Data classification method and system for normal and abnormal temporomandibular movement

Abstract

The application provides a method and system for classifying normal and abnormal temporomandibular joint movement data. The method collects temporomandibular joint movement data from a temporomandibular joint movement visualization platform, which includes normal movement data and abnormal movement data. The temporomandibular joint movement data is converted into structured data features. The structured data features are preprocessed, missing values are filled, and the data is dimensionless to obtain preprocessed data features. The preprocessed data features are divided into a training set and a test set. A temporomandibular joint movement classification model is constructed, which includes a decision tree model, an ensemble learning model, and a support vector machine model. The training set and corresponding classification labels are used to train and optimize the temporomandibular joint movement classification model to obtain an optimal model. The optimal model is used to predict the test set to obtain a classification result of normal or abnormal temporomandibular joint movement. The classification method has high precision and can obtain a high-accuracy abnormal movement detection result. It is objective and can effectively assess the risk of temporomandibular joint disorder (TMD).

Description

Technical Field

[0001] This invention relates to a data classification method and system for normal and abnormal temporomandibular movement, belonging to the field of data classification technology. Background Technology

[0002] Temporomandibular joint disorder (TMD) is a common oral and maxillofacial condition, clinically manifested as pain in the temporomandibular joint area, abnormal mandibular movement, and clicking or chattering sounds when opening and closing the mouth. Abnormal temporomandibular joint function can significantly impact daily life. Research on temporomandibular joint movement can provide a deeper understanding of the anatomical structure and physiological function of the human temporomandibular joint.

[0003] In clinical research, doctors and other researchers often use medical imaging and anatomical methods to empirically observe and statistically analyze oral movements for clinical diagnosis and evaluation. However, these methods suffer from subjectivity, insufficient reproducibility of results, and limited transferability of experience. Furthermore, because the condyle is located deep within the body surface, its movements cannot be directly recorded, increasing the difficulty of research. Currently, there is a lack of a data classification method and system for normal and abnormal temporomandibular movements to achieve objective detection of temporomandibular movement abnormalities.

[0004] The above-mentioned issues should be considered and resolved in the process of classifying data on normal and abnormal temporomandibular movements. Summary of the Invention

[0005] The purpose of this invention is to provide a data classification method and system for normal and abnormal temporomandibular movements, which solves the problem in the prior art that it is difficult to obtain objective and high-precision classification results for normal and abnormal temporomandibular movements.

[0006] The technical solution of this invention is:

[0007] A data classification method for normal and abnormal temporomandibular movement includes the following steps:

[0008] S1. Temporomandibular joint motion data is collected by the temporomandibular joint motion visualization platform. The temporomandibular joint motion data includes normal motion data and abnormal motion data. The temporomandibular joint motion data is converted into structured data features.

[0009] S2. Preprocess the structured data features, fill in missing values, and make the data dimensionless to obtain the preprocessed data features. Divide the preprocessed data features into training set and test set.

[0010] S3. Construct a temporomandibular movement classification model, which includes a decision tree model, an ensemble learning model, and a support vector machine model. After training and optimizing the temporomandibular movement classification model using the training set and corresponding classification labels, the optimal model is obtained.

[0011] S4. Use the optimal model to predict the test set and obtain the classification result of normal or abnormal temporomandibular joint movement.

[0012] Furthermore, in step S1, the structured data features include motion features of the temporomandibular joint range of motion, motion features of the maximum three-dimensional displacement, motion features of the motion repositioning, and motion features of the motion angle.

[0013] Furthermore, in step S2, the structured data features are preprocessed, missing values ​​are filled, and the data is made dimensionless. Specifically,

[0014] S21. For missing data in the structured data features, use the relevant parameters from the corresponding normal motion data and abnormal motion data samples to calculate the mean and fill in the missing values.

[0015] S22. The data is made dimensionless, and the features are transformed into a normal distribution with a mean of 0 and a standard deviation of 1 by using decentralization and scaling transformation.

