Personalized rehabilitation training plan generation system and method based on deep learning

Abstract

The present application provides a personalized rehabilitation training plan generation system and method based on deep learning, which can significantly improve the personalization and accuracy of rehabilitation training plans. Specifically, the rehabilitation training projects and rehabilitation training objects closely related to the rehabilitation plan generation request are first determined, and then a detailed rehabilitation training execution space is constructed based on the prior rehabilitation training execution events of these rehabilitation training objects for the rehabilitation training projects. The deep learning network is used to extract and restore the features of the rehabilitation training execution space, and the space is updated in combination with the thermal feature distribution to generate an updated execution space that reflects the execution probability of each project by the rehabilitation training object. Finally, based on the updated execution space, a personalized rehabilitation training project sequence is extracted for each rehabilitation training object, which not only improves the pertinence of the rehabilitation training plan, but also effectively improves the efficiency and effectiveness of rehabilitation training.

Description

Technical Field

[0001] The present application relates to the field of deep learning technology, and more specifically, to a system and method for generating a personalized rehabilitation training plan based on deep learning. Background Art

[0002] In the field of rehabilitation training, traditional rehabilitation training plans are often based on doctors' experience or general rehabilitation training guidelines, lacking personalized customization for individual rehabilitation needs. With the explosive growth of medical data and the rapid development of artificial intelligence technology, especially the widespread application of deep learning technology, new possibilities have been provided for the personalized generation of rehabilitation training plans.

[0003] However, existing methods for generating rehabilitation training plans based on machine learning mostly focus on simple statistical analysis of historical data, ignoring the dynamic changes and individual differences during rehabilitation training. In addition, these methods have limitations in feature extraction and feature restoration, making it difficult to accurately capture the complex behaviors and potential needs of rehabilitation training subjects when performing different rehabilitation training programs.

[0004] In order to overcome the above shortcomings, there is an urgent need for a method that can comprehensively consider the individual characteristics of rehabilitation training subjects, historical training data, and the characteristics of rehabilitation training projects to automatically generate personalized rehabilitation training plans. Summary of the invention

[0005] In view of the above-mentioned problems, in combination with the first aspect of the present application, an embodiment of the present application provides a method for generating a personalized rehabilitation training plan based on deep learning, the method comprising: Determining a first number of rehabilitation training items to be screened and a second number of rehabilitation training subjects to be assigned in relation to the request for generating a rehabilitation plan; Based on the prior rehabilitation training execution events of the second number of the rehabilitation training subjects for the first number of the rehabilitation training items, generating a second number*first number of rehabilitation training execution spaces, wherein the member of the xth row and yth column in the rehabilitation training execution space represents whether the xth rehabilitation training subject has selected to perform a rehabilitation training action for the yth rehabilitation training item; Determine network loading data of a deep learning network according to the rehabilitation training execution space, and perform space update according to the distribution of thermal characteristics based on a feature extraction unit and a feature restoration unit of the deep learning network to generate a second number*first number of update execution spaces, wherein the member of the xth row and yth column in the update execution space represents the probability that the xth rehabilitation training subject selects to perform a rehabilitation training action for the yth rehabilitation training item; Based on the updated execution space, a sequence of rehabilitation training items corresponding to a second number of the rehabilitation training objects are extracted from the first number of rehabilitation training items, wherein for the x-th rehabilitation training object, the sequence of rehabilitation training items corresponding to the x-th rehabilitation training object is determined based on the probability value represented by the member in the x-th row in the updated execution space.

[0006] In a possible implementation of the first aspect, the method further includes: Based on the execution node sequence of the a priori rehabilitation training execution events of the second number of the rehabilitation training subjects for the first number of the rehabilitation training items, generating a second number*first number of execution node sequence spaces, wherein the member of the xth row and yth column in the execution node sequence space represents the sequence information of the target a priori rehabilitation training execution event in the execution node sequence when the xth rehabilitation training subject selects to execute the target a priori rehabilitation training execution event for the yth rehabilitation training item; The determining of network loading data of the deep learning network according to the rehabilitation training execution space includes: The network loading data of the deep learning network is determined according to the rehabilitation training execution space and the execution node sequence space.

[0007] In a possible implementation of the first aspect, the method further includes: Based on the network loading data, a first graph convolution vector space corresponding to the rehabilitation training execution space and a second graph convolution vector space corresponding to the execution node sequence space are generated using a graph convolution unit of the deep learning network, wherein the spatial parameters of the first graph convolution vector space and the second graph convolution vector space match; The first graph convolution vector space and the second graph convolution vector space are converged based on the nodes where the space members are located to generate a graph convolution vector space to be extracted for the feature extraction unit, wherein the members in the wth row and rth column of the graph convolution vector space to be extracted are obtained based on the members in the wth row and rth column of the first graph convolution vector space and the members in the wth row and rth column of the second graph convolution vector space.

[0008] In a possible implementation manner of the first aspect, the feature extraction unit and the feature restoration unit based on the deep learning network perform a spatial update based on the thermal feature distribution to generate a second number*first number of update execution spaces, including: Determine potential representation parameters corresponding to network loading data based on a feature extraction unit of the deep learning network; By performing feature transformation on the potential representation parameters, a potential representation vector for inputting into the feature restoration unit is generated; The feature restoration unit is used to perform spatial update on the potential characterization vector according to the thermal feature distribution to generate an update execution space of the second quantity*the first quantity.

