Athletic injury rehabilitation scheme generation method and system based on artificial intelligence

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a sports injury rehabilitation scheme generation method and system based on artificial intelligence, and the method comprises the steps: marking a patient whose hospitalization rehabilitation data is most similar to the hospitalization rehabilitation data of a target patient as a reference patient, and obtaining the individual estimation difference of the target patient and the rehabilitation training difference of the target patient; taking the patient with the maximum similarity with the hospitalization rehabilitation data of the target patient and the minimum rehabilitation training difference as a reference patient again; updating the rehabilitation training difference of the target patient for multiple times to obtain a final training difference of the target patient; and taking the rehabilitation scheme of the patient with the maximum similarity with the hospitalization rehabilitation data of the target patient and the minimum final training difference as a generated rehabilitation scheme. The rehabilitation scheme suitable for the target patient is further generated according to the hospitalization rehabilitation data and the discharge rehabilitation data at the same time.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method and system for generating sports injury rehabilitation programs based on artificial intelligence. Background Technology

[0002] Intelligent generation of rehabilitation plans for sports injury patients after discharge is an important technical aspect of artificial intelligence. The conventional approach is to use a dataset composed of a large amount of historical patient rehabilitation data to train a generative or predictive neural network to generate or predict post-discharge rehabilitation plans.

[0003] However, both rehabilitation data collected during hospitalization and those collected upon discharge contain individual differences between patients. These differences introduce errors and interference into the data, meaning that rehabilitation plans for patients with the most similar data may not be suitable for the target patient. Furthermore, compared to data collected during hospitalization, data collected at discharge not only exhibits individual differences but also variations in the rehabilitation training process. For example, a lack of professional supervision and guidance during post-discharge rehabilitation training, differences in training equipment or their usage, or incomplete or erroneous data collection can lead to significant differences in discharge data, even if patients have the same rehabilitation training plan. Consequently, even for the target patient with the most similar data, the post-discharge rehabilitation training plan may not be applicable. This results in high noise levels in the dataset when directly using historical patient rehabilitation data to train neural networks, making it impossible to directly train a neural network that accurately generates rehabilitation training plans, or the generated plans may be unsuitable for the target patient.

[0004] In summary, neither rehabilitation data collected during hospitalization nor data collected during discharge can be used to generate a rehabilitation plan suitable for the target patient. Summary of the Invention

[0005] To address the above problems, this invention provides a method and system for generating sports injury rehabilitation programs based on artificial intelligence.

[0006] The method and system for generating sports injury rehabilitation programs based on artificial intelligence of the present invention adopts the following technical solution: One embodiment of the present invention provides a method for generating sports injury rehabilitation programs based on artificial intelligence, the method comprising the following steps: Record the inpatient rehabilitation data and discharge rehabilitation data of each patient. Any patient is recorded as the target patient. Use autoencoder technology to obtain the N1 patients whose inpatient rehabilitation data is most similar to the target patient's inpatient rehabilitation data and record them as reference patients. N1 is the first preset value. D1: Obtain the individual predicted difference A of the target patient. The difference between A and the inpatient rehabilitation data of the reference patient is positively correlated. The difference between the discharge rehabilitation data of the reference patient is denoted as B. The difference between B and A is denoted as the rehabilitation training difference of the target patient. D2: Obtain N2 patients with the greatest similarity to the target patient's inpatient rehabilitation data and the smallest difference in rehabilitation training, and use them again as reference patients. N2 is the second preset value, and N2>N1. D3: Update the rehabilitation training difference of the target patient by repeatedly executing D1 and D2. Each time the value of N2 is updated, it is greater than the value set by N2 in the previous update. The update stops when the ratio of the rehabilitation training difference of the target patient to N2 is the smallest. The obtained rehabilitation training difference of the target patient is recorded as the final training difference. The patient with the greatest similarity to the target patient's inpatient rehabilitation data and the smallest difference in final training is designated as the final reference patient. The sports injury rehabilitation program of the final reference patient after discharge is used as the rehabilitation program generated for the target patient.

[0007] Preferably, the specific steps for obtaining the similarity of the inpatient rehabilitation data are as follows: The inpatient rehabilitation data of each patient is transformed into a feature vector using autoencoder technology. This feature vector is denoted as the encoding vector of the inpatient rehabilitation data. The cosine similarity between the encoding vectors of any two patients is used as the similarity of the inpatient rehabilitation data.

[0008] Preferably, the specific steps for obtaining the individual predicted difference A of the target patient are as follows: For the similarity of the target patient's inpatient rehabilitation data with all reference patients, the Otsu threshold segmentation algorithm is used to divide all the similarities into two categories. The category with the largest mean similarity is denoted as the first category, and the category with the smallest mean similarity is denoted as the second category. The difference between the mean similarity in the first category and the mean similarity in the second category is denoted as b. b is used as the value of the individual predicted difference A of the target patient.

[0009] Preferably, the specific formula for recording the difference between B and A as the difference in rehabilitation training for the target patient is: C=BA; where C represents the difference in rehabilitation training for the target patient.

[0010] Preferably, the specific steps for obtaining N2 patients with the greatest similarity to the target patient's inpatient rehabilitation data and the smallest difference in rehabilitation training, and using them again as reference patients, are as follows: For any patient F other than the target patient, obtain the similarity a1 between the inpatient rehabilitation training data of patient F and the target patient, and record the difference in rehabilitation training of patient F as b1. Let a1-y×b1 be the assessment index for patient F; the N2 patients with the largest assessment index are re-selected as reference patients; where y is a preset value.

[0011] Preferably, the step of stopping the update when the ratio of the difference in rehabilitation training of the target patient to N2 is the minimum includes the following specific steps: After each update, obtain the ratio Q1 of the difference in rehabilitation training of the target patient to N2. If the obtained Q1 is the smallest after multiple consecutive updates, then stop updating.