[0016] Further, in step S3, the temporomandibular movement classification model is trained and optimized using the training set and corresponding classification labels to obtain the optimal model. Specifically,

[0017] The optimization methods used to train and optimize the decision tree model include random branching during model construction, limiting the minimum number of samples per node, limiting the maximum depth, and limiting the number of branch features selected.

[0018] Optimization methods for training and optimizing ensemble learning models include limiting the number of base learners and limiting the optimization step size.

[0019] Optimization methods for training and optimizing support vector machine models include adjusting kernel function parameters and regularization parameters.

[0020] Set performance evaluation metrics for the models, and select the best-performing model among the trained and optimized decision tree model, ensemble learning model, and support vector machine model as the optimal model.

[0021] A system for implementing the data classification method for normal and abnormal temporomandibular movement as described in any of the preceding claims includes a data structuring module, a data preprocessing module, a model training and optimization module, and a classification module.

[0022] Data structuring module: The temporomandibular joint motion data is collected by the temporomandibular joint motion visualization platform. The temporomandibular joint motion data includes normal motion data and abnormal motion data. The temporomandibular joint motion data is converted into structured data features.

[0023] Data preprocessing module: preprocesses the structured data features, fills in missing values, and makes the data dimensionless to obtain preprocessed data features, and divides the preprocessed data features into training set and test set;

[0024] Model training and optimization module: Construct a temporomandibular movement classification model, which includes a decision tree model, an ensemble learning model, and a support vector machine model. After optimizing and training the temporomandibular movement classification model using the training set and corresponding classification labels, the optimal model is obtained.

[0025] Classification module: Uses the optimal model to predict the test set and obtains the classification result of normal or abnormal temporomandibular joint movement.

[0026] A computer storage medium storing a computer program, which, when executed by a processor, implements the steps of a data classification method for normal and abnormal temporomandibular movements.

[0027] An electronic device includes a memory and one or more processors, the memory being used to store one or more programs; when the one or more programs are executed by the one or more processors, a data classification method for normal and abnormal temporomandibular movement is implemented.

[0028] The beneficial effects of this invention are as follows: This method and system for classifying normal and abnormal temporomandibular movements eliminates the adverse effects of differences between feature data on the results through various data processing techniques, obtains normal and abnormal movement data, and then obtains classification results through an optimal model. The classification accuracy is high, and it can obtain abnormal movement detection results with a high accuracy rate. It is objective and can effectively and intelligently assess the risk of temporomandibular movement abnormalities or temporomandibular joint disorder (TMD), which has certain clinical significance in the field of oral medicine. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating the data classification method for normal and abnormal temporomandibular movement according to an embodiment of the present invention.

[0030] Figure 2 This is a schematic diagram of a visualization example of modeled point cloud data in the embodiment, wherein (a) is a schematic diagram of modeled point cloud data of normal motion data, and (b) is a schematic diagram of modeled point cloud data of abnormal motion data;

[0031] Figure 3 This is a schematic diagram illustrating the classification of temporomandibular joint motion data in an embodiment.

[0032] Figure 4The diagram shows the confusion matrices of the decision tree model DT and two ensemble learning-based random forest (RF) and GBDT in the embodiment. (a) is a diagram of the confusion matrix of the decision tree model DT, (b) is a diagram of the confusion matrix of the random forest (RF) based on ensemble learning, and (c) is a diagram of the confusion matrix of the gradient boosting decision tree (GBDT) based on ensemble learning.

[0033] Figure 5 These are confusion matrix diagrams of the Support Vector Machine (SVM) models under different kernel functions in the embodiments. Among them, (a) is the confusion matrix diagram of the SVM model based on the linear kernel function, (b) is the confusion matrix diagram of the SVM model based on the polynomial kernel function, (c) is the confusion matrix diagram of the SVM model based on the Gaussian radial basis kernel function, and (d) is the confusion matrix diagram of the SVM model based on the Sigmoid kernel function.

[0034] Figure 6 This is a schematic diagram comparing the ROC curves of various models, including DT, RF, GBDT, and RBF, in the embodiments.

