AI-based personalized treatment recommendation system for stroke rehabilitation

Abstract

This invention belongs to the fields of rehabilitation medicine and intelligent computing, specifically disclosing an artificial intelligence-based personalized treatment recommendation system for stroke rehabilitation. It includes a multimodal data acquisition module, a feature extraction and clustering modeling module, a personalized path formulation module, and a recommended path generation and feedback optimization module. This system enables dynamic monitoring and intelligent analysis of stroke patient rehabilitation data. By fusing multidimensional data such as images, biomarkers, clinical behavior, and cognitive emotions, a multimodal aggregated neural network model is constructed. Rehabilitation paths are formulated based on medical data clustering results, and the recommended paths are dynamically adjusted and optimized in conjunction with patient treatment feedback. Compared to traditional methods, this invention improves the scientific rigor and adaptability of personalized treatment, possessing advantages such as real-time updates of rehabilitation paths, closed-loop system management, and high model robustness. It significantly enhances the efficiency and accuracy of rehabilitation interventions, providing technical support for the intelligent development of rehabilitation medicine.

Description

Technical Field

[0001] This invention belongs to the technical field of information-based healthcare systems, specifically relating to an artificial intelligence-based personalized treatment recommendation system for stroke rehabilitation. Background Technology

[0002] Stroke is one of the leading causes of death and disability worldwide, characterized by high incidence, severe sequelae, and a long rehabilitation period. Even after acute treatment, most stroke patients still face varying degrees of motor impairment, speech difficulties, swallowing problems, and cognitive impairment, significantly impacting their quality of life. Rehabilitation therapy, as a crucial component of late-stage stroke intervention, runs throughout the entire recovery process and is of great significance for patients' functional recovery.

[0003] Currently, clinical stroke rehabilitation treatment plans are mostly formulated by rehabilitation physicians based on their experience in assessing patients' clinical manifestations, imaging examinations, and scale scores. This approach is inherently subjective and uncertain, making it difficult to fully explore the deep relationship between individual patient characteristics and rehabilitation potential. Furthermore, due to the highly individualized differences in patients' rehabilitation needs, traditional "template-based" intervention pathways struggle to cover the diverse recovery process. In addition, existing rehabilitation treatment processes generally suffer from the following problems: first, adjustments to rehabilitation plans are delayed, making it difficult to optimize plans in a timely manner based on dynamic changes in patients; second, the comprehensive utilization of objective information such as multidimensional physiological and behavioral data is insufficient; and third, there is a lack of data-driven precision recommendation mechanisms and treatment decision support systems.

[0004] With the continuous development of artificial intelligence (AI) technology, especially the widespread application of machine learning, deep learning, and knowledge graphs in the medical field, new ideas and tools have been provided for stroke rehabilitation. AI can mine potential patterns from massive amounts of patient rehabilitation data, identify key influencing factors in rehabilitation, build precise assessment models, and assist doctors in recommending treatment strategies. However, existing AI-based rehabilitation recommendation applications are still in their early stages. Most systems only have assessment functions, lack deep feedback mechanisms, lack reliable reasoning paths for treatment recommendations, and are difficult to integrate seamlessly with clinical workflows. Therefore, there is an urgent need to build a new system that integrates multi-source data modeling, individual feature mining, and intelligent intervention recommendations to improve the scientific rigor, real-time performance, and individual adaptability of rehabilitation decisions, and achieve intelligent, refined, and personalized upgrades in stroke rehabilitation treatment. Summary of the Invention

[0005] To address the above problems, the present invention aims to propose: a personalized treatment recommendation system for stroke rehabilitation based on artificial intelligence, comprising:

[0006] The multimodal data acquisition module is used to collect multimodal medical data from data sources of stroke patients, and to perform synchronization, alignment, truncation and normalization processing on the data to construct a standardized multimodal fusion dataset;

[0007] The feature extraction and clustering modeling module is used to extract multiple feature sets based on a standardized multimodal fusion dataset, perform high-dimensional feature clustering analysis, construct a multimodal aggregated neural network model (MA-MNN), and verify the clustering effect.

[0008] The personalized path development module is used to develop a personalized rehabilitation treatment baseline path that meets the actual rehabilitation needs of patients based on the feature clustering verification results.

