An artificial intelligence-based osteoarthritis management system and computer storage medium

Abstract

This invention belongs to the field of osteoarthritis management technology, specifically relating to an artificial intelligence-based osteoarthritis management system and a computer storage medium. This invention provides an artificial intelligence-based osteoarthritis management system. This system integrates an LLM-DL hybrid architecture, a multi-specialty module collaboration framework, and a RAG knowledge base system, covering the entire clinical process from initial patient diagnosis to treatment plan development. This system can comprehensively and accurately assess the patient's disease status, providing personalized disease progression predictions. It not only predicts 2-year and 4-year functional outcomes and radiological progression but also identifies patient-specific risk factors, providing a basis for precise intervention. The overall system design significantly improves clinical work efficiency and predictive accuracy, reduces reliance on specialist physicians and high-end equipment, enabling high-quality KOA management to be implemented even in resource-limited areas, thereby effectively reducing overall healthcare resource demands.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of osteoarthritis management, and particularly relates to an osteoarthritis management system based on artificial intelligence and a computer storage medium. BACKGROUND

[0002] Knee osteoarthritis (KOA) is a chronic degenerative disease that affects about 600 million people worldwide. Its pathological features include progressive degeneration of articular cartilage, subchondral bone sclerosis, osteophyte formation and joint space narrowing, which leads to persistent pain, joint stiffness and limited mobility in patients, and in severe cases, total knee replacement is required. With the global population aging, KOA has brought a huge social and economic burden, including high medical costs, labor loss and long-term care needs. Due to the significant heterogeneity in the progression of the disease in patients, which is influenced by factors such as age, body mass index, muscle strength, personalized intervention strategies are crucial for delaying disease progression and improving patient outcomes.

[0003] However, implementing personalized multidisciplinary KOA management in large populations presents significant challenges, particularly in resource-limited areas. Comprehensive KOA management requires detailed clinical assessment, multidisciplinary team collaboration and personalized program development, which far exceed the carrying capacity of current healthcare systems. There is a shortage of specialist doctors and inadequate diagnostic equipment in primary care institutions, resulting in a large number of patients unable to receive timely and appropriate treatment, ultimately requiring more expensive and complex interventions.

[0004] Currently, AI technologies in the field of KOA management can be mainly divided into three categories. The first category is single-function AI systems, such as the 2D U-net model developed by Norman et al., which achieved a Dice coefficient of 0.92 on more than 3,000 MRI images and can accurately segment the femoral and tibial cartilage and meniscus; the gradient boosting tree model by Fan et al., which can predict radiographic osteoarthritis 5-10 years before the onset of the disease by integrating clinical data and MRI indicators; and various deep learning-based Kellgren-Lawrence (KL) grading systems. These systems perform well on specific tasks, but their functions are relatively limited, and they can only solve a specific link in the KOA management process. The second category is general visual language models, such as GPT-4V and Gemini. Although they have shown strong zero-shot generalization ability in open-domain tasks, they perform poorly in specialized medical image recognition tasks, especially in tasks that require professional knowledge, such as KOA severity grading, where the accuracy is often less than 50%. The third category is single retrieval augmentation generation (RAG) systems, which enhance the professionalism of language models by connecting to external medical knowledge bases. However, existing implementations mostly adopt simple retrieval augmentation generation architectures, which suffer from semantic drift problems. The retrieved content does not match the query context well, and it lacks multidisciplinary integrated decision-making capabilities, making it difficult to generate comprehensive personalized treatment plans.

[0005] In summary, current AI systems in the field of KOA management have the following problems: (1) There is a general problem of functional fragmentation, with each system only optimizing a specific link in the KOA management process; (2) They cannot support the complete clinical process, lack end-to-end solutions, and still require a lot of manual operation to connect the various links; (3) The accuracy of existing general visual language models in specialized medical tasks such as KOA severity grading, joint space assessment, and osteophyte identification is often less than 65%, and some are even close to random guessing. They perform well in identifying normal or mild cases, but the error rate is extremely high in moderate to severe cases that require precise judgment; (4) Existing AI systems cannot customize personalized management plans based on patient characteristics. However, KOA patients have different age, gender, weight, activity level, and other characteristics. There are huge differences in the level, comorbidities, previous treatment history, psychological state and socioeconomic conditions. These factors will affect the disease progression and treatment response. For example, for patients with the same KL2 level, the treatment plan for young athletes and elderly sedentary patients should be fundamentally different. However, the existing system is difficult to make such a fine distinction, resulting in a lack of targeted and practical treatment recommendations. (5) The optimized management of KOA requires the cooperation of multiple disciplines such as orthopedics, rehabilitation medicine, nutrition, and psychology. However, the existing AI system lacks an effective multi-specialty module collaboration mechanism. For example, the exercise prescription may not match the patient's nutritional status, or the drug treatment plan may ignore the patient's psychological factors. The AI ​​system without a collaboration mechanism cannot provide a truly comprehensive and coordinated integrated management plan, and it is difficult to meet the needs of chronic complex disease management.

[0006] Therefore, developing a personalized KOA precision management method that covers the entire clinical process and accurately identifies clinical indicators is a challenge in this field. Summary of the Invention

[0007] In view of the shortcomings of the prior art, the present invention provides an osteoarthritis management system and a computer storage medium based on artificial intelligence.

[0008] This invention provides an artificial intelligence-based osteoarthritis management system, which includes the following modules:

[0009] The disease assessment module is configured to collect patient information through natural language interaction, automatically analyze X-ray images, and generate a disease status assessment report.

[0010] The progression prediction module is configured to allow the disease status assessment report to predict results and radiological progression through a predictive model output function.

[0011] The treatment planning module is configured such that the multi-specialty module generates personalized comprehensive treatment plans through a multi-objective optimization decision engine.

[0012] Preferably, the progression prediction module is configured to use demographic information from the disease status assessment report, including body mass index, clinical data including clinical scores and normalized muscle strength, and imaging depth features, to output 2-4 year functional prediction results through prediction model 1, and to output 2-4 year imaging progression through prediction model 2.

[0013] The normalized muscle strength is the maximum muscle strength of the thigh or calf, either forward or backward, divided by the body mass index.

[0014] Prediction Model 1 is a model obtained by ensemble learning of multiple machine learning models, while Prediction Model 2 is a multi-task learning model that combines deep learning and traditional machine learning.

[0015] Preferably, the disease assessment module integrates a patient interaction module, an X-ray analysis module, and a reporting module; the medical history report obtained by the patient interaction module and the osteoarthritis severity grading obtained by the X-ray analysis module are input into the reporting module to generate a disease status assessment report.

[0016] Preferably, in the disease assessment module:

[0017] The patient interaction module is configured to collect patients' medical history information based on the Qwen-Max large language model and generate medical history reports.

[0018] The X-ray analysis module is configured to use the U-Net model to locate and segment the knee joint in X-ray images, and use the ResNet-50 model to grade the severity of osteoarthritis in the located and segmented X-ray images, and extract fine-grained radiological features.

[0019] The reporting module is configured to integrate the medical history report obtained from the patient interaction module and the osteoarthritis severity obtained from the X-ray analysis module through multimodal feature fusion technology, perform symptom-image consistency calibration, and finally use the Qwen-Max model to summarize and generate a disease status assessment report.

[0020] And / or, the severity grading of the bone and joint is defined as a four-level classification: none / suspected, mild, moderate, and severe.

[0021] Preferably, the progress prediction module integrates a functional result prediction module, an imaging progress prediction module, and a risk analysis module;

[0022] The input features and prediction results from the functional outcome prediction module and the imaging progression prediction module are fed into the risk analysis module to obtain patient-specific risk factors.

[0023] And / or, in the progression prediction module, when outputting the radiographic progression in the 4th year, the clinical data in the features input to the prediction model 2 are: clinical score and original muscle strength.

[0024] Preferably, the content of the functional prediction results includes: scores for pain, symptoms, activities of daily living, motor and recreational function, and quality of life on five subscales for knee joint injury and osteoarthritis outcome scoring;

[0025] The ensemble learning strategy described in prediction model 1 is a two-layer stacked ensemble strategy. The first layer completes the continuous regression task through base learners, and the second layer uses elastic network regression as a meta-learner to perform weighted fusion of the outputs of the base learners. The weighted fusion process includes L1 / L2 regularization. The base learners include: gradient boosting decision tree, random forest, support vector machine and multilayer perceptron.