[0035] Figure 7 This is a schematic diagram illustrating the data classification system for normal and abnormal temporomandibular movements in an embodiment. Detailed Implementation

[0036] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0037] The embodiment provides a data classification method for normal and abnormal temporomandibular movement, such as Figure 1 and Figure 3 This includes the following steps:

[0038] S1. Temporomandibular joint motion data is collected by the temporomandibular joint motion visualization platform. The temporomandibular joint motion data includes normal motion data and abnormal motion data. The temporomandibular joint motion data is converted into structured data features.

[0039] In step S1, the temporomandibular joint motion data is modeled using point cloud data, and the visualization example is as follows: Figure 2 Temporomandibular joint movement data includes normal movement data such as Figure 2 (a) and abnormal motion data such as Figure 2 (b) By selecting a reference point and calculating the motion characteristics of the reference point, the temporomandibular joint motion data is converted into structured data features. The structured data features include the motion characteristics of the temporomandibular range of motion, the motion characteristics of the maximum three-dimensional displacement, the motion characteristics of the motion repositioning, and the motion characteristics of the motion angle.

[0040] S2. Preprocess the structured data features, fill in missing values, and make the data dimensionless to obtain the preprocessed data features. Divide the preprocessed data features into training set and test set.

[0041] In step S2, the structured data features are preprocessed, missing values ​​are filled, and the data is made dimensionless. Specifically,

[0042] S21. For missing data in the structured data features, use the relevant parameters from the corresponding normal motion data and abnormal motion data samples to calculate the mean and fill in the missing values.

[0043] S22. The data is made dimensionless, and the features are transformed into a normal distribution with a mean of 0 and a standard deviation of 1 by using decentralization and scaling transformation.

[0044] In step S2, some temporomandibular joint motion data involves angle calculations and cosine calculations. During the programming solution process, data loss may occur due to issues such as "division by zero," so it is necessary to fill in the missing values ​​of this part of the data. A feature selection imputation method is used, where missing values ​​for normal and abnormal motion data are filled by selecting relevant parameters from the normal and abnormal motion data samples, respectively, and calculating the mean. Simultaneously, since there are many types of motion features involved, each with different units of measurement, there are significant differences in calculation scales and measurement standards. This embodiment employs dimensionless data processing to eliminate the negative impact of feature scale variations and unit inconsistencies. This is mainly achieved through decentralization and scale transformation, converting the features in the data to an approximate range so that the features are not affected by dimensions and units. The main methods for dimensionless data processing include: data normalization, scaling normalization, logarithmic transformation, and data standardization. The data processing procedures and principles are as follows:

[0045] To obtain standardized data features, subtract the mean μ of each feature data point x from the structured data's feature data, and then divide by the standard deviation σ of that feature data. After centering by the minimum value in the current feature of the data, the data is scaled according to the difference between the extreme values ​​of the current feature in a mathematical sense, resulting in the scaled data feature: By dividing the original data by the magnitude of the vector, the data is scaled to a unit length, resulting in normalized data features. Ultimately, the data converges to a selected range, typically between 0 and 1 or between -1 and 1. The converged data follows a normal distribution.

[0046] S3. Construct a temporomandibular movement classification model, which includes a decision tree model, an ensemble learning model, and a support vector machine model. After training and optimizing the temporomandibular movement classification model using the training set and corresponding classification labels, the optimal model is obtained.

[0047] In step S3, a decision tree model is constructed using the CART algorithm. Node and branch divisions are based on impurity, which represents the degree to which a data sample within a node belongs to the normal or abnormal classification. The parameters for evaluating impurity include the Gini coefficient (Gini(t)) and the information entropy (Entropy(t)).

[0048]

[0049]

[0050] Where P(n|t) represents the proportion of the classification label n, such as normal sample or abnormal sample, at node t.

[0051] The node partitioning process when constructing a decision tree model using the CART algorithm is as follows:

[0052] Step 1: Select a mandibular movement feature f and a classification threshold h to binary classify the structured dataset into two subsets, G1 and G2, which are either normal or abnormal, and each subset belongs to a different label category. The threshold is selected based on the continuity of the data features, using either the nearest neighbor average or an enumeration method, to find the optimal threshold.

