Knee osteoarthritis grading prediction method based on gait analysis and deep learning

Abstract

This invention discloses a method for predicting the severity of knee osteoarthritis (KOA) based on gait analysis and deep learning, belonging to the fields of biomechanics, computer-aided diagnosis, and machine learning. Addressing the core shortcomings of existing KOA assessments, such as reliance on static imaging, high subjectivity, and disconnect from patient functional status, as well as insufficient depth of gait analysis feature mining, poor model interpretability, and failure to resolve sample heterogeneity caused by biomechanical compensation, this invention simultaneously collects gait data from subjects using an optical motion capture system and a force platform. After standardized preprocessing, a personalized lower limb musculoskeletal model is constructed to extract multi-dimensional deep biomechanical features of joints and muscles. An optimized database retaining only samples from the most severely affected side is built, and six algorithms are integrated to screen core features. A dual-stream adaptive neural network model is constructed to achieve automatic grading and prediction of KOA severity. This invention is non-invasive and radiation-free, with an average model accuracy of 90.7%, enabling dynamic functional quantitative assessment of KOA. It is suitable for early screening, precise clinical diagnosis, and rehabilitation follow-up of KOA.

Description

Technical Field

[0001] This invention belongs to the fields of biomechanics, computer-aided diagnosis, and machine learning, specifically relating to a method for predicting the grading of knee osteoarthritis (KOA) based on gait analysis and deep learning. This method is applicable to early screening of KOA, accurate clinical assessment, and quantitative evaluation of mobility-aiding medical devices. Background Technology

[0002] Knee osteoarthritis (KOA) is a degenerative disease characterized by cartilage degeneration and osteophyte formation at the joint margins, which severely impacts patients' quality of life. Current clinical assessment primarily relies on static imaging techniques such as X-rays and magnetic resonance imaging (MRI) (e.g., KL classification).

[0003] However, the current lack of quantifiable objective standards presents challenges: judgments may differ between doctors, and even among the same doctor at different times; doctors' interpretation of images is time-consuming and laborious, becoming an efficiency bottleneck in large-scale physical examinations, epidemiological surveys, or busy clinical work. With the development of machine learning technology, powerful tools have been provided for extracting information from complex data. However, existing techniques for KOA prediction and grading have many shortcomings:

[0004] (1) Limitations of static assessment: Imaging findings (structural changes) often do not fully match the patient’s subjective pain and actual motor ability (functional impairment), and cannot capture the biomechanical response of joints under dynamic load.

[0005] (2) Insufficient feature mining depth: Existing gait analysis focuses on surface spatiotemporal parameters or external kinematic data, lacking simulation analysis of deep dynamic features such as internal joint contact force, joint torque and muscle activation.

[0006] (3) Lack of interpretability of models: Most deep learning models are “black box” structures, and their decision-making process lacks biomechanical basis, making it difficult to guide targeted clinical treatment.

[0007] Therefore, there is an urgent need for a comprehensive evaluation scheme that can penetrate surface motion data and integrate the internal load characteristics of musculoskeletal simulation with high-performance machine learning. Summary of the Invention

[0008] To overcome the aforementioned shortcomings, this invention provides a method for predicting the severity of knee osteoarthritis (KOA) based on gait analysis and deep learning. It acquires radiation-free and non-invasive gait data simultaneously through optical motion capture and a force platform. Based on the AnyBody musculoskeletal modeling system, it completes personalized model construction and inverse dynamics calculations, penetrating surface motion data to extract deep dynamic features within the joint. Through comparative optimization of three database schemes, it addresses the feature heterogeneity problem caused by biomechanical compensation in the patient's bilateral lower limbs. It integrates six feature selection algorithms to screen core features highly correlated with KOA grading, ultimately constructing a dual-stream adaptive neural network model with clear biomechanical basis. This achieves objective, quantitative, and high-precision automatic grading prediction of KOA severity, comprehensively overcoming the core deficiencies of traditional static imaging assessment.

[0009] The technical solution adopted by this invention to solve its technical problem is as follows: A method for predicting the grading of knee osteoarthritis based on gait analysis and deep learning includes the following steps: Gait data of S1 subjects were collected simultaneously: The optical motion capture system and force platform are used to simultaneously collect three-dimensional kinematic trajectory data of the lower limbs and ground reaction force data of the subject in natural gait. S2 gait data normalization preprocessing: The raw data collected in step S1 is subjected to trajectory loss repair, standardized naming of marker points, and ground reaction signal filtering and noise reduction to obtain standardized gait data that meets the requirements of musculoskeletal modeling. S3 Personalized Musculoskeletal Modeling and Inverse Dynamics Feature Extraction: The standardized gait data obtained in step S2 is imported into the musculoskeletal modeling system. A personalized lower limb musculoskeletal model matching the subject's anatomical features is constructed by personalizing the model scaling and correcting muscle strength. Inverse dynamics simulation is run to extract multi-dimensional biomechanical features of the lower limb joints and muscles during the gait cycle. S4 Feature Set and Optimized Sample Library Construction: A KOA severity labeling system was established by combining the KL clinical grading information of the subjects. The biomechanical features extracted in step S3 and the static clinical features of the subjects were integrated to construct a multi-dimensional full feature set. A SevereOnly sample library was constructed, which only retains the gait cycle samples of the most severe side of the subjects' KOA. The sample library data was standardized and cleaned. S5 multi-algorithm fusion for high-correlation feature selection: By integrating multiple feature selection algorithms, a subset of core features highly correlated with KOA severity rating is selected from the full feature set; S6 KOA hierarchical prediction deep learning model construction: A dual-stream adaptive neural network model is constructed, which sets up parallel static feature branches and dynamic temporal feature branches. The two branches perform adaptive feature extraction on static clinical features and dynamic temporal biomechanical features in the core feature subset, respectively. After multi-source feature fusion is completed by the feature fusion layer, the KOA severity classification result is output through the classification output layer. The hierarchical cross-validation strategy is used to complete the model training and optimization, and the trained KOA classification prediction model is obtained. S7 KOA classification prediction and results output: After the gait data of the subjects to be evaluated are processed in steps S2-S5, they are input into the KOA grading prediction model trained in step S6, and the KOA severity grading results are automatically output.

