Exercise training recommendation method and system for disability risk population

Abstract

The invention discloses an exercise training recommendation method and system for disability risk crowds. According to the method, a multi-level portrait system of a patient is constructed, multi-dimensional data such as disease types, clinical features, demographic information and motion behavior features of the patient are fused, role portraits are generated, and static features are obtained. After the first exercise training scheme is pushed based on the static features, structural data, text type data and image type data of the patient in the training period are collected, a subsequent exercise training scheme is generated through a collaborative filtering model, and iterative optimization is conducted on the dynamic features of the patient through a recurrent neural network model. The exercise training scheme can be dynamically adjusted according to the individual features and training feedback of the patient, and the compliance and the training effect of the patient are improved.

Description

Technical Field

[0001] The present invention relates to a method for recommending exercise training for people at risk of disability, and also relates to a corresponding exercise training recommendation system, belonging to the technical field of healthcare informatics. Background Art

[0002] According to the dual coding theory in cognitive theory, it is generally believed that humans have two main information processing systems: the verbal system and the non-verbal system. These two systems have different coding methods. Some studies have shown that the brain has better memory effect and memory speed for image materials than semantic memory, which indicates that the non-verbal system may have advantages in processing image and spatial information. Although the two systems are functionally independent, they can also activate and influence each other.

[0003] There is a theory called constructivism in the cognitive field, whose core is that knowledge is constructed through the interaction between individuals and the environment. Through communication and cooperation with others, individuals can construct and revise their understanding; when individuals encounter new information that conflicts with existing knowledge, cognitive conflicts will occur. This conflict prompts individuals to re-evaluate and adjust their cognitive structures to adapt to the new information. Moreover, knowledge construction is an evolving process. As individuals' experience increases and their cognitive abilities develop, their knowledge structures will also continuously evolve. Knowledge is constructed in specific situations, so learning in actual situations can better promote the understanding and application of knowledge.

[0004] However, in the current treatment of people at risk of disability, there is no exercise training method that integrates the information of the verbal system and the non-verbal system. Moreover, for people at risk of disability, individualized and targeted exercise training programs need to be developed to improve or alleviate the decline of their cognitive, emotional, motor, verbal and other abilities. Summary of the Invention

[0005] The primary technical problem to be solved by the present invention is to provide a method for recommending exercise training for people at risk of disability.

[0006] Another technical problem to be solved by the present invention is to provide a system for recommending exercise training for people at risk of disability.

[0007] To achieve the above technical objectives, the present invention adopts the following technical solutions:

[0008] According to the first aspect of the embodiments of the present invention, there is provided a method for recommending exercise training for people at risk of disability, including the following steps:

[0009] Step 1: Obtain the disease type, clinical characteristics, demographic information, exercise survey information, and exercise behavior characteristics of the patient to fuse and construct a role portrait and obtain static features; wherein, the patient is a population at risk of disability.

[0010] Step 2: Generate and push the first exercise training plan based on the role portrait.

[0011] Step 3: Collect structured data, text data, and image data during the exercise training.

[0012] Step 4: Input the data obtained in Step 3 into a collaborative filtering model, and combine it with the role portrait to generate and push the next exercise training plan.

[0013] Step 5: Collect structured data, text data, and image data during the exercise training in the same way as in Step 3 multiple times to obtain exercise training cycle data.

[0014] Step 6: Determine whether the exercise training plan is the Nth one. If so, go to Step 7; if not, return to Step 4 to generate the next exercise training plan according to the features in Step 5, where N is a positive integer.

[0015] Step 7: Based on the data collected from the patient in Steps 3 and 5, jointly input them into a recurrent neural network model to iteratively optimize the exercise training plan; wherein, two LSTM networks are introduced into the recurrent neural network model; at each time step, one of the LSTM networks combines the dynamic feature vector of the patient at the previous time step and the static feature vector at the current time step to update the dynamic feature vector of the patient; the other LSTM network combines the dynamic feature vector of the exercise at the previous time step and the static feature vector at the current time step to update the dynamic feature vector of the exercise.

[0016] Preferably, the recurrent neural network model is optimized using the following function:

[0017]

[0018] where θ represents the parameter to be learned. is the set of observed tuples in the training set, and the tuple includes the patient, exercise training plan, and time series; r ij|t represents the score of patient i for exercise j at time t. is the predicted value of r ij|t ; R represents the regularization function.

[0019] Preferably, the recurrent neural network model adopts an alternating subspace descent strategy. Assuming that the motion state is fixed, the gradient is not propagated to the motion training program sequence, and the parameters of the patient sequence are updated by backpropagating the gradients of all patient scores. Subsequently, it alternates between updating the patient sequence and updating the motion training program sequence.

[0020] Preferably, in step four, based on the structured data, the text data, and the image data, multi-modal data fusion is achieved to obtain a user latent vector and an item latent vector, which are then input into an attribute fusion convolutional layer to recommend the next motion training program.

[0021] Preferably, step four includes the following sub-steps:

[0022] 1) Determine key variables as one-dimensional data, two-dimensional data, and three-dimensional data;

[0023] 2) Create a data matrix according to the key variables;

[0024] 3) Add lag features as four-dimensional data;

[0025] 4) Use embedding technology to transform into feature vectors;

[0026] 5) Generate an item latent vector and a user latent vector based on the feature vectors for collaborative filtering learning to recommend a motion training program.

[0027] Preferably, in sub-step 4, the Bayesian TransR knowledge graph embedding method is used to transform entities and relationships into vectors; the Bayesian sparse autoencoder embedding method is used to transform text data into text vectors; and the Bayesian sparse convolutional autoencoder is used to extract the features of images and transform them into image vectors.

[0028] Preferably, the multi-modal features and role portrait features extracted for the patient's condition are used as independent variables, and the doctor's clinically prescribed motion prescription suggestions are used as dependent variables and input into the collaborative filtering model. The loss function of the collaborative filtering model is the cosine contrast loss, which is used to maximize the similarity between positive sample pairs and minimize the similarity between negative sample pairs under the margin constraint.

[0029] Preferably, the motion training recommendation method further includes:

[0030] Evaluate the training effect based on the data in the previous motion training program. If the effect is improved, produce and output a motion training report or directly end. If the effect is not improved, return to step seven to regenerate the motion training program.

[0031] According to the second aspect of the embodiments of the present invention, a motion training recommendation system for people at risk of disability is provided, including a processor and a memory. The memory is coupled to the processor and is used to store one or more programs. When the programs are executed by the processor, the processor implements the motion training recommendation method for people at risk of disability as described above.