[0053] Step 2: Calculate the impurity of the partitioned subsets, and then calculate the average after weighting the impurities. The weighted averaging method for the Gini coefficient of the feature f threshold h of category G is as follows:

[0054]

[0055] Where |G1| and |G2| represent the number of samples in category G1 and category G2, respectively.

[0056] Information entropy weighted averaging method for category G, feature f, and threshold h:

[0057]

[0058] Where |G1| and |G2| represent the number of samples in category G1 and category G2, respectively.

[0059] Step 3: Select the node partitioning criteria based on the corresponding Gini coefficient or information entropy for different features.

[0060] Step 4: Recursively complete steps 1-3 until the classification decision tree for normal and abnormal temporomandibular movement is constructed.

[0061] The decision tree model is optimized by methods such as random branching during model construction, limiting the minimum number of samples per node, limiting the maximum depth, and limiting the number of branch features selected.

[0062] Gradient boosting trees in an ensemble learning model are used to determine the binary classification result between normal and abnormal temporomandibular movement, with the log-likelihood function chosen as the loss model. The specific model construction process is as follows: using the training set, calculate or select a constant with distributional and statistical significance, and set it as the predicted value for all training sets. Calculate the residual, i.e., the negative gradient error. Where, r i It is the residual of the i-th sample, which is also the objective to be minimized during the iteration process, y i It is the actual label value, f(x) i ) is the model's predicted value for the i-th sample. This represents the current estimate of the model or the current prediction of the model.

[0063] Set the negative gradient error as the target value, train a CART tree as the new base estimator, and compute the best negative gradient fit for each leaf node: Where j represents the leaf node region of the tree, y i It is the actual label value, f(x) i Let be the model's prediction for the i-th sample, and c be the adjusted prediction value caused by the specific feature segmentation point considered. Update the model and continue iterating until the end, using the weighted sum of all base estimators as the final model. Optimize the ensemble learning model by limiting the number of base learners and the optimization step size.

[0064] Support vector machines are used to classify jawbone data, with two labels set for the categories of normal and abnormal temporomandibular movement. The classifier settings are as follows: Among them, y i Represents the actual category label, ω T x+b represents the classification condition. Different kernel functions are used to optimize the support vector machine. x represents structured data points, y represents the classification label of the data, and K represents different kernel functions, specifically the following four types: First, the linear kernel function: K(x,y)=x T y, primarily using linearly separable datasets. The second method uses a polynomial kernel function: K(x,y)=(ax T y+c) d Where a, c, and d represent different adjustable parameters, typically using linearly separable datasets. The third type is the Gaussian radial basis function kernel: σ represents the standard deviation of the feature data, which is mainly based on the Euclidean distance between samples to measure the discreteness. It performs well in various complex situations, especially in nonlinearly separable cases, where its performance is superior to most kernel functions. The fourth type is the Sigmoid kernel function: K(x,y)=tanh(ax) T y+c), where a and c represent different adjustable parameters. Its principle is to perform mapping through hyperbolic tangent. Its shape is like "S" and it has a similar function to biological neurons. It has nonlinear response characteristics to input signals. It is mainly good at nonlinearly separable datasets and is often used in neural networks.

[0065] In step S3, the decision tree model, ensemble learning model, and support vector machine model are optimized respectively, and performance evaluation metrics for the models are set to find the optimal model.

[0066] S4. Use the optimal model to predict the test set and obtain the classification result of normal or abnormal temporomandibular joint movement.

[0067] This method for classifying normal and abnormal temporomandibular movements eliminates the adverse effects of differences between feature data on the results through various data processing techniques. It obtains normal and abnormal movement data, and then uses an optimal model to obtain classification results. The classification accuracy is high, and it can obtain abnormal movement detection results with a high accuracy rate. It is objective and can effectively and intelligently assess the risk of temporomandibular movement abnormalities or temporomandibular joint disorders (TMD), which has certain clinical significance in the field of oral medicine.