[0012] Preferably, the patient whose inpatient rehabilitation data has the greatest similarity to the target patient and has the smallest final training difference is recorded as the final reference patient, and the specific steps include the following: Obtain several other patients whose inpatient rehabilitation data are most similar to the target patient, and select the patient with the smallest final training difference from these other patients as the final reference patient.

[0013] Preferably, the specific steps for evaluating the indicators of patient F are as follows: The reference patient corresponding to patient F is denoted as reference patient F1; the mean similarity between the inpatient rehabilitation data of reference patient F1 and target patient is calculated and denoted as y1, and y = exp(-y1); a1-y×b1 is denoted as the evaluation index of patient F, and exp() represents an exponential function with the natural constant as the base.

[0014] Preferably, updates stop when the number of updates exceeds P, where P is the preset maximum number of updates.

[0015] Another embodiment of the present invention provides an artificial intelligence-based sports injury rehabilitation program generation system. The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-described artificial intelligence-based sports injury rehabilitation program generation method.

[0016] The beneficial effects of the technical solution of the present invention are: This invention uses self-encoding technology to analyze and process inpatient rehabilitation data and discharge rehabilitation data separately. First, based on the inpatient rehabilitation data, individual differences among patients are preliminarily analyzed and assessed. Then, based on the discharge rehabilitation data of patients with assessed individual differences, further differences in rehabilitation training are evaluated. This process initially analyzes and processes inpatient and discharge rehabilitation data separately, reasonably combining them to avoid the following problems: when inpatient and discharge rehabilitation data are analyzed together or separately, the interference from individual differences and differences in rehabilitation training leads to large errors in the rehabilitation data, making the generated rehabilitation plan unsuitable for the subsequent rehabilitation needs of the target patient after discharge.

[0017] Then, based on the preliminary assessment of rehabilitation training differences, the inpatient rehabilitation data and the discharged rehabilitation data are analyzed and processed together again. In this process, the individual differences and rehabilitation training differences initially assessed are updated by making patients with more similar inpatient rehabilitation data have as small a rehabilitation training difference as possible. This process avoids the inaccuracies of simply increasing a large amount of rehabilitation data to mine and extract individual difference features (and further obtain rehabilitation training differences).

[0018] Ultimately, by extracting sufficient inpatient and discharge rehabilitation data, the minimum difference in rehabilitation training (i.e., the final training difference) was accurately identified among patients after discharge. The rehabilitation training plan generated based on this final training difference avoids the problem of the generated plan being unsuitable for the target patient due to individual differences and variations in the rehabilitation training process within the rehabilitation data (i.e., inpatient and discharge rehabilitation data). Furthermore, this invention analyzes inpatient and discharge rehabilitation data together, considering both inpatient and discharge situations, thus achieving full utilization and in-depth extraction of rehabilitation data and ensuring that the generated rehabilitation training plan is as applicable as possible to the target patient. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 The flowchart illustrates the steps of a method for generating a sports injury rehabilitation program based on artificial intelligence, as provided in an embodiment of the present invention. Detailed Implementation

[0021] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the artificial intelligence-based sports injury rehabilitation program generation method and system proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0022] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0023] The following description, in conjunction with the accompanying drawings, details the specific scheme of the AI-based sports injury rehabilitation program generation method and system provided by this invention.

[0024] Example 1: Please see Figure 1 The diagram illustrates a flowchart of a method for generating a sports injury rehabilitation plan based on artificial intelligence, according to an embodiment of the present invention. The method includes the following steps: Step S101: Obtain inpatient rehabilitation data and discharge rehabilitation data for each patient.

[0025] Sports injuries include fractures, muscle or tendon strains, and sprains of ligaments around joints. This example uses a fracture as an example. Severe sports injuries require long-term, appropriate rehabilitation training to recover. The recovery process is divided into inpatient and post-discharge rehabilitation. During inpatient rehabilitation, patients receive care from professional medical staff. During post-discharge rehabilitation, patients need to perform rehabilitation training independently.

[0026] Compared to the rehabilitation process during hospitalization, the rehabilitation process after discharge is more challenging because there is no professional nursing supervision or guidance (or hiring professional rehabilitation instructors is costly), making it difficult for patients to develop reasonable follow-up rehabilitation plans based on their current recovery status. Therefore, this embodiment utilizes the rehabilitation status of all patients throughout the entire rehabilitation process (including both hospitalization and discharge phases) to generate rehabilitation plans.

[0027] First, rehabilitation data for each patient throughout the entire rehabilitation process is recorded. Specifically, for inpatient rehabilitation, this data includes medication data (e.g., dosage and intervals of anticoagulants, anti-inflammatory drugs, and analgesics), diagnostic data recorded by medical staff, and rehabilitation training data (e.g., weight-bearing data during weight-bearing training, duration of each training session, etc.). The rehabilitation data recorded differs depending on the type of sports injury. For example, for other types of sports injuries, rehabilitation data in other embodiments includes the duration and frequency of myostatic strength training for patients with muscle strains, and the duration and frequency of manual muscle strengthening training for patients with upper limb injuries. In this embodiment, inpatient rehabilitation data is simply referred to as inpatient rehabilitation data.

[0028] For the rehabilitation process during the discharge stage, rehabilitation data includes medication data (such as dosage and interval of analgesics, anti-inflammatory drugs, etc.), online consultation data on rehabilitation status, and rehabilitation training data (such as weight and training duration during weight-bearing training). In this embodiment, the rehabilitation data during the discharge stage will be referred to as discharge rehabilitation data.