[0009] In a possible implementation of the first aspect, the deep learning network training step includes: Based on the a priori rehabilitation training execution events corresponding to different a priori training periods, an example rehabilitation training execution space matching the a priori training period is obtained, wherein the member of the kth row and the hth column of the example rehabilitation training execution space represents whether the kth example rehabilitation training subject has selected to perform a rehabilitation training action for the hth example rehabilitation training item in the associated a priori training period; Determine network loading data of a basic deep learning network according to the example rehabilitation training execution space, and perform space update according to the thermal feature distribution based on a basic feature extraction unit and a basic feature restoration unit of the basic deep learning network to generate a temporary update execution space, wherein the temporary update execution space matches the space parameters of the example rehabilitation training execution space, and the member of the kth row and hth column in the temporary update execution space represents the probability that the kth example rehabilitation training subject selects to perform a rehabilitation training action for the hth example rehabilitation training item; A first training error value is generated based on the characteristic distance between the example rehabilitation training execution space and the temporary update execution space, and the weight and bias information of the basic deep learning network are updated according to the first training error value to generate the deep learning network.

[0010] In a possible implementation manner of the first aspect, in the process of the basic feature extraction unit and the basic feature restoration unit based on the basic deep learning network performing spatial updating according to the thermal feature distribution, the method further includes: Obtaining example potential representation parameters generated by the basic feature extraction unit; generating a second training error value based on a characteristic distance between the example potential representation parameter and a priori potential representation parameter, the priori potential representation parameter being determined using the example rehabilitation training execution space; The updating of the weight and bias information of the basic deep learning network according to the first training error value to generate the deep learning network includes: Update the weight and bias information of the basic deep learning network according to the first training error value and the second training error value to generate the deep learning network; The member values ​​in the example rehabilitation training execution space follow a standard thermal distribution, the example potential characterization parameters include an expected mean and a degree of dispersion, and the a priori potential characterization parameters include a priori mean and a priori degree of dispersion.

[0011] In a possible implementation manner of the first aspect, updating the weight and bias information of the basic deep learning network according to the first training error value and the second training error value to generate the deep learning network includes: generating a global training error value according to the first training error value and the second training error value, wherein a weighting coefficient of the second training error value is smaller than a weighting coefficient of the first training error value in the global training error value; The weight and bias information of the basic deep learning network are updated based on the global training error value to generate the deep learning network.

[0012] In a possible implementation of the first aspect, the method further includes: Based on the example execution node sequences of the a priori rehabilitation training execution events corresponding to the different a priori training periods, an example execution node sequence space matching the example rehabilitation training execution space is obtained, The member of the kth row and hth column of the example execution node sequence space represents the sequence information of the example priori rehabilitation training execution event in the example execution node sequence when the kth example rehabilitation training subject selects to execute the example priori rehabilitation training execution event for the hth example rehabilitation training item in the associated priori training period; The determining of network loading data of a basic deep learning network according to the example rehabilitation training execution space includes: The network loading data of the basic deep learning network is determined according to the example rehabilitation training execution space and the example execution node sequence space.

[0013] In a possible implementation of the first aspect, determining a first number of rehabilitation training items to be screened and a second number of rehabilitation training subjects to be assigned related to the rehabilitation plan generation request includes: Based on a request timestamp corresponding to the rehabilitation plan generation request, determining a target priori training period closest to the request timestamp; According to the a priori rehabilitation training execution events in the target a priori training period, a first number of the rehabilitation training items and a second number of the rehabilitation training subjects to be allocated are determined.

[0014] On the other hand, an embodiment of the present application also provides a personalized rehabilitation training plan generation system based on deep learning, including a processor and a machine-readable storage medium, the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0015] Based on the above aspects, the embodiments of the present application can significantly improve the personalization and accuracy of the rehabilitation training plan. Specifically, the rehabilitation training projects and rehabilitation training objects closely related to the rehabilitation plan generation request are first determined, and then a detailed rehabilitation training execution space is constructed based on the prior rehabilitation training execution events of these rehabilitation training objects for the rehabilitation training projects. The deep learning network is used to extract and restore the features of the space, and the space is updated in combination with the thermal feature distribution to generate an updated execution space that reflects the execution probability of each project by the rehabilitation training object. Finally, based on the updated execution space, a personalized sequence of rehabilitation training projects is extracted for each rehabilitation training object, which not only improves the pertinence of the rehabilitation training plan, but also effectively improves the efficiency and effectiveness of rehabilitation training. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic diagram of the execution flow of the method for generating a personalized rehabilitation training plan based on deep learning provided in an embodiment of the present application.

[0017] Figure 2 This is a schematic diagram of the hardware architecture of a personalized rehabilitation training plan generation system based on deep learning provided in an embodiment of the present application. DETAILED DESCRIPTION

[0018] The present application will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of a method for generating a personalized rehabilitation training plan based on deep learning provided by an embodiment of the present application. The method for generating a personalized rehabilitation training plan based on deep learning is introduced in detail below.

[0019] Step S110 , determining a first number of rehabilitation training items to be screened and a second number of rehabilitation training subjects to be assigned that are related to the rehabilitation plan generation request.

[0020] In this embodiment, in this step, the server first needs to receive a rehabilitation plan generation request, which may come from a doctor, a rehabilitation therapist or the patient himself. The rehabilitation plan generation request contains some basic information, such as the patient's physical condition, rehabilitation goals, available rehabilitation resources, etc.

[0021] Based on this rehabilitation plan generation request, the server needs to determine two key pieces of information: the rehabilitation training items to be screened and the rehabilitation training subjects to be assigned, which is usually achieved by analyzing the timestamp in the rehabilitation plan generation request. The server will find the prior training period closest to the target timestamp, and the rehabilitation training execution events within this prior training period will be used as a reference.