[0068] This novel data classification method for normal and abnormal temporomandibular movements is the first to employ a combination of models, including decision tree, ensemble learning, and support vector machine, to establish a model for classifying normal and abnormal temporomandibular movements. Through comprehensive comparison using multiple evaluation metrics, the optimal model for the classification task was determined, achieving high classification accuracy.

[0069] This data classification method for normal and abnormal temporomandibular joint movements is the first to establish structured data of temporomandibular joint movement, consisting of range of motion, maximum three-dimensional displacement, movement repositioning, and movement angle. It classifies normal and abnormal temporomandibular joint movements, which can effectively assist medical practitioners in intelligently assessing the risk of having TMD.

[0070] like Figure 7 The embodiments also provide a system for implementing the data classification method for normal and abnormal temporomandibular movement as described in any of the preceding claims, including a data structuring module, a data preprocessing module, a model training and optimization module, and a classification module.

[0071] Data structuring module: The temporomandibular joint motion data is collected by the temporomandibular joint motion visualization platform. The temporomandibular joint motion data includes normal motion data and abnormal motion data. The temporomandibular joint motion data is converted into structured data features.

[0072] Data preprocessing module: preprocesses the structured data features, fills in missing values, and makes the data dimensionless to obtain preprocessed data features, and divides the preprocessed data features into training set and test set;

[0073] Model training and optimization module: Construct a temporomandibular movement classification model, which includes a decision tree model, an ensemble learning model, and a support vector machine model. After optimizing and training the temporomandibular movement classification model using the training set and corresponding classification labels, the optimal model is obtained.

[0074] Classification module: Uses the optimal model to predict the test set and obtains the classification result of normal or abnormal temporomandibular joint movement.

[0075] This data classification system for normal and abnormal temporomandibular joint movements constructs a temporomandibular joint movement classification model to categorize temporomandibular joint movements as normal or abnormal. The input temporomandibular joint movement data sequentially passes through a data structuring module, a preprocessing module, and a model training and optimization module to obtain the binary classification result. It achieves a high classification accuracy.

[0076] The experimental verification of this method and system for classifying normal and abnormal temporomandibular movements in the embodiment is as follows:

[0077] The models in the examples were evaluated based on confusion matrix, accuracy, recall ROC curve, AUC, etc., and the trained temporomandibular movement classification models were compared and analyzed.

[0078] Table 1 shows the confusion matrix of the temporomandibular movement classification model for normal and abnormal temporomandibular movements. In the confusion matrix, the sum of the four values ​​TN, FN, TM, and FM represents the number of training samples. True and False represent the correctness of the prediction results for normal and abnormal movements, respectively, while Positive and Negative represent normal or abnormal samples, respectively.

[0079] Table 1. Confusion matrix of the temporomandibular movement classification model

[0080]

[0081] The implementation example sets the accuracy to verify the classification precision between individuals with temporomandibular movement disorders and those with normal temporomandibular movement in the dataset. Accuracy is calculated as the ratio of the number of correctly predicted results to the total number of samples. Precision verifies that among the individuals at high risk of oral dysfunction, there are also individuals who actually have oral dysfunction. Recall is the proportion of individuals with temporomandibular movement disorders detected by the classification model to the actual number of individuals with temporomandibular movement disorders. The F1-score is a parameter used to comprehensively evaluate precision and recall; its value is the harmonic mean of the two.

[0082]

[0083]

[0084]

[0085]

[0086] The experimental code used in this example was deployed on a physical machine with an AMD Ryzen 75800X CPU, 32GB of memory, and Windows 10 Professional operating system. The program code was primarily written in Python, specifically Python 3.7.

[0087] The data used in the examples were all collected from temporomandibular joint movement visualization software, which included mouth opening movement data of temporomandibular joint abnormalities and normal individuals. As shown in Table 2, there were a total of 238 sets of experimental data, of which 156 sets of data came from normal individuals and the remaining 82 sets of movement came from patients with oral dysfunction. Seventy percent of the dataset was used as the training set and the remaining thirty percent was used as the test set to verify the training results. The training_test_split function was used to randomly split the dataset, and the training set ratio parameter test_size was set to 0.3.

