Methods, equipment, and media for predicting the postoperative rehabilitation stage of osteoarthritis

Abstract

This invention relates to the field of rehabilitation data management technology, specifically to a method, device, and medium for predicting postoperative rehabilitation stages for osteoarthritis. The method includes: acquiring multi-dimensional prognostic factors corresponding to the patient; outputting a personalized postoperative rehabilitation stage for the patient through a pre-trained machine learning model; periodically updating physical function parameters; updating psychosocial parameters; and updating the personalized postoperative rehabilitation stage based on the updated parameters of the multi-dimensional prognostic factors. By integrating multi-dimensional prognostic factors such as basic physical characteristics, intraoperative parameters, functional status, and psychosocial factors, this invention overcomes the limitations of traditional template-based rehabilitation stages, generating a rehabilitation stage prediction tailored to each patient's individual circumstances, thus achieving personalized assessment of the patient's rehabilitation stage.

Description

Technical Field

[0001] This invention relates to methods, equipment, and media for predicting the postoperative rehabilitation stage of osteoarthritis, and belongs to the technical field of rehabilitation data management. Background Technology

[0002] Osteoarthritis is a common joint disease, and platelet-rich plasma (PRP) injection and extracorporeal shock wave therapy (ESWT) have become common treatments for this condition. However, the rehabilitation management after treatment faces serious challenges.

[0003] In traditional approaches, postoperative rehabilitation typically requires patients to frequently visit the hospital for check-ups to understand their current recovery stage. However, for osteoarthritis patients, especially those with knee osteoarthritis, frequent visits increase the burden on them. Furthermore, many osteoarthritis patients are elderly and have limited mobility, which can lead to decreased patient compliance and difficulty in obtaining timely professional guidance.

[0004] With the development of technology, online medical consultation services have emerged. While these services can provide convenience to some extent, the shortage of doctors makes it difficult to provide continuous and in-depth guidance to each patient. At the same time, the guidance and suggestions generally have the limitation of being one-size-fits-all, lacking consideration for individual patient differences and making it difficult to make personalized adjustments based on the patient's actual situation. This results in the generation of standardized and templated postoperative recovery stages, which may not match the actual situation of the patient.

[0005] For example, the postoperative recovery phase for platelet-rich plasma typically includes two stages: the initial reaction phase and the repair phase. During the initial reaction phase, temporary local pain, swelling, or inflammation may occur; during the repair phase, growth factors continue to act, regulating inflammation, promoting cell proliferation and angiogenesis, and the entire process can last from several weeks to several months.

[0006] For extracorporeal shock wave therapy, the postoperative recovery phase typically includes two stages: the initial reaction stage and the cumulative effect stage. In the initial reaction stage, redness, swelling, soreness, or bruising may occur at the treatment site, which usually subsides on its own within a few days. In the cumulative effect stage, the biological effects of the shock waves (such as promoting angiogenesis, disrupting pain receptors, and stimulating healing) can last for weeks or even months after treatment.

[0007] Therefore, it can be seen that the postoperative recovery stage given in the traditional plan is too routine and templated, and it is difficult to provide sufficient reference for patients and clinicians. Summary of the Invention

[0008] This invention addresses the shortcomings of existing technologies by providing a method, device, and medium for predicting the postoperative rehabilitation stage of osteoarthritis.

[0009] The technical solution of this invention to solve the above-mentioned technical problems is as follows: a method for predicting the postoperative rehabilitation stage of osteoarthritis, comprising:

[0010] Obtain multidimensional prognostic factors corresponding to the patient; the multidimensional prognostic factors include basic physical characteristic parameters, intraoperative treatment parameters, physical function parameters, and psychosocial parameters;

[0011] Based on the aforementioned multidimensional prognostic factors, a pre-trained machine learning model outputs the patient's corresponding personalized postoperative rehabilitation stage.

[0012] Based on the medical sensors worn by the patient, the physical function parameters are updated periodically; and based on the analysis results corresponding to the interaction content with the patient, the psychosocial parameters are updated.

[0013] The personalized postoperative rehabilitation stage is updated by updating the parameters of the multidimensional prognostic factors.

[0014] The beneficial effects of this invention are:

[0015] 1. By integrating multiple prognostic factors such as basic physical characteristics, intraoperative parameters, functional status, and psychosocial factors, it breaks through the limitations of traditional template-based rehabilitation stages and can generate rehabilitation stage predictions that fit each patient's individual situation, thus achieving personalized assessment of the patient's rehabilitation stage.

[0016] 2. By regularly updating physical function data through medical sensors and combining this with interactive content analysis to update psychosocial parameters, the rehabilitation phase predictions can be dynamically adjusted along with the patient's recovery process, maintaining the accuracy of the assessment and thus possessing dynamic update capabilities.

[0017] 3. It reduces reliance on frequent in-person follow-up examinations, enabling remote and continuous monitoring of changes in key patient indicators. This not only alleviates the burden of travel for patients, especially the elderly with limited mobility, but also helps maintain their enthusiasm and compliance with rehabilitation training, effectively improving the efficiency and compliance of rehabilitation management.

[0018] 4. Through automated data collection and model analysis, this solution can provide clinicians with continuous and objective decision support, reducing the burden on doctors to continuously track each patient, thereby allowing valuable medical resources to focus more on dealing with complex situations and optimizing the allocation of medical resources.

[0019] Based on the above technical solution, the present invention can be further improved as follows:

[0020] Furthermore, the model architecture of the machine learning model is based on the multimodal time-aware Transformer architecture, and the machine learning model includes an input layer, an encoding layer, a fusion layer, and an output layer;

[0021] The input layer receives the multi-dimensional prognostic factors;

[0022] The coding layer includes a static coding module, a temporal coding module, and a psychosocial module in parallel;

[0023] The static encoding module performs static encoding on the basic body feature parameters and the intraoperative treatment parameters, and outputs a static representation vector.

[0024] The timing coding module dynamically encodes the body function parameters and outputs a timing context vector;

[0025] The psychosocial module introduces personalized variation coefficients based on the psychosocial parameters and outputs a smoothed perturbation sequence.

[0026] The fusion layer fuses the static representation vector, the temporal context vector, and the smoothed perturbation sequence, and generates a personalized postoperative rehabilitation stage for the patient based on the fusion result. The personalized postoperative rehabilitation stage includes multiple sub-stages, each of which includes a corresponding duration, corresponding symptoms, and degree of recovery.

[0027] The output layer outputs the personalized postoperative rehabilitation stage.

[0028] The beneficial effects of adopting the above-mentioned further solutions are:

[0029] The machine learning model based on the multimodal temporal-aware Transformer architecture achieves accurate prediction of personalized postoperative rehabilitation stages by effectively processing and integrating multi-dimensional prognostic factors. Compared with traditional template-based methods, it can better meet the individual needs of patients, improve the quality and effectiveness of rehabilitation management, and also provide clinicians with a more scientific and objective basis for decision-making.