[0022] For example, if the request to generate a rehabilitation plan is made on May 15, 2023, the server can find all rehabilitation training execution events from May 1, 2023 to May 14, 2023. The server then analyzes these rehabilitation training execution events to determine which rehabilitation training items are most commonly used and which rehabilitation training subjects (such as patients or patient groups) are most active.

[0023] The results of the hypothesis analysis showed that the five most commonly used rehabilitation training items were walking training, balance training, strength training, flexibility training and aerobic training. At the same time, the three most active rehabilitation training subjects were patient A, patient B and patient C.

[0024] Therefore, the server determines that the first number of rehabilitation training items to be screened are these five items, and the second number of rehabilitation training subjects to be assigned are these three patients.

[0025] Step S120, based on the prior rehabilitation training execution events of the second number of the rehabilitation training subjects for the first number of the rehabilitation training items, generate a second number*first number of rehabilitation training execution spaces, wherein the member in the xth row and yth column of the rehabilitation training execution space represents whether the xth rehabilitation training subject has chosen to perform the rehabilitation training action for the yth rehabilitation training item.

[0026] In this step, the server needs to use the prior rehabilitation training execution event to generate a rehabilitation training execution space, which is a two-dimensional matrix, whose number of rows is equal to the number of rehabilitation training objects (the second number), and whose number of columns is equal to the number of rehabilitation training items (the first number). Taking the previous example, the server will generate a matrix with 3 rows and 5 columns (because the second number is 3 and the first number is 5). Each element in the matrix represents whether a rehabilitation training object has chosen to perform a rehabilitation training action for a rehabilitation training item.

[0027] Specifically, if patient A chooses to perform walking training during the prior training period, the element in the first row and first column of the matrix will be set to 1 (or a value indicating "yes"), otherwise it will be set to 0 (or a value indicating "no"). Similarly, if patient B does not choose to perform balance training, the element in the second row and second column of the matrix will be set to 0.

[0028] In this way, the server can fill the entire rehabilitation training execution space so that it can reflect the rehabilitation training subjects' choices and preferences for rehabilitation training items.

[0029] Step S130, determining the network loading data of the deep learning network based on the rehabilitation training execution space, and performing a spatial update based on the thermal feature distribution based on the feature extraction unit and the feature restoration unit of the deep learning network to generate a second number * a first number of updated execution spaces, wherein the member of the xth row and yth column in the updated execution space represents the probability that the xth rehabilitation training subject chooses to perform the rehabilitation training action for the yth rehabilitation training item.

[0030] In this embodiment, in this step, the server needs to use the deep learning network to update the rehabilitation training execution space. First, the server determines the network loading data according to the rehabilitation training execution space, and the network loading data will be used to initialize the deep learning network.

[0031] Then, the server will use the feature extraction unit of the deep learning network to extract features in the rehabilitation training execution space. These features may include the rehabilitation training subjects' selection preferences, selection frequency, selection patterns, etc. for rehabilitation training items.

[0032] Next, the server will use the feature restoration unit to perform space update based on these features. The goal of the update is to generate a new execution space, that is, to update the execution space, which reflects the probability that the rehabilitation training subject chooses to perform rehabilitation training actions for the rehabilitation training items.

[0033] During the update process, the server considers the thermal feature distribution, which means that it adjusts the probability value in the update execution space according to the heat and frequency of the rehabilitation training items selected by the rehabilitation training subjects. For example, if a rehabilitation training subject often chooses to perform a rehabilitation training item, the corresponding probability value in the update execution space will be relatively high.

[0034] Finally, the server generates an updated execution space of the same size as the rehabilitation execution space, but the elements in it represent probability values ​​instead of binary values ​​of whether or not to select.

[0035] Step S140, based on the updated execution space, extracting rehabilitation training item sequences corresponding to the second number of the rehabilitation training objects from the first number of rehabilitation training items, wherein for the x-th rehabilitation training object, the rehabilitation training item sequence corresponding to the x-th rehabilitation training object is determined based on the probability value represented by the member in the x-th row in the updated execution space.

[0036] In this embodiment, in this step, the server needs to generate a rehabilitation training item sequence for each rehabilitation training object according to the updated execution space, and the rehabilitation training item sequence is determined according to the probability value represented by each row in the updated execution space.

[0037] Specifically, for each rehabilitation training object (i.e., each row in the update execution space), the server sorts the rehabilitation training items from high to low according to the probability value. Then, a rehabilitation training item sequence is generated based on this sorting.

[0038] For example, if the row corresponding to patient A in the update execution space shows that walking training has the highest probability, followed by strength training, then the sequence of rehabilitation training items for patient A will be: walking training, strength training, ... (other items are sorted by probability).

[0039] In this way, the server can generate a personalized sequence of rehabilitation training items for each rehabilitation training subject, which is customized according to their selection preferences and probabilities. Ultimately, these rehabilitation training item sequences can be used to develop personalized rehabilitation plans to meet the specific needs and goals of different rehabilitation training subjects.

[0040] Based on the above steps, the embodiment of the present application can significantly improve the personalization and accuracy of the rehabilitation training plan. Specifically, the rehabilitation training projects and rehabilitation training objects closely related to the rehabilitation plan generation request are first determined, and then a detailed rehabilitation training execution space is constructed based on the prior rehabilitation training execution events of these rehabilitation training objects for the rehabilitation training projects. The deep learning network is used to extract and restore the features of the space, and the space is updated in combination with the thermal feature distribution to generate an updated execution space that reflects the execution probability of each project by the rehabilitation training object. Finally, based on the updated execution space, a personalized sequence of rehabilitation training projects is extracted for each rehabilitation training object, which not only improves the pertinence of the rehabilitation training plan, but also effectively improves the efficiency and effectiveness of rehabilitation training.