[0010] The beneficial effects of this invention are as follows: This invention addresses the core shortcomings of existing knee osteoarthritis (KOA) grading and assessment technologies by systematically innovating in six dimensions: assessment paradigm, modeling accuracy, sample processing, feature mining, model architecture, and clinical application. Compared to traditional imaging assessment methods and existing gait analysis methods, it possesses the following significant advantages, all of which are supported by experimental data and a complete technical solution: (1) Breaking through the limitations of the traditional static assessment paradigm, a complete closed-loop system for dynamic functional assessment is constructed to fundamentally address the pain point of the disconnect between imaging assessment and clinical function. This invention completely breaks away from the traditional assessment framework of static imaging such as X-rays and MRI, which can only evaluate joint structural degeneration. It pioneers a complete technical system encompassing "non-invasive gait acquisition - standardized preprocessing - personalized musculoskeletal simulation - multi-dimensional feature mining - interpretable deep learning grading - clinical quantitative report output." It uses the dynamic biomechanical response of the subject under natural gait as the core basis for KOA grading, while simultaneously considering both joint structural degeneration and motor function impairment. This completely solves the industry pain point of mismatch between imaging findings and patients' subjective pain and actual motor function impairment in existing technologies. Furthermore, this solution is based entirely on optical motion capture and force table acquisition, with no ionizing radiation and no invasive operation. A single effective acquisition session takes only 15-20 minutes, requiring no fasting or special preparation from the patient. It is suitable for both precise clinical diagnosis and large-scale epidemiological screening and long-term follow-up assessment of KOA patients, with clinical versatility far exceeding existing technologies.

[0011] (2) The innovative personalized musculoskeletal modeling and dual-layer correction strategy significantly improves the accuracy of dynamic simulation and solves the problem of poor applicability of general models in high-risk groups of KOA. To address the shortcomings of existing musculoskeletal modeling, which simply scales height and weight without considering the diverse composition of high-risk groups like the middle-aged and elderly, those with high BMI, and those with low muscle mass, leading to distorted muscle strength estimation and simulation results deviating from true physiological states, this invention constructs a personalized modeling scheme that combines anatomical and physiological matching. Firstly, it employs the Length-Mass-Fat (LMF) scaling law, dynamically determining joint center positions through a least-squares optimization algorithm, effectively reducing the impact of soft tissue artifacts on computational accuracy. Secondly, it introduces the Deurenberg formula to calculate subject body fat percentage and pioneers a muscle strength correction strategy based on physiological cross-sectional area (PCSA) of lean body mass. This avoids overestimation of muscle strength in high-BMI patients and ensures the biological rationality of simulation results for high-risk groups such as obesity and low muscle mass. This scheme significantly improves the simulation accuracy of musculoskeletal models, ensuring the authenticity and reliability of subsequent feature extraction from the data source, and solving the industry problem of poor applicability of general musculoskeletal models in the target population of KOA.

[0012] (3) Innovative sample processing for KOA pathological features, breaking through the sample heterogeneity bottleneck caused by biomechanical compensation, and achieving a significant improvement in model classification performance. To address the shortcomings of existing gait-based KOA grading models, which commonly use mixed left and right leg data and indiscriminate input of affected / healthy side data, completely ignoring the contralateral biomechanical compensation caused by unilateral KOA lesions, resulting in highly heterogeneous sample features and severely degrading model accuracy and robustness due to compensatory noise, this invention designs three differentiated construction schemes for KOA compensatory pathological features: a unified database, a two-tiered database, and a SevereOnly database (only the most severe side). Through systematic comparison and validation with six machine learning models, including Random Forest and XGBoost, the SevereOnly scheme, which retains only the periodic samples of the most severe side of the patient, was ultimately determined to be the optimal scheme, eliminating compensatory noise interference from the data source. This innovative design directly achieves a significant improvement in model performance: taking the Random Forest model as an example, the SevereOnly scheme achieves a grading accuracy of up to 93.96%, which is more than 10 percentage points higher than the unified database scheme. At the same time, the model's precision, recall, and F1 score are all significantly improved, representing a breakthrough solution in the field for the problem of KOA sample heterogeneity.

[0013] (4) Construct a deep dynamic feature mining and multi-algorithm fusion screening system to penetrate the limitations of surface motion data and achieve a unity of high feature correlation and strong interpretability. Existing gait analysis methods can only extract superficial kinematic features such as gait speed and joint angles, failing to capture deep internal joint dynamic features directly related to KOA degeneration. Furthermore, single-algorithm feature selection is prone to feature redundancy and lacks clear biomechanical significance. This invention, through personalized musculoskeletal modeling and inverse dynamics simulation, extracts 64 multidimensional biomechanical features, including lower limb joint reaction forces, joint torques, muscle strength and activation, and joint range of motion. This penetrates superficial kinematic data to directly capture the core mechanical mechanisms of KOA lesions. Simultaneously, it integrates six feature selection algorithms—random forest, XGBoost, mutual information, LASSO, recursive feature elimination, and tree model importance—to systematically select a subset of core features highly correlated with KOA grading. All selected features possess clear physiological and mechanical significance. This invention achieves a comprehensive upgrade in feature dimensions, significantly reducing feature redundancy and laying a core foundation for high model accuracy. It also solves the "black box" problem of machine learning models from the source of features, allowing the selected core features to directly guide the formulation of personalized clinical treatment plans, possessing clinical guidance value unattainable by existing technologies.