[0032] Compared with the prior art, the present invention has the following technical advantages by formulating an individualized and targeted motion training plan:

[0033] 1) A motion recommendation framework that combines the advantages of multiple model algorithms is proposed. On the one hand, this framework combines collaborative filtering and autoencoder learning techniques to construct a fusion analysis system for multi-modal data, which can effectively integrate multi-source information into the recommendation model, thus enriching the analysis source and providing more comprehensive data support for motion recommendation. On the other hand, a recurrent neural network (RNN) is introduced to effectively calculate temporal data. By retrospectively analyzing the motion data of patients at different time periods, the coherence and sustainability of motion recommendations are ensured, thereby improving the compliance of patients.

[0034] 2) A training cycle data system is constructed. The motion training process is divided into annual cycles, large cycles, medium cycles, and small cycles, and corresponding time series data sets are generated for different cycles. When these data are introduced into the model for motion recommendation calculation, the temporal dimension features are increased. This enables the model to make more full and regular use of historical motion data, thereby assisting in generating periodic motion recommendation plans.

[0035] 3) A motion push system based on deep learning is constructed. This system integrates the multi-dimensional motion training push logic into a complete process, and can continuously iterate and update the motion push content based on the multi-dimensional motion data of the patient himself. In this way, it is ensured that the training plan always matches the current state of the patient individual, thereby improving the training effect.

[0036] 4) A multi-level patient portrait system is constructed. This system gradually deconstructs and evaluates the attribute characteristics, behavior characteristics, and comprehensive characteristics (including cognition, emotion, motion, speech, etc.) of patients, ensuring that the individual label system has a fine granularity. This provides a large amount of effective sequence information for subsequent input of patient data into the model for accurate motion recommendation, further improving the accuracy and personalization of the recommendation. Brief Description of the Drawings

[0037] Figure 1 It is a schematic flow chart of the motion training recommendation method for people at risk of disability in the first embodiment of the present invention;

[0038] Figure 2Schematic diagram of multimodal data fusion in the first embodiment of the present invention;

[0039] Figure 3 Schematic flow chart of generating a motion training plan based on a collaborative filtering model in the first embodiment of the present invention;

[0040] Figure 4 Schematic diagram of the structure of a recurrent neural network model in the first embodiment of the present invention;

[0041] Figure 5 Schematic diagram of the structure of a motion training recommendation system for people at risk of disability in the second embodiment of the present invention. Detailed implementation manners

[0042] The technical content of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] The technical concept of the present invention is: combining a content-based recommendation method with a collaborative filtering model, so as to integrate the feature learning of patients or items and the item recommendation process into a unified framework. First, use various deep learning models to learn the latent features of patients or items, and combine a collaborative filtering model to construct a unified optimization function for parameter training. Then, use the trained model to obtain the final latent vectors of patients and items, and further realize personalized motion training recommendations for patients.

[0044] First embodiment

[0045] As Figure 1 shown, the motion training recommendation method for people at risk of disability provided in the first embodiment of the present invention includes at least the following steps:

[0046] Step 1: Obtain the disease type, clinical features, demographic information, sports research information, and sports behavior characteristics of the patient to fuse and construct a role portrait and obtain static features; wherein, the patient is a person at risk of disability;

[0047] Using the basic model in the motion training recommendation for people at risk of disability, the core portrait can be constructed using the disease type and clinical features of the patient; the basic portrait can be constructed using the demographic information, sports research information, and sports behavior characteristics of the patient; then, by fusing the core portrait and the basic portrait, the role portrait of the patient can be constructed. In this process, static features are obtained.

[0048] It should be noted that for the construction method of the above basic model, reference can be made to the prior patent application with the application number 202411764001.3 and the title "Portrait construction method, motion training recommendation method and system for people at risk of disability", which is briefly described as follows:

[0049] The basic model is constructed based on the data of multiple patients through the following steps:

[0050] S1: Determine the key elements of image mapping according to the pathological mechanism;

[0051] S2: Construct the core image based on the key elements;

[0052] S3: Construct the basic image according to demographic information, exercise research information, and exercise behavior characteristics;

[0053] S4: Construct the role image based on the core image set and the basic image set.

[0054] Among them, the key elements include cognitive cortex damage, physical function damage, and emotional control damage; the attributes of the core image of disability diseases include cognitive cortex damage, physical function damage, and emotional control damage. Different attributes include different labels, and the label data comes from the patient's diagnosis information, research information, etc. The exercise research information includes exercise preferences, etc.

[0055] Such as Figure 2 shown, cognitive cortex damage can lead to cognitive impairment, manifested in aspects such as complex attention, executive function, learning and memory, and language. Physical function damage can lead to movement disorders, manifested as movement disorders in muscle strength, muscle tone, coordination ability, balance ability, etc. Emotional control damage can lead to mental and behavioral disorders, manifested as hallucinations / delusions or depression / anxiety in the psychological and mental aspects, and also manifested as aggression / hoarding or depression / wandering in the behavioral aspects. By evaluating the health status (disorders or diseases) of these patients, the labels of the patients can be mined.

[0056] The labels of type A cognitive cortex damage at least include:

[0057] A1. Frontal lobe damage: Patients may have manifestations of mental disorders such as memory loss and apathy, and may also have manifestations such as epilepsy, monoparesis, motor aphasia, high fever, excessive sweating, visual and olfactory abnormalities, etc.

[0058] A2. Parietal lobe damage: Patients may have contralateral limb complex sensory disorders, body image disorders, calculation and other functional disorders, etc.

[0059] A3. Temporal lobe damage: Patients may have sensory aphasia, nominal aphasia, olfactory hallucinations, epilepsy, mental abnormalities, memory disorders, visual field changes, etc.

[0060] A4. Occipital lobe damage: Patients may have visual disorders.

[0061] A5. Insular lobe damage: Patients may have disorders of visceral movement and sensation such as increased salivation, nausea, and fullness.

[0062] A6. Injury to the limbic lobe: Patients may have mental abnormalities such as emotional and memory disorders, hallucinations, and visceral activity disorders.