[0041] In a possible implementation, the method further includes: Step A110, based on the execution node sequence of the a priori rehabilitation training execution events of the second number of the rehabilitation training subjects for the first number of the rehabilitation training items, generate a second number * first number of execution node sequence spaces, wherein the member of the xth row and yth column in the execution node sequence space represents the sequence information of the target a priori rehabilitation training execution event in the execution node sequence when the xth rehabilitation training subject selects to execute the target a priori rehabilitation training execution event for the yth rehabilitation training item.

[0042] Step S130 includes: determining the network loading data of the deep learning network according to the rehabilitation training execution space and the execution node sequence space.

[0043] In this embodiment, in this step, the server needs to further analyze the execution order of the rehabilitation training items by the rehabilitation training subject. For example, patient A may first perform walking training and then balance training. This execution order is very important for formulating a personalized rehabilitation plan.

[0044] Continuing with the previous example, the server will analyze all rehabilitation training execution events from May 1, 2023 to May 14, 2023, not only focusing on which items were executed, but also on their execution order. Then, the server will generate an execution node sequence space with 3 rows and 5 columns (because the second quantity is 3 and the first quantity is 5). In this space, each element represents the execution order of a rehabilitation training object for a rehabilitation training item.

[0045] For example, if patient A performed walking training first and then balance training during the prior training period, the element in the first row and first column of the execution node sequence space will be 1 (indicating that walking training was performed first), and the element in the first row and second column will be 2 (indicating that balance training was performed second). If patient B did not perform strength training, the element in the second row and third column will be set to a special value, indicating that the item was not performed.

[0046] Therefore, the server needs to combine the rehabilitation training execution space and the execution node sequence space to determine the network loading data of the deep learning network. These network loading data will contain rich information about the rehabilitation training subjects' selection preferences, selection frequency, selection patterns and execution order of rehabilitation training items.

[0047] Specifically, the server will integrate the data in the rehabilitation training execution space and the execution node sequence space to form a comprehensive data set, which will be used as the input of the deep learning network to initialize the network and guide its subsequent learning process. In this way, the deep learning network will be able to more accurately capture the rehabilitation training subjects' choices and preferences for rehabilitation training items, as well as their execution sequence habits.

[0048] That is, through the expanded approach, the server is able to more comprehensively analyze the behaviors and preferences of rehabilitation training subjects, including their choice of rehabilitation training items, selection frequency, selection mode, and execution order. This information is crucial for developing personalized rehabilitation plans because they can ensure that the plans meet the specific needs and goals of each rehabilitation training subject.

[0049] In a possible implementation, the method further includes: Step B110, based on the network loading data, using the graph convolution unit of the deep learning network to generate a first graph convolution vector space corresponding to the rehabilitation training execution space, and a second graph convolution vector space corresponding to the execution node sequence space, the spatial parameters of the first graph convolution vector space and the second graph convolution vector space match.

[0050] Step B120, converge the first graph convolution vector space and the second graph convolution vector space based on the nodes where the space members are located, and generate a graph convolution vector space to be extracted for the feature extraction unit, wherein the member in the wth row and rth column of the graph convolution vector space to be extracted is obtained based on the member in the wth row and rth column of the first graph convolution vector space and the member in the wth row and rth column of the second graph convolution vector space.

[0051] In this embodiment, the server first loads the necessary data sets from the storage system, which contain information about various rehabilitation training execution spaces and execution node sequence spaces used to train the deep learning network. For example, these data may include the patient's rehabilitation training action sequence, the execution time of each action, the execution quality, etc.

[0052] The server processes the loaded rehabilitation training execution space data through the graph convolution unit in the deep learning network. The graph convolution unit encodes each node in the data (representing a rehabilitation training action or state) and its relationship (such as the order or dependency between actions) to generate a high-dimensional vector space, the first graph convolution vector space, which can capture the complex patterns and features in the rehabilitation training execution process.

[0053] Similarly, the server also uses the graph convolution unit to process the data in the execution node sequence space to generate a second graph convolution vector space, which focuses on capturing the order and dependencies between execution nodes (i.e., individual actions or steps in rehabilitation training).

[0054] The server ensures that the two graph convolutional vector spaces have the same dimensions and parameter settings so that they can be effectively compared and fused in subsequent steps, which usually involves adjusting the dimensions of the two spaces to align them and ensure that their coordinate axes represent the same features or attributes.

[0055] The server converges the corresponding nodes in the first graph convolution vector space and the second graph convolution vector space, which means that for each node in the two spaces (representing a specific rehabilitation training action or state), the server combines their information to form a new node containing richer information. This process is achieved by performing some form of fusion (such as weighted averaging, splicing, etc.) on the members in the same position in the two spaces (i.e., members in the wth row and rth column).

[0056] In the process of generating the graph convolution vector space to be extracted, the server ensures that each member (i.e., each node) in the new space is calculated based on the members at the corresponding positions in the original two spaces. For example, if the server chooses to use weighted average as the fusion strategy, then the member in the wth row and rth column of the graph convolution vector space to be extracted will be the weighted average of the member in the wth row and rth column of the first graph convolution vector space and the member in the wth row and rth column of the second graph convolution vector space. In this way, the graph convolution vector space to be extracted contains rich information from the two original spaces, which can be used for subsequent feature extraction and optimization of rehabilitation training strategies.

[0057] In a possible implementation, step S130 includes: Step S131, based on the feature extraction unit of the deep learning network, determine the potential representation parameters corresponding to the network loading data.

[0058] Step S132: generating a latent representation vector for inputting into the feature restoration unit by performing feature transformation on the latent representation parameter.