[0030] Furthermore, the static encoding module performs static encoding on the basic body feature parameters and the intraoperative treatment parameters, outputting a static representation vector, specifically including:

[0031] The static encoding module obtains the data vector corresponding to each data point by standardization and embedding retrieval processing based on the basic body feature parameters, numerical features and categorical features corresponding to each data point in the intraoperative treatment parameters;

[0032] A static representation vector is obtained by performing feature cross-processing on each data vector through a cross-network.

[0033] The static representation vector is linearly transformed to output the mapped static representation vector.

[0034] The beneficial effects of adopting the above-mentioned further solutions are:

[0035] By standardizing and embedding basic body feature parameters and intraoperative treatment parameters, different types of data can be effectively transformed into a unified data vector form, facilitating subsequent feature cross-processing. The use of cross-networks can uncover the potential relationships between data vectors, generating more expressive static representation vectors. Furthermore, linear transformation of these static representation vectors can further adjust their feature space, making the mapped static representation vectors more suitable for subsequent processing by machine learning models, thereby improving the model's ability to process static information from multi-dimensional prognostic factors.

[0036] Furthermore, the timing encoding module dynamically encodes the bodily function parameters and outputs a timing context vector, specifically including:

[0037] The temporal coding module performs one-dimensional convolution processing on the body function parameters and outputs a surface temporal feature map.

[0038] Feature extraction is performed on the surface temporal feature map using a recurrent neural network to output a deep temporal feature map;

[0039] The deep temporal feature map is subjected to attention weighting processing through an attention mechanism to output a temporal context vector.

[0040] The time context vector is linearly transformed to output the mapped time context vector.

[0041] The beneficial effects of adopting the above-mentioned further solutions are:

[0042] One-dimensional convolution processing can extract local surface temporal features from bodily function parameters, capturing short-term variation patterns in the data. Recurrent neural networks (RNNs) can further mine these surface temporal features, taking into account the temporal order of the data to output deep temporal feature maps, thus better capturing the long-term trends of bodily function parameters. Attention mechanisms can weight the deep temporal feature maps according to the importance of features at different times, highlighting key information and making the output temporal context vector more accurately reflect the dynamic changes of bodily function parameters. Finally, a linear transformation of the temporal context vector adjusts its feature space, making the mapped temporal context vector more suitable for subsequent processing by the machine learning model. This improves the model's ability to process and analyze dynamic information in multi-dimensional prognostic factors, thereby more accurately predicting the personalized postoperative rehabilitation stage of patients.

[0043] Furthermore, the psychosocial module introduces personalized variation coefficients based on the psychosocial parameters and outputs a smoothed perturbation sequence, specifically including:

[0044] The psychosocial module performs sentiment analysis, keyword extraction, and scale scoring quantification on the psychosocial parameters to obtain corresponding sentiment feature vectors, keyword feature vectors, and scale score vectors.

[0045] The emotional feature vector, the keyword feature vector, and the scale score vector are concatenated to obtain a structured psychosocial indicator tuple.

[0046] Encode the psychosocial indicator tuples to obtain a psychological state representation vector;

[0047] The psychological state representation vector is processed by a predefined weight matrix to obtain a psychosocial raw score; the psychosocial raw score is then subjected to bit transformation to obtain a psychosocial standard score; the psychosocial standard score is then subjected to range-adapted linear mapping, nonlinear probability transformation, and clinical range-adapted linear mapping in sequence to scale the psychosocial standard score to obtain the corresponding personalized coefficient of variation.

[0048] The mental state representation vector is hashed to obtain the corresponding integer hash coefficient; the integer hash coefficient is used as the initial state of the pseudo-random number generator to generate a random number sequence that conforms to the distribution pattern of the mental state representation vector.

[0049] Based on the personalized variation coefficient, each random number contained in the random number sequence is scaled; the scaled random number sequence is then smoothed using a filter to obtain a smoothed perturbation sequence.

[0050] Furthermore, before generating a random number sequence that conforms to the distribution of the psychological state representation vector by using the integer hash coefficient as the initial state of the pseudo-random number generator, the method further includes:

[0051] The mental state representation vector is mapped using a trained nonlinear activation function to obtain the patient's stability score; and the patient's extremity score is obtained based on the distance between the mental state representation vector and the pre-generated baseline representation vector.

[0052] Based on the stability score and the extremity score, select the corresponding distribution mode of the psychological state representation vector from a set of preset distribution modes;

[0053] If the stability score is higher than the first stability threshold and the extremity score is lower than the first extreme threshold, then a narrow-peak Gaussian distribution is selected from among the preset multiple distribution methods.

[0054] If the stability score is lower than the second stability threshold, or the extremity score is higher than the second extreme threshold, then the heavy-tailed distribution is selected from the preset multiple distribution methods.

[0055] Otherwise, select the Gaussian mixture distribution from among the preset distribution options.

[0056] The beneficial effects of adopting the above-mentioned further solutions are:

[0057] By conducting multi-dimensional analysis and quantification of psychosocial parameters, the psychosocial characteristics of patients can be comprehensively and accurately captured. Introducing personalized coefficients of variation fully considers the impact of individual patient psychological differences on the rehabilitation stage, making the rehabilitation stage prediction more closely aligned with the patient's actual situation. Furthermore, selecting an appropriate random number distribution based on the patient's stability and extremity scores can more reasonably simulate the uncertainty and changing patterns of the patient's psychological state. The generation of smoothed perturbation sequences further enhances the model's ability to handle patients' psychosocial factors, enabling the model to better integrate psychosocial factors when predicting personalized postoperative rehabilitation stages, improving the accuracy and personalization of rehabilitation stage predictions, providing patients with rehabilitation guidance more suited to their actual situations, and also providing strong support for clinicians to develop more scientific and effective rehabilitation plans.

[0058] Furthermore, the personalized postoperative rehabilitation stage is updated using the updated parameters of the multi-dimensional prognostic factors, specifically including:

[0059] Determine the first updated parameter for the physical function parameter and the second updated parameter for the psychosocial parameter among the multidimensional prognostic factors;

[0060] Based on the first update parameter and the second update parameter, the machine learning model re-outputs the updated personalized postoperative rehabilitation stage corresponding to the patient.

[0061] If the degree of difference between the updated personalized postoperative rehabilitation stage and the previous historical personalized postoperative rehabilitation stage is higher than the first preset degree, then for the current second update parameter, determine the degree of psychological stability weakening compared to the second update parameter at the time of the previous interaction.

[0062] If the degree of weakening of psychological stability is higher than the second preset degree, then the stage description conversion process is performed for the updated personalized postoperative rehabilitation stage, and an online communication channel between the patient and the clinician is established.