[0063] A7……

[0064] The labels for type B physical function injuries at least include:

[0065] B1. Headache

[0066] B2. Nausea or vomiting

[0067] B3. Fatigue or drowsiness

[0068] B4. Tremors, stiffness, and bradykinesia

[0069] B5. Dizziness or loss of balance

[0070] B6. Sensory problems, such as blurred vision, tinnitus, bad breath, or changes in smell

[0071] B7. Sensitivity to light or sound

[0072] B8……

[0073] The labels for type C emotional control injuries at least include:

[0074] C1. Emotional changes or mood swings

[0075] C2. Depression or anxiety

[0076] C3. Difficulty falling asleep

[0077] C4. Sleepier than usual

[0078] C5……

[0079] Quantify and calibrate the attribute features of the core portrait to construct the core portrait. The attributes of the core portrait can correspond one-to-one with the core key elements, or can be more than the core key elements. Multiple labels of each attribute constitute the label library, and different label libraries represent different disease courses, injury types, and disability characteristics. In this step, determine the core portrait labels of the patient according to the diagnosed disease type and clinical characteristics of the patient.

[0080] Obtain demographic information, exercise research information, exercise behavior characteristics, and social participation information through research or medical records, etc., and obtain the labels of the patient from them to construct the basic portrait.

[0081] Demographic information includes gender, age, education level, etc.; sports research information includes sports preferences and activity abilities, etc. Among them, sports preferences include sports types (such as rope skipping, running, etc.), sports venues (such as gyms, parks, etc.), sports atmosphere (quiet, dynamic), etc. Activity abilities include BADL (basic activities of daily living) and IADL (instrumental activities of daily living).

[0082] Sports behavior characteristics include sports frequency, sports cycle, etc.

[0083] Social participation information includes participation activities within the family and participation activities in society. Participation activities within the family include family decision-making, raising children, etc.; participation activities in society include work and study, cultural and entertainment activities, etc.

[0084] The basic portrait is not the core factor affecting the occurrence of disability, but as a basic feature, it has a synergistic influence on the information of the core portrait such as the patient's disease course, injury type, disability characteristics, etc. Therefore, it is included in the portrait system and participates in the analysis as a potential influencing variable.

[0085] The attributes of the basic portrait at least include:

[0086] Natural attributes: The gender attribute is a widely used label. People of different genders have obvious differences in their preferences for different contents; moreover, through natural attribute labels such as age, region, education level, occupation, marital status, and children status, it is relatively easy to analyze the basic proportion of the patient group;

[0087] Vertical attributes: Reflect the sports needs of patients, such as different types of sports preferences, etc.;

[0088] Training attributes: Training attributes are also an important attribute category, which helps to judge the patient's sports intention, sports cycle, and sports frequency;

[0089] Label attributes: When a patient starts using the system and generates the first piece of data, the system can assign it the first label - new user; afterwards, as the number of product patients accumulates, it is gradually possible to distinguish low-frequency patients, active patients, and high-frequency patients.

[0090] Quantify and calibrate the basic attribute characteristics to form a basic attribute portrait set.

[0091] Through the previous steps, a core portrait set and a basic portrait set have been constructed for each patient. Next, it is necessary to construct a role portrait based on the core portrait set and the basic portrait set.

[0092] The core portrait and the attribute portrait are in a parallel relationship. The core portrait represents the core characteristics of an individual, and the attribute portrait represents the basic characteristics of an individual. The role portrait has the basic attribute portrait as its one-dimensional structure and the core portrait as its two-dimensional structure. As time progresses, the characteristics of the core portrait at different time points (different stages of disease development) form a three-dimensional structure. In this step, weight differences are added to the portrait set for superposition (union) to form the role portrait of the individual.

[0093] Merge the portrait sets constructed based on different characteristics according to predetermined rules. For example, different weights are given according to the importance of different characteristics, and then weighted superposition is performed to merge the characteristics in these two portrait sets. Here, the weight of the characteristics in the core portrait is greater than the weight of the characteristics in the basic portrait. Specifically, the factors to be considered in weight allocation include the feature coverage of each portrait set, the consistency of data, the rationality of weight allocation, etc., to ensure that the superimposed portrait can accurately reflect the comprehensive characteristics of different patient groups and will not lose balance due to overemphasis on certain characteristics.

[0094] Take the basic portrait set as the coarse-grained characteristics of the individual and the core portrait as the fine-grained characteristics of the individual for merging and superposition. By merging characteristics of different granularities, the feature information of the user can be captured more comprehensively.

[0095] Coarse-grained characteristics refer to relatively broad and general characteristics that provide general information about an individual or object. These characteristics include some basic attributes such as age, gender, occupation, etc., which are very helpful for understanding the basic profile of an individual or patient. This feature is mainly applied to scenarios for quickly positioning the overall profile of an individual, serving as basic data classification feature labels. These characteristics help with preliminary screening and classification, providing directions for subsequent in-depth analysis.

[0096] Fine-grained characteristics, on the other hand, refer to more specific and detailed characteristics that provide deeper information about an individual or object. These characteristics include some specific behaviors, interests, habits, etc., which are very important for understanding the uniqueness and personalized needs of an individual or object. Fine-grained characteristics are applicable to fields that require in-depth understanding of the uniqueness of an individual or object, such as personalized recommendation systems, precise training, and behavioral feedback analysis. These characteristics can capture subtle differences and provide more accurate calculation and push support.

[0097] Therefore, coarse-grained characteristics provide general information about the user, while fine-grained characteristics provide more specific details. This combination of coarse and fine characteristics can improve the accuracy of portrait mapping; the fine-grained feature selection module can extract key local and fine-grained discriminative features in the object, while the coarse-grained feature selection module can obtain coarse-grained diverse features that provide context information, improving the discriminative power of the model; by combining characteristics of different granularities, the model can better generalize to different scenarios and conditions.

[0098] Specifically, the basic image set (coarse-grained features) and the core image set (fine-grained features) are merged and superimposed. In this process, different weights need to be assigned to different features. When different weights are given according to the importance of different features, the following are satisfied: 1) The feature weights in the core image are greater than the feature weights in the basic image; 2) The feature coverage, data consistency, weight assignment rationality, etc. of each feature in the core image; 3) The feature coverage, data consistency, weight assignment rationality, etc. of each feature in the basic image.

[0099] The core attribute features are quantified and calibrated one by one to form a core attribute image set.

[0100] The role image is a three-dimensional mapping system that combines the features of the basic image and the core image. The attributes of the core image at least include cognitive cortex damage, physical function damage, and emotional control damage. When obtaining the role image, a three-dimensional time slice based on the basic image and the core image needs to be constructed. This three-dimensional time slice is obtained through the following steps:

[0101] First, stage division is carried out. According to the progression of different types of disability diseases in the elderly, the disease course is divided into several stages, including the early stage, the middle stage, and the late stage. Each main stage can be further divided into more specific sub-stages to understand the development of the disease course more meticulously.