[0059] Step S133, utilizing the feature restoration unit to perform spatial update on the potential representation vector according to the thermal feature distribution, and generating an update execution space of the second quantity*the first quantity.

[0060] In this embodiment, the server uses the feature extraction unit in the deep learning network to process the data previously loaded and processed by the graph convolution unit. The feature extraction unit is responsible for analyzing the data and extracting key potential representation parameters from it. These parameters are high-level, abstract representations of the data that can capture the core features and patterns in the data. For example, in the scenario of rehabilitation training, these potential representation parameters may represent the patient's rehabilitation progress, the fluency of action execution, or the quality of a specific action.

[0061] The server further processes and transforms the extracted latent representation parameters into a format suitable for input to the feature restoration unit, namely, a latent representation vector. This process may include dimensionality adjustment, normalization, or other preprocessing steps to ensure that the latent representation vector matches the input requirements of the feature restoration unit. For example, the server may map the latent representation parameters from a high-dimensional space to a low-dimensional space, or normalize them to better perform subsequent space updates.

[0062] The server uses the feature restoration unit to perform spatial updates on the latent representation vector. This process is based on the thermal feature distribution, which means that the server adjusts the position or distribution of the latent representation vector in space according to a certain thermal or energy model. This update can be to optimize a certain goal, such as improving the efficiency or effectiveness of rehabilitation training. Through spatial updates, the server generates a new execution space whose number is the second number multiplied by the first number. The new execution space may represent a more optimized rehabilitation training plan or strategy, which contains more diverse and personalized action sequences and execution plans. For example, in the context of rehabilitation training, this new execution space may contain customized training plans for different patients, different rehabilitation stages, and different training goals.

[0063] In a possible implementation, the step of training the deep learning network includes: Step S101, based on the prior rehabilitation training execution events corresponding to different prior training time periods, obtain an example rehabilitation training execution space matching the prior training time period, wherein the member of the kth row and the hth column of the example rehabilitation training execution space represents whether the kth example rehabilitation training subject has chosen to perform a rehabilitation training action for the hth example rehabilitation training item in the associated prior training time period.

[0064] Step S102, determining the network loading data of the basic deep learning network according to the example rehabilitation training execution space, and performing a spatial update based on the thermal feature distribution based on the basic feature extraction unit and the basic feature restoration unit of the basic deep learning network to generate a temporary update execution space, wherein the temporary update execution space matches the spatial parameters of the example rehabilitation training execution space, and the member of the kth row and the hth column in the temporary update execution space represents the probability that the kth example rehabilitation training object chooses to perform the rehabilitation training action for the hth example rehabilitation training item.

[0065] Step S103, generating a first training error value based on the characteristic distance between the example rehabilitation training execution space and the temporary update execution space, and updating the weight and bias information of the basic deep learning network according to the first training error value to generate the deep learning network.

[0066] In this embodiment, the server first accesses the storage system to retrieve data of a priori rehabilitation training execution events related to different a priori training time periods, which records whether the rehabilitation training subject has performed specific rehabilitation training actions in different time periods.

[0067] Based on these data, the server constructs an example rehabilitation training execution space, which is a two-dimensional matrix, in which the member of the kth row and hth column indicates whether the kth example rehabilitation training subject in the hth example rehabilitation training project has chosen to perform a rehabilitation training action within the associated prior training period. For example, if a subject performs an action within a certain period, the corresponding matrix member may be set to 1, otherwise it is 0.

[0068] Next, the server uses the data in the example rehabilitation training execution space as the input of the basic deep learning network, that is, the network loads the data. The server processes this data through the basic feature extraction unit of the basic deep learning network to extract potential representation parameters that capture the key features and patterns in the data. Then, the server uses the basic feature restoration unit to spatially update these potential representation parameters and adjust them according to the thermal feature distribution. This process may involve optimizing a certain goal, such as improving prediction accuracy.

[0069] Through the space update, the server generates a temporary update execution space, which matches the space parameters of the example rehabilitation training execution space. The member of the kth row and hth column in the temporary update execution space represents the probability that the kth example rehabilitation training subject chooses to perform the rehabilitation training action for the hth example rehabilitation training item.

[0070] The server calculates the characteristic distance between the example rehabilitation training execution space and the temporary update execution space, which generally involves comparing the differences between members at corresponding positions in the example rehabilitation training execution space and the temporary update execution space. Based on this characteristic distance, the server generates a first training error value that quantifies the accuracy of the current deep learning network prediction. The server uses this first training error value to update the weights and bias information of the basic deep learning network, which is an iterative process designed to reduce the prediction error and improve the performance of the network.

[0071] Through continuous training and weight updates, the server eventually generated a trained deep learning network that can more accurately predict the probability of rehabilitation training subjects performing actions in different items and time periods.

[0072] In a possible implementation manner, during the process of the basic feature extraction unit and the basic feature restoration unit based on the basic deep learning network performing spatial updating according to the thermal feature distribution, the method further includes: Step C110, obtaining example potential representation parameters generated by the basic feature extraction unit.

[0073] Step C120 , generating a second training error value based on a characteristic distance between the example potential representation parameter and a priori potential representation parameter, wherein the priori potential representation parameter is determined using the example rehabilitation training execution space.

[0074] Step S103 includes: updating the weight and bias information of the basic deep learning network according to the first training error value and the second training error value to generate the deep learning network.

[0075] The member values ​​in the example rehabilitation training execution space follow a standard thermal distribution, the example potential characterization parameters include an expected mean and a degree of dispersion, and the a priori potential characterization parameters include a priori mean and a priori degree of dispersion.