[0063] Furthermore, the description transformation process is performed on the updated personalized postoperative rehabilitation stage, specifically including:

[0064] For the first updated parameter, its corresponding time-series characteristics are determined by a sliding window;

[0065] The multimodal transformation features corresponding to the patient are obtained by fusing the correlation between the first update parameter and the second update parameter, the pre-set composite risk index, the comparison results between the temporal features and the personalized postoperative rehabilitation stage;

[0066] Based on the multimodal transformation features, corresponding prompt words are generated, and a large language model is used to perform descriptive transformation processing on the updated personalized postoperative rehabilitation stage.

[0067] The beneficial effects of adopting the above-mentioned further solutions are:

[0068] Updating personalized postoperative rehabilitation stages by incorporating multidimensional prognostic factors ensures that rehabilitation stage predictions closely align with the patient's actual condition, improving the accuracy and effectiveness of rehabilitation management. During the update process, differentiating between updating physical function parameters and psychosocial parameters allows for targeted consideration of different aspects of the patient, resulting in more accurate updated personalized postoperative rehabilitation stages output by the model.

[0069] On the other hand, this application also proposes a device for predicting the postoperative rehabilitation stage of osteoarthritis, comprising:

[0070] At least one processor; and,

[0071] A memory communicatively connected to the at least one processor; wherein,

[0072] The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform a postoperative rehabilitation stage prediction method for osteoarthritis as described in any of the above examples.

[0073] On the other hand, this application also proposes a non-volatile computer storage medium storing computer-executable instructions configured to implement the postoperative rehabilitation stage prediction method for osteoarthritis as described in any of the above examples. Attached Figure Description

[0074] Figure 1 This is a flowchart illustrating the method for predicting the postoperative rehabilitation stage of osteoarthritis in an embodiment of this application.

[0075] Figure 2 This is a schematic diagram of the model architecture of a machine learning model in one scenario of this application embodiment;

[0076] Figure 3This is a schematic diagram of the static coding module architecture in one scenario of this application embodiment;

[0077] Figure 4 This is a schematic diagram of the architecture of the timing coding module in one scenario of this application embodiment;

[0078] Figure 5 This is a schematic diagram of the architecture of the psychosocial module in one scenario of this application embodiment;

[0079] Figure 6 This is a schematic diagram of a postoperative rehabilitation stage prediction device for osteoarthritis in an embodiment of this application. Detailed Implementation

[0080] The specific embodiments of the present invention will be described in detail below. The present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed.

[0081] 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. The terminology used is for describing particular embodiments only and is not intended to limit the invention.

[0082] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0083] like Figure 1 As shown in the embodiments of this application, a method for predicting the postoperative rehabilitation stage of osteoarthritis is provided, including:

[0084] S101: Obtain the patient's corresponding multidimensional prognostic factors; the multidimensional prognostic factors include basic physical characteristic parameters, intraoperative treatment parameters, physical function parameters, and psychosocial parameters.

[0085] Prognostic factors refer to a series of specific indicators or characteristics that can predict the progression of a patient's disease after surgery. Multidimensional prognostic factors refer to multiple pre-set dimensions, including basic physical characteristics parameters, intraoperative treatment parameters, physical function parameters, and psychosocial parameters.

[0086] Specifically, basic physical characteristics parameters mainly refer to the patient's basic physical parameters, such as the patient's age, weight, gender, physical condition and comorbidities (e.g., whether it includes diabetes, cardiopulmonary diseases, and other pre-set diseases), nutritional status, and lifestyle (e.g., whether the patient smokes or drinks alcohol).

[0087] Intraoperative treatment parameters mainly refer to relevant parameters during the surgical process. For example, they may include the type and scope of surgery (e.g., platelet-rich plasma injection, extracorporeal shock wave therapy, total knee replacement, etc.), operation time and blood loss, anesthesia method (e.g., general anesthesia, regional block anesthesia, etc.), and unexpected intraoperative situations (e.g., vascular and nerve injury, etc.).

[0088] Body function parameters primarily represent parameters related to musculoskeletal function, such as range of motion, muscle strength, and pain levels. These parameters include both preoperative and postoperative data, which can be collected together and combined to obtain the final body function parameters. However, it is usually difficult to collect postoperative body function parameters immediately after surgery. In such cases, these parameters can be temporarily withheld until the body has recovered to within acceptable limits.

[0089] Psychosocial parameters primarily represent the patient's psychological thoughts and the psychological thoughts of those with close social relationships with them (such as parents, children, siblings, etc.). For example, they can include the patient's mental state (e.g., whether they are healthy, whether they experience anxiety, depression, or catastrophic thinking), self-efficacy (the patient's confidence in their recovery), and social support (the level of support from family and friends, living environment, etc.).

[0090] Data can be collected in various ways. For example, basic physical characteristic parameters and physical function parameters can be obtained manually or by exporting data from relevant hospital records and outpatient medical records. Intraoperative treatment parameters can be obtained through transmission from relevant medical sensors, in addition to hospital records. Psychosocial parameters are mainly obtained through questionnaires conducted with patients and relevant personnel.

[0091] Of course, during the data acquisition process, it is necessary to ensure that the patient's privacy is not compromised, and data collection can only proceed after obtaining the consent of the patient and relevant personnel.

[0092] S102: Based on the multidimensional prognostic factors, output the personalized postoperative rehabilitation stage corresponding to the patient through a pre-trained machine learning model.

[0093] The machine learning model is pre-trained, and the acquired multi-dimensional prognostic factors are used as input to output the personalized postoperative rehabilitation stage corresponding to the patient.

[0094] Of course, this personalized postoperative rehabilitation stage serves only as an auxiliary guide for doctors and patients. It is mainly used to help doctors develop corresponding recovery strategies based on this personalized postoperative rehabilitation stage, such as whether rest is needed, whether exercise is needed, and what precautions should be taken. It can also inform patients about the general process of subsequent postoperative rehabilitation based on their needs, helping them to make corresponding psychological expectations and prepare in advance.

[0095] Specifically, such as Figure 2 As shown, the machine learning model architecture is based on the multimodal temporally aware Transformer (MT-Transformer) architecture. The machine learning model includes an input layer, an encoding layer, a fusion layer, and an output layer. MT-Transformer refers to a modification of the original Transformer architecture, enabling the resulting machine learning model to be suitable for multi-task learning. In this embodiment, multiple tasks are executed in parallel within the encoding layer.

[0096] The input layer receives multi-dimensional prognostic factors and performs corresponding data preprocessing on these factors (e.g., data format standardization, outlier removal, etc.).

[0097] The coding layer includes parallel static coding modules, temporal coding modules, and psychosocial modules, each of which performs different tasks.

[0098] The static encoding module performs static encoding on basic body feature parameters and intraoperative treatment parameters, outputting static representation vectors. Since the data for basic body feature parameters and intraoperative treatment parameters is static and unchanging, it can be encoded using the static encoding module.