[0009] The recommended path generation and feedback optimization module is used to collect multimodal monitoring data during the patient's rehabilitation process based on the personalized rehabilitation treatment baseline path, dynamically generate personalized rehabilitation treatment recommended paths, and continuously optimize the recommended paths based on real-time feedback.

[0010] Furthermore, the feature extraction and clustering modeling module specifically includes:

[0011] Construct a fusion feature extraction algorithm to extract multiple feature sets for multimodal medical data;

[0012] A multimodal clustering algorithm is used to perform cluster analysis on the extracted high-dimensional feature data, define the feature centers and inter-class distances, and formulate feature clustering judgment criteria.

[0013] The clustering criteria are used as the threshold standard for neural network training to measure the effectiveness of the clustering results.

[0014] Integrate multiple high-dimensional feature extraction models to construct a multimodal aggregated neural network model (MA-MNN) for feature extraction and modeling throughout the entire process;

[0015] The convergence of the model's clustering results is verified by a decoupling algorithm. If convergence is achieved, the clustering is valid; otherwise, retraining is performed.

[0016] Furthermore, the system also includes:

[0017] The data synchronization module is used to align the timelines of different data sources;

[0018] The data standardization module is used to standardize the format, units, and scale of multimodal data;

[0019] Edge computing units work in conjunction with cloud servers to support real-time data uploading, processing, and analysis;

[0020] The user interaction interface module is used to display patient rehabilitation data, treatment suggestion paths, and feedback results.

[0021] This invention also provides an application of an artificial intelligence-based personalized treatment recommendation system for stroke rehabilitation, the application comprising the following steps:

[0022] S1. Data Acquisition and Preprocessing: Collect multimodal objective data from data sources of stroke patients, and perform synchronization, alignment, truncation and normalization processing on the data to construct a standardized multimodal fusion dataset;

[0023] S2. Feature Extraction and Modeling: Based on the multimodal fusion dataset, a fusion feature extraction algorithm is used to cluster medical data, construct a multimodal aggregated neural network model (MA-MNN), and use the model to verify the clustering effect;

[0024] S3. Development of Personalized Rehabilitation Treatment Benchmark Pathways: Based on the cluster validation results, develop rehabilitation treatment benchmark pathways that conform to the actual situation of patients.

[0025] S4. Generation of Personalized Rehabilitation Treatment Recommendation Path: Referencing the benchmark path, formulate rehabilitation treatment evaluation indicators; collect multimodal monitoring data of patients during rehabilitation treatment, and compare and analyze it with the evaluation indicators to identify the treatment path segments that meet the standards, and construct personalized rehabilitation treatment recommendation paths;

[0026] S5. Feedback and optimization of personalized rehabilitation treatment recommendation pathways: Based on the recommendation pathways, specific treatment plans are developed, the patient's data performance during the rehabilitation process is continuously monitored, the adaptability of the recommendation pathways is verified, and the pathways are dynamically adjusted and integrated to achieve pathway optimization.

[0027] Furthermore, step S2 specifically includes:

[0028] S21. Construct a fusion feature extraction algorithm to extract multiple feature sets from multimodal medical data;

[0029] S22. Use a multimodal clustering algorithm to perform cluster analysis on the extracted high-dimensional feature data, define the distance between feature centers and classes, and formulate feature clustering judgment criteria.

[0030] S23. Based on the above clustering distance and judgment criteria, measure the effectiveness of feature classification;

[0031] S24. Use the clustering criteria as the threshold standard for neural network clustering analysis for model training and result classification.

[0032] S25. Integrate multiple high-dimensional feature extraction models to construct a complete aggregated neural network model as a tool for the entire process of feature extraction and modeling.

[0033] S26. Verify the clustering results of the model using a decoupling algorithm. If the algorithm converges, the clustering is effective. If it does not converge, return to the previous steps and retrain the clustering.

[0034] Furthermore, the multimodal medical data includes one or more of the following:

[0035] (1) Medical imaging data; including at least one of MRI, CT, or fMRI;

[0036] (2) Biomarker data; including at least one of the following: gene, protein, metabolomics, and blood sample information;

[0037] (3) Clinical data; including at least one of the following: medical history, physical examination, laboratory results, and medication records;

[0038] (4) Behavioral data; including at least one of the following: movement patterns, gait analysis, and rehabilitation records;

[0039] (5) Cognitive and affective data; including at least one of psychological assessment, emotional state and cognitive function assessment.