[0088] Table 2. Data related to mouth opening movements in individuals with temporomandibular movement disorders and those with normal temporomandibular movement.

[0089]

[0090]

[0091] like Figure 4 The confusion matrices represent the classification performance of three decision tree models (DT, RF, and GBDT) on the test set after training on open-ended data. Figure 4The number of TP indicators in the confusion matrices of (a), (b), and (c) and the recall indicators show that the trained DT algorithm detected more patients with mandibular movement abnormalities than the RF and GBDT models based on ensemble learning, and had the highest recall rate of 86.2%, indicating that the DT model has a stronger ability to assess the risk of oral dysfunction.

[0092] Table 3 shows the relevant performance indicators calculated based on the confusion matrix of the DT, RF, and GBDT models for classifying normal and abnormal temporomandibular movements. Evaluating the three models from the perspectives of accuracy and precision, it can be found that the ensemble learning models RF and GBDT have higher accuracy and precision than the DT model. Among them, the GBDT model has the highest accuracy and precision in both metrics, at 86.1% and 82.8% respectively. This indicates that the GBDT model has the highest classification accuracy among the three models for normal and abnormal temporomandibular movements, thus reducing the possibility of misclassifying patients with normal temporomandibular movements as having a high risk of oral dysfunction in the assessment of oral dysfunction risk. The F1-score is a model evaluation metric that combines precision and recall, thus possessing good evaluation capabilities for imbalanced binary classification problems, especially in the medical field for disease assessment. Since the cost of misclassification for diseases is high, a comprehensive evaluation of the model's ability to misclassify diseases is necessary. From the F1-score, GBDT scores the highest at 0.828, indicating that while maintaining accuracy, the GBDT model has the lowest risk rate in the comprehensive assessment of normal individuals and oral functional disorders. In conclusion, the trained GBDT model achieves a classification accuracy of 86.1% and an F1-score of 0.828, making it most suitable for the medical field, assisting doctors in intelligently assessing mandibular movement abnormalities.

[0093] Table 3 Performance metrics for the three models: DT, RF, and GBDT

[0094] DT 0.819 0.735 0.862 0.794 RF 0.833 0.815 0.759 0.786 GBDT 0.861 0.828 0.828 0.828

[0095] Figure 5 This is a confusion matrix diagram of the Support Vector Machine (SVM) model under different kernel functions in the embodiments. Figure 5 (a) is the confusion matrix diagram of the Support Vector Machine (SVM) model based on the Linear kernel function. Figure 5 (b) is the confusion matrix diagram of the support vector machine model SVM based on the polynomial kernel function Poly. Figure 5 (c) is the confusion matrix diagram of the support vector machine model SVM based on the Gaussian radial basis function kernel function RBF. Figure 5Table (d) shows the confusion matrix of the SVM model based on the Sigmoid kernel function. Table 4 shows the relevant performance indicators calculated based on the confusion matrices of the classification performance of SVM models based on four different kernel functions for normal individuals and those with mandibular movement disorders. As can be seen from Table 4, the SVM models based on the Linear and Sigmoid kernel functions have poor risk assessment capabilities for oral functional disorders, with recall rates below 50%. This indicates that SVM models based on these two kernel functions are not suitable for auxiliary risk assessment, and also suggests that the feature distribution of the data is neither completely linearly separable nor completely nonlinearly separable. Furthermore, the SVM model based on the Poly kernel function has a slightly higher recall rate of 63.6%, but it is still much lower than the DT model and the ensemble learning model. Meanwhile, the F1-scores of the SVM models based on these three kernel functions are similar, at 0.6 ± 0.36, indicating that the SVM models based on these three kernel functions have low classification capabilities for normal individuals and those with movement disorders, and are not suitable for auxiliary risk assessment.