[0102] Secondly, feature extraction is carried out. In each stage, the key features of the patient are extracted, including cognitive function, emotional state, and motor ability, etc. These features can be measured through standardized assessment tools, such as the Mini-Mental State Examination (MMSE), Montreal Cognitive Assessment (MoCA), and Berg Balance Scale, etc.

[0103] Finally, the construction of the time axis is carried out. The time axis starting from the discovery of the disease (such as every month, every 2 months, every year, etc.) is used to represent the development of the disease course over time. On the time axis (as the time dimension), the starting point and ending point of each stage, as well as the key features of each stage, can be marked. The time axis can be linear or non-linear to reflect the complexity and uncertainty of the disease course. Such a time axis helps to more intuitively understand the progression of the disease course and provides guidance for treatment and intervention.

[0104] According to industry consensus, quantify the time series dimension features in the role portrait for portrait fusion. Incorporate knowledge graph information such as clinicians' experience, industry diagnostic criteria, and domain expert consensus. Use the disease development stage as the time dimension division node to associate different disease development periods, quantified clinical features, and other dimension information in the role portrait. Taking Alzheimer's disease as an example, according to industry consensus, the disease development can be divided into the early stage (the stage of asymptomatic cerebral amyloidosis: SCD, MCI stage), the middle stage (the stage of amyloid - positive + synaptic dysfunction and / or neurodegeneration), and the late stage (amyloid - positive + evidence of neurodegeneration + severe cognitive decline). For different disease development periods, the corresponding exercise training programs are different. The content of the exercise training program is as Figure 2 shown, including cognitive function rehabilitation, motor function rehabilitation, rehabilitation of mental and behavioral symptoms, rehabilitation of activities and participation, comprehensive rehabilitation, etc.

[0105] Associate different disease development periods, quantified clinical features, and other dimension information in the role portrait as the multi - dimensional portrait feature input for the pushed exercise training program, preparing for the precise push of exercise training.

[0106] Step 2: Generate and push the first exercise training program based on the role portrait.

[0107] Automatically generate the first exercise training program based on the role portrait. The first exercise training program can be generated according to a pre - trained model (such as the collaborative filtering model mentioned later) or designed according to the doctor's experience.

[0108] Step 3: Collect structured data, text - type data, and image - type data during exercise training.

[0109] Collect the structured data, text - type data, and image - type data of the patient performing exercises according to the first exercise training program. Use the structured data, text - type data, and image - type data (imaging data or image - type data) as the independent variables input to the collaborative filtering model (deep learning model). Specifically, it includes but is not limited to:

[0110] ■ Structured data: exercise diagnosis - type, life ability - type assessment scales, demographic characteristic data, etc.;

[0111] Text - type data: clinical diagnoses and case report - type data in the past two years, etc. (including the patient's exercise diagnosis data obtained through cases, test reports, etc., such as upper limb motor function impairment);

[0112] ■ Image - type data: magnetic resonance data, CT imaging data, etc.;

[0113] Step 4: Input the data obtained in Step 3 into the collaborative filtering model, and combine it with the role portrait to generate and push the next exercise training plan.

[0114] On the one hand, multiple autoencoders and collaborative filtering networks are adopted in the collaborative filtering model. The graph embedding technology is used to map the entities (such as patients, exercise types, physiological indicators, etc.) and relationships in the graph structure into a low-dimensional vector space. This step is realized through deep learning embedding technologies such as the Bayesian embedding model (TransR), the Bayesian stacked denoising autoencoder (SDAE), and the Bayesian stacked convolutional autoencoder (SCAE), which respectively transform structured data, text data, and image data into structure vectors, text vectors, and image vectors.

[0115] Specifically, the autoencoder is a layer-by-layer unsupervised learning model, mainly including two processes: decoding and encoding, which are used to process high-dimensional data and serve as a tool for feature extraction, enabling the collaborative filtering model to learn the low-dimensional representations (latent vectors) of users and items. These latent vectors can capture the internal connections between users and items, thereby improving the performance of the recommendation system. Here, the autoencoder fuses three types of information (structured data, text data, and image data) to achieve the fusion analysis of multi-modal information and obtain item latent vectors. Moreover, based on the role portrait data, the autoencoder uses feature extraction to obtain user latent vectors.

[0116] On the other hand, to fully learn the auxiliary information features, the collaborative filtering model respectively uses the Bayesian stacked denoising autoencoder (SDAE) to learn the vector representation of text information, the Bayesian stacked convolutional autoencoder (SCAE) to learn the vector representation of image information, and the Bayesian embedding model (TransR) to learn the vector representation of structural information. SDAE is a deep learning model that extracts high-level features of data through multi-layer unsupervised learning and can capture semantic and structural information in text, which is particularly useful for dealing with noise and uncertainty in text data. SCAE is used to learn the vector representation of image information because convolutional neural networks (CNNs) have advantages in processing image data and can capture the spatial hierarchical structure and local features of images. By stacking multiple convolutional layers and pooling layers, SCAE can learn more abstract and high-level image features, which helps the model make more robust predictions when facing the diversity and complexity of image data. TransR is used to learn structural information because it is an embedding model specifically designed for knowledge graphs and can handle complex relationships in knowledge graphs. By mapping entities and relationships to the same vector space, TransR can effectively capture the multi-dimensional relationships between entities and can handle incomplete information and noise in knowledge graphs. In addition, the flexibility and scalability of TransR enable it to adapt to different structured data, thereby improving the accuracy and robustness of the model when processing structured data. By converting text, images, and structured data into a unified vector representation, it can be ensured that data from different sources are comparable in the feature space.

[0117] After feature extraction, graph embedding technology is beneficial for constructing an individual data space, integrating multi-dimensional data such as time, motion preferences, physiological and psychological characteristics, and post-training status together to form a four-dimensional data structure. Furthermore, it integrates the structure vector, text vector, and image vector to form an item latent vector. The item latent vector synthesizes the features of multiple data types and can represent the features of the item more comprehensively.