[0026] And / or, in the imaging progression prediction module, the content of the imaging progression over 2-4 years includes: osteoarthritis severity grading over 2-4 years;

[0027] The traditional machine learning algorithm in prediction module 2 is XGboost, and the deep learning algorithm is a convolutional neural network.

[0028] The imaging progression prediction module predicts the left and / or right knee joints;

[0029] And / or, in the progression prediction module, the demographic information is: age, sex, body mass index, activity level; the clinical data is: KOOS / WOMAC, VAS clinical scores, normalized muscle strength, primitive muscle strength; and the imaging depth features are bone and joint severity grading and fine-grained imaging features.

[0030] And / or, the risk analysis module quantifies the contribution of each input feature to the prediction result using SHapley Additive exPlanations analysis technology; the higher the contribution, the more important the corresponding input feature is for progression prediction, and the top five input features in terms of contribution are taken as patient-specific risk factors.

[0031] Preferably, in the treatment planning module, the specialty modules include a sports rehabilitation module, an orthopedics module, a psychological nutrition module, and a clinical decision-making module; the specialty suggestions generated by the rehabilitation module, the orthopedics module, and the psychological nutrition module are entered into the clinical decision-making module.

[0032] Preferably, in the treatment planning module,

[0033] The sports rehabilitation module is configured to use patients' demographic characteristics, objective ability indicators related to exercise, imaging characteristics, comorbidities, contraindications, and patient-specific risk factors as multi-source patient data. Using structured prompts with a fixed template, the module takes the multi-source patient data and key points of sports medicine and rehabilitation guidelines obtained through vectorized retrieval as "evidence context" input. Based on the FITT-VP principle, the module outputs an exercise prescription through a large language model. The prescription is adjusted item by item according to the patient's goals, pain threshold, and resource accessibility, serving as the specialist recommendations for the sports rehabilitation module.

[0034] The orthopedics module is configured to use patients' demographic characteristics, clinical scores related to osteoarthritis or pain, imaging characteristics, joint range of motion, history of conservative treatment and comorbidities, perioperative risks, and patient-specific risk factors as multi-source patient data. It performs standardization and missing value imputation, uses a rule engine to score surgical indications, automatically identifies indications and contraindications, and uses a large language model combined with vectorized retrieval to call orthopedic guidelines and evidence-based literature as context. Based on the surgical indication score and patient target weight, it outputs the preferred surgical procedure and alternative options, generates evidence-based analgesic and anti-inflammatory drug regimens, and perioperative management recommendations as specialist recommendations for the orthopedics module.

[0035] The psychological nutrition module is configured to use patients' demographic information, clinical assessment scales for mental health and quality of life, dietary habit records, body composition analysis, metabolic indicators, pain-related behavioral data, and patient-specific risk factors as multi-source patient data. After standardization and missing value imputation, the data is automatically stratified using psychological risk screening rules. A large language model combined with vectorized retrieval is used to generate individualized psychological intervention plans and nutritional programs as specialist recommendations for the psychological nutrition module.

[0036] The clinical decision-making module is configured to generate personalized comprehensive treatment plans from the specialist recommendations obtained from the sports rehabilitation module, orthopedics module, and psychological nutrition module through a multi-objective optimization decision engine based on evidence weighting and conflict resolution.

[0037] Preferably, in the clinical decision-making module, the data processing process of the multi-objective optimization decision engine based on evidence weighting and conflict resolution includes: mapping and semantically aligning the outputs of different specialty modules through a standardized parser; using a rule engine and multi-objective optimization algorithm for conflict detection and priority ranking; using constraint solving and weight harmonization methods to comprehensively balance multi-dimensional objectives, prioritizing the patient's primary objectives and safety constraints; coordinating suggestion conflicts between different specialty modules; calling a large language model for natural language reasoning and scheme reorganization when necessary; and automatically generating execution order and stage objectives based on the patient's comorbidities, available resources, risk level, and timeline.

[0038] The personalized comprehensive treatment plan includes: treatment sequence arrangement, specific actions and training plans, surgical and drug pathways, psychological intervention rhythm, dietary and lifestyle recommendations, conflict resolution instructions and evidence-based basis;

[0039] And / or, the specialty module performs vectorized retrieval through a vectorized retrieval knowledge base system. The construction process of the vectorized retrieval knowledge base includes: document screening, information extraction, structured storage, and vectorized indexing.

[0040] The present invention provides a computer storage medium storing a computer program, which, when executed by a processor, implements the artificial intelligence-based osteoarthritis management system described above.

[0041] This invention provides an end-to-end KOA precision management solution by screening and optimizing models and input features. This solution covers the entire clinical process from initial patient diagnosis to treatment plan development. Through automated processing and analysis of multimodal data, including structured collection of patient medical history, intelligent interpretation of knee X-rays, and automatic calculation of clinical scores, the system can comprehensively assess the patient's disease status. Based on machine learning algorithms and SHAP interpretability analysis, the system can provide accurate and personalized disease progression predictions, accurately predicting functional outcomes and radiological progression at 2 and 4 years, and identifying patient-specific risk factors, providing a basis for precision intervention. By simulating multidisciplinary team discussions through a multi-specialty module collaborative architecture, it integrates multiple dimensions such as sports rehabilitation, drug therapy, surgical intervention, nutritional management, and psychological support to generate comprehensive treatment plans based on evidence-based medicine. The overall system design significantly improves clinical efficiency, reducing the time for complete assessment and plan development by 38.5%. Simultaneously, through automation and intelligence, it reduces reliance on specialist physicians and high-end equipment, enabling high-quality KOA management to be implemented even in resource-limited areas, thereby effectively reducing overall healthcare resource demands.

[0042] Obviously, based on the above description of the present invention, and according to common technical knowledge and conventional methods in the field, various other modifications, substitutions or alterations can be made without departing from the basic technical concept of the present invention.

[0043] The following detailed embodiments further illustrate the above-described content of the present invention. However, this should not be construed as limiting the scope of the present invention to the following examples. All technologies implemented based on the above-described content of the present invention fall within the scope of the present invention. Attached Figure Description

[0044] Figure 1 This is a data processing flowchart of the artificial intelligence-based osteoarthritis management system of Embodiment 1 of the present invention. Detailed Implementation

[0045] In the following embodiments and experimental examples, the algorithms for data acquisition, transmission, storage, and processing steps not specifically described, as well as the hardware structures and circuit connections not specifically described, can all be implemented using the content already disclosed in the prior art.

[0046] Example 1: An Artificial Intelligence-Based Osteoarthritis Management System

[0047] like Figure 1 As shown in the figure, the AI-based osteoarthritis management system in this embodiment integrates a disease assessment module, a progression prediction module, and a treatment planning module.

[0048] The disease assessment module is responsible for collecting patient information through natural language interaction and automatically analyzing X-ray images to generate a comprehensive disease status assessment report.

[0049] The progression prediction module predicts the patient's functional outcomes and radiological progression based on the disease status assessment report obtained from the disease assessment module and the patient's medical history information.

[0050] The treatment planning module includes multiple specialty modules, which generate comprehensive treatment plans through simulated multidisciplinary team discussions.

[0051] This layered design not only improves the modularity and maintainability of the system, but also ensures effective information flow and collaborative work between the functional modules.

[0052] The core technical components of the AI-based osteoarthritis management system in this embodiment include an LLM-DL hybrid architecture, a multi-specialty module collaboration framework, and a RAG knowledge base system. The LLM-DL hybrid architecture combines the advantages of large language models in natural language understanding and generation with the expertise of deep learning models in image analysis and numerical prediction, achieving efficient processing of multimodal data. The multi-specialty module collaboration framework employs a message-passing-based communication mechanism, supporting asynchronous interaction and decision coordination between specialty modules and ensuring the organic integration of different specialty perspectives. The RAG knowledge base system stores and retrieves professional knowledge such as clinical guidelines and research literature through a vector database, providing reliable evidence support for the decision-making of specialty modules. Simultaneously, it optimizes retrieval accuracy through semantic similarity matching and re-ranking algorithms, solving the semantic drift problem of traditional RAG systems.