[0096] Table 4 Performance metrics of Support Vector Machine (SVM) models based on four different kernel functions

[0097] Linear 0.750 0.824 0.483 0.609 Poly 0.833 0.636 0.636 0.636 RBF 0.861 0.828 0.828 0.828 Sigmoid 0.736 0.778 0.483 0.596

[0098] Depend on Figure 4 and Figure 5 As shown in Tables 3 and 4, the confusion matrix and four correlation indicators of the RBF-based SVM model are basically the same as those of the DT model. This indicates that after training, the RBF-based SVM model, when used to classify normal samples and individuals with mandibular movement abnormalities in the training set, exhibits similar classification ability and robustness to the DT model, and is superior to other models. In conclusion, given a low misdiagnosis rate, the RBF-based SVM model and the DT model are the most suitable for assisting in the assessment of the risk of oral functional disorders, while the DT and RF models are within the scope of applicability.

[0099] like Figure 6 In the example, ROC curves of four models were plotted based on the true positive rate and false positive rate under different classification thresholds, and the AUC index was calculated based on the ROC curves of the four models, as shown in Table 5.

[0100] Table 5 shows the AUC indices for four models: DT, RF, GBDT, and SVM (RBF).

[0101] AUC 0.826 0.895 0.900 0.854

[0102] from Figure 6As shown in Table 5, the ROC curve of the GBDT algorithm based on ensemble learning exhibits high classification performance for both normal individuals and those with mandibular movement abnormalities across various classification thresholds. Its AUC index is the highest among the four classification models. Furthermore, for the test set, the AUC index of the SVM model, which has the same confusion matrix and correlation performance index, is significantly lower than that of the GBDT model. This indicates that among the four models, the GBDT model, with an AUC index of 0.900, is the most suitable for assisting in the assessment of the risk of having TMD.

[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A data classification method for normal and abnormal temporomandibular movement, characterized in that: Includes the following steps, S1. Temporomandibular joint motion data is collected by the temporomandibular joint motion visualization platform and converted into structured data features. The temporomandibular joint motion data includes normal motion data and abnormal motion data. The structured data features include motion features of temporomandibular range of motion, motion features of maximum three-dimensional displacement, motion features of motion repositioning, and motion features of motion angle. S2. Preprocess the structured data features, imput missing values, and dimensionlessly convert the data to obtain preprocessed data features. Then, divide the preprocessed data features into training and testing sets; specifically, S21. For missing data in the structured data features, use the relevant parameters from the corresponding normal motion data and abnormal motion data samples to calculate the mean and fill in the missing values. S22. The data is made dimensionless, and the features are transformed into a normal distribution with a mean of 0 and a standard deviation of 1 by using decentralization and scaling transformation. S3. Construct a temporomandibular movement classification model, which includes a decision tree model, an ensemble learning model, and a support vector machine model. After training and optimizing the temporomandibular movement classification model using the training set and corresponding classification labels, the optimal model is obtained. The node partitioning process when using the CART algorithm to construct a decision tree model is as follows: Step 1: Select a mandibular movement feature f and a classification threshold h to binary classify the structured dataset into two subsets G1 and G2, which are either normal or abnormal and belong to different label categories. The threshold is selected based on the continuity of the data features, either by choosing the nearest average or by enumeration, and the optimal threshold is chosen. Step 2: Calculate the impurity of the partitioned subsets. The parameters for evaluating impurity include the Gini coefficient and information entropy. Step 3: Select the node partitioning criteria based on the corresponding Gini coefficient or information entropy for different features; Step 4: Recursively complete steps 1-3 until the classification decision tree for normal and abnormal temporomandibular movement is constructed; optimize the decision tree model by performing random branching during model construction, limiting the minimum number of samples per node, limiting the maximum depth, and limiting the number of branch features selected. Gradient boosting trees in the ensemble learning model are used to determine the binary classification result between normal and abnormal temporomandibular movement, and the log-likelihood function is selected as the loss model. Support vector machines were used to classify jawbone data. Two labels were set in the classification of normal temporomandibular movement and abnormal temporomandibular movement. The support vector machines were optimized using different kernel functions, including linear kernel function, polynomial kernel function, Gaussian radial basis kernel function and Sigmoid kernel function. S4. Use the optimal model to predict the test set and obtain the classification result of normal or abnormal temporomandibular joint movement.