[0099] The temporal coding module dynamically encodes bodily functional parameters and outputs a temporal context vector. Bodily functional parameters undergo gradual recovery and change after surgery, which is a dynamic process. Therefore, the temporal coding module encodes these functional changes over time.

[0100] The psychosocial module introduces personalized coefficients of variation based on psychosocial parameters and outputs a smoothed perturbation sequence. Patients' psychological thoughts and social relationships are unstable factors for subsequent recovery, especially for patients with poor psychological states. Their lack of cooperation may lead to significant discrepancies between the postoperative recovery stage and expectations. Therefore, these are treated as personalized coefficients of variation and introduced into the psychosocial module to smooth the predicted personalized recovery stage.

[0101] The fusion layer fuses the static representation vector, the temporal context vector, and the smoothed perturbation sequence, and generates a personalized postoperative rehabilitation stage for the patient based on the fusion result. The personalized postoperative rehabilitation stage includes multiple sub-stages, each of which includes the corresponding duration, corresponding symptoms, and degree of recovery.

[0102] The fusion layer fuses the obtained vectors and uses the fusion results to predict the personalized postoperative rehabilitation stage. The personalized postoperative rehabilitation stage divides the postoperative rehabilitation process into multiple sub-stages, listing the duration, corresponding symptoms, and recovery level for each patient in each sub-stage. These durations, symptoms, and recovery levels are customized for each patient and are not generic sub-stage content, allowing both the attending physician and the patient to have a more accurate psychological assessment.

[0103] The output layer outputs the personalized postoperative rehabilitation stage in a preset text format.

[0104] S103: Based on the medical sensors worn by the patient, the physical function parameters are updated periodically; and based on the analysis results corresponding to the interaction content between the patient and the patient, the psychosocial parameters are updated.

[0105] Once a personalized postoperative rehabilitation stage is generated for each patient, clinicians can provide corresponding rehabilitation suggestions based on this stage, such as which sub-stages require rest and which sub-stages allow for appropriate exercise.

[0106] However, due to some special circumstances (such as resistance to exercise, failure to take medication as prescribed, etc.), the patient's recovery may not be satisfactory and may not meet the description of the personalized postoperative rehabilitation stage, or the recovery may exceed expectations.

[0107] Based on this, patients can use medical sensors to collect and update their bodily function parameters at regular intervals, such as daily. Specifically, only certain bodily function parameters can be collected based on the patient's individual circumstances. For example, range of motion of joints can be collected using wearable inertial sensors equipped with inertial measurement units and electronic goniometers; muscle strength can be collected using dynamometers; and joint pain levels can be self-assessed using appropriate pain level scales.

[0108] Simultaneously, a corresponding application is provided, which contains a large language model that can be trained in the medical field. Patients can interact with the application by asking questions such as how to use medical sensors or about medication. The application can also proactively communicate with patients periodically through questionnaires to obtain relevant information, thereby facilitating the updating of psychosocial parameters based on the interaction content.

[0109] S104: Update the personalized postoperative rehabilitation stage using the updated parameters of the multidimensional prognostic factors.

[0110] The update frequency of personalized postoperative rehabilitation stages can be consistent with the update frequency of multidimensional prognostic factors, or it can be consistent in the early postoperative sub-stages, while the update frequency can be appropriately reduced in the later sub-stages after the patient's recovery is better.

[0111] Specifically, we determine the first updated parameter for physical function parameters and the second updated parameter for psychosocial parameters among the multidimensional prognostic factors. Updated parameters refer to the updated parameters of physical function parameters and psychosocial parameters, which are categorized and described here for the parameter processing process described below.

[0112] Based on the first and second updated parameters, a machine learning model is used to re-output the updated personalized postoperative rehabilitation stage for the patient. In the machine learning model, the first and second updated parameters are used as supplementary inputs, combined with the original physical function parameters and psychosocial parameters to generate new parameters, which are then used by the machine learning model to re-output the personalized postoperative rehabilitation stage (here referred to as the updated personalized postoperative rehabilitation stage).

[0113] If the updated personalized postoperative rehabilitation stage differs more from the previous historical personalized postoperative rehabilitation stage than the first preset level, then for the current second update parameter, determine the degree of psychological stability weakening compared to the second update parameter at the time of the previous interaction.

[0114] The previous personalized postoperative rehabilitation phase will be referred to as the historical personalized postoperative rehabilitation phase. The content of the personalized postoperative rehabilitation phase is basically the same, including the duration, corresponding symptoms, and degree of recovery. Each part can be quantified here. The duration has been quantified, while the corresponding symptoms can be quantified based on the severity of different preset symptoms, and the degree of recovery can be quantified based on relevant indicators in the body function parameters.

[0115] At this point, the quantitative results of each part of the updated personalized postoperative rehabilitation stage and the historical personalized postoperative rehabilitation stage are weighted and summed to obtain a comparison result. If the difference in this comparison result is high, the degree of weakening of the patient's psychological stability is further assessed.

[0116] The degree of weakened psychological stability can be obtained by weighting the stability score and the extremity score. For ease of calculation, the stability score can be kept constant, and the patient's non-extreme score can be obtained by subtracting the extremity score from 1. The two scores are then weighted and summed to obtain the final comprehensive score of the second updated parameter. The degree of weakened psychological stability is then obtained by subtracting the historical comprehensive score from the current comprehensive score. A negative score indicates that the degree of weakened psychological stability has been reached, and the higher the absolute value of the negative value, the greater the degree of weakened psychological stability.

[0117] If the degree of weakened psychological stability exceeds the second preset level, it is considered that the patient may be experiencing psychological fluctuations due to failure to meet the pre-defined content of the personalized postoperative rehabilitation stage. In this case, a stage description conversion process is implemented to update the personalized postoperative rehabilitation stage, and an online communication channel is established between the patient and the clinician.

[0118] The phase description conversion process is primarily used to transform existing descriptions, making it easier for patients to accept the updated, personalized postoperative recovery phases and preventing emotional problems. Simultaneously, it pre-stores the contact information of both parties, establishing online communication channels (such as voice and video calls), allowing clinicians to quickly understand the patient's recovery progress.

[0119] Furthermore, during the description transformation process, for the first updated parameter, its corresponding time-series feature is determined using a sliding window. This time-series feature mainly targets bodily function parameters; for example, updated data of bodily function parameters collected over 7 days are used as the corresponding time-series feature through a sliding window.

[0120] By integrating the correlation between the first and second update parameters, the pre-set composite risk indicators, and the comparison results between temporal features and personalized postoperative rehabilitation stages, the multimodal transformation features corresponding to the patient are obtained.

[0121] The correlation degree is mainly used to calculate whether the update direction and update degree of the first update parameter and the second update parameter are consistent. As mentioned above, both the first update parameter and the second update parameter can be quantified. Therefore, the update direction and update degree of the first update parameter and the second update parameter can be determined by the difference and ratio between the current quantified score and the previous quantified score. Thus, the two can be compared to determine the correlation degree between them.