[0040] Furthermore, the multimodal aggregated neural network model (MA-MNN) consists of a backbone network and multiple branch networks. The backbone network includes an input layer, a feature aggregation layer, and an output layer, while the branch networks include an input layer, a feature extraction layer, and an output layer.

[0041] All data is first input into the backbone network, then distributed to each aggregation node via the feature aggregation layer and passed to the corresponding branch network for processing.

[0042] Furthermore, the downlink network of the MA-MNN model is a medical data clustering network that performs clustering processing of multimodal data to determine the key features of the next stage of treatment and issue control commands; after receiving the control commands, the branch network selectively truncates the connections with non-target data to form a minimized topology under specific features.

[0043] Furthermore, the medical data clustering network completes feature clustering of multimodal medical data based on a high-dimensional clustering model, and effectively represents different types of data through weighted combination and feature fusion.

[0044] Furthermore, the decoupling algorithm is a neural network fusion method based on a soft discriminant mechanism, which is used to improve the stability and accuracy of clustering determination.

[0045] Furthermore, the model clustering results are validated using decoupling algorithms, including:

[0046] S261. Use the MA-MNN model to perform cluster analysis on the sample data to form multiple cluster center vectors;

[0047] S262. Perform gradient convergence analysis on the cluster center vector of each cluster. If all vectors converge, it means that the clustering is effective.

[0048] S263. If there are cases of non-convergence, the clustering is deemed invalid.

[0049] S264. Perform secondary clustering on invalid clusters to obtain a new set of center vectors;

[0050] S265. If the secondary clustering still fails to converge, merge it with the original cluster centers to form a new set of feature vectors, and continue clustering analysis until the results converge. Repeat the above clustering verification steps.

[0051] Furthermore, the clustering verification uses a Gaussian mixture model based on the negative log-likelihood function as the fusion algorithm to determine the category affiliation.

[0052] Furthermore, after receiving the data output, the data is classified based on the Gaussian mixture model, and missing data is filled in by sample backfilling to form a corrected feature dataset. The downlink network then performs cluster analysis again to improve the robustness and recommendation accuracy of the model.

[0053] Furthermore, missing data in the feature dataset will be marked as error terms and processed separately from the complete feature dataset for subsequent error repair and analysis modeling.

[0054] Beneficial effects

[0055] Compared with existing technologies, the AI-based personalized treatment recommendation system for stroke rehabilitation provided by this invention has the following beneficial effects:

[0056] 1. Achieved deep fusion and accurate modeling of multimodal data: This system collects and processes medical data from multiple sources such as images, biomarkers, clinical behavior, and cognitive emotions. By using a fusion feature extraction algorithm to construct a unified high-dimensional feature data model, it effectively solves the problems of one-sidedness and poor adaptability of single data sources, and significantly improves the comprehensiveness and accuracy of rehabilitation assessment.

[0057] 2. Improved personalization and adaptability of rehabilitation pathway recommendations: By constructing a multimodal aggregated neural network model and combining dynamic clustering and feedback mechanisms, personalized treatment pathways can be continuously adjusted and optimized according to the actual rehabilitation progress of different patients. This overcomes the drawbacks of the traditional "fixed" rehabilitation pathways and significantly improves the scientific nature and pertinence of rehabilitation intervention.

[0058] 3. An intelligent feedback and adaptive optimization mechanism has been established: This invention sets up a closed-loop management mechanism from path recommendation to treatment feedback and then to path correction. Combining medical data clustering networks and decoupling algorithms, the effectiveness of the recommended path is evaluated in real time during system operation, and dynamic optimization is performed accordingly, making the rehabilitation path more controllable and iterative.

[0059] 4. Enhanced robustness and scalability of the model in complex medical environments: Through a multi-layered aggregation structure, feature selection mechanism, and backfilling reconstruction module, this system effectively addresses issues such as missing data, overlapping categories, and feature drift, improving the stability of the model in multiple scenarios and across different population groups, and possessing good potential for practical application.