[0014] (5) A dual-stream adaptive neural network architecture adapted to the KOA classification is proposed, achieving a dual breakthrough in high diagnostic accuracy and clinical interpretability. To address the shortcomings of existing KOA hierarchical deep learning models, which are mostly general "black box" structures with no biomechanical basis for decision-making, low clinical acceptance, and lack of a dedicated architecture for the multi-source heterogeneous data characteristics of KOA ("static clinical features + dynamic temporal features"), leading to overfitting and poor generalization ability, this invention constructs a dual-stream adaptive neural network architecture that processes static clinical features and dynamic gait sequences in parallel. The two branches adaptively extract features from static clinical data (age, gender, BMI, etc.) and dynamic temporal biomechanical data, respectively, specifically adapting to the characteristics of multi-source heterogeneous data. A feature fusion layer achieves efficient integration of multi-source information. Combined with optimization strategies such as five-fold hierarchical cross-validation, warmup + cosine annealing learning rate scheduling, and a smooth early stopping mechanism, excellent generalization performance is achieved even in small clinical sample scenarios. Furthermore, all input features are quantitative indicators with clear biomechanical significance, making the decision-making process traceable and interpretable. Ultimately, this model achieved an excellent performance of 90.7% average accuracy, 91.9% precision, and 92.8% F1 score in five-fold cross-validation, far exceeding the 65%-75% accuracy of traditional clinical grading methods. At the same time, it completely broke the "black box" barrier of deep learning models, obtaining a decision basis that clinicians can understand and trust, and achieving a dual breakthrough in diagnostic accuracy and clinical applicability.

[0015] (6) Achieve standardized and quantitative assessment of the entire KOA process, completely solve the problem of subjective differences in traditional assessment, and significantly improve diagnostic consistency and clinical work efficiency. Addressing the shortcomings of traditional imaging interpretation, which relies on physicians' subjective judgment, leading to significant discrepancies in diagnoses between different doctors and even among the same doctor at different times, and is time-consuming and labor-intensive, creating an efficiency bottleneck in large-scale physical examinations and epidemiological surveys, this invention addresses these issues. Through a standardized process of data collection, preprocessing, feature extraction, and model inference, it automatically outputs objective and quantifiable KOA grading results and predicted confidence levels. Simultaneously, it generates a standardized assessment report including core biomechanical indicators for patients, comparisons with reference values ​​for healthy individuals, and visualized trend charts. This invention completely avoids the subjective judgment differences inherent in manual image interpretation, exhibits excellent repeatability, significantly reduces the workload of clinicians, and substantially improves diagnostic efficiency and consistency. It can be widely applied in various scenarios, including early KOA screening in primary hospitals, large-scale epidemiological surveys, quantitative evaluation of patient treatment outcomes, and clinical efficacy evaluation of assistive medical devices, demonstrating high clinical promotion and industrial application value. Attached Figure Description

[0016] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0017] Figure 1 This is a schematic diagram of the overall technical route of the present invention; Figure 2 This is a schematic diagram illustrating the ranking of feature importance and a performance comparison of different machine learning models in an embodiment of the present invention. Figure 3 This is a schematic diagram of the accuracy and loss variation curves during the training process of a deep learning model in an embodiment of the present invention. Figure 4 This is a schematic diagram of an electronic device structure according to an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments and accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] This invention relates to a method for predicting the grading of knee osteoarthritis based on gait analysis and deep learning. It is a workflow that combines individualized musculoskeletal biomechanical data with a temporal deep learning model, as described above. Figure 1This method involves performing inverse dynamics on collected patient gait data, followed by signal processing (low-pass filtering, peak detection, etc.) and gait segmentation. Algorithms such as random forest, XGBoost, mutual information, LASSO, recursive feature elimination, and tree model importance are used to select highly relevant features, constructing a hierarchical prediction model. Hierarchical cross-validation and early stopping mechanisms are employed to optimize training. Model performance is evaluated using multiple metrics including accuracy, precision, recall, F1 score, and confusion matrix. This enables automatic hierarchical diagnosis of knee osteoarthritis (KOA). Specific implementation steps include... Step 1: Data Source The study was conducted from September 2024 to June 2025, enrolling patients treated in the Department of Orthopedics at the People's Hospital of Ningxia Hui Autonomous Region. Inclusion criteria included: ① diagnosed with knee osteoarthritis without a history of other joint diseases; ② ability to independently complete a full gait; ③ KL grade 2 or higher on the most severely affected side; ④ informed consent and voluntary participation in the study. Simultaneously, several age-, height-, and weight-matched healthy volunteers without a history of knee pain were recruited as a control group (details are shown in Table 1).

[0020] During the experiment, eight Qualisys (Oqus700) optical motion capture lenses and two Kistler (Type 9281EA) sensors were deployed to simultaneously collect kinematic and ground reaction data. After lens calibration, subjects, wearing loose clothing, had 36 reflective marker balls affixed to their pelvis, thighs, calves, and feet (such as the anterior superior iliac spine and lateral epicondyle of the femur) using the CAST marker system. Static calibration was then performed with arms and legs extended. After this, eight markers around the bilateral patellae and ankle joints were removed. Subjects walked back and forth along a 10m path with a natural gait. Data was collected three times per group, and gait data with relatively uniform pace and no pauses were selected for subsequent analysis.

[0021] A total of 22 sets of effective gait data from healthy volunteers and 161 sets of effective gait data from patients with knee osteoarthritis were collected, covering the entire range of KL grades 0 to 4 on both sides of the affected side. Based on the current small sample size and the difficulty in distinguishing between grades 1 and 2, in this embodiment, grades 1 and 2 are classified as mild, grade 3 as moderate, and grade 4 as severe.