[0076] In this embodiment, during the space update process, the server first obtains example potential representation parameters from the basic feature extraction unit. These example potential representation parameters are obtained by processing and analyzing the data in the example rehabilitation training execution space, and they capture the key features and patterns in the data. The example potential representation parameters may include a series of values ​​that represent the representation of different features and attributes in the latent space.

[0077] The server then calculates a characteristic distance between the example potential representation parameters and the prior potential representation parameters. The prior potential representation parameters are a set of parameters determined in advance using data in the example rehabilitation training execution space, which represent expectations or prior knowledge of the potential representation before training. The characteristic distance may be obtained by calculating the Euclidean distance, cosine similarity, or other appropriate metric between the two sets of parameters. Based on this characteristic distance, the server generates a second training error value that quantifies the difference between the example potential representation parameters and the prior potential representation parameters.

[0078] During the training process, the server considers not only the characteristic distance between the example rehabilitation training execution space and the temporary update execution space (i.e., the first training error value), but also the characteristic distance between the example potential representation parameters and the prior potential representation parameters (i.e., the second training error value). The server uses these two training error values ​​to update the weights and bias information of the basic deep learning network, which is an iterative process that aims to reduce prediction errors and improve network performance by adjusting the parameters of the network. Through continuous training and weight updates, the server eventually generates a trained deep learning network that can more accurately predict the probability of rehabilitation training subjects performing actions in different items and time periods, and its potential representation is more consistent with prior knowledge and expected thermal distribution.

[0079] In this scenario, the member values ​​in the example rehabilitation training execution space are designed to follow a standard thermodynamic distribution, which means that each member value in the space has a corresponding probability density that follows the principles of thermodynamics. The example latent representation parameters contain a set of expected means and dispersions that describe the distribution of the latent representation in the space. Similarly, the prior latent representation parameters also contain a set of prior means and prior dispersions that represent the expected distribution of the latent representation before training. By comparing the example latent representation parameters and the prior latent representation parameters, the server can evaluate the performance of the network and make adjustments and optimizations as needed.

[0080] In a possible implementation, updating the weight and bias information of the basic deep learning network according to the first training error value and the second training error value to generate the deep learning network includes: Step S1031: Generate a global training error value according to the first training error value and the second training error value, in which a weighting coefficient of the second training error value is smaller than a weighting coefficient of the first training error value.

[0081] Step S1032: Update the weight and bias information of the basic deep learning network based on the global training error value to generate the deep learning network.

[0082] In this embodiment, during the training of the deep learning network, the server first calculates two training error values: a first training error value and a second training error value. The first training error value is generated based on the characteristic distance between the example rehabilitation training execution space and the temporary update execution space, which quantifies the accuracy of the network in predicting the selection of rehabilitation training actions. The second training error value is generated based on the characteristic distance between the example potential representation parameters and the prior potential representation parameters, which quantifies the accuracy of the network in capturing the potential representation.

[0083] In order to comprehensively consider the impact of these two training error values ​​on network performance, the server generates a global training error value, which is the weighted sum of the first training error value and the second training error value. In the weighting process, the server assigns a larger weighting coefficient to the first training error value and a smaller weighting coefficient to the second training error value, because the server believes that the accuracy of predicting the selection of rehabilitation training actions (i.e., the first training error value) is more important than the accuracy of capturing the potential representation (i.e., the second training error value).

[0084] Once the server calculates the global training error value, it can use this global training error value to update the weights and bias information of the basic deep learning network. This is an iterative process, and the server will repeatedly adjust the parameters of the network to minimize the global training error value. In each iteration, the server will calculate a new global training error value and update the parameters of the network based on this value. This process will continue until the global training error value reaches an acceptable level or reaches a predetermined number of iterations.

[0085] Through continuous training and weight updates, the server eventually generated a trained deep learning network that can more accurately predict the probability of rehabilitation training subjects performing actions in different projects and time periods, and its potential representation is more consistent with prior knowledge and expected thermal distribution. At the same time, because the server comprehensively considers the accuracy of prediction and the accuracy of potential representation during the training process, the generated network shows good performance in both aspects.

[0086] In a possible implementation, the method further includes: Based on the example execution node sequence of the prior rehabilitation training execution events corresponding to different prior training time periods, an example execution node sequence space matching the example rehabilitation training execution space is obtained, wherein the member of the kth row and hth column of the example execution node sequence space represents the sequence information of the example prior rehabilitation training execution event in the example execution node sequence when the kth example rehabilitation training subject selects to execute the example prior rehabilitation training execution event for the hth example rehabilitation training item in the associated prior training time period.

[0087] Step S102 includes: determining network loading data of the basic deep learning network according to the example rehabilitation training execution space and the example execution node sequence space.

[0088] In this embodiment, the server first accesses the storage system to retrieve data of a priori rehabilitation training execution events related to different a priori training periods, which not only records whether the rehabilitation training subject has performed specific rehabilitation training actions, but also records the execution order of these actions.

[0089] Based on these data, the server constructs an example execution node sequence space, which is a two-dimensional matrix that matches the example rehabilitation training execution space. Each member in the example execution node sequence space represents the position information of an action in the execution sequence when an example rehabilitation training subject performs a rehabilitation training action on an example rehabilitation training item during a prior training period.

[0090] After determining the sample rehabilitation training execution space and the sample execution node sequence space, the server combines the data of the sample rehabilitation training execution space and the sample execution node sequence space as the network loading data of the basic deep learning network. The network loading data not only contains information about whether the rehabilitation training subject has performed specific rehabilitation training actions, but also contains information about the execution order of these actions. Such data is more comprehensive and can provide richer training information for the deep learning network.