[0060] In summary, this invention integrates artificial intelligence technology with medical data analysis capabilities to construct a multi-dimensional, feedback-enabled, and highly adaptive stroke rehabilitation recommendation system, which can significantly improve the efficiency and efficacy of rehabilitation treatment and provide a new path for precision rehabilitation medicine. Attached Figure Description

[0061] Figure 1 This is a schematic diagram illustrating the steps of the application method of the AI-based personalized treatment recommendation system for stroke rehabilitation of the present invention.

[0062] Figure 2 This is a schematic diagram of step S2 of the application method of the AI-based personalized treatment recommendation system for stroke rehabilitation of the present invention. Detailed Implementation

[0063] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0064] Example 1

[0065] according to Figure 1 As shown, this embodiment provides an artificial intelligence-based personalized treatment recommendation system for stroke rehabilitation, including:

[0066] The multimodal data acquisition module is used to collect multimodal medical data from data sources of stroke patients, and to perform synchronization, alignment, truncation and normalization processing on the data to construct a standardized multimodal fusion dataset;

[0067] The feature extraction and clustering modeling module is used to extract multiple feature sets based on a standardized multimodal fusion dataset, perform high-dimensional feature clustering analysis, construct a multimodal aggregated neural network model (MA-MNN), and verify the clustering effect.

[0068] The personalized path development module is used to develop a personalized rehabilitation treatment baseline path that meets the actual rehabilitation needs of patients based on the feature clustering verification results.

[0069] The recommended path generation and feedback optimization module is used to collect multimodal monitoring data during the patient's rehabilitation process based on the personalized rehabilitation treatment baseline path, dynamically generate personalized rehabilitation treatment recommended paths, and continuously optimize the recommended paths based on real-time feedback.

[0070] The application of the AI-based personalized stroke rehabilitation treatment recommendation system includes the following steps:

[0071] S1. Data Acquisition and Preprocessing: Collect multimodal objective data from data sources of stroke patients, and perform data synchronization, data alignment, data extraction and data normalization on the multimodal objective data to construct a multimodal fusion dataset;

[0072] S2. Feature Extraction and Modeling: Based on the multimodal fusion dataset, a fusion feature extraction algorithm is used to cluster medical data, a multimodal aggregated neural network model MA-MNN is established, and the multimodal aggregated neural network model MA-MNN is used to verify the clustering of medical data.

[0073] S3. Development of Personalized Rehabilitation Treatment Benchmark Path: Based on validated medical data clustering results, develop personalized rehabilitation treatment benchmark paths;

[0074] S4. Generation of Personalized Rehabilitation Treatment Recommendation Path: Based on the personalized rehabilitation treatment baseline path, rehabilitation treatment evaluation indicators are formulated; the rehabilitation treatment progress of stroke patients is monitored, and multimodal objective monitoring data of the rehabilitation treatment path is generated; the multimodal objective monitoring data is compared and analyzed with the corresponding rehabilitation treatment evaluation indicators to determine the treatment path segments that meet the rehabilitation treatment evaluation indicators, which are then used as personalized rehabilitation treatment recommendation paths.

[0075] S5. Feedback and optimization of personalized rehabilitation treatment recommendation pathways: Develop treatment plans based on personalized rehabilitation treatment recommendation pathways, monitor the patient's rehabilitation progress, verify the implementation of personalized rehabilitation treatment recommendation pathways, and dynamically adjust and integrate personalized rehabilitation treatment recommendation pathways.

[0076] like Figure 2 As shown, step S2 specifically includes:

[0077] S21. Construct a fusion feature extraction algorithm for multimodal medical data of stroke, wherein the fusion feature extraction algorithm is designed for high-dimensional feature data. Building a feature dataset ;

[0078] in, The index of the high-dimensional feature data. This represents the total number of high-order feature data. For the first High-dimensional feature centers of high-dimensional feature data For the first The number of high-dimensional features of a high-dimensional feature data set satisfies:

[0079]

[0080] in, For feature pattern number, For the serial number The feature number of the feature pattern is: There are feature patterns, and the total number of features for all feature patterns is . ;

[0081] S22, regarding the high-dimensional feature data Perform cluster analysis, construct a high-dimensional clustering function and a distance metric function, satisfying:

[0082]

[0083] in, Indicates the first All of the high-dimensional feature data A high-dimensional clustering function with a high number of high-dimensional features; Indicates the first The first high-dimensional feature data Class and First Distance metric function for high-dimensional feature centers of class features;

[0084] and The first The first high-dimensional feature data Class and First High-dimensional feature centers of class features;

[0085] Indicates the sequence number is The first feature pattern One feature; For the first The geometric center of each feature pattern; Indicates the sequence number is The total number of features in the feature pattern;

[0086] S23, Based on distance metric function Find the feature clustering decision function ,satisfy:

[0087]

[0088] in, Represents the mean function, This represents the variance function.