[0022] Table 1 Subject Information

[0023] Step Two: Data Processing This embodiment utilizes the Qualisys 3D motion capture system to acquire the subject's motion trajectory data. For trajectory gaps caused by occlusion or missing markers, a cubic spline interpolation algorithm is used for repair to ensure the continuity and integrity of the temporal data. The repaired gait trajectory is exported in standard C3D format.

[0024] The data was then imported into Mokka (v0.6.2) software for further processing. Renaming protocol: All markers are renamed in accordance with the naming conventions of the AnyBody (v8.0) modeling system to ensure full compatibility with the topology of downstream computational models.

[0025] Dynamic filtering: The synchronously acquired ground reaction force (GRF) signal was denoised using a fourth-order bidirectional low-pass Butterworth filter with a cutoff frequency of 10Hz. (This cutoff frequency was chosen to maintain consistency with the spatial filtering of the kinematic data, thereby minimizing residual loads caused by signal frequency mismatch in inverse dynamics calculations while preserving the main dynamic characteristics of the gait support phase.) The preprocessed data was archived into the AnyBody lower limb model library as standard input for model scaling and dynamic analysis.

[0026] The simulation process is based on the anatomical lower limb model provided by AnyBody Managed Model Repository (AMMR). This model has high anatomical fidelity and includes major skeletal segments such as the pelvis, femur, tibia-fibula complex, talus, and metatarsals.

[0027] Degrees of freedom definition: The system defines approximately 12 degrees of freedom. The hip joint is simplified as a three-degree-of-freedom ball-and-socket joint; the knee joint is defined as a two-degree-of-freedom joint including flexion / extension and adduction / abduction directions; the ankle joint and subtalar joint are modeled as single-degree-of-freedom rotator joints (Hinge Joint).

[0028] Muscle-driven: The model is driven by more than 100 muscle-tendon units that follow the Hill-type constitutive relation, which can accurately simulate muscle force transmission and joint dynamic response during dynamic activities.

[0029] To achieve a high degree of matching between the model and the subject's anatomical features, this embodiment adopts the Length-Mass-Fat (LMF) scaling law.

[0030] Geometric Parameter Identification: Unlike traditional functional calibrations (such as SCoRE / SARA), this embodiment optimizes bone segment dimensions through a parameter identification process. Using pelvic width, thigh and calf lengths as design variables, a least-squares optimization algorithm minimizes the residuals between virtual markers and experimental trajectories, thereby dynamically determining the joint center position and effectively reducing the impact of soft tissue artifacts (STA) on computational accuracy.

[0031] Muscle strength scaling strategy: Using the Deurenberg formula, BMI is converted into body fat percentage.

[0032] Where BMI is weight (kg) / height^2 (m^2), Age is the subject's age, and Sex is gender, with 1 for males and 0 for females.

[0033] To avoid the effects of allometric growth on high BMI subjects This leads to an overestimation of muscle strength. To address this, this embodiment introduces a physiological cross-sectional area (PCSA) correction strategy based on lean body mass (Lean Body Mass). Its strength scaling factor... Follow the following relationship:

[0034] Where M represents mass, BF represents body fat percentage, and L represents bone segment length. This method, by incorporating body fat parameters, ensures the biological validity of muscle output intensity when simulating obese or low-muscle-mass subjects.

[0035] Step 3: Model Correlation Study After completing the data accumulation, the system uses regular expressions to parse the patient's KL grade information, determine the bilateral severity, and distinguish between the "affected side" and the "better side", thus establishing a severity labeling system.

[0036] Next, by combining an improved algorithm of low-pass filtering and peak detection, continuous gait cycles are identified in the vertical ground reaction force, and temporal features such as the number of cycles, duration, proportion of standing period, and gait regularity are automatically extracted.

[0037] Then, the statistical moment characteristics (maximum, mean, standard deviation, skewness, kurtosis, etc.) of the target biomechanical signal are calculated at the periodic level, and together with the gait temporal characteristics, they form a unified feature set.

[0038] To achieve personalized severity characterization, this embodiment designs three database construction schemes: (1) A unified database was used, without distinguishing between left and right legs, and the left and right legs of patients were predicted independently; (2) A two-tiered database was used, distinguishing between the affected side and the better side, and the severity and side information were labeled respectively; (3) Only the periodic samples of the most severe side were retained. After all samples were processed by stratified missing value imputation, IQR tailing and robust standardization, they entered the statistical analysis and machine learning stage. A systematic analysis framework was used to explore the association between gait biomechanical characteristics and disease severity in patients with knee osteoarthritis (KOA). First, the correlation between each feature and severity was evaluated by the Mann-Whitney U test and a linear mixed-effects model (with severity score as a fixed effect and patient ID as a random intercept), and the Benjamini-Hochberg method was used for multiple comparison correction. In the feature selection stage, prior knowledge and data-driven methods (Spearman correlation, mutual information and random forest feature importance) were integrated to construct the final feature set.

[0039] Based on this, Lasso, Ridge, ElasticNetwork, RandomForest, XGBoost, and LightGBM models were trained and compared on the affected and contralateral data, respectively, using five-fold cross-validation with R... 2 The performance of the models was evaluated using MSE and MAE. The classification performance of the six machine learning classification methods is compared in Table 2.

[0040] Table 2 Comparison of classification performance of six algorithm models

[0041] Based on the performance comparison of the six machine learning classification methods in the table above, we can conclude that: In most models, the SevereOnly (retaining only periodic samples from the most severely affected side) scheme achieved the highest overall classification performance, particularly in RF and XGBoost, reaching approximately 94% and 91% accuracy respectively, demonstrating the best performance. The Affected / Unaffected (two-tiered database) scheme achieved high prediction accuracy for the affected side (e.g., 91% for the affected side in RF), but its prediction performance for the better side decreased significantly, reflecting a significant difference in the feature distribution between the two sides. The Unified (unified database) scheme generally performed worse than the other two schemes overall.