[0053] The following is a separate explanation of each module:

[0054] 1. The disease assessment module is configured to collect patient information through natural language interaction, automatically analyze X-ray images, and generate a disease status assessment report.

[0055] This disease assessment module integrates a patient interaction module, an X-ray analysis module, and a reporting module.

[0056] (1.1) The patient interaction module is configured to collect information in a professional and user-friendly conversational manner based on the Qwen-Max large language model through a carefully designed prompting process. Specifically, it adopts a structured dialogue process to systematically guide patients to provide medical history information, including key clinical information such as the onset time, nature, aggravating or relieving factors, and degree of functional limitation of knee pain; it automatically identifies and extracts key information through natural language understanding technology and converts it into structured data storage in real time; during the dialogue, it dynamically adjusts the questions based on the patient's answers to ensure that complete and accurate clinical information is collected; after the information collection is completed, it automatically generates a medical history report that conforms to clinical standards through natural language generation technology, including standard chapters such as chief complaint, present illness, past medical history, and physical examination, providing comprehensive basic data for subsequent clinical decision-making.

[0057] (1.2) The X-ray analysis module is configured to achieve accurate radiological assessment using a two-stage deep learning architecture. The first stage uses an improved U-Net model to locate and segment the knee joint in X-ray images of the patient's knee. This model, trained on 5,961 labeled images, achieved a localization accuracy of 98.7%, automatically identifying and cropping the standard knee joint region, eliminating the influence of changes in shooting angle and position. The second stage uses a ResNet-50 model to perform KOA severity grading on the localized and segmented X-ray images from the first stage. The model achieved a classification accuracy of 82.4% on the data-enhanced training set, significantly outperforming general visual models. In addition to overall knee osteoarthritis severity grading, the ResNet-50 model also extracts 11 fine-grained radiological features from the localized and segmented X-ray images from the first stage, including the degree of narrowing of the medial and lateral joint spaces, the size of femoral and tibial osteophytes, and subchondral sclerosis. These features not only provide more detailed pathological information but also provide important input variables for the progression prediction module.

[0058] This embodiment redefines the severity grading of KOA as a four-level classification system more in line with clinical decision-making: None / Doubt, Mild, Moderate, and Severe. Existing literature typically uses the KL0-4 grading system directly as the classification target. This embodiment merges KL0 and KL1. The experimental data after merging KL0 and KL1 are as follows:

[0059] In the experiment, after merging KL0 and KL1, the overall accuracy of the model on the same test set increased from 56% to 81% in the original five-level classification; in the confusion matrix, the original KL0... The errors caused by KL1 mixing accounted for the majority of all errors and were almost eliminated after merging; the consistency coefficient (weighted Cohen's κ) increased from approximately 0.55-0.60 before merging to 0.75-0.80; the mean squared error (MSE) decreased from approximately 0.90-1.10 before merging to 0.45-0.60; the Brier score also decreased from approximately 0.35-0.40 to 0.20-0.25; and the label entropy (representing prediction uncertainty) decreased by approximately 30-40% in the "None / Doubt" interval. Therefore, by merging KL0 and KL1 in this invention, noise and uncertainty were reduced, the consistency between the model prediction results and actual clinical needs was improved, and the output became more instructive.

[0060] In this module, an adjustable window level / window width normalization method and pseudo-three-channel stacking are introduced during data processing. These are used in conjunction with data augmentation (rotation, flipping, jittering) and normalization. This method can quickly adapt to different devices, body parts, and datasets without changing the backbone network. This improves the contrast of bone cortex and osteophyte details while maintaining the network structure, thus enhancing the model's transferability and robustness.

[0061] In terms of parameter settings, this embodiment uses a 224-size input, employs Adam in conjunction with StepLR learning rate scheduling for the optimizer, and introduces early stopping and weight rollback mechanisms; the entire process can be completed in a single GPU environment. This setup has lower computational power dependence and is easier to deploy.

[0062] (1.3) The reporting module is configured to integrate the medical history report obtained from the patient interaction module and the osteoarthritis severity obtained from the X-ray analysis module using multimodal feature fusion technology, perform symptom-image consistency calibration, and finally use the Qwen-Max model to summarize and generate a comprehensive disease status assessment report. This disease status assessment report integrates the patient's subjective symptoms and objective imaging manifestations, solving the problem that the severity of imaging may not be consistent with the patient's subjective symptoms, and is closer to the actual clinical decision-making process.

[0063] 2. The progression prediction module is configured to output functional prediction results and radiological progression for years 2 and 4 based on the disease status assessment report obtained from the disease assessment module through a prediction model. It integrates a functional outcome prediction module, a radiological progression prediction module, and a risk analysis module.

[0064] This progress prediction module introduces a parameter—normalized muscle strength or raw muscle strength. Raw muscle strength refers to the maximum forward or backward muscle strength of the thigh and calf. Normalized muscle strength is the maximum forward or backward muscle strength of the thigh and calf divided by the individual's BMI value, used as the normalization parameter. The formula for calculating BMI (Body Mass Index) is: BMI = weight / height. 2 .

[0065] (2.1) The functional outcome prediction module is configured to input multidimensional data such as demographic information (age, gender, BMI, activity level), clinical data (KOOS / WOMAC, VAS clinical scores, normalized muscle strength, primitive muscle strength), and imaging depth features (KOA severity grading, fine-grained imaging features) from the disease status assessment report obtained by the reporting module into the trained prediction model 1, and output the functional outcome predictions and confidence intervals for the second and fourth years. The content of the functional outcome prediction is based on the five subscales of KOOS (knee injury and osteoarthritis outcome score) - pain, symptoms, activities of daily living, motor and recreational function, quality of life scores, feature importance (including permutation_importance as a fallback), residual plot, and regression scatter plot to ensure the interpretability of the model.

[0066] The prediction model 1 of this module employs a multimodal feature-driven ensemble learning framework. It uses multidimensional data such as patient demographics, clinical data, and imaging depth features as input, integrating four types of base learners: Gradient Boosting Decision Tree (GBDT), Random Forest (RF), Support Vector Machine (SVM), and Multilayer Perceptron (MLP). A two-layer stacking strategy is used: the first layer of base learners predicts each component of the KOOS task (a continuous regression task), and the second layer uses ElasticNet as a meta-learner to weight and fuse the outputs of the base learners. ElasticNet achieves feature sparsity and model robustness through L1 / L2 regularization, ultimately resulting in stronger generalization ability and lower prediction error. Experiments show that this strategy reduces the mean absolute error by approximately 12–15% and increases the consistency correlation coefficient (CCC) to 0.73, significantly outperforming single models.

[0067] The prediction model 1 of this module is a continuous outcome prediction (regression) used for multi-objective regression of the KOOS five-dimensional (Pain, Symptoms, ADL, Sport / Rec, QOL, etc.) functional scores at multiple time points (Year 2 of V01 / Year 4 of V06).

[0068] (2.2) The imaging progression prediction module is configured to input demographic information (age, sex, BMI, activity level), clinical data (KOOS / WOMAC, VAS clinical scores, normalized muscle strength, primitive muscle strength), and imaging depth features (KOA severity grading, fine-grained imaging features) from the disease status assessment report obtained by the reporting module into the trained prediction model 2, and output the knee osteoarthritis severity grading for the 2nd and 4th years.

[0069] Prediction Model 2 is a multi-task learning model that combines deep learning (convolutional neural network CNN) with traditional machine learning (such as XGboost), which improves the model's generalization ability by sharing underlying feature representations.

[0070] The prediction model 2 in this module is a multi-class classification model for grading the severity of knee osteoarthritis, targeting different time points and patient sides (V01 / V06, left / right). For data balancing and robustness assessment, the `balance_dataset` algorithm is used: equal resampling is performed on each class (a combination of undersampling and oversampling, aiming for 1000 samples per class). The performance of prediction model 2 is evaluated using the `evaluate_model` algorithm, which evaluates through multiple rounds of sampling (default 100 rounds), with a maximum of 100 samples randomly selected from each class in each round. Finally, the confusion matrix is ​​accumulated, and the average accuracy, precision, recall, F1 score, and AUC (Area Under the ROC Curve) are calculated. Prediction model 2 also outputs ROC curves, a confusion matrix (multi-color scheme), and a comparison bar chart. The following model families were used during training: XGBoost, LightGBM, RandomForest, GradientBoosting, AdaBoost, SVM (with probability=True and class_weight='balanced'), KNN, and MLP.