[0091] The server inputs these network loading data into the basic deep learning network for subsequent training and feature extraction operations.

[0092] Through the above steps, the server can obtain the example execution node sequence space containing the rehabilitation training action execution sequence information, and combine it with the data of the example rehabilitation training execution space as the network loading data of the basic deep learning network. Such data can provide more comprehensive training information for the deep learning network, which helps to improve the network's prediction accuracy and performance.

[0093] In a possible implementation, step S110 includes: Step S111 : determining a target priori training period closest to the request timestamp based on the request timestamp corresponding to the rehabilitation plan generation request.

[0094] Step S112, determining a first number of the rehabilitation training items and a second number of the rehabilitation training subjects to be allocated according to the a priori rehabilitation training execution events in the target a priori training period.

[0095] In this embodiment, the server receives a rehabilitation plan generation request, which includes a request timestamp indicating the specific time when the request occurs. The server accesses a storage system, which stores a large amount of prior rehabilitation training data, which are organized according to different training periods.

[0096] The server searches the storage system for the target a priori training period closest to the request timestamp, which usually involves comparing the request timestamp with the start and end timestamps of each a priori training period, and finding the one with the smallest time difference as the target a priori training period.

[0097] Once the target a priori training period is determined, the server accesses the data of a priori rehabilitation training execution events within the period. The server analyzes the data and extracts the rehabilitation training projects executed within the period and the rehabilitation training subjects participating in the projects. Based on the analysis results, the server determines a list containing a first number of rehabilitation training projects and a list containing a second number of rehabilitation training subjects to be assigned, and these rehabilitation training projects and subjects will serve as the basis for the subsequent generation of a rehabilitation plan.

[0098] Through the above steps, the server can find the closest prior training period according to the request timestamp in the rehabilitation plan generation request, and extract relevant rehabilitation training items and rehabilitation training objects from the period to provide basic data for subsequent rehabilitation plan generation.

[0099] Figure 2 The hardware structure of the personalized rehabilitation training plan generation system 100 based on deep learning provided in the embodiment of the present application for implementing the above-mentioned personalized rehabilitation training plan generation method based on deep learning is shown as follows: Figure 2 As shown, the deep learning-based personalized rehabilitation training plan generation system 100 may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0100] In one possible design, the personalized rehabilitation training plan generation system 100 based on deep learning can be a single server or a server group. The server group can be centralized or distributed (for example, the personalized rehabilitation training plan generation system 100 based on deep learning can be a distributed system). In some embodiments, the personalized rehabilitation training plan generation system 100 based on deep learning can be local or remote. For example, the personalized rehabilitation training plan generation system 100 based on deep learning can access information and / or data stored in the machine-readable storage medium 120 via a network. For another example, the personalized rehabilitation training plan generation system 100 based on deep learning can be directly connected to the machine-readable storage medium 120 to access the stored information and / or data. In some embodiments, the personalized rehabilitation training plan generation system 100 based on deep learning can be implemented on a personalized rehabilitation training plan generation system based on deep learning. As an example only, the personalized rehabilitation training plan generation system based on deep learning can include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, etc. or any combination thereof.

[0101] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data acquired from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions that the deep learning-based personalized rehabilitation training plan generation system 100 uses to execute or use to complete the exemplary method described in this application.

[0102] During the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the personalized rehabilitation training plan generation method based on deep learning in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0103] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned deep learning-based personalized rehabilitation training plan generation system 100. The implementation principles and technical effects are similar, and this embodiment will not be repeated here.

[0104] In addition, an embodiment of the present application also provides a readable storage medium, in which computer executable instructions are preset. When the processor executes the computer executable instructions, the above-mentioned method for generating a personalized rehabilitation training plan based on deep learning is implemented.

[0105] It should be noted that in order to simplify the description disclosed in the present application and thereby facilitate the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, various features are sometimes grouped into one embodiment, drawing or description thereof.

Claims

1. A method for generating a personalized rehabilitation training plan based on deep learning, characterized in that: The method comprises: Determining a first number of rehabilitation training items to be screened and a second number of rehabilitation training subjects to be assigned in relation to the request for generating a rehabilitation plan; Based on the prior rehabilitation training execution events of the second number of the rehabilitation training subjects for the first number of the rehabilitation training items, generating a second number*first number of rehabilitation training execution spaces, wherein the member of the xth row and yth column in the rehabilitation training execution space represents whether the xth rehabilitation training subject has selected to perform a rehabilitation training action for the yth rehabilitation training item; Determine network loading data of a deep learning network according to the rehabilitation training execution space, and perform space update according to the distribution of thermal characteristics based on a feature extraction unit and a feature restoration unit of the deep learning network to generate a second number*first number of update execution spaces, wherein the member of the xth row and yth column in the update execution space represents the probability that the xth rehabilitation training subject selects to perform a rehabilitation training action for the yth rehabilitation training item; Based on the updated execution space, a sequence of rehabilitation training items corresponding to a second number of the rehabilitation training objects are extracted from the first number of rehabilitation training items, wherein for the x-th rehabilitation training object, the sequence of rehabilitation training items corresponding to the x-th rehabilitation training object is determined based on the probability value represented by the member in the x-th row in the updated execution space.

2. The method for generating a personalized rehabilitation training plan based on deep learning according to claim 1, characterized in that: The method further comprises: Based on the execution node sequence of the a priori rehabilitation training execution events of the second number of the rehabilitation training subjects for the first number of the rehabilitation training items, generating a second number*first number of execution node sequence spaces, wherein the member of the xth row and yth column in the execution node sequence space represents the sequence information of the target a priori rehabilitation training execution event in the execution node sequence when the xth rehabilitation training subject selects to execute the target a priori rehabilitation training execution event for the yth rehabilitation training item; The determining of network loading data of the deep learning network according to the rehabilitation training execution space includes: The network loading data of the deep learning network is determined according to the rehabilitation training execution space and the execution node sequence space.