[0042] The above studies indicate that, due to the influence of biomechanical compensation, even within the same disease grade, there is a high degree of heterogeneity in the kinematic / dynamic characteristics between the affected leg and the better leg of different patients, and even between the "affected leg" of different patients.

[0043] Therefore, this embodiment will select scheme 3 for study.

[0044] Based on the aforementioned scheme 3, six algorithms are used to obtain a ranking of feature relevance scores for each of the 64 different dimensions of anybody features, resulting in the following: Figure 2 The top five most relevant key features for each algorithm are shown in the figure. The left side of the figure displays the top five important features and their importance scores obtained based on various feature selection algorithms. Among them, the knee proximal-distal force has the highest importance (0.95), followed by the ankle medial-latal force (0.85) and the vastus lateralis force (0.75), indicating that deep dynamic features play a key role in KOA grading. The middle part compares the accuracy of six machine learning models (random forest, XGBoost, LightGBM, LASSO, Ridge, and elastic network) in the grading prediction task. Random forest and XGBoost performed best (accuracies of 0.90 and 0.80, respectively).

[0045] Step 4: Gait Feature Deep Learning Hierarchy To further improve model performance and interpretability, this embodiment constructs a deep learning model based on key features selected using random forest and XGBoost algorithms, targeting gait data from the side with the most severe condition in the database. The overall technical approach mainly includes the following modules: Model Architecture and Training Strategy: To address the challenge of small sample sizes in the grading diagnosis of knee osteoarthritis (KOA), this embodiment constructs a two-stream adaptive neural network architecture. The model design employs a two-branch structure that processes static clinical features and dynamic gait sequences in parallel, achieving efficient integration of multi-source information through a feature fusion layer. The network adaptively learns multi-dimensional biomechanical indicators such as joint mechanics, muscle activity, kinematic angles, and temporal features, effectively capturing key discriminative features related to KOA grading in gait patterns.

[0046] Training Optimization and Validation Framework: The model training employs a stratified 5-fold cross-validation strategy, with each fold comprising 80% of the training set and 20% of the validation set. Considering the relatively small sample size of the healthy and mild groups, a class weight balancing strategy is introduced into the training loss function to further compensate for the potential bias of the classification decision boundary caused by moderate sample imbalance. This strategy guides the algorithm from 'pursuing overall accuracy' to 'balancing the sensitivity of each level' by assigning a higher penalty for misclassification to the minority classes (healthy and mild), thereby minimizing the differences in data representation caused by uneven distribution of subjects, improving the robustness of the model in identifying early lesion features, and ensuring a balanced distribution of KOA levels. The optimizer used is Adam, with an initial learning rate of 0.001 (reduced to 0.0005 for small sample scenarios), and a warmup (15 epochs) and cosine annealing learning rate scheduling strategy. To enhance the model's generalization ability, gradient clipping (clipnorm=1.0), L2 regularization (λ=0.001), Dropout (0.1-0.3), and batch normalization were introduced during training. Overfitting was controlled using a smooth early stopping mechanism based on exponential moving average (SmoothEarlyStopping, patience=30, min_delta=0.0005), terminating training when the validation loss showed no significant improvement for 30 consecutive epochs; model parameter configurations are shown in Table 3.

[0047] Table 3 Deep Learning Model Parameter Configuration

[0048] Model performance evaluation: Refer to Figure 3 The fusion model demonstrated excellent convergence and classification performance in five-fold cross-validation. The average accuracy reached 90.7% (standard deviation ± 3.2%), with precision, recall, and F1 score of 91.9%, 88.1%, and 92.8%, respectively. In the graph, the horizontal axis represents the number of training epochs, and the vertical axes represent accuracy and loss, respectively. As the number of training epochs increased, the model accuracy gradually increased from an initial 0.29, reaching 0.85 after 40 epochs, stabilizing above 0.91 after 60 epochs, and finally converging to 0.93 after 80-100 epochs; the loss value decreased from 0.28 to 0.89 and then plateaued. These curves demonstrate the effectiveness and convergence of the model training, and no overfitting was observed.

[0049] Compared to traditional clinical grading methods (with an accuracy rate typically of 65-75%), the model proposed in this embodiment achieves a significant improvement in diagnostic performance, fully validating the important value of multi-dimensional gait biomechanical characteristics in objective KOA grading. As gait data becomes more complete and readily available, the prediction accuracy will further improve in later stages.

[0050] The foregoing embodiments fully disclose the complete technical solution of this invention, from non-invasive gait data acquisition, standardized data preprocessing, personalized musculoskeletal modeling and inverse dynamics calculation, multi-dimensional feature engineering optimization, to the construction of a dual-stream adaptive neural network model and automatic grading and prediction of KOA severity. The core operational specifications, key parameter configurations, technical innovation logic, and laboratory validation results for each stage are clearly defined. To enable those skilled in the art to more clearly understand the actual implementation methods, clinical application boundaries, and real-world application value of this invention, and to fully reproduce the technical solution of this invention, the application process of this invention is divided into four stages below. Each stage has clear operational steps and technical content. Further detailed explanations and supplementary descriptions are provided regarding the scenario-based operational details, case validation effects, and application advantages compared to traditional assessment methods in actual clinical applications of this invention.

[0051] Phase 1: Data Collection – Getting Patients to "Take a Few Steps" Patients do not need to fast or have a special diet; they only need to wear close-fitting clothing. Based on the CAST marker system, 36 reflective marker balls are attached to specific anatomical locations on the patient's pelvis, both thighs, lower legs, and feet. Key locations include: the anterior superior iliac spine, posterior superior iliac spine, lateral epicondyle of the femur, medial epicondyle of the femur, tibial tuberosity, medial malleolus, lateral malleolus, first metatarsal head, fifth metatarsal head, and heel.