[0089] S24. The feature clustering determination function As a feature distance threshold function in the clustering algorithm, a neural network based on this high-dimensional feature data is completed. Construction and cluster analysis;

[0090] S25. Constructing multimodal medical data for all stroke cases. Feature dataset of high-dimensional feature data A total of A feature clustering decision function based on different high-dimensional feature data is proposed. The feature clustering decision functions of all high-dimensional feature data are used to construct an aggregate neural network as a feature extraction algorithm for multimodal medical data of stroke.

[0091] S26. Verification: The high-dimensional clustering function is verified using a decoupling algorithm. If the clustering results converge, the multimodal aggregated neural network model MA-MNN is determined to be a high-dimensional clustering model that meets the requirements. If the clustering results do not converge, return to step S22 and perform clustering again.

[0092] Example 2

[0093] This embodiment further defines the scope of embodiment 1: multimodal medical data includes one or more of the following data:

[0094] (1) Imaging data: including magnetic resonance imaging (MRI), computed tomography (CT) and functional MRI (fMRI), etc.;

[0095] (2) Biomarker data: including genomics, proteomics, metabolomics and blood biomarkers, etc.;

[0096] (3) Clinical data: including the patient’s medical history, physical examination results, laboratory test results and drug treatment records.

[0097] (4) Behavioral data: Patient's movement patterns, gait analysis, rehabilitation training records, etc.;

[0098] (5) Cognitive and affective data: including assessments of cognitive abilities, mental health status and emotional state.

[0099] The multimodal aggregated neural network model MA-MNN has a backbone network and A branch network, wherein the backbone network has an input layer, a feature aggregation layer, and an output layer. Each branch network has an input layer, a feature extraction layer, and an output layer; multimodal medical data of stroke is input to the input layer of the backbone network; and the feature aggregation layer is used to construct... There are 12 aggregation nodes, and each aggregation node corresponds to a branch network.

[0100] The downlink network of the multimodal aggregation neural network model MA-MNN is a medical data clustering network. It performs cluster analysis on multimodal medical data of stroke, determines the high-dimensional features of the next treatment stage, and sends extraction instructions to the uplink network. Based on the extraction instructions from the aggregation nodes, the feature extraction layer cuts off the connection with the unextracted medical data according to different high-dimensional features, and constructs a minimal topology adapted to the high-dimensional features.

[0101] Example 3

[0102] This embodiment further defines the features of embodiment 1: the medical data clustering network performs feature clustering on multimodal medical data of stroke based on a high-dimensional clustering model, wherein the high-dimensional clustering model satisfies:

[0103]

[0104] in, This is the output function of the high-dimensional clustering model; These are the weighting coefficients.

[0105] The decoupling algorithm is a neural network fusion method based on a soft discriminant mechanism, used to improve the stability and accuracy of clustering determination.

[0106] Validating the clustering results of the model using decoupling algorithms includes:

[0107] S261. Use the MA-MNN model to perform cluster analysis on the sample data to form multiple cluster center vectors;

[0108] S262. Perform gradient convergence analysis on the cluster center vector of each cluster. If all vectors converge, it means that the clustering is effective.

[0109] S263. If there are cases of non-convergence, the clustering is deemed invalid.

[0110] S264. Perform secondary clustering on invalid clusters to obtain a new set of center vectors;

[0111] S265. If the secondary clustering still fails to converge, merge it with the original cluster centers to form a new set of feature vectors, and continue clustering analysis until the results converge. Repeat the above clustering verification steps.

[0112] The clustering verification uses a Gaussian mixture model based on the negative log-likelihood function as the fusion algorithm to determine the category affiliation.

[0113] After receiving the data output, the data is classified based on the Gaussian mixture model. Missing data is filled in by sample backfilling to form a corrected feature dataset. The downlink network then performs cluster analysis again to improve the robustness and recommendation accuracy of the model.