[0071] (2.3) The risk analysis module is configured to quantify the contribution of each input feature in prediction model 1 and prediction model 2 to the prediction results through SHAP (SHapley Additive exPlanations) analysis technology, i.e., the SHAP value of each input feature. The higher the SHAP value, the more important the input feature is to the progression prediction. The top five input features with the highest SHAP values ​​are considered as patient-specific risk factors.

[0072] At the individual level, post-hoc SHAP analysis of prediction results for individual patients can identify the top five contributing characteristics as the patient's primary individualized risk factors and related influencing factors. Notably, this module not only reveals the importance of each risk factor in prediction but also provides directional guidance for clinical intervention: for characteristics with positive contributions, an increase in their value often indicates improved disease recovery or prognosis, suggesting that clinical intervention should be strengthened; conversely, for characteristics with negative contributions, a decrease in their value helps inhibit disease progression, indicating that this input should be controlled or reduced to delay disease progression. The risk analysis module significantly improves the interpretability of the prediction model, making the prediction results more transparent and actionable. This interpretability analysis not only helps clinicians understand the basis of the predictions but also identifies patient-specific key risk factors. For example, for some patients, medial joint space narrowing may be the most significant predictor of progression, while for others, osteophyte formation or BMI may be more critical. This personalized risk factor identification and directional analysis provides scientific and feasible guidance for developing targeted intervention strategies.

[0073] This progress prediction module constructs a dual-channel, time-spanning, multi-task prediction framework around "functional continuous outcomes (regression) + imaging grade outcomes (multi-category)," and proposes improvements in process robustness and clinically interpretable characteristics.

[0074] 3. The treatment planning module is configured to output personalized comprehensive treatment plans through multi-specialty modules. It integrates four specialty modules: sports rehabilitation, orthopedics, psychology and nutrition, and clinical decision-making. Each specialty module has been trained and optimized with specific domain knowledge.

[0075] (3.1) The exercise rehabilitation module is configured to directly generate personalized exercise prescriptions by the Large Language Model (LLM) under structured cue words and rule constraints, including muscle strength training, joint range of motion exercises, aerobic exercise and functional training, without the need for an additional traditional optimizer.

[0076] The specific process is as follows:

[0077] Patient demographics (age, sex, BMI, activity level), clinical scales (KOOS / WOMAC, VAS), objective ability indicators (such as joint range of motion, muscle strength testing, gait analysis), imaging summaries (KL classification, osteophytes, joint space width), comorbidities and contraindications, and patient-specific risk factors were used as multi-source patient data. A structured prompt with a fixed template was used, combining the multi-source patient data with key points from sports medicine and rehabilitation guidelines obtained through vectorized retrieval as "evidence context" input. LLMs were required to directly output prescriptions based on the FITT-VP principles (frequency, intensity, time, type, volume, progression), and to make individualized adjustments to patient goals, pain thresholds, and resource accessibility. The prompts explicitly specified safety rules such as limiting weight-bearing during acute synovitis, limiting target heart rate zones for high-risk cardiopulmonary conditions, and avoiding high-impact treatments for high-risk NSAID and gastrointestinal conditions. The prompts were required to return the movement, sets, intensity, frequency, progression plan, pain cessation rules, and precautions in structured JSON format, along with the rationale for generation and evidence-based sources.

[0078] To achieve personalization, firstly, the training type, intensity, and progression speed are dynamically adjusted based on the patient's goal weight (e.g., prioritizing pain reduction over muscle gain, weight management, or returning to running). Secondly, the program's focus is automatically adjusted by calculating the symptom-image inconsistency index; if the inconsistency is significant, analgesia and psychological / behavioral interventions are strengthened, initial intensity is reduced, or deload and force alignment are emphasized. Thirdly, the open-chain / closed-chain training mode, load starting point (e.g., 40–60% of 1RM), and a progressive regimen of 5–10% per week are determined by combining ROM and muscle strength baseline. Perioperative optimization and training safety restrictions are automatically added for comorbidities (VTE, OSA, diabetes, etc.). The system also provides alternatives based on available resources, such as recommending bodyweight or isometric training when equipment is unavailable, and walking or water exercise when space is unavailable. After 2–4 weeks of follow-up, the training intensity is automatically adjusted up or down based on KOOS / VAS / ROM retest results, and a new prescription is generated by calling LLM. To enhance stability, the system also uses similar case retrieval (KNN or cosine similarity) to call historical successful solutions as few-shot examples, helping LLM generate outputs that are more clinically logical. All generated content undergoes dual checks by a rule validator and a physician review interface before publication to ensure that the recommended results are safe, interpretable, and clinically usable.

[0079] (3.2) The orthopedics module is configured to focus on assessing surgical indications, recommending appropriate surgical procedures (such as arthroscopic debridement, high tibial osteotomy, unicompartmental knee replacement or total knee replacement), and generating perioperative management and drug treatment recommendations.

[0080] The procedure uses multi-source structured patient data as input, including demographic characteristics (age, sex, BMI, occupation, and activity level), clinical scores (KOOS / WOMAC, VAS), imaging indicators (KL grade, alignment deviation, joint space width, osteophyte presence), joint range of motion, history of conservative treatment and comorbidities, perioperative risks (ASA grade, VTE risk, cardiopulmonary function, etc.), and patient-specific risk factors. All data are first standardized and imputed for missing values, and then a rule engine is used to score surgical indications, automatically identifying indications and contraindications. For example, patients with multi-compartment lesions who have received adequate conservative treatment but still experience severe pain, or those with a KL grade ≥3, are considered candidates for total knee replacement; patients with unicompartment lesions whose alignment can be corrected are considered candidates for osteotomy or unicompartmental knee replacement; and surgery is postponed if there is active infection or high risk of anesthesia.

[0081] In the recommendation phase, the module employs a Large Language Model (LLM) combined with Representational Imagery (RAG) to access orthopedic guidelines and evidence-based literature as context. Based on surgical indication scoring and patient target weights, it outputs the preferred surgical procedure and alternative options, explaining the rationale and potential risks. For drug therapy, the module generates evidence-based analgesic and anti-inflammatory drug regimens, including oral NSAIDs, topical formulations, and intra-articular injections. It automatically screens for contraindications or recommends alternatives based on the patient's renal function, gastrointestinal risk, anticoagulation status, and allergy history. Perioperative management recommendations cover preoperative optimization (nutrition, weight, blood glucose, infection screening), intraoperative antimicrobial prophylaxis, blood loss management, analgesia pathways, and early postoperative rehabilitation and VTE prevention plans, and indicate situations requiring multidisciplinary consultation.

[0082] Personalization is reflected in multidimensional dynamic adjustment: recommended pathways are adjusted based on patient goals (pain reduction, functional recovery, and delaying joint replacement); age and activity level influence surgical procedure selection (knee-preservation strategies are prioritized for younger, high-needs patients); joint alignment and stability determine whether osteotomy or unicompartmental replacement is recommended; comorbidities trigger perioperative optimization and anesthesia risk warnings; and medication recommendations are automatically matched with individualized dosages and safety windows. The module output uses structured JSON and readable reports for physicians and patients, including recommended protocols, explanations of indications and contraindications, medication dosage recommendations, perioperative optimization checklists, and follow-up plans. All results undergo rule validation and physician review before taking effect, ensuring that recommendations are safe, interpretable, and clinically usable.

[0083] (3.3) The psychological nutrition module is configured to integrate knowledge of pain psychology and clinical nutrition, focusing on identifying psychological problems related to chronic pain and developing intervention plans, while providing individualized nutritional recommendations based on the patient's BMI, body composition, and metabolic status. This module first receives multi-source data from the patient, including demographic information (age, gender, BMI, occupation and lifestyle), clinical scales (Pain Catastrophism Scale PCS, Depression and Anxiety Scale PHQ-9 / GAD-7, Sleep Quality Index PSQI), dietary habit records, body composition analysis (body fat percentage, skeletal muscle mass), metabolic indicators (fasting blood glucose, HbA1c, lipid profile), pain-related behavioral data (activity level, sleep duration, medication use), and patient-specific risk factors.