[0052] The patient stands in the center of the data acquisition area with arms and legs naturally extended, maintaining this static posture for approximately 3-5 seconds. The system records the positions of static markers for anatomical reference during subsequent skeletal model creation. Eight markers around the bilateral patellae and ankle joints are removed (these markers are prone to dislodgement during dynamic data acquisition and do not affect core data). The patient walks back and forth at a natural and comfortable pace on a 10-meter-long walkway. Technicians observe the patient's gait, ensuring it is relaxed and natural, avoiding any deliberate alteration of gait patterns. Each patient is recorded 3-5 effective walks, each containing multiple complete gait cycles (from heel strike on one side to heel strike on the same side again).

[0053] During the data collection process, eight Qualisys infrared cameras captured the three-dimensional coordinates of the marker points at a frequency of 120Hz, while two Kistler force tables simultaneously recorded the reaction force of the foot contacting the ground. The entire data collection process took approximately 15-20 minutes, requiring only normal walking by the patient, with no radiation exposure or invasive procedures.

[0054] Phase Two: Backend Data Processing – Automated “Deep Computation” (1) Data repair and preprocessing Trajectory repair: If individual marker data is temporarily lost due to occlusion, a cubic spline interpolation algorithm is used to automatically fill in the missing data, ensuring the continuity and integrity of the time series data.

[0055] Filtering and noise reduction: The ground reaction force signal collected by the force measuring station is processed by a fourth-order bidirectional low-pass Butterworth filter with a cutoff frequency of 10Hz to eliminate high-frequency noise and retain the main dynamic characteristics of the gait support phase.

[0056] (2) Personalized musculoskeletal modeling Model scaling: The Length-Mass-Fat (LMF) scaling law is adopted to adjust the general model into a personalized model based on the patient's actual height, weight, age, and gender. Through the parameter recognition process, pelvic width, thigh length, and calf length are used as design variables. Least squares optimization is used to minimize the error between the virtual marker point and the actual marker point trajectory, so as to accurately determine the center position of the hip, knee, and ankle joints.

[0057] Muscle strength correction: The Deurenberg formula is introduced to calculate the patient's body fat percentage, and then a physiological cross-sectional area correction strategy based on lean body mass is used to calculate the muscle strength scaling factor. This avoids the overestimation of muscle strength in patients with high BMI and ensures that the simulation results of obese or low muscle mass populations are consistent with biological reality.

[0058] (3) Inverse dynamics calculation Run inverse dynamics simulation to simulate how muscles exert force and joints are stressed during the patient's walking process.

[0059] The calculations include: joint reaction forces: contact forces (proximal-distal force, anterior-posterior shear force, medial-lateral force) of the hip, knee, and ankle joints in three directions; joint torques: flexion-extension, adduction-abduction, and internal-external rotation torques of each joint; muscle activation: activation level and force contribution of more than 100 muscles, including the quadriceps, hamstrings, and gastrocnemius; and joint angles and range of motion: kinematic parameters of each joint during the gait cycle.

[0060] (4) Feature extraction and screening A total of 64 multidimensional features were automatically extracted from various parts of the lower limbs based on the simulation results. These features include: Statistical characteristics: peak value, trough value, mean, variance, skewness, and kurtosis of each indicator. Timing characteristics: Waveform data of the entire gait cycle Symmetry characteristics: Indicators of difference between the affected side and the healthy side Clinical characteristics: age, sex, BMI, etc. The system incorporates six feature selection algorithms (random forest, XGBoost, mutual information, LASSO, recursive feature elimination, and tree model importance) to automatically filter out the high-value feature subset most relevant to KOA grading. Based on the aforementioned research, deep dynamic features such as proximal-distal knee joint force, medial-lateral ankle joint force, and vastus lateralis muscle force contribute the most to grading.

[0061] (5) Deep learning model inference The filtered features are input into a trained dual-stream adaptive neural network model. This model contains two parallel branches: one processes static clinical features (age, BMI, etc.), and the other processes dynamic temporal features (joint force sequences, muscle activation waveforms, etc.), and finally fuses and outputs the grading results.

[0062] Phase Three: Results Output – Doctors receive a “quantitative report” The system automatically generates a visually appealing evaluation report, containing the following core elements: (1) Basic patient information: Name, age, gender, height, weight, BMI, and collection date. (2) KOA Severity Classification Grading results: Mild / Moderate / Severe (corresponding to KL grading levels 1-2, 3, and 4) Confidence level: For example, "Moderate KOA, model confidence level 91.9%". Grading criteria: Briefly explain the key indicators on which the model's judgment is based. (3) Key biomechanical indicators Joint load: Peak contact force of the knee joint (expressed as a multiple of body weight, such as "3.2 times body weight"), compared with the reference value for healthy people of the same age (such as "15% higher than the normal range"). Muscle function: Activation levels of key muscles such as the quadriceps and hamstrings to assess muscle function status. Gait symmetry: Comparison of gait parameters on the left and right sides to identify abnormal gait patterns. Abnormal indicators: List indicators that are outside the normal range, such as "significantly increased medial load on the knee joint" and "insufficient activation of the vastus lateralis muscle".

[0063] (4) Visual charts Joint force curve: Shows the change in knee joint contact force during the gait cycle, and marks the peak position. Muscle activation diagram: Activation timing and intensity of key muscles during the gait cycle Radar chart: A comprehensive comparison of multiple indicators, intuitively showing the differences between patients and healthy individuals. Trend chart: If multiple assessments have been conducted, it shows the changing trends of key indicators. Phase Four: Clinical Application – Physicians make decisions based on this information. Scenario Example: Assisting Diagnosis – Making Diagnosis Data-Driven Traditional dilemma: Doctors interpret X-rays to determine the KL classification, but the judgments of different doctors, and even the same doctor at different times, may differ, making it highly subjective.