[0118] Specifically, as Figure 3 shown, using the collaborative filtering model to generate and push the next motion training plan includes the following sub-steps:

[0119] 1) Determine the key variables as one-dimensional data, two-dimensional data, three-dimensional data

[0120] First, it is necessary to determine which variables are important for evaluating and predicting the patient's recovery process. By combining cognitive tasks (such as memory and attention tasks) and motor tasks (such as walking and balance training), the patient's motor function and balance ability can be improved. Through exercise, the patient's cerebral blood flow and blood oxygen level are improved, dormant neurons are activated, the growth and connection of neurons are promoted, specific neural conduction patterns are formed, and neurotransmitter levels are regulated; while the improvement of cognitive function helps to improve the patient's motor skills and efficiency, enhance motivation and self-control ability (control ability over emotions, speech, etc.), and optimize motor strategies.

[0121] In summary, for patients with different diseases (different damaged brain regions), different key independent variables need to be determined.

[0122] For all patients with cognitive diseases, the important common key independent variables include:

[0123] Time (T)

[0124] Type of Exercise

[0125] Intensity

[0126] Duration

[0127] Frequency

[0128] Patient Status

[0129] Physiological Indicators (such as heart rate, blood pressure, etc.)

[0130] Psychological Indicators (such as anxiety level, depression level, etc.)

[0131] Recovery Progress

[0132] For patients with different diseases, different key variables are determined, for example, including:

[0133] Patients with Alzheimer's disease:

[0134] ROI indicators (Region of Interest) of magnetic resonance imaging for brain injury, hippocampus, medial temporal lobe, parietal lobe, and frontal lobe

[0135] Cognitive Ability Scores, such as memory decline, spatial disorientation, inattention, and executive dysfunction, etc.

[0136] Parkinson's disease patients:

[0137] Brain injury MRI ROI indicators (Region of Interest) Substantia nigra, striatum, prefrontal cortex and parietal lobe

[0138] Activity of Daily Living Indicators Motor disorders, cognitive decline, sensory abnormalities, etc.

[0139] Hormone Level Score Dopamine level, etc.

[0140] 2) Create a data matrix based on the key variables

[0141] Create a data matrix according to the key variables of each patient, where each row represents the data record of a time point and each column represents a variable. For example:

[0142] One-dimensional data: [Time (T)]

[0143] Two-dimensional data (exercise preference): [Type of Exercise

[0144] Intensity

[0145] Duration

[0146] Frequency]

[0147] Three-dimensional data (physiological and psychological characteristics): [Heart Rate

[0148] Blood Pressure

[0149] Anxiety Level

[0150] Depression Level]

[0151]

[0152] 3) Add lag features

[0153] To capture the dependencies in the time series, lag features can be added.

[0154] Four-dimensional: [Patient Status after training Recovery Progress]

[0155] For example:

[0156] Time Patient status after training Rehabilitation progress T1 (e.g., 10:00 AM, October 3, 2024) Good 5％ T2 (e.g., 3:00 PM, October 3, 2024) Good 5％ T3 (e.g., 10:00 AM, October 4, 2024) Good 5％ T4 (e.g., 3:00 PM, October 4, 2024) Improved 10％ ... ... ...

[0157] Based on the aforementioned one-dimensional data, two-dimensional data, three-dimensional data, and four-dimensional data, a four-dimensional data structure is formed. That is, the one-dimensional data is used as the W axis; the two-dimensional data is used as the Y axis, the three-dimensional data is used as the X axis, and the four-dimensional data is used as the Z axis to construct an individual data space. The one-dimensional data includes a time variable, and the two-dimensional data includes multiple variables describing motion preferences; the three-dimensional data includes variables describing physical and psychological characteristics; the four-dimensional data includes post-training state data.

[0158] 4) Use embedding technology to transform it into a feature vector.

[0159] Use the hidden layer of the deep learning model for internal processing and fusion, and convert the four-dimensional data into a feature vector representation through embedding technology.

[0160] As Figure 3 shown, through structured embedding, the structured data in the aforementioned individual data space of each patient is converted into a structure vector. For example, using the Bayesian TransR knowledge graph embedding method, entities and relationships can be converted into vectors, and these vectors can capture the complex relationships between entities.

[0161] Through text embedding, text-type data is converted into text vectors. For example, using the Bayesian sparse autoencoder can learn the low-dimensional representation of text data and capture the semantic information of the text.

[0162] Through image embedding, image-type data is converted into image vectors. For example, using the Bayesian sparse convolutional autoencoder can extract the features of the image and convert them into vectors.

[0163] Using the aforementioned Bayesian TransR knowledge graph embedding method, Bayesian sparse autoencoder, and Bayesian sparse convolutional autoencoder, the structure vector, text vector, and image vector are integrated into a project latent vector. This project latent vector is a vector that synthesizes the features of multiple data types and can more comprehensively represent the characteristics of different types of patients and is a vector that can represent the project characteristics.

[0164] 5) Based on the feature vector, generate a project latent vector and a user latent vector for collaborative filtering learning to recommend exercise training programs.

[0165] Use the Attribute Fusion Graph Convolutional Network (AF-GCN) to fuse the feature vectors from different modal data, and use the attention mechanism to weight the importance of different modal data, and gradually layer the feature vectors into latent vectors.

[0166] The embodiment of the present invention adopts a collaborative filtering model that combines graph structure and deep learning technology. The graph structure is used to represent the knowledge base, which includes entities (such as sports categories, user attributes, user behaviors) and their relationships. Therefore, the graph structure can capture the interaction data between users and items, learn user preferences and item features.

[0167] Taking the above-mentioned multi-modal features (item latent vectors) and role portrait features (user latent vectors) extracted according to the patient's condition as independent variables, and the first sports training plan in the previous step (for example, the sports prescription clinically prescribed by a doctor) as the dependent variable, input them into the collaborative filtering model to generate the next sports training plan. The loss function of the collaborative filtering model is the Cosine Contrastive Loss, which is used to maximize the similarity between positive sample pairs and minimize the similarity between negative sample pairs under the margin constraint.

[0168] As described above, through graph embedding technology, each entity and relationship in the graph structure of the individual data space are mapped into a low-dimensional vector space, which enables them to be calculated and compared, so that the collaborative filtering model can analyze the interaction data between users and items, capture the complex non-linear relationship between patients and items, learn user preferences and item characteristics, and thus predict the items that users may be interested in.