[0029] This completes the acquisition of rehabilitation data for each patient throughout the entire rehabilitation process, including inpatient and discharge rehabilitation data. It should be noted that both inpatient and discharge rehabilitation data include basic information such as the patient's gender, weight, height, and age, as well as the time of each data collection. Furthermore, the inpatient and discharge rehabilitation data include various data types, such as text data (e.g., medical records kept by medical staff), numerical data (e.g., time-series data of weight-bearing training at different times), and even video image data (e.g., CT data, patient physical and gait videos). In this embodiment, these data are stored in a data table within the hospital's MySQL database. In other embodiments, inpatient and discharge rehabilitation data are also stored in a wide table format. Discharge rehabilitation data is uploaded by the patient through a data collection platform provided by the hospital. Since different data storage methods and data collection platforms are all conventional technologies, this embodiment will not elaborate on them.

[0030] This embodiment provides a comparative example for generating a sports injury rehabilitation program, including: Any patient can be designated as the target patient. It should be noted that the patients mentioned in all embodiments of the present invention refer to patients who have recovered or are recovering within a preset time period (e.g., within 24 weeks from the start of hospitalization) during the entire rehabilitation training process, and must have discharge rehabilitation data. Patients who have not fully recovered within the preset time period or who do not have discharge rehabilitation data will not be considered.

[0031] Obtain the rehabilitation data (including inpatient rehabilitation data and discharge rehabilitation data) of other patients most similar to the target patient, and use the rehabilitation training program of the discharged patient as the sports injury rehabilitation program generated for the target patient.

[0032] However, the above comparative embodiments have the following problems: Compared to inpatient and discharge rehabilitation data for all patients, inpatient rehabilitation data is more complete, detailed, and abundant, and better reflects the patient's recovery status and condition. In contrast, discharge rehabilitation data may be incomplete, and may even contain data that does not accurately reflect the patient's recovery status. This is because, on the one hand, the rehabilitation training process after discharge lacks professional guidance and supervision, the selection and use of training equipment, and the rehabilitation training environment varies from patient to patient. Even when different patients use the same rehabilitation training program, there are differences in rehabilitation training, resulting in incomplete, incomplete, abnormal, and interfered discharge rehabilitation data. On the other hand, in addition to differences in rehabilitation training, there are also individual differences among patients (such as differences in underlying diseases).

[0033] In summary, the rehabilitation data of all patients, consisting of inpatient rehabilitation data and discharge rehabilitation data, contains significant error interference, leading to the inaccuracy of the above comparative examples.

[0034] Step S102: Obtain the N1 patients whose inpatient rehabilitation data is most similar to the target patient's inpatient rehabilitation data, and record them as reference patients. N1 is the first preset value.

[0035] The greater the similarity of inpatient rehabilitation data, the more likely different patients have the same or similar rehabilitation status during hospitalization. Reference patients refer to other patients whose rehabilitation status is similar or close to that of the target patient during hospitalization. This embodiment uses N=15 as an example; in other embodiments, N can be set to other values, and this embodiment does not impose specific limitations.

[0036] As an example, the method for obtaining similarity in inpatient rehabilitation data is as follows: The inpatient rehabilitation data of each patient is transformed into a feature vector (e.g., a 2048-dimensional feature vector) using autoencoder technology. This feature vector is denoted as the encoding vector of the inpatient rehabilitation data. The cosine similarity between the encoding vectors of any two patients is used as the similarity of the inpatient rehabilitation data.

[0037] Autoencoder technology is a well-known technique. It utilizes the encoder and decoder of an autoencoder neural network to encode and decode input data. Its general working principle is as follows: each patient's inpatient rehabilitation data is input into the encoder of the autoencoder neural network. The encoder outputs a feature vector (i.e., the encoded vector mentioned above). This feature vector is then input into the decoder, which outputs decoded data. This decoded data has minimal difference from each patient's inpatient rehabilitation data. The specific working principle of autoencoder technology will not be elaborated in this embodiment, nor will this embodiment specifically limit the results of the autoencoder neural network.

[0038] It should be noted that, since each patient's inpatient rehabilitation data is multimodal, including text data, numerical (or time-series) data, video image data, etc., this embodiment uses autoencoder technology to obtain encoding vectors for each type of data. For example, the autoencoder technology is used to obtain encoding vectors for all text-type inpatient rehabilitation data of each patient, the autoencoder technology is used to obtain encoding vectors for all numerical-type inpatient rehabilitation data of each patient, and the autoencoder technology is used to obtain encoding vectors for all video image-type inpatient rehabilitation data of each patient. The vector formed by concatenating these encoding vectors end to end is used as the final encoding vector of the inpatient rehabilitation data. The cosine similarity between the final encoding vectors corresponding to any two patients is used as the similarity of the inpatient rehabilitation data.

[0039] As another example, the method for obtaining similarity of inpatient rehabilitation data is as follows: To reduce computational load (e.g., to reduce the enormous computational cost of autoencoder technology), video image data is removed from each patient's inpatient rehabilitation data. Furthermore, using word2vec technology (or Universal Sentence Encoder technology in other embodiments), all text-type data in each patient's inpatient rehabilitation data is encoded into text vectors. Then, using autoencoder technology, all numerical data for each patient is encoded into feature vectors. These text vectors and feature vectors are concatenated to form the encoded vector of the inpatient rehabilitation data. The cosine similarity between the encoded vectors of any two patients is used as the similarity of the inpatient rehabilitation data.

[0040] In some examples, the method for obtaining similarity of inpatient rehabilitation data is as follows: To further reduce computational load, all numerical data types for each patient are saved in text format (each numerical value is treated as a single character). Then, word2vec technology (or UniversalSentence Encoder technology in other embodiments) is used to encode all text data types in each patient's inpatient rehabilitation data into text vectors. The cosine similarity between the text vectors corresponding to any two patients is used as the inpatient rehabilitation data similarity.

[0041] At this point, a reference patient for each patient has been obtained.