[0064] The solution of this invention is that doctors receive an objective and quantitative grading report as a reference for diagnosis.

[0065] Case: A 55-year-old female patient's X-ray showed mild narrowing of the joint space, making it difficult for the doctor to determine whether she had reached moderate KOA. Based on the gait analysis report of this invention, her peak knee contact force reached 3.5 times her body weight (far exceeding the 2.8 times of healthy women of the same age), indicating insufficient quadriceps activation. The model output "Moderate KOA, confidence level 92%". Based on this result, the doctor diagnosed her with moderate KOA and recommended rehabilitation treatment.

[0066] The embodiments of the present invention also relate to a knee osteoarthritis grading prediction system based on gait analysis and deep learning. This system is used to implement the prediction method of the aforementioned embodiments, including: The subject gait data synchronous acquisition module is used to synchronously acquire the subject's lower limb three-dimensional kinematic trajectory data and ground reaction force data under natural gait through an optical motion capture system and a force measuring platform; The gait data standardization preprocessing module is used to perform trajectory missing repair, standardized naming of marker points, and ground reaction signal filtering and noise reduction on the collected raw data to obtain standardized gait data that meets the requirements of musculoskeletal modeling. The personalized musculoskeletal modeling and inverse dynamics feature extraction module is used to import the obtained standardized gait data into the musculoskeletal modeling system, and construct a personalized lower limb musculoskeletal model that matches the anatomical features of the subject through personalized model scaling and muscle strength correction; and run inverse dynamics simulation to extract multi-dimensional biomechanical features of lower limb joints and muscles during the gait cycle. The feature set and optimized sample library construction module is used to establish a KOA severity labeling system by combining the KL clinical grading information of subjects' knee osteoarthritis, and to integrate the extracted biomechanical features and the subjects' static clinical features to construct a multi-dimensional full feature set; and to construct a SevereOnly sample library that retains only the gait cycle samples of the most severe KOA side of the subjects, and to perform standardized cleaning processing on the sample library data. The high-relevance feature selection module, which integrates multiple algorithm fusions, is used to select a subset of core features that are highly correlated with the severity level of KOA from the full feature set. The KOA severity prediction deep learning model construction module is used to build a two-stream adaptive neural network model. The model sets up parallel static feature branches and dynamic temporal feature branches. The two branches perform adaptive feature extraction on static clinical features and dynamic temporal biomechanical features in the core feature subset, respectively. After multi-source feature fusion is completed by the feature fusion layer, the KOA severity grading result is output through the classification output layer. The hierarchical cross-validation strategy is used to complete the model training and optimization, resulting in a trained KOA grading prediction model. The KOA grading prediction and result output module is used to process the gait data of the subjects to be evaluated through the aforementioned modules, input it into the trained KOA grading prediction model, and automatically output the KOA severity grading results.

[0067] Figure 4 This is a schematic diagram of an electronic device 10 provided in another embodiment of the present invention. (See diagram below.) Figure 4 As shown, the electronic device 10 of this embodiment includes: a processor 11, a memory 12, and a computer program 13 stored in the memory 12 and executable on the processor 11, such as a program for a knee osteoarthritis grading prediction method based on gait analysis and deep learning. When the processor 11 executes the computer program 13, it implements the steps in the above embodiment of the knee osteoarthritis grading prediction method based on gait analysis and deep learning, for example... Figure 1 The process steps are shown.

[0068] For example, computer program 13 may be divided into one or more modules / units, one or more of which are stored in memory 12 and executed by processor 11 to complete the present invention. One or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 13 in electronic device 10.

[0069] Electronic device 10 can be a desktop computer, laptop, handheld computer, cloud server, or other computing device. Electronic device 10 may include, but is not limited to, a processor 11 and a memory 12. Those skilled in the art will understand that... Figure 4 This is merely an example of electronic device 10 and does not constitute a limitation on electronic device 10. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device 10 may also include input / output devices, network access devices, buses, etc.

[0070] The processor 11 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0071] The memory 12 can be an internal storage unit of the electronic device 10, such as a hard disk or RAM of the electronic device 10. The memory 12 can also be an external storage device of the electronic device 10, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the electronic device 10. Furthermore, the memory 12 can include both internal and external storage units of the electronic device 10. The memory 12 is used to store computer programs and other programs and data required by the electronic device 10. The memory 12 can also be used to temporarily store data that has been output or will be output.

[0072] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0073] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments or adapt to the prior art.