[0122] Composite risk indicators are mainly used to mark risk indicators that require extra attention and may cause special situations, such as high pain level + low joint mobility + negative psychological state, as risk indicators.

[0123] The comparison between temporal characteristics and personalized postoperative rehabilitation stages is used to describe the difference between the patient's actual recovery level and the estimated recovery level.

[0124] Based on multimodal transformation features, corresponding prompt words are generated, and a large language model is used to perform descriptive transformation processing on the updated personalized postoperative rehabilitation stages. The prompt words incorporate the multimodal transformation features and the resulting updated personalized postoperative rehabilitation stages, along with explanations. This makes the descriptions of the original updated personalized postoperative rehabilitation stages more tactful and adheres to the relevant features in the multimodal transformation features to ensure the rigor of the descriptions. This makes the descriptions more acceptable to patients and helps them maintain a stable mindset during the rehabilitation phase.

[0125] 1. By integrating multiple prognostic factors such as basic physical characteristics, intraoperative parameters, functional status, and psychosocial factors, it breaks through the limitations of traditional template-based rehabilitation stages and can generate rehabilitation stage predictions that fit each patient's individual situation, thus achieving personalized assessment of the patient's rehabilitation stage.

[0126] 2. By regularly updating physical function data through medical sensors and combining this with interactive content analysis to update psychosocial parameters, the rehabilitation phase predictions can be dynamically adjusted along with the patient's recovery process, maintaining the accuracy of the assessment and thus possessing dynamic update capabilities.

[0127] 3. It reduces reliance on frequent in-person follow-up examinations, enabling remote and continuous monitoring of changes in key patient indicators. This not only alleviates the burden of travel for patients, especially the elderly with limited mobility, but also helps maintain their enthusiasm and compliance with rehabilitation training, effectively improving the efficiency and compliance of rehabilitation management.

[0128] 4. Through automated data collection and model analysis, this solution can provide clinicians with continuous and objective decision support, reducing the burden on doctors to continuously track each patient, thereby allowing valuable medical resources to focus more on dealing with complex situations and optimizing the allocation of medical resources.

[0129] In one embodiment, a detailed explanation is provided regarding the data processing procedures, model training procedures, and other related content of each module included in the machine learning model.

[0130] Example 1: The static coding module mainly processes static data such as basic physical characteristic parameters and intraoperative treatment parameters. Its core function is to encode the patient's inherent characteristics that do not change over time into a unified representation vector.

[0131] The static encoding module obtains the data vector corresponding to each data point by standardizing and embedding retrieval processes based on the numerical and categorical features corresponding to each data point in the basic body feature parameters and intraoperative treatment parameters.

[0132] For basic physical characteristics and intraoperative treatment parameters, the static data they contain mainly includes two types: numerical features and categorical features. For example, basic physical characteristics such as age and weight, as well as intraoperative treatment parameters such as operation time and medication dosage, belong to numerical features; while basic physical characteristics such as gender and corresponding symptoms, as well as intraoperative treatment parameters such as surgery type and anesthesia method, belong to categorical features.

[0133] like Figure 3 As shown, the static encoding module includes a standardization layer and an embedding layer. The standardization layer can use a Z-score standardizer to process the input numerical features, obtaining their corresponding data vectors. The embedding layer can use multiple pre-generated embedding lookup tables to map each categorical feature to its corresponding data vector.

[0134] The concatenation layer concatenates the obtained data vectors into a comprehensive feature vector, merging different types of features into a unified representation.

[0135] A static representation vector is obtained by performing feature crossing processing on each data vector through a cross-network. The cross-network includes a feature crossing layer, a self-attention layer, and a pooling layer.

[0136] The feature cross layer contains fully connected layers and ReLU activation layers, which can automatically learn the explicit interaction relationships between features, mine combined features, and perform feature cross processing on the input comprehensive feature vector to obtain the cross feature representation.

[0137] The self-attention layer incorporates a multi-head self-attention mechanism, which captures global dependencies between features based on the cross-feature representations of the input, evaluates the importance of each feature, and outputs attention-weighted feature representations.

[0138] Pooling layers can use global average pooling to pool the attention-weighted feature representations and output the pooled feature vector, which can then be used as a static representation vector.

[0139] The static representation vector is linearly transformed to output the mapped static representation vector. The linear projection layer contains a fully connected layer without an activation function, which maps the pooled static representation vector to the target dimension, forming the final static representation vector.

[0140] Example 2: The time-series coding module mainly performs dynamic coding of bodily function parameters. Its core function is to extract and condense key dynamic rehabilitation mode information from the time-series data of bodily function parameters.

[0141] like Figure 4 As shown, the temporal coding module performs one-dimensional convolution processing on the body function parameters, outputting a surface temporal feature map. Before the one-dimensional convolution processing, data preprocessing can be performed through a data preprocessing layer. This includes slicing the input body function parameters into windows using a sliding window (referred to as sliding window slicing), dividing the continuous data stream into fixed-length analysis windows and outputting regular temporal data segments. Missing value imputation (e.g., through linear interpolation or forward padding) is then performed to address data loss, ensuring temporal continuity and outputting complete temporal data segments.

[0142] One-dimensional convolution is performed through one-dimensional convolutional layers to extract short-term patterns within a local time window from a complete input time series data segment, ensuring temporal causality and outputting a surface-level temporal feature map (also known as a local feature map).

[0143] Feature extraction from the surface temporal feature map is performed using a recurrent neural network (RNN) to output a deep temporal feature map. RNNs are artificial neural networks with internal loop connections, capable of processing sequential data, and can employ either an LSTM or Transformer architecture.

[0144] The recurrent neural network includes a GLU activation layer and an LSTM / Transformer encoder. The GLU activation layer controls the information flow through a gating mechanism, enhancing nonlinear expressive power and transforming the input surface temporal feature map into gated features. The LSTM / Transformer encoder is used to capture long-term dependencies and overall trends in the temporal data, outputting a deep temporal feature map through the input gated features.

[0145] An attention mechanism is implemented through a temporal attention layer to perform attention-weighted processing on deep temporal feature maps and output a temporal context vector. This temporal attention layer assigns importance weights to different time points, thereby focusing on key information and outputting the temporal context vector.

[0146] The temporal context vector is linearly transformed to output the mapped temporal context vector. Similar to the static coding module, the temporal coding module also includes a linear projection layer, which contains a fully connected layer without an activation function. This layer maps the temporal context vector to the target dimension, forming the final temporal context vector.

[0147] Example 3: The psychosocial module mainly targets psychosocial parameters. Its core function is to quantify the psychological state from the patient's interaction content and generate controlled random perturbations for personalized rehabilitation pathways.

[0148] like Figure 5 As shown, the psychosocial module performs sentiment analysis, keyword extraction, and scale scoring quantification on psychosocial parameters, resulting in corresponding sentiment feature vectors, keyword feature vectors, and scale score vectors. The psychosocial parameters can include interactive texts with the patient and their relatives, questionnaires completed by the patient, and other relevant content.