[0042] Step S103: Obtain the individual predicted difference A of the target patient. The difference between A and the inpatient rehabilitation data of the reference patient is positively correlated.

[0043] Based on step S101 above, this step first provides another comparative example: The rehabilitation plans generated for the target patients were generated by using all the sports injury rehabilitation plans of the N1 reference patients after their discharge.

[0044] This comparative embodiment generates rehabilitation plans based solely on inpatient rehabilitation data. Its advantage lies in avoiding interference introduced by discharge rehabilitation data with greater noise error (i.e., interference with the results obtained from the reference patients mentioned above). However, its disadvantage is that the lack of discharge rehabilitation data results in incomplete rehabilitation data for each patient and fails to reflect the rehabilitation status of the target patient after discharge. Consequently, the generated rehabilitation plan cannot be applied to the subsequent rehabilitation needs of the target patient after discharge.

[0045] Even when incorporating discharge rehabilitation data in other embodiments—for example, using the rehabilitation plan for the sports injury of the patient whose discharge rehabilitation data is most similar to that of the target patient among N reference patients as the rehabilitation plan generated for the target patient—the following problems still exist: Firstly, individual differences exist between different patients, which means that even the target patient with the most similar inpatient rehabilitation data may not be suitable for their post-discharge rehabilitation training plan. Secondly, in addition to individual differences, the lack of professional supervision and guidance during post-discharge training, differences in training equipment or the methods of using the equipment, and incomplete or incorrect collection of rehabilitation data further lead to the following: even the target patient with the most similar inpatient rehabilitation data may not be suitable for their post-discharge rehabilitation training plan.

[0046] To address the aforementioned issues, this embodiment first requires analyzing the individual differences among different patients; in this embodiment, the individual predicted difference A is used to describe the individual differences between the target patient and the reference patient.

[0047] The greater the discrepancy between the inpatient rehabilitation data of the reference patients, the more significant the differences between them, even though the reference patients' data are most similar to those of the target patients. These differences are primarily due to individual variations (such as underlying diseases and differences in metabolic function), resulting in a larger individual predicted difference A. Conversely, the smaller the discrepancy between the inpatient rehabilitation data of the reference patients, the more similar they are to the target patients in terms of rehabilitation training methods and physical recovery, leading to a smaller individual predicted difference A.

[0048] As an example, obtaining the individual predicted difference A of the target patient, where A is positively correlated with the difference between the inpatient rehabilitation data of the reference patient, includes the following methods: Let a be the mean of the similarity between the inpatient rehabilitation data of the target patient and all reference patients. The individual predicted difference A is negatively correlated with a, for example, A=1 / a. In other embodiments, A=exp(-a) can also be set, where exp() represents an exponential function with the natural constant as the base.

[0049] In this example, the larger 'a' is, the smaller the differences between the inpatient rehabilitation data of the reference patients, and the smaller the individual predicted difference A is; the smaller 'a' is, the larger the differences between the inpatient rehabilitation data of the reference patients, and the larger the individual predicted difference A is.

[0050] However, this example may have the following problem: when the similarity of the target patient's inpatient rehabilitation data with all reference patients has a large local distribution difference, that is, when the target patient's inpatient rehabilitation data is highly similar to that of some reference patients, but lowly similar to that of other reference patients, the obtained individual predicted difference A is not accurate enough.

[0051] As another example, obtaining the individual predicted difference A for the target patient, where A is positively correlated with the difference between the inpatient rehabilitation data of the reference patient, includes the following methods: For the similarity of the target patient's inpatient rehabilitation data with all reference patients, the Otsu threshold segmentation algorithm is used to divide these similarities into two categories. The category with the largest mean similarity is denoted as the first category, and the category with the smallest mean similarity is denoted as the second category. The difference between the mean similarity in the first category and the mean similarity in the second category is denoted as b. b is used as the value of the individual predicted difference A of the target patient.

[0052] The advantage of this example is that when the similarity between the inpatient rehabilitation data of the target patient and all reference patients has a large local distribution difference, the individual predicted difference A can be obtained based on the difference in the corresponding similarity between different reference patients. The larger b is, the more different the similarity between the inpatient rehabilitation data of different reference patients and the target patient is (that is, the similarity is not concentrated, but has a large local difference), and the larger the individual predicted difference A is.

[0053] In other embodiments, the two examples above are combined to obtain the individual prediction difference A, for example, A = b / a.

[0054] At this point, the individual predicted difference A for each patient is obtained using the method described in this step.

[0055] Step S104: The difference between the discharge rehabilitation data of the reference patients is denoted as B, and the difference between B and A is denoted as the rehabilitation training difference of the target patients.

[0056] The above-mentioned individual predicted difference A for each patient is obtained. Further, this embodiment needs to introduce discharge rehabilitation data. However, the discharge rehabilitation data contains more serious errors and interferences. These interferences and errors mainly come from the differences between individual patients, as well as from the differences in the rehabilitation training process (such as differences caused by the lack of professional supervision and guidance during training, differences in training instruments or the methods of using instruments, and differences in the completeness of rehabilitation data collection). This embodiment uses rehabilitation training differences to describe the differences in the rehabilitation training process of the target patient at the discharge stage, as reflected in the discharge rehabilitation data, compared with the reference patient.

[0057] In this embodiment, the difference in rehabilitation training for the target patient is equal to the difference between B and A.

[0058] Specifically, a larger difference (B) between the discharge rehabilitation data of reference patients indicates that patients with the most similar inpatient rehabilitation data during hospitalization show significant differences in their discharge rehabilitation data during the post-discharge rehabilitation training process. Conversely, a smaller individual predicted difference (A) suggests that the differences in discharge rehabilitation data among reference patients are mainly due to differences in the rehabilitation training process, resulting in greater differences in rehabilitation training among target patients. Similarly, a smaller difference (B) between the discharge rehabilitation data of reference patients indicates that patients with the most similar inpatient rehabilitation data during hospitalization show relatively similar differences in their discharge rehabilitation data during the post-discharge rehabilitation training process. Furthermore, a larger individual predicted difference (A) suggests that the differences in discharge rehabilitation data among reference patients are mainly due to differences in individual predicted differences, resulting in relatively smaller differences in rehabilitation training among target patients.