[0169] It should be noted that the collaborative filtering model provided by the embodiment of the present invention itself has a cold start problem. That is, for new patients, due to the lack of their historical behavior data, it is impossible to calculate the similarity between patients according to the conventional collaborative filtering model, and it is difficult to provide personalized recommendations for them. This results in that when the collaborative filtering model just enters the hospital and starts to be used, it is very difficult for the model to provide accurate personalized recommendations for users. However, for structured data, image data, text data, etc., the collaborative filtering model can analyze the behaviors of other similar patients or similar items to predict the preferences of new users or new items. Only by using a small amount of data of sports training plans with clinical authoritative diagnoses, it can achieve the recommendation of fast sports training plans for people with similar characteristics (cold start function) without a large amount of training. This is because in the collaborative filtering model, based on the role portrait data and multi-modal data of new patients, it is very easy to find similar patients and similar items.

[0170] That is, the advantages of role portrait and multi-modal data fusion are exploited in this collaborative filtering model. For example, for patients with cognitive cortex damage, specific abnormal signals in certain brain regions will be shown in their MRI images, and similar manifestations of cognitive dysfunction will also be mentioned in clinical diagnosis reports. There are cases with relatively high feature similarity in multi-modal data related to cognitive cortex damage and physical function damage. Through the collaborative filtering model, the commonalities and regularities of different patients in key features can be mined. This model can effectively utilize the similarity features in multi-modal data, thereby improving the accuracy of feature similarity judgment. Another example is that although two patients are different in terms of emotional control damage, if they are highly similar in cognitive cortex damage and physical function damage, then when making exercise training recommendations, these two patients are more likely to be classified as similar patients, and then a suitable exercise plan can be recommended for new patients. Using the fine-grained core portrait enables the collaborative filtering model to accurately grasp the key needs of new patients and find similar patients and projects.

[0171] Based on the rich information in the role portrait, matching and screening are carried out from multiple levels to find a group of patients who are similar to the new patient in multiple dimensions, and then a more personalized exercise training plan that matches the comprehensive characteristics of the new patient can be recommended for the new patient, improving the personalization degree of the recommendation. Another example is that for patients with more severe cognitive cortex damage, exercise plans that focus more on cognitive training and brain function activation are recommended; while for patients with prominent physical function damage, training that focuses on the recovery and enhancement of limb motor function is emphasized. This personalized recommendation method can better solve the cold start problem compared with the traditional collaborative filtering model that only relies on historical behavior data, and can provide accurate and effective exercise training recommendations for new patients to help them improve their abilities in aspects such as cognition, emotion, movement, and speech.

[0172] In addition, in the process of fusing the patient status sequence data generated based on basic portraits such as medical history and demographic characteristics and core portraits such as examination results and doctor's suggestions with different modal data such as imaging data (image data) and scale data (text data), using the collaborative filtering model can more fully grasp the associations between different media modalities through multi-layer attribute fusion, improving the accuracy and generalization ability of exercise plan recommendations.

[0173] Step 5: Collect structured data, text data, image data, and clinical diagnosis data during exercise training multiple times in the way of Step 3 to obtain exercise training cycle data.

[0174] The patient exercises multiple times according to the exercise training plan, and accumulatively collects the data (structured data, text data, image data) in multiple exercises in the way of Step 3, and then the exercise training cycle data can be obtained.

[0175] The training cycle data in the sports training plan are generally divided into: multi-year cycle data, macro-cycle data, meso-cycle data, and micro-cycle data.

[0176] ■ Multi-year cycle: Due to the large individual differences among the elderly, specific problems need to be analyzed specifically. For example, (1) the physical conditions and recovery abilities of each elderly person are different, so the sports rehabilitation cycle will vary from person to person. Generally speaking, the elderly with better physical conditions and less disability may recover faster and can take a 2-year cycle. (2) The more severe the disability, the longer the recovery time usually is. The elderly who are completely bedridden take an 8-year cycle, and those who can take care of themselves partially take a 4-year cycle. During the rehabilitation process, it is necessary to continuously monitor the physical conditions and rehabilitation progress of the elderly and make adjustments as needed. This helps to ensure the effectiveness of the rehabilitation plan and shorten the rehabilitation cycle as much as possible.

[0177] ■ The macro-cycle generally takes this year or one year as the time limit to formulate a half-year or annual training plan. There are also cases where a year is divided into three macro-cycles, which is generally related to the classification of sports events.

[0178] ■ In training practice, the meso-cycle is called the training of phases and months, generally lasting 4 - 8 weeks, and a phase or monthly training plan is formulated based on this.

[0179] ■ In training practice, the micro-cycle is called the training of weeks, usually with a calendar week as the deadline to formulate a weekly training plan. It is also possible to arrange a micro-cycle of 4 to 10 days. For example, if a small recovery and adjustment can be completed in 4 days, a 4-day recovery micro-cycle can be arranged.

[0180] Step Six: Determine whether this sports training plan is the Nth one. If it is, go to Step Seven; if not, return to Step Four and generate the next sports training plan according to the characteristics in Step Five.

[0181] N is the preset number of times, such as 3 times. Through this step, the dynamic characteristics of the patient during training according to multiple sports training plans can be obtained, and the characteristics with a temporal dimension are obtained. Such a design is conducive to the Recurrent Residual Network learning more complex dynamic evolution representations.

[0182] Step Seven: Based on the data collected by the patient in Steps Three and Five, jointly input them into the recurrent neural network model to iteratively optimize the sports training plan.

[0183] In this step, use such as Figure 4The shown recurrent neural network model can better learn more complex dynamic evolution representations. In the recurrent neural network model, the historical interaction information between the patient and the movement is the key data that drives the changes in the patient's preferences and movement states. Therefore, by using a co-evolution model, the evolutionary latent representations of the patient and the movement can be captured. Here, using the training cycle vectors input into the recurrent neural network model, the recurrent neural network model uses the recurrent neural network to learn the dynamic feature representations of the patient and the movement, and obtains dynamic evolution representations with temporal relationships. The training cycle vector refers to the vector obtained by converting each attribute (such as time, frequency, intensity, etc.) in the training cycle data into numerical features and then through embedding or encoding techniques.

[0184] As Figure 4 shown, the recurrent neural network model uses two long short-term memory networks (LSTM) on the basis of the traditional recurrent neural network (RNN) to separately learn the temporal changes in the patient's movement preferences and the long-term (such as seasonal) evolution of the movement. In the figure, y i is learned through an LSTM network, which represents the dynamic feature vector of patient i at time series (moments t, t + 1, etc.); y j is learned through another LSTM network, which represents the dynamic feature vector of movement j at time series. Moreover, both y i and y j are affected by the static feature u i of the patient and the static feature m j of the movement. This is to consider the patient's long-term movement preferences and the static attributes of the movement, so that the recurrent neural network model can learn the static latent representations of the patient and the movement simultaneously.