[0149] The system comprises three parallel layers: a sentiment analysis layer, a keyword extraction layer, and a scale scoring analysis layer, to perform corresponding data processing actions in parallel. The sentiment analysis layer can employ a combination of the BERT model and a sentiment classifier to analyze the input psychosocial parameters and output a sentiment feature vector. The keyword extraction layer can use algorithms such as TF-IDF or TextRank to identify key themes and concerns related to rehabilitation, outputting a set of keywords and forming a keyword feature vector. The scale scoring analysis layer can parse the scores of the standardized scales corresponding to the questionnaire by establishing appropriate rules, thereby quantifying psychological state and outputting a scale score vector.

[0150] By concatenating the emotional feature vector, keyword feature vector, and scale score vector, a structured psychosocial indicator tuple (also known as a comprehensive psychosocial indicator) is obtained, which can integrate multi-source psychosocial data.

[0151] The psychosocial indicator tuples are encoded to obtain the psychological state representation vector. The psychological state encoder, which includes a fully connected layer and a Tanh activation layer, encodes the psychosocial indicators into a low-dimensional dense vector through dimensionality reduction, noise reduction, and information fusion, which is referred to here as the psychological state representation vector.

[0152] It also includes a parallel random seed generator and a perturbation distribution encoder.

[0153] The random seed generator primarily maps psychological states deterministically to random number seeds. The psychological state representation vector is processed using a predefined weight matrix to obtain the original psychosocial score. This predefined weight can be determined based on factor analysis results from standard psychological resilience scales such as CD-RISC. The original psychosocial score is obtained by multiplying this predefined weight by the psychological state representation vector.

[0154] The original psychosocial scores are transformed to obtain the standard psychosocial scores; for example, they are transformed to percentiles to map the original psychosocial scores to the standard psychosocial scores in the range [0,100].

[0155] The psychosocial standard scores are successively subjected to range-fitted linear mapping, non-linear probability transformation, and clinical range-fitted linear mapping to scale the psychosocial standard scores and obtain the corresponding personalized coefficient of variation. At this point, the personalized coefficient of variation can be used as a random seed.

[0156] The conversion range adaptation linear mapping uses a sigmoid function scaling component to linearly map the psychosocial standard scores to the [-6, 6] interval, obtaining a scaled score that fits the input range of the sigmoid function. The nonlinear probability transformation can be processed using a logistic function, thus mapping the score output by the conversion range adaptation linear mapping to the [0, 1] interval. The clinical range adaptation linear mapping, based on clinical experience, maps the score output by the nonlinear probability transformation to the [0.1, 0.3] interval again, obtaining the final personalized coefficient of variation.

[0157] For the perturbation distribution encoder, it is mainly used to learn the perturbation distribution conditioned on psychological state. Generally speaking, patients are stable for most of the time during the recovery process, with occasional small fluctuations. The characteristic of Gaussian distribution is that most values ​​will be concentrated around the mean, and a few values ​​will be distributed at both ends. Therefore, the default value of the perturbation distribution can be set to Gaussian distribution, which can simulate the real situation of most patients.

[0158] Personalized variation coefficients are generated through a random seed generator, and a perturbation distribution is learned through a perturbation distribution encoder. Then, the mental state representation vector is hashed through a perturbation generation layer to obtain the corresponding integer hash coefficient. For example, the SHA-256 hash function can be used to obtain a uniquely determined integer hash coefficient.

[0159] Integer hash coefficients are used as the initial state of the pseudo-random number generator to generate a sequence of random numbers that conform to the distribution of the psychological state representation vector. Typically, the default value for the distribution of the psychological state representation vector is a Gaussian distribution. This random number sequence is not a pre-generated sequence, but rather a deterministic, Gaussian-distributed sequence of random numbers randomly generated based on the patient's actual condition (reflected by their psychological state representation vector).

[0160] Based on personalized variation coefficients, the random numbers contained in the random number sequence are scaled; for example, each random number in the random number sequence is multiplied by the personalized variation coefficient, thereby ensuring that the amplitude of the entire perturbation sequence is limited to a range determined by psychological resilience.

[0161] A smoothed perturbation sequence is obtained by smoothing the scaled random number sequence using a filter. Directly using the scaled random number sequence introduces high-frequency jitter, making the recovery path appear unnatural and difficult to execute. Therefore, smoothing can be performed using a moving average filter or a Gaussian filter to eliminate spikes and generate a smooth, slowly changing perturbation curve, which can be used as the smoothed perturbation sequence.

[0162] Furthermore, when selecting the corresponding distribution method, the perturbation distribution encoder can choose the most appropriate distribution method based on the different states of different patients, in addition to selecting the default Gaussian distribution method.

[0163] A stability score for each patient is obtained by mapping the psychological state representation vector to a trained nonlinear activation function. The nonlinear activation function can be the ReLU function, which maps the vector to a scalar value and then compresses it to the [0,1] interval using a Sigmoid function. A higher stability score indicates a more stable psychological state.

[0164] Meanwhile, during training, the parameters of the ReLU and Sigmoid functions can be adjusted through indirect supervised learning. For example, in the training set, if a patient's daily functional parameters (such as pain scores and range of motion) fluctuate less, their stability score should be higher; conversely, it should be lower.

[0165] The extreme score for each patient is obtained based on the distance between their mental state representation vector and a pre-generated baseline representation vector. The baseline representation vector can be obtained by averaging the mental state representation vectors of all patients in the training set. The distance can be calculated using cosine distance, and after normalization, the extreme score is obtained.

[0166] Based on the stability score and extreme score, the corresponding distribution method for the psychological state representation vector is selected from several preset distribution methods. The stability score and extreme score represent the patient's psychological state from two dimensions; therefore, the distribution method can be selected based on these two dimensions to match the patient's actual situation.

[0167] Specifically, if the stability score is higher than the first stability threshold (e.g., set to 0.7) and the extrema score is lower than the first extreme threshold (e.g., set to 0.3), the patient's psychological state is considered stable, and the recovery process is expected to be smooth. Among several preset distribution methods, a narrow-peak Gaussian distribution is selected. The narrow-peak Gaussian distribution (which has the characteristics of a mean of 0 and small variance) ensures that the generated perturbations are close to zero most of the time, simulating a "stable for most of the time" pattern with occasional small fluctuations.

[0168] If the stability score is below the second stability threshold (e.g., set to 0.4), or the extremism score is above the second extreme threshold (e.g., set to 0.6), the patient's psychological state is considered to fluctuate greatly or deviate from the norm. The recovery process may have a short period of stability and be prone to large fluctuations. Among several preset distribution methods, a heavy-tailed distribution (e.g., Student's t-distribution) is selected. The heavy-tailed distribution allows for larger perturbations, simulating the real situation of "occasional large fluctuations".