[0059] The smaller the difference in rehabilitation training among the target patients, the more reliable the rehabilitation training program of the reference patient at the discharge stage can be as the rehabilitation program generated for the target patient (that is, the comparative embodiment described in step S103 is more reliable).

[0060] As an alternative example, the method for obtaining the differences (i.e., B) between the discharge and rehabilitation data of reference patients is as follows: Following the method in step S102, the discharge rehabilitation data of the reference patients are converted into encoding vectors, and the Euclidean distance between the encoding vectors of the discharge rehabilitation data of any two reference patients is obtained. The mean of the Euclidean distances between all reference patients is denoted as B.

[0061] As a preferred example, the method for obtaining the differences (i.e., B) between the discharge and rehabilitation data of reference patients is as follows: According to the method in step S102, the discharge rehabilitation data of the reference patients are converted into encoding vectors. For all reference patients of the target patient, the encoding vectors of the discharge rehabilitation data of all reference patients are clustered. In this embodiment, the K-Means clustering algorithm is used, and the number of clusters K is set to one-fifth (rounded up) of the number of all reference patients. In other embodiments, K can be set to other values. This embodiment does not make specific limitations.

[0062] The average value of all encoded vectors in each cluster is calculated to obtain the average encoded vector. The two clusters with the largest Euclidean distance between their average encoded vectors are then selected, and the Euclidean distance between the average encoded vectors of these two clusters is used as the value of B.

[0063] This preferred example can describe the difference B between the discharge rehabilitation data of the reference patient from the local distribution (i.e., multiple categories) and the overall distribution (i.e., the two categories with the largest Euclidean distance) between the discharge rehabilitation data, making the calculation result of B relatively reliable.

[0064] As an example, the formula for calculating the difference in rehabilitation training among target patients includes: C=BA, where C represents the difference in rehabilitation training among target patients.

[0065] In other embodiments, the formula for calculating the difference in rehabilitation training for the target patient includes: C=k×BA, where k is a preset scaling factor, such as k=2.0. In other embodiments, k can also be set to other values ​​greater than 0.

[0066] Thus, this embodiment has obtained the differences in rehabilitation training for each patient. In this embodiment, the differences in rehabilitation training for all patients are linearly normalized.

[0067] Step S105: Obtain N2 patients who have the greatest similarity to the target patient's inpatient rehabilitation data and the smallest difference in rehabilitation training, and use them again as reference patients. N2 is a second preset value, and N2>N1.

[0068] Reviewing all the steps above, this embodiment considers that the patient's inpatient rehabilitation data is more complete and accurate, and the rehabilitation training is more standardized. Compared with discharge rehabilitation data, the differences between patients with the same or similar inpatient rehabilitation data are mainly affected by individual differences. Therefore, obtaining the individual predicted difference A (i.e., step S103) by utilizing the differences between the inpatient rehabilitation data of reference patients is relatively accurate. However, if we want to further evaluate more accurate individual differences (and rehabilitation training differences), it is actually necessary to perform data mining on a large amount of the same or similar inpatient rehabilitation data (and their corresponding discharge rehabilitation data). (That is, N needs to be set to a large value in step S102, and then the obtained individual predicted difference A will be relatively accurate, thus ensuring the accuracy of rehabilitation training differences.) However, there are some target patients whose number of other patients with the most similar inpatient rehabilitation data is insufficient, or when N is set too large, the inpatient rehabilitation data of some of the reference patients will differ significantly from the inpatient rehabilitation data of the target patients. Especially when there is less patient rehabilitation data stored in the hospital, the accuracy of obtaining rehabilitation training differences will be even lower.

[0069] In summary, it is not possible to further assess more accurate individual differences simply by increasing the value of N. At the same time, if the value of N is too small, it is also impossible to comprehensively assess individual differences, which leads to low accuracy of rehabilitation training differences calculated based on individual differences (i.e., step S104).

[0070] When the assessment of individual differences and differences in rehabilitation training at the discharge stage is inaccurate, it is impossible to further resolve the problems existing in the comparative embodiment described in step S103.

[0071] This embodiment further considers that, for a patient (let's call him patient F), the more similar patient F's inpatient rehabilitation data is to the target patient, and the smaller the difference in discharge rehabilitation status between patient F and the patient whose inpatient data is most similar to his / hers (i.e., the smaller the difference in rehabilitation training), even if the difference in rehabilitation training is not the most accurate, it is still more important to generate a rehabilitation training plan based on patient F's inpatient rehabilitation data and discharge rehabilitation data than to simply increase the value of N to further evaluate individual differences and differences in rehabilitation training. This would make the generated rehabilitation training plan suitable for the target patient.

[0072] This step involves identifying N² patients who have the greatest similarity to the target patient's inpatient rehabilitation data and the smallest difference in rehabilitation training, and these N² patients are then used as reference patients again. N² is set to be greater than N. This embodiment uses N² = (w + 1) × N as an example; in other embodiments, N² can be set to other values, and this embodiment does not impose specific limitations. Additionally, in this embodiment, w = 0.5; in other embodiments, w can also be set to other values.

[0073] The newly acquired reference patients not only contain more similarities to the target patients' inpatient rehabilitation data, but also have smaller differences in discharge rehabilitation data between these reference patients and the patients whose inpatient data is most similar to theirs. Generating rehabilitation training programs based on the inpatient and discharge rehabilitation data of these reference patients can avoid the problem of inaccurate prediction of individual differences and inaccurate acquisition of rehabilitation training differences, resulting in rehabilitation training programs that are not suitable for the target patients.