[0084] After data standardization and missing value imputation, the module automatically stratifies data using psychological risk screening rules: if PCS ≥ 30 or PHQ-9 ≥ 10, it is marked as high psychological risk, and cognitive behavioral intervention and pain education programs are generated first; if anxiety and insomnia are significant, relaxation training, mindfulness exercises, or breathing regulation guidance are added. The nutrition component assesses energy balance using BMI, body fat percentage, and metabolic indicators, and formulates macronutrient allocation and calorie targets based on patient goals (weight loss, maintenance, muscle gain), adjusting dietary restrictions or recommending alternative foods according to comorbidities (diabetes, gout, kidney disease, etc.).

[0085] The module employs a Large Language Model (LLM) combined with Retrievable Vectorized Search (RAG) to retrieve evidence-based guidelines and nutritional databases, directly generating individualized psychological intervention plans and nutritional protocols based on the patient's psychological and metabolic profile. The prompts explicitly request the LLM to output the intervention type, frequency, duration, and specific operational suggestions, such as the number of cognitive behavioral therapy modules per week, mindfulness practice scripts, recommended daily energy intake and macronutrient ratios, specific dietary examples, and progress goals (e.g., a 2–4% weight loss per month).

[0086] Personalization is reflected in multidimensional regulation: intervention priorities are selected based on psychological assessment results (mainly cognitive restructuring, exposure therapy, or relaxation training); energy deficits or supplementation levels are determined based on the patient's BMI and metabolic status; protein and trace element intake are adjusted based on activity levels and comorbidities; and changes in psychological scales and body composition are dynamically tracked. The intervention plan is automatically updated after retesting every 4–8 weeks, prompting LLM re-editing to generate a new version of the plan.

[0087] The final output is presented in structured JSON and a readable report for patients / doctors, including the psychological intervention plan, dietary structure, daily or weekly goals, behavioral reminders and monitoring indicators, along with the rationale for generation and evidence-based sources. All results must be reviewed by a rule validator and professionals (psychologists, nutritionists) before taking effect to ensure the safety, scientific validity, and operability of the intervention.

[0088] (3.4) The Clinical Decision Module is configured as a multidisciplinary coordinator to process prescriptions, surgical or drug recommendations, and psychological and nutritional intervention plans generated by the Exercise Rehabilitation Module, Orthopedics Module, and Psychological Nutrition Module in a unified manner, and output personalized comprehensive treatment plans. This module first receives the structured output (JSON format) from each specialty module, including exercise prescription parameters, recommended surgical procedures and perioperative management lists, drug treatment recommendations, psychological intervention plans, and nutritional plans. A standardized parser performs field mapping and semantic alignment on the outputs of different modules, and a rule engine and multi-objective optimization algorithm are used for conflict detection and priority ranking. For example, when the Orthopedics Module recommends preoperative arthroscopic debridement while the Exercise Module proposes high-load strength training, the system automatically reduces the training intensity and adjusts the rehabilitation start time to ensure safety and evidence-based practice.

[0089] This module employs a constraint-based solution and weighted reconciliation approach to comprehensively balance multidimensional objectives: prioritizing the patient's primary goals (pain control, functional recovery, weight management) and safety constraints, coordinating conflicting recommendations between different modules, and, when necessary, invoking Large Language Modeling (LLM) for natural language reasoning and protocol reorganization. This ensures that the output comprehensive treatment plan retains the core value of each specialty's recommendations while meeting clinical operability and patient compliance requirements. The module also automatically generates execution sequences and phased goals based on the patient's comorbidities, available resources, risk level, and timeline. For example, it might first complete preoperative optimization and elective surgery, then connect early rehabilitation and psychological intervention, and gradually introduce high-intensity training and dietary adjustments.

[0090] The final comprehensive output includes treatment timing, specific actions and training plans, surgical and pharmacological pathways, psychological intervention rhythm, dietary and lifestyle recommendations, along with conflict resolution instructions and evidence-based support. All results are presented in structured JSON and visual treatment pathway diagrams for physician review and patient implementation. The system supports dynamic updates; when patient follow-up data (KOOS / VAS, ROM, psychological scales, body composition) changes significantly, the module automatically re-integrates and optimizes the treatment plan, ensuring that the plan remains personalized, consistent, and in line with clinical best practices throughout the entire treatment cycle.

[0091] This clinical decision module acts as a coordinator, integrating recommendations from various specialty modules, resolving potential conflicts, and ensuring the consistency and feasibility of the final plan.

[0092] Each specialty module uses the RAG knowledge base system for retrieval. The RAG knowledge base is built upon 4,017 high-quality articles, covering the five intervention categories of KOA management: exercise and physical therapy (1,245 articles), pharmacological therapy (987 articles), surgical intervention (876 articles), nutritional management (521 articles), and psychosocial support (388 articles). The knowledge base construction process includes steps such as literature screening, information extraction, structured storage, and vectorized indexing. All articles have undergone rigorous quality assessment, prioritizing the inclusion of systematic reviews, meta-analyses, and high-quality randomized controlled trials. Knowledge is stored in the form of triplets, including the intervention, indication, and strength of evidence, facilitating rapid retrieval and reasoning within specialty modules. By employing a hybrid retrieval strategy combining dense and sparse searches, the system can improve retrieval precision while maintaining recall, ensuring that decisions within specialty modules are based on the most relevant and reliable clinical evidence.

[0093] Example 2: An Artificial Intelligence-Based Osteoarthritis Management System

[0094] Similar to the AI-based osteoarthritis management system in Example 1, the difference is:

[0095] In the imaging progression prediction module, when predicting the severity grade of knee osteoarthritis in the 4th year, the input features are: demographic information (age, sex, BMI, activity level) from the disease status assessment report obtained by the reporting module, clinical data (KOOS / WOMAC, VAS clinical score, primitive muscle strength), and imaging depth features (KOA severity grade, fine-grained imaging features).

[0096] The system in this embodiment has predictive accuracy comparable to that of Embodiment 1.

[0097] Example 3: An Artificial Intelligence-Based Osteoarthritis Management System

[0098] Similar to the AI-based osteoarthritis management system in Example 1, the difference is:

[0099] The features input into the functional outcome prediction module are: demographic information (age, gender, activity level), clinical data (KOOS / WOMAC, VAS clinical scores, normalized muscle strength), and imaging depth features (KOA severity grading, fine-grained imaging features) and other multidimensional data from the disease status assessment report obtained by the reporting module.

[0100] The features input into the imaging progression prediction module are: demographic information (age, sex, activity level), clinical data (KOOS / WOMAC, VAS clinical scores, normalized muscle strength), and imaging depth features (KOA severity grading, fine-grained imaging features) from the disease status assessment report obtained by the reporting module.

[0101] Normalized muscle strength is calculated by dividing the maximum forward or backward muscle strength of the thigh and calf by the individual's BMI value. The system in this embodiment also exhibits high predictive accuracy.

[0102] The technical solution of the present invention will be further explained through experiments below.

[0103] Experiment Example 1: Experiment Predicting the Progression of Knee Osteoarthritis (OA)

[0104] The goal of this study was to predict the progression of the Kellgren-Lawrence (KL) grade of the knee joint, as well as changes in the five subscales of the Knee Injury and Osteoarthritis Outcome Score (KOOS), using imaging data (X-rays) and clinical data (such as raw muscle strength, BMI, and normalized muscle strength obtained by different normalization methods). The study involved data from multiple time points (e.g., year 2 and year 4) and employed various machine learning models for prediction.

[0105] I. Experimental Design

[0106] 1. Goals and Tasks

[0107] (1) Imaging progress prediction: Predict the progression of XRKL (Kellgren–Lawrence) (left and right knees, V01 / V06 time points), multi-classification task.

[0108] Evaluation metrics: Accuracy, Precision, Recall, F1, AUC-ROC, Brier.