[0169] Otherwise, in other situations, the patient's condition is considered complex, possibly stable overall but occasionally distracted. Among the preset distribution options, a Gaussian mixture distribution is selected. The Gaussian mixture distribution is composed of two Gaussian distributions: a principal distribution (with high weights and low variance) to simulate a stable baseline, and an auxiliary distribution (with low weights and high variance) to simulate occasional fluctuations. The mixture weights can be dynamically adjusted by the stability score and the extreme score.

[0170] In one embodiment, a specific example is given to describe the prediction and updating process of personalized postoperative rehabilitation stages. To protect patient privacy, only some of the parameters are given here.

[0171] This patient suffers from knee joint disease, and its multidimensional prognostic factors are shown in Tables 1-4:

[0172] Table 1: Basic Physical Characteristic Parameters

[0173] age weight BMI physical condition Nutritional status lifestyle 58 85KG 28.7 (overweight) High blood pressure, no diabetes Moderate, insufficient protein intake Light smoker, office worker

[0174] Table 2: Intraoperative Treatment Parameters

[0175] Surgical type Surgery duration Bleeding Anesthesia Intraoperative complications Arthroscopic partial meniscectomy and chondroplasty 90 minutes 50ml spinal anesthesia none

[0176] Table 3: Body Function Parameters Table

[0177] Range of motion Muscle strength Pain level Flexion 0-115 degrees, extension -5 degrees Quadriceps strength grade 4 / 5, hamstring strength grade 4+ / 5 VAS score 6 (while walking), WOMAC pain subscale score 12 / 20

[0178] Table 4: Psychosocial Parameter Table

[0179] Psychological state Self-efficacy Social support Moderate anxiety (GAD-7 score 10) Low (ASES score 45 / 100) His wife is very supportive, but his children live in another city, so he spends a lot of time alone.

[0180] Based on this, a personalized postoperative rehabilitation phase was generated on the first day after surgery, as shown in Table 5:

[0181] Table 5: Personalized Postoperative Rehabilitation Stages

[0182] sub-stage time Corresponding symptoms degree of recovery Phase 1: Protection Period Weeks 1-2 1. Moderate pain (VAS 4-6) 2. Significant swelling (2+) 3. Significant joint stiffness 4. Nighttime pain affecting sleep 1. Pain control: VAS score reduced to below 4. 2. Swelling: Reduced to grade 1+. 3. Range of motion: Flexion up to 90 degrees. Phase 2: Early Recovery Period Weeks 3-6 1. Mild to moderate pain (VAS 2-4 points) 2. Swelling after activity 3. Muscle atrophy begins to appear 4. Poor walking endurance 1. Range of motion: Flexion reaches 120 degrees. 2. Muscle strength: Quadriceps strength recovered to grade 4 / 5. 3. Function: Can walk short distances without assistance. Phase 3: Functional Enhancement Period Weeks 7-12 1. Occasional pain (VAS 0-2 points) 2. Difficulty climbing stairs 3. Fatigue after prolonged walking 1. Range of motion: Fully recovered. 2. Muscle strength: Recovered to grade 4+ / 5. 3. Function: Can climb stairs normally.

[0183] On the 7th day after surgery, the updated parameters for physical function and psychosocial parameters are shown in Tables 6 and 7 below:

[0184] Table 6: Update Table of Body Function Parameters

[0185] Range of motion Muscle strength Pain level Flexion 0-75 degrees (slow progress) Quadriceps muscle strength grade 2 / 5 (significantly inhibited) VAS score: 5 (resting), 7 (active), and swelling level: 2+ (no significant improvement).

[0186] Table 7: Update Table of Psychosocial Parameters

[0187] Psychological state Self-efficacy Social support Worsening anxiety (GAD-7 score 14), with depressive mood. Significant decrease (ASES score 32 / 100) No change

[0188] Therefore, it can be seen that the patient did not reach the initial recovery status of the personalized postoperative rehabilitation stage due to certain circumstances. At this point, the personalized postoperative rehabilitation stage needs to be updated, as shown in Table 8 below:

[0189] Table 8: Update Table of Personalized Postoperative Rehabilitation Stages

[0190] sub-stage time Corresponding symptoms degree of recovery Phase 1: Protection Period Weeks 1-3 (extended by 1 week) 1. Moderate pain (VAS 5-7) (improving VAS score) 2. Significant muscle inhibition (new) 3. Low mood affecting rehabilitation motivation (new) 1. Pain control: VAS score decreased to below 5 (improving VAS score) 2. Range of motion: flexion reached 75 degrees (reducing flexion) 3. Anxiety score decreased to below 10 (new addition)

[0191] Since the patient is currently in stage 1, updates can be made only for stage 1. Adaptive adjustments can be made for stages 2 and 3, or a reassessment can be conducted after stage 1 is completed. The personalized postoperative rehabilitation stage can also be updated based on the patient's updated parameters, which is similar to the update method in Table 8, so it will not be elaborated here.

[0192] like Figure 6 As shown in the embodiment of this application, a device for predicting the postoperative rehabilitation stage of osteoarthritis is also provided, including:

[0193] At least one processor; and,

[0194] A memory communicatively connected to the at least one processor; wherein,

[0195] The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the postoperative rehabilitation stage prediction method for osteoarthritis as described in any of the above embodiments.

[0196] This application also provides a non-volatile computer storage medium storing computer-executable instructions configured to implement the postoperative rehabilitation stage prediction method for osteoarthritis as described in any of the above embodiments.

[0197] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device and medium embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the description of the method embodiments.

[0198] The devices and media provided in this application are one-to-one with the methods. Therefore, the devices and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.

[0199] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are exhaustively listed. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0200] For those skilled in the art, various modifications and improvements can be made without departing from the concept of the present invention, and these modifications and improvements are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the appended claims.