[0074] As an optional example, obtaining N2 patients with the greatest similarity to the target patient's inpatient rehabilitation data and the smallest difference in rehabilitation training includes the following steps: For any patient F other than the target patient, obtain the similarity a1 between patient F and the target patient's inpatient rehabilitation training data, and record the difference in rehabilitation training of patient F as b1 (obtained from steps S101 to S104). Record a1-b1 as the evaluation index of patient F. For the N2 patients with the largest evaluation index, these patients have the largest similarity to the target patient's inpatient rehabilitation data and the smallest difference in rehabilitation training. Record these N2 patients as reference patients.

[0075] As another optional example, obtaining N2 patients who have the greatest similarity to the target patient's inpatient rehabilitation data and the smallest difference in rehabilitation training includes the following steps: Let a1-y×b1 be the assessment index for patient F, where y is a preset value. The larger y is, the less attention is paid to the differences in rehabilitation training of patient F (equivalent to paying more attention to similarity a1). That is, when y is larger, even if the differences in rehabilitation training of patient F are relatively small, the assessment index of patient F will not be very large (at this time, patient F is less likely to be regarded as a reference patient). Conversely, the smaller y is, the less attention is paid to the differences in rehabilitation training of patient F. That is, when y is smaller, even if the differences in rehabilitation training of patient F are relatively large, the assessment index of patient F is also relatively large (at this time, patient F may also be regarded as a reference patient).

[0076] This embodiment uses y=2 as an example for description. In other embodiments, y can be set to other values ​​greater than or equal to 0.

[0077] Specifically, when y=0, it is equivalent to simply increasing the value of N to further assess individual differences and differences in rehabilitation training.

[0078] Furthermore, the N2 patients with the highest evaluation indicators were designated as reference patients.

[0079] As a preferred example, obtaining N2 patients with the greatest similarity to the target patient's inpatient rehabilitation data and the smallest difference in rehabilitation training includes the following steps: Obtain the reference patients corresponding to patient F (similar to step S102), denoted as reference patients F1. Calculate the similarity between these reference patients F1 and the target patients (see step S102 for the similarity calculation method), and denote the mean of these similarities as y1. Let y = exp(-y1). Denote a1 - y × b1 as the evaluation index of patient F, and denote the N2 patients with the largest evaluation index as reference patients.

[0080] In this preferred example, a larger y1 (i.e., a smaller y) indicates that the patient with the most similar inpatient rehabilitation data to patient F (i.e., reference patient F1) also has significant similarity to the target patient's inpatient rehabilitation data. In this case, even if patient F's rehabilitation training differences are relatively large, patient F is still considered a reference patient for the target patient. This avoids interference caused by an excessive pursuit of minimizing rehabilitation training differences when the current differences are inaccurate. Conversely, a smaller y1 (i.e., a larger y) indicates that the patient with the most similar inpatient rehabilitation data to patient F (i.e., reference patient F1) has relatively small similarity to the target patient's inpatient rehabilitation data. In this case, even if patient F's rehabilitation training differences are relatively small, patient F is not considered a reference patient for the target patient. This is because among patients with significantly different inpatient rehabilitation data, even if patient F's rehabilitation training differences are small, it may not mean that the target patient's rehabilitation training differences are also small. This avoids situations where using patient F as a reference patient for the target patient would prevent the subsequent acquisition of more accurate rehabilitation training differences for the target patient.

[0081] In summary, this preferred example, by further analyzing the relationship between patient F and the target patient in inpatient rehabilitation data and dynamically setting the value of y, obtains a more accurate reference patient compared to the optional examples mentioned above.

[0082] Step S106: Update the rehabilitation training difference of the target patient, wherein the value of N2 is greater than the value set by N2 in the previous update each time; stop updating when the ratio of the rehabilitation training difference of the target patient to N2 is the smallest; record the obtained rehabilitation training difference of the target patient as the final training difference.

[0083] After obtaining the reference patient for the target patient again, steps S103, S104, and S105 are repeated. Each time S103, S104, and S105 are repeated, the individual predicted difference A and the rehabilitation training difference of the target patient are updated, and the reference patient is obtained again. Note that in this embodiment, the value of N2 is greater than the value set in the previous update. For example, the value of N2 in each update is equal to 1.2 times the value set in the previous update.

[0084] The above-described update process enables the gradual mining and extraction of individual differences between the target patient and other patients (i.e., reference patients) from a larger pool of patients with the most similar inpatient rehabilitation data (i.e., gradually updating the individual predicted difference A). This allows the gradual mining and extraction of rehabilitation training differences of the target patient from their discharge rehabilitation data based on a larger pool of patients with known individual differences (as described in step S104).

[0085] Furthermore, updates are stopped when the ratio of the rehabilitation training difference of the target patient to N2 is minimized; the obtained rehabilitation training difference of the target patient is recorded as the final training difference.

[0086] The update stops when the ratio of the rehabilitation training difference of the target patient to N2 is minimized. This process ensures that sufficient inpatient rehabilitation data and discharge rehabilitation data are used to accurately extract the minimum rehabilitation training difference (i.e., the final training difference) when the patient undergoes rehabilitation training after discharge. As described in step S104: the smaller the rehabilitation training difference, the more likely the rehabilitation training plan of the reference patient at the discharge stage can be used as the rehabilitation plan generated for the target patient.

[0087] As an example, updating is stopped when the ratio of the difference in rehabilitation training for the target patient to N2 is minimized. This includes methods such as: After each update, the ratio Q1 of the difference in rehabilitation training for the target patient to N2 is obtained. When, after several consecutive updates (e.g., after three consecutive updates), the obtained Q1 is the smallest among all obtained Q1 values, the update process stops. Additionally, updates also stop when the number of updates is greater than or equal to P, indicating that Q1 has reached its minimum value. This is to avoid endless updates and also to save computational resources. P is the preset maximum number of updates; this embodiment uses P=20 as an example. Other values ​​can be set in other embodiments, and this embodiment does not impose specific limitations.