[0074] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0075] In the embodiments provided by this invention, it should be understood that the disclosed devices / electronic devices and methods can be implemented in other ways. For example, the device / electronic device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0076] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0077] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0078] If integrated modules / units are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0079] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A knee osteoarthritis grading prediction method based on gait analysis and deep learning, characterized in that, Includes the following steps: Gait data of S1 subjects were collected simultaneously: The optical motion capture system and force platform are used to simultaneously collect three-dimensional kinematic trajectory data of the lower limbs and ground reaction force data of the subject in natural gait. S2 gait data normalization preprocessing: The raw data collected in step S1 is subjected to trajectory loss repair, standardized naming of marker points, and ground reaction signal filtering and noise reduction to obtain standardized gait data that meets the requirements of musculoskeletal modeling. S3 Personalized Musculoskeletal Modeling and Inverse Dynamics Feature Extraction: The standardized gait data obtained in step S2 is imported into the musculoskeletal modeling system. A personalized lower limb musculoskeletal model matching the subject's anatomical features is constructed by personalizing the model scaling and correcting muscle strength. Inverse dynamics simulation is run to extract multi-dimensional biomechanical features of the lower limb joints and muscles during the gait cycle. S4 Feature Set and Optimized Sample Library Construction: A KOA severity labeling system was established by combining the KL clinical grading information of the subjects. The biomechanical features extracted in step S3 and the static clinical features of the subjects were integrated to construct a multi-dimensional full feature set. A SevereOnly sample library was constructed, which only retains the gait cycle samples of the most severe side of the subjects' KOA. The sample library data was standardized and cleaned. S5 multi-algorithm fusion for high-correlation feature selection: By integrating multiple feature selection algorithms, a subset of core features highly correlated with KOA severity rating is selected from the full feature set; S6 KOA hierarchical prediction deep learning model construction: A dual-stream adaptive neural network model is constructed, which sets up parallel static feature branches and dynamic temporal feature branches. The two branches perform adaptive feature extraction on static clinical features and dynamic temporal biomechanical features in the core feature subset, respectively. After multi-source feature fusion is completed by the feature fusion layer, the KOA severity classification result is output through the classification output layer. The hierarchical cross-validation strategy is used to complete the model training and optimization, and the trained KOA classification prediction model is obtained. S7 KOA classification prediction and results output: After the gait data of the subjects to be evaluated are processed in steps S2-S5, they are input into the KOA grading prediction model trained in step S6, and the KOA severity grading results are automatically output.

2. The gait analysis and deep learning based knee osteoarthritis grading prediction method according to claim 1, characterized in that, In step S1, before data acquisition, 36 reflective marker balls are attached to designated anatomical locations on the subject's pelvis, thigh, calf and foot according to the CAST marker system. After static calibration is completed, 8 markers around the bilateral patella and ankle joint are removed. Subjects walked back and forth on a 10m trail with a natural gait, and three-dimensional kinematic trajectory data and ground reaction force data were collected simultaneously at a frequency of 120Hz. At least three valid walking data were collected for each group. 3.The gait analysis and deep learning based knee osteoarthritis grading prediction method according to claim 1, characterized in that, In step S2, the trajectory missing repair uses a cubic spline interpolation algorithm to fill in the missing marker point data caused by occlusion; The standardized naming of marker points was completed according to the naming conventions of the musculoskeletal modeling system. The ground reaction force signal filtering and noise reduction was achieved using a fourth-order bidirectional low-pass Butterworth filter with a cutoff frequency of 10Hz.

4. The method for predicting the grading of knee osteoarthritis based on gait analysis and deep learning according to claim 1, characterized in that, In step S3, the musculoskeletal modeling system adopts the AnyBody musculoskeletal modeling system, and a personalized model is constructed based on the AMMR lower limb anatomical model. The personalized scaling of the model adopts the Length-Mass-Fat LMF scaling law, and minimizes the residual between the virtual marker point and the experimental trajectory through the least squares optimization algorithm to dynamically determine the joint center position; The muscle strength correction is calculated by the Deurenberg formula to determine the subject's body fat percentage, and a physiological cross-sectional area (PCSA) correction strategy based on lean body mass is introduced to complete the personalized correction of the muscle strength scaling factor. The multidimensional biomechanical features extracted by the inverse dynamics simulation include 64 multidimensional features such as lower limb joint reaction force, joint torque, muscle activation, muscle strength, and the peak, mean, standard deviation, skewness, and kurtosis of the joint range of motion.

5. The method for predicting the grading of knee osteoarthritis based on gait analysis and deep learning according to claim 1, characterized in that, In step S4, the KOA severity labeling system is divided into three levels: mild KOA, moderate KOA, and severe KOA, corresponding to KL levels 1-2, 3, and 4. The static clinical characteristics include the subject's age, sex, height, weight, BMI, and walking speed; Before constructing the SevereOnly sample library, the model classification performance of three schemes—unified database, dual-layered database, and SevereOnly database—was compared, and the SevereOnly scheme was ultimately determined to be the optimal sample library construction scheme. The standardized cleaning process includes stratified missing value filling, IQR tail reduction, and robust standardization.

6. The method for predicting the grading of knee osteoarthritis based on gait analysis and deep learning according to claim 1, characterized in that, In step S5, the multiple feature selection algorithms specifically include six algorithms: Random Forest, XGBoost, Mutual Information, LASSO, Recursive Feature Elimination, and Tree Model Importance. By fusing and filtering these six algorithms, a subset of core features with the highest correlation to the KOA severity rating is obtained.

7. The method for predicting the grading of knee osteoarthritis based on gait analysis and deep learning according to claim 1, characterized in that, In step S6, the static feature branch of the dual-stream adaptive neural network model is used to process static clinical features including age, gender, and BMI, while the dynamic temporal feature branch is used to process dynamic temporal biomechanical features including joint force and muscle activation. The model training adopts a five-fold hierarchical cross-validation strategy, with each fold dividing the training set and validation set in an 8:2 ratio. The optimizer uses the Adam optimizer, combined with warmup and cosine annealing learning rate scheduling strategies; During training, gradient clipping, L2 regularization, Dropout, batch normalization, and smooth early stopping mechanisms are used to suppress overfitting.

8. The method for predicting the grading of knee osteoarthritis based on gait analysis and deep learning according to claim 1, characterized in that, In step S7, while outputting the KOA severity grading results, a standardized assessment report is generated simultaneously. The assessment report includes basic information of the subjects, KOA classification results and prediction confidence levels, key biomechanical indicators, comparison results with reference values ​​of healthy individuals, and visual analysis charts.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it implements the steps of the knee osteoarthritis grading prediction method based on gait analysis and deep learning as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by the processor, it can implement the steps of the knee osteoarthritis grading prediction method based on gait analysis and deep learning as described in any one of claims 1 to 8.

Images