Claims

1. A method of estimating a postoperative rehabilitation phase for osteoarthritis, characterized in that, The method comprises the following steps: acquiring a plurality of dimensional prognostic factors corresponding to a patient; the plurality of dimensional prognostic factors comprise basic physical characteristic parameters, intraoperative treatment parameters, physical function parameters, and psychosocial parameters; outputting, by a pre-trained machine learning model, a personalized postoperative rehabilitation stage corresponding to the patient according to the plurality of dimensional prognostic factors; updating the physical function parameters on a regular basis based on a medical sensor worn by the patient, and updating the psychosocial parameters based on analysis results corresponding to interactive content with the patient; updating the personalized postoperative rehabilitation stage by updated parameters of the plurality of dimensional prognostic factors; a model architecture of the machine learning model is based on a multi-modal time series perception Transformer architecture, and the machine learning model comprises an input layer, an encoding layer, a fusion layer, and an output layer; the input layer receives the plurality of dimensional prognostic factors; the encoding layer comprises a static encoding module, a time series encoding module, and a psychosocial module arranged side by side; the static encoding module statically encodes the basic physical characteristic parameters and the intraoperative treatment parameters, and outputs a static representation vector; the time series encoding module dynamically encodes the physical function parameters, and outputs a time series context vector; the psychosocial module introduces a personalized variation coefficient according to the psychosocial parameters, and outputs a smoothed disturbance sequence; the fusion layer fuses the static representation vector, the time series context vector, and the smoothed disturbance sequence, and generates the personalized postoperative rehabilitation stage corresponding to the patient according to a fusion result; the personalized postoperative rehabilitation stage comprises a plurality of sub-stages, and each sub-stage comprises a corresponding duration, a corresponding symptom, and a recovery degree; the output layer outputs the personalized postoperative rehabilitation stage; the psychosocial module introduces a personalized variation coefficient according to the psychosocial parameters, and outputs a smoothed disturbance sequence, and specifically comprises: the psychosocial module respectively performs sentiment analysis, keyword extraction, and scale score quantification on the psychosocial parameters, and respectively obtains a corresponding sentiment feature vector, a keyword feature vector, and a scale score vector; wherein the sentiment feature vector is obtained by inputting the psychosocial parameters into a BERT model and a sentiment classification head for sentiment analysis; the keyword feature vector is obtained by identifying key themes and concerns related to rehabilitation based on a TF-IDF algorithm or a TextRank algorithm, and outputting a keyword set; and the scale score vector is obtained by quantifying a psychological state by analyzing scores of a standardized scale corresponding to a questionnaire according to corresponding rules; the sentiment feature vector, the keyword feature vector, and the scale score vector are spliced to obtain a structured psychosocial index tuple; the psychosocial index tuple is encoded to obtain a psychological state representation vector; wherein a psychological state encoder comprises a fully connected layer and a Tanh activation layer. The method comprises the following steps: setting a parallel random seed generator and a perturbation distribution encoder; the random seed generator determines the psychological state and maps it into a random number seed, generating a personalized variation coefficient; the perturbation distribution encoder is used for learning the perturbation distribution conditioned on the psychological state; The psychological state representation vector is processed through a predefined weight matrix to obtain a psychosocial original score; the psychosocial original score is subjected to bit conversion processing to obtain a psychosocial standard score; the psychosocial standard score is sequentially subjected to conversion range adaptation linear mapping, nonlinear probability conversion and clinical range adaptation linear mapping to scale the psychosocial standard score to obtain a corresponding personalized variation coefficient; The psychological state representation vector is hashed to obtain a corresponding integer type hash coefficient; the integer type hash coefficient is used as the initial state of a pseudo-random number generator to generate a random number sequence conforming to the corresponding distribution mode of the psychological state representation vector; Based on the personalized variation coefficient, each random number contained in the random number sequence is scaled; the scaled random number sequence is smoothed through a filter to obtain a smoothed perturbation sequence; Before the integer type hash coefficient is used as the initial state of the pseudo-random number generator to generate a random number sequence conforming to the corresponding distribution mode of the psychological state representation vector, the method further comprises the following steps: The psychological state representation vector is mapped through a trained nonlinear activation function to obtain a stability score corresponding to the patient; and based on the distance value between the psychological state representation vector and a pre-generated reference representation vector, an extremity score corresponding to the patient is obtained; wherein the reference representation vector is obtained by averaging the psychological state representation vectors of each patient in the training set; According to the stability score and the extremity score, a corresponding distribution mode of the psychological state representation vector is selected from a plurality of preset distribution modes; If the stability score is higher than a first stability threshold and the extremity score is lower than a first extremity threshold, a narrow peak Gaussian distribution is selected from the plurality of preset distribution modes; If the stability score is lower than a second stability threshold or the extremity score is higher than a second extremity threshold, a heavy-tailed distribution is selected from the plurality of preset distribution modes; Otherwise, a Gaussian mixture distribution is selected from the plurality of preset distribution modes.

2. The method of claim 1, wherein, The static encoding module statically encodes the basic body feature parameters and the intraoperative treatment parameters to output a static representation vector, specifically comprising the following steps: The static encoding module respectively processes the numerical features and the categorical features corresponding to each data in the basic body feature parameters and the intraoperative treatment parameters through standardization processing and embedding retrieval processing to obtain a data vector corresponding to each data; The feature cross processing is performed on each data vector through a cross network to obtain a static representation vector; The static representation vector is subjected to linear transformation to output a mapped static representation vector.

3. The method of claim 1, wherein, The time series encoding module dynamically encodes the body function parameters to output a time series context vector, specifically comprising the following steps: The time sequence coding module performs one-dimensional convolution processing on the body function parameter, and outputs a surface time sequence feature map; The surface time sequence feature map is subjected to feature extraction through a recurrent neural network, and a deep time sequence feature map is outputted; The deep time sequence feature map is subjected to attention weighting processing through an attention mechanism, and a time sequence context vector is outputted; The time sequence context vector is subjected to linear transformation, and a mapped time sequence context vector is outputted.

4. The method of claim 1, wherein, The personalized postoperative rehabilitation stage is updated through the updated parameters of the multi-dimensional prognostic factors, specifically including: The first update parameter for the body function parameter and the second update parameter for the psychosocial parameter in the multi-dimensional prognostic factors are determined; The first update parameter and the second update parameter are used to re-output the updated personalized postoperative rehabilitation stage of the patient through the machine learning model; If the difference between the updated personalized postoperative rehabilitation stage and the previous historical personalized postoperative rehabilitation stage is higher than a first preset degree, the current second update parameter is determined to have a psychological stability weakening degree compared with the second update parameter at the last interaction; If the psychological stability weakening degree is higher than a second preset degree, the updated personalized postoperative rehabilitation stage is subjected to stage description conversion processing, and an online communication channel between the patient and the clinician is established.

5. The method of claim 4, wherein, The description conversion processing is performed on the updated personalized postoperative rehabilitation stage, specifically including: The first update parameter is subjected to time sequence feature determination through a sliding window; The multi-modal conversion feature corresponding to the patient is fused through the correlation degree between the first update parameter and the second update parameter, the pre-set composite risk indicator, and the comparison result between the time sequence feature and the personalized postoperative rehabilitation stage; The corresponding prompt word is generated based on the multi-modal conversion feature, and the description conversion processing is performed on the updated personalized postoperative rehabilitation stage through a large language model.

6. An apparatus for estimating a postoperative rehabilitation phase for osteoarthritis, characterized in that, It includes: At least one processor; And The memory is in communication connection with the at least one processor; wherein The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the postoperative rehabilitation stage estimation method for osteoarthritis as claimed in any one of claims 1-5.

7. A non-transitory computer storage medium storing computer-executable instructions that, when executed, cause a computer to perform: The computer executable instructions are configured to implement the postoperative rehabilitation stage estimation method for osteoarthritis as claimed in any one of claims 1-5.

Images