehensive features of the user's spinal rehabilitation status; The rehabilitation assessment module is used to perform state classification and risk prediction processing on the fused feature data through a preset rehabilitation status assessment model to obtain rehabilitation assessment data; wherein, the rehabilitation assessment data includes the current rehabilitation level, degree of functional impairment, and rehabilitation risk index; The initial training plan generation module is used to perform personalized training plan matching processing on the rehabilitation assessment data through a preset training plan generation network to obtain initial training plan data; wherein the training plan generation network includes a generator and a discriminator. The rehabilitation training program generation module is used to dynamically optimize and adaptively adjust the initial training program data to obtain a rehabilitation training program; the rehabilitation training program includes training movements, execution frequency, intensity gradient arrangement and phased goal setting.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6.

Images