[0088] Step S107: The patient whose inpatient rehabilitation data is most similar to that of the target patient and has the smallest final training difference is recorded as the final reference patient. The sports injury rehabilitation plan of the final reference patient after discharge is used as the rehabilitation plan generated for the target patient.

[0089] Specifically, N3 other patients with the greatest similarity to the target patient's inpatient rehabilitation data are obtained, and the patient with the smallest final training difference is selected from these N3 other patients as the final reference patient.

[0090] As one embodiment, N3 = N. In other embodiments, N3 may be set to other values, and this embodiment does not impose specific limitations.

[0091] In other embodiments, multiple (e.g., N4) patients with the smallest final training differences are selected from these N3 other patients as the final reference patients, for example, N4 = half of N3 (rounded to the nearest integer). In these embodiments, the value of N4 is not limited.

[0092] This concludes the example.

[0093] In summary, considering all the embodiments of the present invention described above: This embodiment divides rehabilitation data into discharge rehabilitation data and inpatient rehabilitation data according to the distribution characteristics of the patient's rehabilitation data. Specifically, the distribution characteristics of the rehabilitation data are as follows: Inpatient rehabilitation data is more complete, specific, and richer, and it can better reflect the patient's rehabilitation status and condition. However, inpatient rehabilitation data is greatly affected by individual differences among patients. In addition, inpatient rehabilitation data mainly focuses on rehabilitation training during the inpatient rehabilitation phase and cannot be directly used as a reference for subsequent rehabilitation training after discharge. Inpatient rehabilitation data may be incomplete or even inaccurate. Furthermore, in addition to being affected by individual patient differences, it is also influenced by variations in the rehabilitation training process (described in step S101), resulting in significant noise interference in the inpatient rehabilitation data. This means that even identical or very similar discharge rehabilitation data cannot be used as a reference for rehabilitation training plans, or conversely, even dissimilar discharge rehabilitation data may serve as a reference for rehabilitation training plans. However, discharge rehabilitation data contains information about the rehabilitation training process after discharge, which is crucial for patients' subsequent rehabilitation training.

[0094] When a rehabilitation training program is generated by directly using a large amount of patient rehabilitation data (that is, by analyzing and processing inpatient rehabilitation data and discharge rehabilitation data together) (as in the comparative example provided in step S101), or when a rehabilitation training program is generated solely based on a large amount of inpatient rehabilitation data or discharge rehabilitation data (as in the comparative example provided in step S103), it is impossible to obtain a more reliable rehabilitation training program that is suitable for the patient's rehabilitation condition.

[0095] In the above embodiments, inpatient rehabilitation data and discharge rehabilitation data are analyzed and processed separately: First, individual differences among patients are preliminarily analyzed and assessed based on the inpatient rehabilitation data. Then, based on the discharge rehabilitation data of patients with assessed individual differences, differences in rehabilitation training are further evaluated. This process initially analyzes and processes inpatient rehabilitation data and discharge rehabilitation data together, and reasonably combines inpatient rehabilitation data and discharge rehabilitation data based on the above distribution characteristics, thus initially avoiding the above-mentioned problems (i.e., the problems existing in the comparative embodiments provided in steps S101 and S103). That is, when inpatient rehabilitation data and discharge rehabilitation data are analyzed together or separately, the rehabilitation data error interference caused by individual differences and differences in rehabilitation training is large, which makes the generated rehabilitation plan unsuitable for the subsequent rehabilitation needs of the target patient after discharge.

[0096] Then, based on the preliminary assessment of rehabilitation training differences, the inpatient rehabilitation data and the discharged rehabilitation data are analyzed and processed together again. In this process, the individual differences and rehabilitation training differences initially assessed are updated by making patients with similar inpatient rehabilitation data have as small a rehabilitation training difference as possible. This process avoids the inaccuracy of simply increasing a large amount of rehabilitation data to mine and extract individual difference features (and further obtain rehabilitation training differences) (that is, it avoids the problem of low accuracy when simply setting N to a large value and then using the comparative examples provided in steps S101 and S103).

[0097] Ultimately, from a sufficient amount of patient inpatient and discharge rehabilitation data, the minimum rehabilitation training difference (i.e., the final training difference) was accurately extracted when patients underwent rehabilitation training after discharge. When generating a rehabilitation training plan based on this final training difference, the problem of the generated rehabilitation training plan being unsuitable for the target patient due to individual differences and differences in the rehabilitation training process (i.e., inpatient and discharge rehabilitation data) is avoided. Based on this, this embodiment divides inpatient and discharge rehabilitation data together, considering both inpatient and discharge situations, achieving full utilization and in-depth extraction of rehabilitation data, and ensuring that the generated rehabilitation training plan is as applicable as possible to the target patient.

[0098] Example 2: The above embodiment one is followed by the following steps: For each discharged patient, the sports injury rehabilitation plan implemented after discharge is obtained and denoted as the sample plan. The sports injury rehabilitation plan generated for each patient is obtained using the method described in Example 1 and denoted as the tag plan.

[0099] Each patient's sample plan and label plan constitute a sample, and all patient samples form a dataset. This dataset is used to train a neural network to generate rehabilitation plans for sports injuries. Specifically, this includes: A neural network for generating sports injury rehabilitation programs is constructed. In this embodiment, the GPT4 large language network model is used. Other neural network models, such as DeepSeek, can be used in other embodiments. This embodiment does not limit the specific structure of the network, nor does it limit the number of meta-parameters of the network.