[0109] (2) Functional outcome prediction: Predict the five dimensions of KOOS (Pain / KPR, Symptoms / YMR, ADL / YML, Sport / Rec / KPL, QOL) and the regression task.

[0110] Evaluation metrics: macro-AUC, Acc, macro-F1, weighted-F1, Brier, ECE, R².

[0111] 2. Feature grouping (ablation / combination design)

[0112] G0_base: Basic features (base_categorical + base_cont).

[0113] G1: G0 + BMI.

[0114] G2: G0 + Raw muscle strength (RAW).

[0115] G3: G0 + *normalized muscle strength*_norm_BMI (defined as muscle strength / BMI).

[0116] G4: G0 + RAW + BMI.

[0117] G5: G0 + RAW + BMI + *_norm_BMI.

[0118] G6: G0 + Normalized muscle strength_norm_WT (defined as muscle strength / body weight)

[0119] We progressively examine the marginal contributions and interaction effects of BMI, RAW, RAW / BMI (*_norm_BMI), and _norm_WT.

[0120] 3. Learning device

[0121] (1) Prediction of progress in imaging: CNN is combined with models constructed by the following learners: XGBoost, LightGBM, RandomForest, GradientBoosting, AdaBoost, SVM, KNN, NeuralNetwork.

[0122] (2) Functional result prediction: ElasticNet, RandomForest, GradientBoosting, XGBoost, LightGBM, and SVR were used as meta-learners to perform weighted fusion of the outputs of the base learners (GBDT, RF, SVM, and MLP).

[0123] It covers linear / sparse, tree models, kernel methods, and shallow neural networks, avoiding method bias.

[0124] 4. Dataset

[0125] The dataset consists of 4275 patient data points (8550 knee joint data points), with a training set, validation set, and test set ratio of 8:1:1.

[0126] II. Main Experimental Results

[0127] Generally, a higher Macro-AUC value indicates better model performance for each class in multi-class classification problems; a closer Acc value to 1 indicates higher model accuracy; a higher Macro-F1 value, closer to 1, indicates a better balance between precision and recall; a closer Weighted-F1 value to 1 indicates better overall model performance for most classes in multi-class classification problems; a lower Brier score indicates more accurate probability predictions; and a lower ECE value indicates that the predicted probabilities are closer to the actual frequencies, meaning better model calibration. In this experiment, the dataset was large, numerous, and complex. Under such circumstances, even small performance improvements are very difficult to achieve.

[0128] 1. Advances in Radiology (XRKL) – Multi-classification

[0129] (1) V01 (Predicted data from the second year of follow-up)

[0130] (1.1) Left knee (V01XRKL_Left)

[0131] The prediction performance of the XGBoost model under different features is shown in Table 1.

[0132] Table 1 Performance metrics of XGboost in predicting radiographic progression of the left knee

[0133]

[0134] Table Note: In the table, ↑ indicates an improvement compared to group G0, and ↓ indicates a decrease compared to group G0.

[0135] As shown in Table 1, even with an already high overall AUC, the combination of G3 (muscle strength / BMI) and G5 (RAW+BMI+muscle strength / BMI) features still delivers stable and meaningful performance gains. Compared to baseline G0, G3 improves AUC by 1.5%, Acc by 1.5%, and macro-F1 by 2.4%, while significantly reducing the Brier score (-8.1%) and ECE (-17%), indicating that this feature not only enhances classification discriminative ability but also improves the model's probabilistic calibration performance. G5 further enhances this, increasing AUC by 1.7%, macro-F1 by 2.4%, and showing the largest reductions in Brier and ECE (-9.4% and -20%, respectively), making it the feature combination with the best overall discriminative ability and calibration performance.

[0136] In contrast, G6 (muscle strength / body weight) performance deteriorated significantly: AUC decreased by 1.15%, Acc decreased by 1.60%, and macro-F1 decreased by 1.30%, while Brier and ECE increased by 5.0% and 17.0%, respectively. This indicates that the body weight normalization strategy introduced noise, causing a simultaneous degradation in the model's discriminative performance and calibration ability. Overall, the BMI-based muscle strength normalization feature performed stably and effectively in the XGBoost model, while the body weight-based normalization strategy should be discarded in the final modeling.

[0137] The prediction performance of the RandomForest model under different features is shown in Table 2.

[0138] Table 2 Performance metrics of RandomForest in predicting radiographic progression of the left knee

[0139]

[0140] Table Note: In the table, ↑ indicates an improvement compared to group G0, and ↓ indicates a decrease compared to group G0.

[0141] As shown in Table 2, under the RandomForest model, G2, G3, and G5 all achieved stable performance improvements compared to the baseline (G0). Among them, G5 performed best in AUC and Acc, with an AUC improvement of 1.6%, an Acc improvement of 2.1%, and a macro-F1 improvement of 2.9%, while the Brier score decreased by 11.4%, indicating that both the overall discrimination ability and calibration performance were improved. G5 (RAW+BMI+muscle strength / BMI) achieved the best overall performance, representing the best feature combination in both discrimination and calibration performance.

[0142] In contrast, G6 (muscle strength / body weight) showed significantly worse performance, with AUC decreasing by 1.4%, Acc decreasing by 1.9%, macro-F1 decreasing by 1.7%, and Brier increasing by 5.1%. This indicates that body weight normalization not only failed to improve performance but also weakened the model's stability and calibration capabilities. Overall, BMI normalization (G3, G5) also exhibited stable gains in RandomForest, while the G6 feature combination should be avoided in the final modeling scheme.

[0143] (1.2) Right knee (V01XRKL_ Right)

[0144] When using the XGBoost model to predict radiographic progression of the right knee, compared with G0, the AUC of the G5 group was basically the same but slightly increased, while the Acc improved by nearly 1.9%, which was the highest among all feature groups. This shows that it not only improves the discrimination ability, but also significantly optimizes the model accuracy and probability calibration.

[0145] In summary, BMI-based normalization strategies (G3, G5) can deliver robust and multi-dimensional performance gains for both knees, while weight-based normalization (G6) introduces noise, significantly weakening discrimination and calibration performance, and should be removed in the final modeling.

[0146] (2) V06 (Fourth year follow-up)

[0147] (2.1) Left knee (V06XRKL_Left)

[0148] In experiments using the XGBoost model to predict radiographic progression of the left knee, the RAW+BMI (G4) combination achieved a performance improvement, with the AUC significantly increasing from 0.8187 to 0.8252, far exceeding that of G1 / G2 and other feature combinations. This demonstrates that this combination provides robust and reliable discriminative gain in challenging prediction tasks. In contrast, the AUCs of G3 (muscle strength / BMI) and G5 (RAW+BMI+muscle strength / BMI) decreased by 0.74% and 0.82%, respectively, showing a significant performance decline, suggesting that such feature combinations may interfere with the model's discriminative ability. Comprehensive analysis indicates that G4 performed best across all groups and is the recommended feature combination for the current task, providing a more convincing basis for clinical prediction and individualized decision-making.

[0149] When using the LightGBM model to predict radiographic progression of the left knee, the AUC of the G4 group decreased slightly by 0.62% relative to baseline (G0), suggesting differences in the sensitivity of different tree models to this combination. In contrast, G6 (muscle strength / body weight normalized) performed worse under V06, with the AUC expected to decrease by approximately 0.57% to approximately 0.8135, the Acc decrease exceeding 1.5%, macro-F1 decreasing simultaneously, and Brier and ECE increasing by ≥5% and ≥20%, respectively, showing significant discrimination and calibration degradation.

[0150] In summary, G4 is the best combination for predicting the left knee in V06 and is recommended as the primary feature; although G3 and G5 show a slight decrease, the impact is limited and can be considered for retention based on task requirements; however, G6's performance deteriorates significantly at this stage and should be removed from the final model to ensure prediction stability and calibration quality.

[0151] (2.2) Right knee (V06XRKL_Right)

[0152] In predicting the right knee radiographic progression in the fourth year using two tree models (RandomForest and XGBoost), RAW+BMI (G4) also showed the most robust performance: Regarding the AUC metric, RandomForest improved by 0.0029 (+0.31%), while XGBoost improved from 0.8196 to 0.8244, making it the best feature combination, demonstrating that it can still provide sustained and significant discriminative gains in more complex tasks.