[0185] In other words, at each time step, the recurrent neural network model uses an LSTM network to combine the dynamic feature vector of the patient at the previous time step and the current static feature vector to update the dynamic feature vector; it also uses another LSTM network to combine the dynamic feature vector of the movement at the previous time step and the current static feature vector to update the dynamic feature vector. Such a recurrent neural network model that combines static features and dynamic features can consider both the patient's long-term movement preferences and short-term changes, as well as the long-term and short-term changes in the movement training plan.

[0186] Specifically, the recurrent neural network model uses an LSTM-based recurrent neural network to model the dynamic changes of the patient and the movement training plan. For patient i and movement j (in a movement training plan, including multiple movements), assume that u i represents the static feature vector of patient i, and m j represents the static feature vector of movement j. At time t, assume that uit is the dynamic feature vector of patient i (i.e., Figure 4 y in i，t ), and m jt is the dynamic feature vector of movement j (i.e., Figure 4 y in j，t ). The dynamic features u i，t+1 and m j，t+1 at time t + 1 can be solved sequentially through an LSTM network as follows:

[0187] u i,t+1 = g(u it , {r ij|t})

[0188] m j，t+1 = h(m jt , {r ij|t})

[0189] where g and h are functions to be learned. g represents the function in the LSTM network for updating the dynamic features of the patient; h represents the function in the LSTM network for updating the dynamic features of the exercise training plan; r ij|t represents the score of patient i for movement j at time t. The dynamic feature u i,t+1 of the patient at time t + 1 is calculated through the LSTM network based on the dynamic feature u i,t of the patient at time t and the response r ij|t . The dynamic feature m j,t+1 of the movement at time t + 1 is calculated through the LSTM network based on the dynamic feature m j，t of the movement at time t and the response r ij|t . This indicates that the dynamic features of the patient and the exercise training plan are updated through the LSTM network according to their states at the previous time point and the current response. Therefore, the recurrent neural network model can capture the time-varying dynamic features, thereby better simulating and predicting the patient's response to different exercise training plans.

[0190] Inputting the dynamic and static features of patient i (u i，t+1 , m j，t+1 , u it , m jt ), the recurrent neural network model can predict the patient's current interest (score), and the predicted value of r ij is where f is also a function to be learned.

[0191] The loss function of the recurrent neural network model is the sum of the squared error loss function and the regularization term. Adjusting the parameter θ to minimize the loss function can achieve the overall optimization of the model. That is, the following function is used for optimization:

[0192]

[0193] As can be seen, the optimization objective of the recurrent neural network model is to make the pushed motion training plan more and more in line with the patient's current motion preference, motion level, and motion state, that is, to make the prediction close to the actual through the parameters generated by training. Among them, θ represents the parameters to be learned, is the set of (patient, motion training plan, time series) tuples observed in the training set, and R represents the regularization function.

[0194] Although the objective function and building blocks in the recurrent neural network model are very standard, the simple application of backpropagation cannot easily solve this optimization problem. The key challenge is that each patient's score depends on the patient state and the motion training plan. Backpropagation through two sequences is computationally prohibitive. This problem is alleviated by backpropagating the gradient from the patient's motion feedback. However, each score still depends on the patient state, which in turn acts on the entire sequence of the pushed motion training plan.

[0195] In an embodiment of the present invention, the recurrent neural network model adopts an alternating subspace descent strategy, where it is assumed that the motion state is fixed, the gradient is not propagated into the motion training plan sequence, and the parameters of the patient sequence are updated by backpropagating the gradients of all patient scores; subsequently, it alternates between updating the patient sequence and updating the motion training plan sequence. In this way, only one standard feedforward and backpropagation can be performed for the pushed motion training plan of each patient, and finally the optimization of the recurrent neural network model is achieved.

[0196] As can be seen from the above, the present invention uses a hybrid recommendation model (collaborative filtering model and recurrent neural network model) as the core tool for motion training recommendation. Compared with the traditional content recommendation model, the collaborative filtering model in the hybrid recommendation model can simultaneously learn the representations of multi-source data (structured data, text data, image data, training cycle data), and utilize multi-dimensional information such as structural information, content information, and image information. At the same time, the recurrent neural network in the hybrid recommendation model is used to capture the long-term preference evolution representation and short-term preference representation of the patient. Therefore, by integrating different models, it is possible to effectively evaluate the motion dose, motion state, and motion mode of a specific patient, and finally achieve the push of the optimal motion training plan.

[0197] During each training of the patient, the data collected during the previous motion training is used to output a new motion training plan in the manner of step seven. This is repeated until a predetermined goal is reached, such as the recovery or maintenance of abilities in aspects such as cognition, emotion, motion, and speech.

[0198] As an optional step, Step 8 can be added after Step 7: Evaluate the training effect based on the data in the previous exercise training plan. If the effect has improved, generate and output an exercise training report or directly end; if the effect has not improved, return to Step 7 to regenerate the exercise training plan. The method for evaluating the training effect can be evaluated by conventional methods, including being evaluated by a doctor or comparing and evaluating the data in the last exercise training plan with the data in the previous exercise training plan.

[0199] Compared with the prior art, the present invention has the following technical advantages by formulating an individualized and targeted exercise training plan:

[0200] 1) A motion recommendation framework that combines the advantages of multiple model algorithms is proposed. On the one hand, this framework combines collaborative filtering and autoencoder learning techniques to construct a fusion analysis system for multi-modal data, which can effectively integrate multi-source information into the recommendation model, thus enriching the analysis source and providing more comprehensive data support for motion recommendation. On the other hand, a recurrent neural network is introduced to effectively calculate temporal data. By retrospectively analyzing the motion data of patients at different time periods, the coherence and sustainability of motion recommendation are ensured, thereby improving the compliance of patients.

[0201] 2) A training cycle data system is constructed. The exercise training process is divided into annual cycles, large cycles, medium cycles, and small cycles, and corresponding time series data sets are generated for different cycles. When introducing this data into the model for motion recommendation calculation, the temporal dimension feature is added. This enables the model to make more full and regular use of historical motion data, thereby assisting in generating a periodic motion recommendation plan.

[0202] 3) A motion push system based on deep learning is constructed. This system integrates the multi-dimensional motion training push logic into a complete process, and can continuously iterate and update the motion push content based on the multi-dimensional motion data of the patient himself. In this way, it is ensured that the training plan always matches the current state of the patient individual, thereby improving the training effect.