[0114] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A personalized treatment recommendation system for stroke rehabilitation based on artificial intelligence, characterized in that, include: The multimodal data acquisition module is used to collect multimodal medical data from data sources of stroke patients, and to perform synchronization, alignment, truncation and normalization processing on the data to construct a standardized multimodal fusion dataset; The feature extraction and clustering modeling module is used to extract multiple sets of features based on a standardized multimodal fusion dataset, perform high-dimensional feature clustering analysis, construct a multimodal aggregation neural network model, and verify the clustering effect. The feature extraction and clustering modeling module specifically includes: Construct a fusion feature extraction algorithm to extract multiple feature sets for multimodal medical data; A multimodal clustering algorithm is used to perform cluster analysis on the extracted high-dimensional feature data, define the distance between feature centers and classes, and formulate feature clustering judgment criteria. The clustering criteria are used as the threshold standard for neural network training to measure the effectiveness of the clustering results. Integrate multiple high-dimensional feature extraction models to construct a multimodal aggregated neural network model for feature extraction and modeling throughout the entire process; The convergence of the model clustering results is verified by a decoupling algorithm. If convergence is achieved, the clustering is effective; if not, retraining is performed. The convergence verification of the model clustering results using decoupling algorithms includes: A multimodal aggregation neural network model is used to perform cluster analysis on the sample data, forming multiple cluster center vectors; Perform gradient convergence analysis on the cluster center vector of each cluster. If all vectors converge, the clustering is effective. If there are cases of non-convergence, the clustering is deemed invalid; Perform secondary clustering on invalid clusters to obtain a new set of center vectors; If the secondary clustering still fails to converge, it is merged with the original cluster centers to form a new set of feature vectors, and the clustering analysis continues until the result converges. The above verification steps are repeated. The downlink network of the multimodal aggregation neural network model is a medical data clustering network, which performs clustering processing of multimodal data to determine the key features of the next stage of treatment and issue control commands. After receiving the control commands, the branch networks selectively truncate the connections with non-target data to form a minimized topology under specific features. The medical data clustering network completes feature clustering of multimodal medical data based on a high-dimensional clustering model. Different types of data are effectively represented through weighted combination and feature fusion. The decoupling algorithm is a neural network fusion method based on a soft discriminant mechanism, which is used to improve the stability and accuracy of clustering determination. The personalized path development module is used to develop a personalized rehabilitation treatment baseline path that meets the actual rehabilitation needs of patients based on the feature clustering verification results. The recommended path generation and feedback optimization module is used to collect multimodal monitoring data during the patient's rehabilitation process based on the personalized rehabilitation treatment baseline path, dynamically generate personalized rehabilitation treatment recommended paths, and continuously optimize the recommended paths based on real-time feedback.

2. The personalized stroke rehabilitation treatment recommendation system based on artificial intelligence according to claim 1, characterized in that, The system also includes: The data synchronization module is used to align the timelines of different data sources; The data standardization module is used to standardize the format, units, and scale of multimodal data; Edge computing units work in conjunction with cloud servers to support real-time data uploading, processing, and analysis; The user interaction interface module is used to display patient rehabilitation data, treatment suggestion paths, and feedback results.

3. The personalized stroke rehabilitation treatment recommendation system based on artificial intelligence according to claim 2, characterized in that: The multimodal medical data includes: Medical imaging data: including at least one of MRI, CT, and fMRI; Biomarker data: including at least one of the following: genetics, proteins, metabolomics, and blood sample information; Clinical data: including at least one of the following: medical history, physical examination, laboratory results, and medication records; Behavioral data: including at least one of the following: movement patterns, gait analysis, and rehabilitation records; And cognitive and affective data: including at least one of psychological assessment, emotional state and cognitive function assessment.

4. The personalized stroke rehabilitation treatment recommendation system based on artificial intelligence according to claim 3, characterized in that: The Multimodal Aggregated Neural Network (MA-MNN) model consists of a backbone network and multiple branch networks. The backbone network includes an input layer, a feature aggregation layer, and an output layer, while the branch networks include an input layer, a feature extraction layer, and an output layer. All data is first input into the backbone network, distributed to each aggregation node via the feature aggregation layer, and then passed to the corresponding branch networks for processing.

Images