[0153] Among the other feature groups, G2 brought a slight improvement in the AUC index (+0.12%), while G5 showed a slight decrease (-0.26%) under RandomForest and remained basically flat under XGBoost. This indicates that the positive contribution of BMI-normalized muscle strength to late-stage right knee prediction has weakened, but the overall fluctuation is small and has a limited impact on model performance.

[0154] Based on the trends of V01 and the left knee, it is reasonable to infer that G6 (muscle strength / body weight normalized) will further deteriorate below V06 in the right knee, with AUC possibly decreasing by more than 0.5%, Acc and macro-F1 decreasing by more than 1%, while Brier and ECE significantly increase, showing stronger noise effects and calibration bias. Therefore, it is not recommended to retain G6 in late right knee prediction.

[0155] Based on the results of the above performance indicators, in the early (second year) prediction of radiographic progression of both knees, the combination of features G3 and G5 provides better predictive performance, especially the G5 feature combination. In the late (fourth year) prediction of radiographic progression of both knees, the combination of features G4 provides the best predictive performance. In addition, the AUC index variation when using the G5 feature group in the late (fourth year) prediction of radiographic progression of both knees is smaller, less than 0.5%, which also shows better performance.

[0156] 2. Functional Result Prediction (KOOS Five-Dimensional, ElasticNet)

[0157] R² represents the proportion of variance explained by the model relative to the total variance, and it intuitively reflects how well the model fits the data. A higher R² value indicates that the model explains more variance and fits the data better. In this experiment, the dataset had a large sample size and was rich and complex. Under such circumstances, even a small performance improvement, such as an increase in R² of 0.5%, was very difficult. The specific results are described below:

[0158] (1) V01 (Second Year)

[0159] When using the ElasticNet model, the prediction performance (R²) of different features in the KOOS five-dimensional model. 2As shown in Table 3.

[0160] Table 3. ElasticNet's predictions of KOOS's five-dimensional performance metrics for the second year.

[0161]

[0162] Table Notes: + indicates an increase in value compared to group G0; - indicates a decrease in value compared to group G0; except for group G0, the values ​​in the table represent the change in value and the percentage change compared to group G0. The value for group G0 is R. 2 .

[0163] As shown in the table above, G3 provides a stable improvement of +0.83% to +0.91% in YMR, YML, and KPL dimensions, while G5 provides a stable improvement of +0.83% to +0.91% in KPR and QOL dimensions. This indicates that BMI-normalized muscle strength helps improve early functional prediction performance. G3 and G5 are the preferred feature combinations for predicting the five dimensions of KOOS in the second year.

[0164] In contrast, G6 showed a significant decrease across all dimensions, with the largest decrease in Sport / Rec (−3.40%), indicating that the weight normalization strategy introduced noise, weakening the model's explanatory power and generalization performance.

[0165] (2) V06 (Fourth Year)

[0166] When using the ElasticNet model, the prediction performance (R²) of different features in the KOOS five-dimensional model. 2 As shown in Table 4.

[0167] Table 4. ElasticNet's predictions of KOOS's five-dimensional performance metrics in the fourth year.

[0168]

[0169] Table Notes: + indicates an increase in value compared to group G0; - indicates a decrease in value compared to group G0; except for group G0, the values ​​in the table represent the change in value and the percentage change compared to group G0. The value for group G0 is R. 2 .

[0170] As shown in the table above, even under more challenging prediction conditions, G3 still provides a stable positive gain (approximately +0.6%) in the YMR, YML, and KPL tasks, suggesting that BMI-based normalized muscle strength features can generally be considered a favorable factor. Among them, the feature combination of G5 has the best positive gain.

[0171] G6's performance deteriorated across the board, R 2The decline ranged from -2.5% to -8.2%, with the most severe decreases in Pain and QOL, which significantly impaired the model's predictive performance and robustness.

[0172] Based on the results of the above performance indicators, the combination of G3 and G5 features performs better in predicting the five-dimensional results of KOOS in the early and late stages (second and fourth years).

[0173] Integrating the two tasks of XRKL (Advances in Radiology) and KOOS (Kinetic Outcomes Prediction), this experimental result shows that the normalization method for primitive muscle strength has a significant impact on the prediction results. In the early stage (second year) of XRKL prediction, the feature combination of G3 and G5 provides better predictive performance, especially the G5 feature combination; in the late stage (fourth year) of XRKL prediction, the feature combination of G4 provides the best predictive performance; and in functional outcome prediction, the feature combination of G3 provides the best predictive performance.

[0174] Further optimization revealed that BMI-based normalized muscle strength (G3) and its combined feature (G5) are the best-performing and most robust designs for both prediction tasks. Overall, G3 and G5 can provide stable and multi-dimensional performance gains across multiple tasks and time points, making them the most recommended feature designs for the progression prediction module.

[0175] The osteoarthritis management system provided in Embodiment 1 of the present invention has the following advantages:

[0176] (1) Improved accuracy: The system achieved an accuracy of 77.16% on the KOA severity grading task, which is 76.4% and 30.9% higher than general visual language models such as GPT-4V (43.75%) and Gemini Pro Vision (58.93%), respectively. This significant improvement is due to the deep learning architecture specifically designed for knee X-rays and the training on large-scale labeled data, which enables the system to accurately identify subtle pathological features such as joint space narrowing and osteophyte formation.

[0177] (2) Efficiency improvement: The system reduces the time for a complete KOA assessment and treatment planning from an average of 65 minutes to 40 minutes through automated data acquisition, image analysis and treatment plan generation, improving efficiency by 38.5%. In particular, in the image analysis stage, the automated processing can complete a comprehensive assessment in just 30 seconds, while traditional manual interpretation usually takes 5-10 minutes, which greatly improves the efficiency of clinical work.

[0178] (3) Quality improvement: In clinical validation, the treatment plan formulated by the AI ​​system with the assistance of doctors achieved an expert approval rate of 94.7%, while the approval rate of the purely manual group was only 71.3%, an improvement of 32.8%. This improvement is mainly attributed to the fact that the AI ​​system can comprehensively consider multidisciplinary perspectives, provide evidence-based medicine support, and ensure the integrity and consistency of the plan, effectively reducing human error and experience bias.

[0179] (4) Enhanced personalization: The system generates highly personalized management plans based on the patient's individual characteristics such as age, BMI, activity level, comorbidities, functional scores, and imaging findings. For example, for young and active patients, the system will prioritize conservative treatment and exercise rehabilitation programs; while for elderly obese patients, it will emphasize weight management and joint protection strategies, truly achieving "one prescription per person" precision medicine.

[0180] (5) Interpretability: Through SHAP analysis technology, the system can clearly display the main influencing factors and their contribution to each prediction result, helping doctors understand the decision-making logic of the AI. At the same time, the treatment recommendations are accompanied by corresponding literature evidence and recommendation levels, enabling doctors to trace the scientific basis of each recommendation, thereby enhancing the credibility and clinical acceptance of the AI ​​system.

[0181] In the application of artificial intelligence models for disease prediction, especially when dealing with large, complex datasets, improving the performance of prediction models is challenging. Even an improvement in AUC of less than 1 percentage point is often considered a significant improvement and worthy of recognition in the field. This invention captures features among numerous characteristics that are key factors in osteoarthritis, thereby improving the overall performance of the system. In the second year's imaging progression prediction, the AUC improved by 1.7%, which is a very considerable improvement in this field and has high practical application value.

[0182] As can be seen from the above embodiments, the present invention provides an artificial intelligence-based method and system for osteoarthritis management. This system covers the entire clinical process from initial patient diagnosis to treatment plan development. By automating the processing and analysis of multimodal data, including structured collection of patient medical history information, intelligent interpretation of knee X-rays, and automatic calculation of clinical scores, the system can comprehensively assess the patient's disease status. Based on machine learning algorithms and SHAP interpretability analysis, the system can provide personalized disease progression predictions, not only predicting 2-year and 4-year functional outcomes and radiological progression, but also identifying patient-specific risk factors, providing a basis for precise intervention. Through a multi-specialty module collaborative architecture simulating multidisciplinary team discussions, it integrates multiple dimensions such as exercise rehabilitation, drug therapy, surgical intervention, nutritional management, and psychological support to generate comprehensive treatment plans based on evidence-based medicine. The overall system design significantly improves clinical work efficiency, reducing the time for complete assessment and plan development by 38.5%. Simultaneously, through automation and intelligence, it reduces reliance on specialist physicians and high-end equipment, enabling high-quality KOA management to be implemented even in resource-limited areas, thereby effectively reducing overall medical resource demands.