[0203] 4) A multi-level patient portrait system is constructed. This system gradually deconstructs and evaluates the attribute characteristics, behavior characteristics, and comprehensive characteristics (including cognition, emotion, motion, speech, etc.) of patients, ensuring that the individual label system has a fine granularity. This provides a large amount of effective sequence information for subsequent inputting patient data into the model for accurate motion recommendation, further improving the accuracy and personalization of the recommendation.

[0204] Second Embodiment

[0205] Based on the above-mentioned method for recommending exercise training for people at risk of disability, the second embodiment of the present invention further provides a system for recommending exercise training for people at risk of disability. As Figure 5 shown, the exercise training recommendation system includes one or more processors and a memory. Among them, the memory is coupled to the processor and is used to store one or more programs. When the program is executed by the processor, the processor implements the method for recommending exercise training for people at risk of disability as described in the above embodiment.

[0206] Among them, the processor is used to control the overall operation of the exercise training recommendation system to complete all or part of the steps of the above-mentioned method for recommending exercise training for people at risk of disability. The processor can be a central processing unit (CPU), a graphics processing unit (GPU), a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), a digital signal processing (DSP) chip, etc. The memory is used to store various types of data to support the operation of the exercise training recommendation system. These data can include, for example, instructions for any application program or method operating on the exercise training recommendation system, as well as application program-related data. The memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, etc.

[0207] In another exemplary embodiment, the present invention also provides a computer-readable storage medium including program instructions. When the program instructions are executed by a processor, the steps of the method for recommending exercise training in any of the above embodiments are implemented. For example, the computer-readable storage medium can be the above-mentioned memory including program instructions. The above program instructions can be executed by the processor of the system to complete the method for recommending exercise training for people at risk of disability and achieve the same technical effects as the above method.

[0208] It should be noted that the above-mentioned multiple embodiments are only examples. The technical solutions of each embodiment can be combined, and the order of each step can be changed, all within the protection scope of this patent.

[0209] The above has described in detail the method and system for recommending exercise training for people at risk of disability provided by the present invention. For those of ordinary skill in the art, any obvious changes made without departing from the essence of the present invention will constitute an infringement of the patent right of the present invention and will bear corresponding legal responsibilities.

Claims

1. A method for recommending exercise training for people at risk of disability, characterized in that The steps include: Step 1: Obtain the patient's disease type and clinical characteristics, demographic information, exercise survey information, and exercise behavior characteristics to integrate and construct a role portrait to obtain static characteristics; wherein the patient is a person at risk of disability; Step 2: Generate and push the first sports training plan based on the role portrait; Step 3: Collect structured data, text data, and image data during sports training; Step 4: Input the data obtained in step 3 into the collaborative filtering model and combine it with the role portrait to generate and push the next sports training plan; Step 5: Collect structured data, text data, and image data during the sports training multiple times in the manner of step 3 to obtain sports training cycle data; Step 6: Determine whether the exercise training program is the Nth one. If yes, proceed to step 7; if not, return to step 4 and generate the next exercise training program according to the features in step 5, where N is a positive integer; Step seven: Based on the data collected from the patient in step three and step five, the data are jointly input into a recurrent neural network model to iteratively optimize the exercise training plan, wherein the recurrent neural network model introduces two LSTM networks based on the recurrent neural network; at each time step, one LSTM network combines the patient's dynamic feature vector of the previous time step and the static feature vector of the current time step to update the patient's dynamic feature vector; the other LSTM network combines the dynamic feature vector of the previous time step of the movement and the static feature vector of the current time step to update the dynamic feature vector of the movement.

2. The method for recommending exercise training for people at risk of disability according to claim 1, characterized in that The recurrent neural network model is optimized using the following function: Among them, θ represents the parameters to be learned; is the set of tuples observed in the training set, which includes patients, exercise training programs and time series; r ij|t represents the score of movement j by patient i at time t; For r ij|t The predicted value of ; R represents the regularization function.

3. The method for recommending exercise training for people at risk of disability according to claim 2, characterized in that: The recurrent neural network model adopts an alternating subspace descent strategy, in which the motion state is assumed to be fixed and the gradient is not propagated to the motion training scheme sequence. At the same time, the parameters of the patient sequence are updated by back-propagating the gradients of all patient scores; subsequently, switching is alternately performed between updating the patient sequence and updating the motion training scheme sequence.

4. The method for recommending exercise training for people at risk of disability according to claim 3, characterized in that: In the step 4, multimodal data fusion is implemented based on the structured data, the text data, and the image data to obtain user latent vectors and item latent vectors, which are input into the attribute fusion convolutional layer together to recommend the next sports training plan.

5. The method for recommending exercise training for people at risk of disability according to claim 4, characterized in that The step 4 includes the following sub-steps: 1) Determine the key variables as one-dimensional data, two-dimensional data, and three-dimensional data; 2) Create a data matrix based on key variables; 3) Add hysteresis features as four-dimensional data; 4) Convert it into feature vector using embedding technology; 5) Based on the feature vector, the item latent vector and the user latent vector are generated for collaborative filtering learning to recommend sports training plans.

6. The method for recommending exercise training for people at risk of disability according to claim 5, characterized in that In the sub-step 4, the entities and relationships are converted into vectors using the Bayesian TransR knowledge graph embedding method; the text data is converted into text vectors using the Bayesian sparse autoencoder embedding method; and the Bayesian sparse convolutional autoencoder is used to extract image features and convert them into image vectors.

7. The method for recommending exercise training for people at risk of disability according to claim 6, characterized in that: The multimodal features and role portrait features extracted from the patient's condition are used as independent variables, and the exercise prescription recommendations clinically issued by the doctor are used as dependent variables and are input into the collaborative filtering model.

8. The method for recommending exercise training for people at risk of disability according to claim 7, characterized in that: The loss function of the collaborative filtering model is cosine contrast loss, which is used to maximize the similarity between positive sample pairs and minimize the similarity between negative sample pairs under margin constraints.

9. The method for recommending exercise training for people at risk of disability according to claim 1, characterized in that The following steps are also included: Evaluate the training effect based on the data in the previous exercise training program. If the effect is improved, create and output an exercise training report or end it directly; if the effect is not improved, return to step seven to regenerate the exercise training program.

10. An exercise training recommendation system for people at risk of disability, characterized in that It includes one or more processors and memories, wherein the memory is coupled to the processor and is used to store one or more programs. When the programs are executed by the processor, the processor implements the exercise training recommendation method for people at risk of disability as described in any one of claims 1 to 9.

Images