3. The method for generating a personalized rehabilitation training plan based on deep learning according to claim 2, characterized in that: The method further comprises: Based on the network loading data, a first graph convolution vector space corresponding to the rehabilitation training execution space and a second graph convolution vector space corresponding to the execution node sequence space are generated using a graph convolution unit of the deep learning network, wherein the spatial parameters of the first graph convolution vector space and the second graph convolution vector space match; The first graph convolution vector space and the second graph convolution vector space are converged based on the nodes where the space members are located to generate a graph convolution vector space to be extracted for the feature extraction unit, wherein the members in the wth row and rth column of the graph convolution vector space to be extracted are obtained based on the members in the wth row and rth column of the first graph convolution vector space and the members in the wth row and rth column of the second graph convolution vector space.

4. The method for generating a personalized rehabilitation training plan based on deep learning according to claim 1, characterized in that: The feature extraction unit and the feature restoration unit based on the deep learning network perform space update according to the thermal feature distribution to generate a second number*first number of update execution spaces, including: Determine potential representation parameters corresponding to network loading data based on a feature extraction unit of the deep learning network; By performing feature transformation on the potential representation parameters, a potential representation vector for inputting into the feature restoration unit is generated; The feature restoration unit is used to perform spatial update on the potential characterization vector according to the thermal feature distribution to generate an update execution space of the second quantity*the first quantity.

5. The method for generating a personalized rehabilitation training plan based on deep learning according to any one of claims 1 to 4, characterized in that: The training steps of the deep learning network include: Based on the a priori rehabilitation training execution events corresponding to different a priori training periods, an example rehabilitation training execution space matching the a priori training period is obtained, wherein the member of the kth row and the hth column of the example rehabilitation training execution space represents whether the kth example rehabilitation training subject has selected to perform a rehabilitation training action for the hth example rehabilitation training item in the associated a priori training period; Determine network loading data of a basic deep learning network according to the example rehabilitation training execution space, and perform space update according to the thermal feature distribution based on a basic feature extraction unit and a basic feature restoration unit of the basic deep learning network to generate a temporary update execution space, wherein the temporary update execution space matches the space parameters of the example rehabilitation training execution space, and the member of the kth row and hth column in the temporary update execution space represents the probability that the kth example rehabilitation training subject selects to perform a rehabilitation training action for the hth example rehabilitation training item; A first training error value is generated based on the characteristic distance between the example rehabilitation training execution space and the temporary update execution space, and the weight and bias information of the basic deep learning network are updated according to the first training error value to generate the deep learning network.

6. The method for generating a personalized rehabilitation training plan based on deep learning according to claim 5, characterized in that: In the process of performing spatial updating according to the thermal feature distribution by the basic feature extraction unit and the basic feature restoration unit based on the basic deep learning network, the method further includes: Obtaining example potential representation parameters generated by the basic feature extraction unit; generating a second training error value based on a characteristic distance between the example potential representation parameter and a priori potential representation parameter, the priori potential representation parameter being determined using the example rehabilitation training execution space; The updating of the weight and bias information of the basic deep learning network according to the first training error value to generate the deep learning network includes: Update the weight and bias information of the basic deep learning network according to the first training error value and the second training error value to generate the deep learning network; The member values ​​in the example rehabilitation training execution space follow a standard thermal distribution, the example potential characterization parameters include an expected mean and a degree of dispersion, and the a priori potential characterization parameters include a priori mean and a priori degree of dispersion.

7. The method for generating a personalized rehabilitation training plan based on deep learning according to claim 6, characterized in that: The updating of the weight and bias information of the basic deep learning network according to the first training error value and the second training error value to generate the deep learning network includes: generating a global training error value according to the first training error value and the second training error value, wherein a weighting coefficient of the second training error value is smaller than a weighting coefficient of the first training error value in the global training error value; The weight and bias information of the basic deep learning network are updated based on the global training error value to generate the deep learning network.

8. The method for generating a personalized rehabilitation training plan based on deep learning according to claim 5, characterized in that: The method further comprises: Based on the example execution node sequences of the a priori rehabilitation training execution events corresponding to the different a priori training periods, an example execution node sequence space matching the example rehabilitation training execution space is obtained, The member of the kth row and hth column of the example execution node sequence space represents the sequence information of the example priori rehabilitation training execution event in the example execution node sequence when the kth example rehabilitation training subject selects to execute the example priori rehabilitation training execution event for the hth example rehabilitation training item in the associated priori training period; The determining of network loading data of a basic deep learning network according to the example rehabilitation training execution space includes: The network loading data of the basic deep learning network is determined according to the example rehabilitation training execution space and the example execution node sequence space.

9. The method for generating a personalized rehabilitation training plan based on deep learning according to claim 1, characterized in that: The step of determining a first number of rehabilitation training items to be screened and a second number of rehabilitation training subjects to be assigned that are related to the rehabilitation plan generation request includes: Based on a request timestamp corresponding to the rehabilitation plan generation request, determining a target priori training period closest to the request timestamp; According to the a priori rehabilitation training execution events in the target a priori training period, a first number of the rehabilitation training items and a second number of the rehabilitation training subjects to be allocated are determined.

10. A system for generating personalized rehabilitation training plans based on deep learning, characterized in that: The deep learning-based personalized rehabilitation training plan generation system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the deep learning-based personalized rehabilitation training plan generation method described in any one of claims 1 to 9.

Images