[0100] The network takes a sample scheme for each patient as input and outputs a label scheme. It uses cross-entropy as the loss function and is trained using stochastic gradient descent. The specific training process is well-known and will not be described in detail in this embodiment.

[0101] It should be noted that both the sample scheme and the label scheme use text-based data, which records the specific sports injury rehabilitation process. This includes details such as the start and end times of each training session, the training equipment used, weight data, and the body parts being trained.

[0102] After the neural network for generating rehabilitation programs for sports injuries is trained, for patients who need rehabilitation programs, the rehabilitation training programs that the patients have already performed are input into the network. The network outputs a new rehabilitation training program, which can be used by the patients for reference or to improve the existing rehabilitation training programs.

[0103] Example 3: This embodiment provides an artificial intelligence-based sports injury rehabilitation program generation system. The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements all the steps of the artificial intelligence-based sports injury rehabilitation program generation method included in Embodiments 1 and 2.

[0104] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for generating sports injury rehabilitation programs based on artificial intelligence, characterized in that, The method includes the following steps: Record the inpatient rehabilitation data and discharge rehabilitation data of each patient. Any patient is recorded as the target patient. Use autoencoder technology to obtain the N1 patients whose inpatient rehabilitation data is most similar to the target patient's inpatient rehabilitation data and record them as reference patients. N1 is the first preset value. D1: Obtain the individual predicted difference A of the target patient. The difference between A and the inpatient rehabilitation data of the reference patient is positively correlated. The difference between the discharge rehabilitation data of the reference patient is denoted as B. The difference between B and A is denoted as the rehabilitation training difference of the target patient. D2: Obtain N2 patients with the greatest similarity to the target patient's inpatient rehabilitation data and the smallest difference in rehabilitation training, and use them again as reference patients. N2 is the second preset value, and N2>N1. D3: Update the rehabilitation training difference of the target patient by repeatedly executing D1 and D2. Each time the value of N2 is updated, it is greater than the value set by N2 in the previous update. The update stops when the ratio of the rehabilitation training difference of the target patient to N2 is the smallest. The obtained rehabilitation training difference of the target patient is recorded as the final training difference. The patient with the greatest similarity to the target patient's inpatient rehabilitation data and the smallest difference in final training is designated as the final reference patient. The sports injury rehabilitation program of the final reference patient after discharge is used as the rehabilitation program generated for the target patient.

2. The method for generating sports injury rehabilitation programs based on artificial intelligence according to claim 1, characterized in that, The specific steps for obtaining the similarity of the inpatient rehabilitation data are as follows: The inpatient rehabilitation data of each patient is transformed into a feature vector using autoencoder technology. This feature vector is denoted as the encoding vector of the inpatient rehabilitation data. The cosine similarity between the encoding vectors of any two patients is used as the similarity of the inpatient rehabilitation data.

3. The method for generating sports injury rehabilitation programs based on artificial intelligence according to claim 1, characterized in that, The specific steps involved in obtaining the individual predicted difference A for the target patient are as follows: For the similarity of the target patient's inpatient rehabilitation data with all reference patients, the Otsu threshold segmentation algorithm is used to divide all the similarities into two categories. The category with the largest mean similarity is denoted as the first category, and the category with the smallest mean similarity is denoted as the second category. The difference between the mean similarity in the first category and the mean similarity in the second category is denoted as b. b is used as the value of the individual predicted difference A of the target patient.

4. The method for generating sports injury rehabilitation programs based on artificial intelligence according to claim 1, characterized in that, The specific formula for recording the difference between B and A as the difference in rehabilitation training for the target patient is: C=BA; where C represents the difference in rehabilitation training for the target patient.

5. The method for generating sports injury rehabilitation programs based on artificial intelligence according to claim 1, characterized in that, The specific steps involved in obtaining N² patients who have the greatest similarity to the target patient's inpatient rehabilitation data and the smallest difference in rehabilitation training, and then using them again as reference patients, are as follows: For any patient F other than the target patient, obtain the similarity a1 between the inpatient rehabilitation training data of patient F and the target patient, and record the difference in rehabilitation training of patient F as b1. Let a1-y×b1 be the assessment index for patient F; the N2 patients with the largest assessment index are re-selected as reference patients; where y is a preset value.

6. The method for generating sports injury rehabilitation programs based on artificial intelligence according to claim 1, characterized in that, The specific steps for stopping the update when the ratio of the difference in rehabilitation training of the target patient to N2 is minimized are as follows: After each update, obtain the ratio Q1 of the difference in rehabilitation training of the target patient to N2. If the obtained Q1 is the smallest after multiple consecutive updates, then stop updating.

7. The method for generating sports injury rehabilitation programs based on artificial intelligence according to claim 1, characterized in that, The patient whose inpatient rehabilitation data is most similar to that of the target patient and has the smallest final training difference is designated as the final reference patient. The specific steps involved are as follows: Obtain several other patients whose inpatient rehabilitation data are most similar to the target patient, and select the patient with the smallest final training difference from these other patients as the final reference patient.

8. The method for generating sports injury rehabilitation programs based on artificial intelligence according to claim 5, characterized in that, The specific steps for assessing the indicators of patient F are as follows: The reference patient corresponding to patient F is denoted as reference patient F1; the mean similarity between the inpatient rehabilitation data of reference patient F1 and target patient is calculated and denoted as y1, and y = exp(-y1); a1-y×b1 is denoted as the evaluation index of patient F, and exp() represents an exponential function with the natural constant as the base.

9. The method for generating sports injury rehabilitation programs based on artificial intelligence according to claim 6, characterized in that, Updates will stop when the number of updates exceeds P, where P is the preset maximum number of updates.

10. An artificial intelligence-based sports injury rehabilitation program generation system, the system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for generating sports injury rehabilitation programs based on artificial intelligence as described in any one of claims 1 to 9.

Images