Claims

1. An artificial intelligence-based osteoarthritis management system, characterized in that, It includes the following modules: The disease assessment module is configured to collect patient information through natural language interaction, automatically analyze X-ray images, and generate a disease status assessment report. The progression prediction module is configured to allow the disease status assessment report to predict results and radiological progression through a predictive model output function. The treatment planning module is configured such that the multi-specialty module generates personalized comprehensive treatment plans through a multi-objective optimization decision engine. In the treatment planning module, the specialty modules include a sports rehabilitation module, an orthopedics module, a psychological nutrition module, and a clinical decision-making module; Specialty recommendations generated by the rehabilitation, orthopedics, and psychological nutrition modules are incorporated into the clinical decision-making module. The sports rehabilitation module is configured to use patients' demographic characteristics, objective ability indicators related to exercise, imaging characteristics, comorbidities, contraindications, and patient-specific risk factors as multi-source patient data. Using structured prompts with a fixed template, the module takes the multi-source patient data and key points of sports medicine and rehabilitation guidelines obtained through vectorized retrieval as "evidence context" input. Based on the FITT-VP principle, the module outputs an exercise prescription through a large language model and makes individualized adjustments according to the patient's goals, pain threshold, and resource accessibility, serving as the specialist recommendations for the sports rehabilitation module. The orthopedics module is configured to use patients' demographic characteristics, clinical scores related to osteoarthritis or pain, imaging characteristics, joint range of motion, history of conservative treatment and comorbidities, perioperative risks, and patient-specific risk factors as multi-source patient data. It performs standardization and missing value imputation, uses a rule engine to score surgical indications, automatically identifies indications and contraindications, and uses a large language model combined with vectorized retrieval to call orthopedic guidelines and evidence-based literature as context. Based on the surgical indication score and patient target weight, it outputs the preferred surgical procedure and alternative options, generates evidence-based analgesic and anti-inflammatory drug regimens, and perioperative management recommendations as specialist recommendations for the orthopedics module. The psychological nutrition module is configured to use patients' demographic information, clinical assessment scales for mental health and quality of life, dietary habit records, body composition analysis, metabolic indicators, pain-related behavioral data, and patient-specific risk factors as multi-source patient data. After standardization and missing value imputation, the data is automatically stratified using psychological risk screening rules. A large language model combined with vectorized retrieval is used to generate individualized psychological intervention plans and nutritional programs as specialist recommendations for the psychological nutrition module. The clinical decision-making module is configured to generate personalized comprehensive treatment plans from the specialist recommendations obtained from the sports rehabilitation module, orthopedics module, and psychological nutrition module through a multi-objective optimization decision engine based on evidence weighting and conflict resolution.

2. The artificial intelligence-based osteoarthritis management system according to claim 1, characterized in that: The progression prediction module is configured to use demographic information from the disease status assessment report, including body mass index, clinical data including clinical scores and normalized muscle strength, and imaging depth features, to output 2-4 year functional prediction results through prediction model 1, and to output 2-4 year imaging progression through prediction model 2. The normalized muscle strength is the maximum muscle strength of the thigh or calf, forward or backward, divided by the body mass index. Prediction Model 1 is a model obtained by ensemble learning of multiple machine learning models, while Prediction Model 2 is a multi-task learning model that combines deep learning and traditional machine learning.

3. The artificial intelligence-based osteoarthritis management system according to claim 1, characterized in that: The disease assessment module integrates a patient interaction module, an X-ray analysis module, and a reporting module; the medical history report obtained from the patient interaction module and the osteoarthritis severity grading obtained from the X-ray analysis module are input into the reporting module to generate a disease status assessment report.

4. The artificial intelligence-based osteoarthritis management system according to claim 3, characterized in that, In the disease assessment module: The patient interaction module is configured to collect patients' medical history information based on the Qwen-Max large language model and generate medical history reports. The X-ray analysis module is configured to use the U-Net model to locate and segment the knee joint in X-ray images, and use the ResNet-50 model to grade the severity of osteoarthritis in the located and segmented X-ray images, and extract fine-grained radiological features. The reporting module is configured to integrate the medical history report obtained from the patient interaction module and the osteoarthritis severity obtained from the X-ray analysis module through multimodal feature fusion technology, perform symptom-image consistency calibration, and finally use the Qwen-Max model to summarize and generate a disease status assessment report. And / or, the severity grading of the bone and joint is defined as a four-level classification: none / suspected, mild, moderate, and severe.

5. The artificial intelligence-based osteoarthritis management system according to claim 2, characterized in that: The progress prediction module integrates a functional outcome prediction module, an imaging progress prediction module, and a risk analysis module; The input features and prediction results from the functional outcome prediction module and the imaging progression prediction module are fed into the risk analysis module to obtain patient-specific risk factors. And / or, in the progression prediction module, when outputting the radiographic progression in the 4th year, the clinical data in the features input to the prediction model 2 are: clinical score and original muscle strength.

6. The artificial intelligence-based osteoarthritis management system according to claim 5, characterized in that, The functional prediction results include scores for pain, symptoms, activities of daily living, motor and recreational function, and quality of life on five subscales for knee injury and osteoarthritis outcome scoring. The ensemble learning strategy described in prediction model 1 is a two-layer stacked ensemble strategy. The first layer completes the continuous regression task through base learners, and the second layer uses elastic network regression as a meta-learner to perform weighted fusion of the outputs of the base learners. The weighted fusion process includes L1 / L2 regularization. The base learners include: gradient boosting decision tree, random forest, support vector machine and multilayer perceptron. And / or, in the imaging progression prediction module, the content of the imaging progression over 2-4 years includes: osteoarthritis severity grading over 2-4 years; The traditional machine learning algorithm in prediction module 2 is XGboost, and the deep learning algorithm is a convolutional neural network. The imaging progression prediction module predicts the left and / or right knee joints; And / or, in the progression prediction module, the demographic information is: age, sex, body mass index, activity level; the clinical data is: KOOS / WOMAC, VAS clinical scores, normalized muscle strength, primitive muscle strength; and the imaging depth features are: bone and joint severity grading, fine-grained imaging features. And / or, the risk analysis module quantifies the contribution of each input feature to the prediction result using SHapley Additive exPlanations analysis technology; the higher the contribution, the more important the corresponding input feature is for progression prediction, and the top five input features in terms of contribution are taken as patient-specific risk factors.

7. The artificial intelligence-based osteoarthritis management system according to claim 1, characterized in that: In the clinical decision module, the data processing process of the multi-objective optimization decision engine based on evidence weighting and conflict resolution includes: mapping and semantically aligning the outputs of different specialty modules through a standardized parser, and using a rule engine and multi-objective optimization algorithm to perform conflict detection and priority ranking. The constraint solving and weight harmonization method is used to comprehensively balance multi-dimensional objectives, prioritizing the patient's main objectives and safety constraints, and coordinating the suggestion conflicts between different specialty modules. The personalized comprehensive treatment plan includes: treatment sequence arrangement, specific actions and training plans, surgical and drug pathways, psychological intervention rhythm, dietary and lifestyle recommendations, conflict resolution instructions and evidence-based basis; And / or, the specialty module performs vectorized retrieval through a vectorized retrieval knowledge base system. The construction process of the vectorized retrieval knowledge base includes: document screening, information extraction, structured storage, and vectorized indexing.

8. The artificial intelligence-based osteoarthritis management system according to claim 7, characterized in that, In the clinical decision-making module, the methods for coordinating conflicting recommendations between different specialty modules include: calling a large language model for natural language reasoning and scheme reorganization, and automatically generating execution order and phase goals based on the patient's comorbidities, available resources, risk level, and timeline.

9. A computer storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the artificial intelligence-based osteoarthritis management system according to any one of claims 1-8.

Images