2. The data classification method for normal and abnormal temporomandibular movement as described in claim 1, characterized in that: The method for making the data dimensionless includes: data normalization, scaling normalization, logarithmic transformation, and data standardization. The data processing procedure is as follows: Subtract the mean of each feature data point x from the structured data's feature data. Then divide by the standard deviation of that feature data. This yields standardized data features; After centering by the minimum value in the current feature of the data, the data is scaled according to the difference between the extreme values ​​of the current feature in a mathematical sense, and the scaled data feature is obtained. By dividing the original data by the magnitude of the vector, the data is scaled to a unit length, resulting in normalized data features. Ultimately, the data converges to a selected range, which is 0 to 1 or -1 to 1; the converged data follows a normal distribution.

3. The data classification method for normal and abnormal temporomandibular movement as described in claim 2, characterized in that: In step S3, the temporomandibular movement classification model is trained and optimized using the training set and corresponding classification labels to obtain the optimal model. Specifically, The decision tree model is trained and optimized using optimization methods, including random branching during model construction, limiting the minimum number of samples per node, limiting the maximum depth, and limiting the number of branch features selected. The ensemble learning model is trained and optimized using optimization methods, including limiting the number of base learners and limiting the optimization step size. The support vector machine model is trained and optimized using optimization methods including adjusting kernel function parameters and regularization parameters. Set performance evaluation metrics for the models, and select the best-performing model among the trained and optimized decision tree model, ensemble learning model, and support vector machine model as the optimal model.

4. The data classification method for normal and abnormal temporomandibular movement as described in claim 3, characterized in that: A decision tree model is constructed using the CART algorithm, relying on impurity to divide nodes and branches, i.e., the degree to which the data samples within a node belong to the normal or abnormal classification; the evaluation parameters for impurity include the Gini coefficient. and information entropy They are respectively: Where P(n | t) represents the proportion of the classification label n, such as normal sample or abnormal sample, at node t; It is the predicted adjustment value caused by the specific feature segmentation point considered; Among them, the Gini coefficient weighted averaging method for feature f threshold h of category G: Where |G1| and |G2| represent the number of samples in category G1 and category G2, respectively; Information entropy weighted averaging method for category G, feature f, and threshold h: Where |G1| and |G2| represent the number of samples in category G1 and category G2, respectively.

5. The data classification method for normal and abnormal temporomandibular movement as described in claim 4, characterized in that: The specific model building process is as follows: using the training set, calculate or select a constant with distributional and statistical significance, and set it as the predicted value for all training sets; Calculate the residual, i.e., the negative gradient error: in, It is the residual of the i-th sample, which is also the objective to be minimized during the iteration process. It is the actual tag value. It is the model's predicted value for the i-th sample. This represents the current estimate or prediction of the model. Set the negative gradient error as the target value, train a CART tree as the new base estimator, and compute the best negative gradient fit for each leaf node: Where j represents the leaf node region of the tree. It is the actual tag value. It is the model's predicted value for the i-th sample. The predicted adjustment value is caused by the specific feature segmentation point considered; the model is updated and iterated until the end, and the weighted sum of all base estimators is used as the final model; the ensemble learning model is optimized by limiting the number of base learners and the optimization step size.

6. A system for implementing the data classification method for normal and abnormal temporomandibular movement as described in any one of claims 1-5, characterized in that: It includes a data structuring module, a data preprocessing module, a model training and optimization module, and a classification module. Data structuring module: The temporomandibular joint motion data is collected by the temporomandibular joint motion visualization platform. The temporomandibular joint motion data includes normal motion data and abnormal motion data. The temporomandibular joint motion data is converted into structured data features. Data preprocessing module: preprocesses the structured data features, fills in missing values, and makes the data dimensionless to obtain preprocessed data features, and divides the preprocessed data features into training set and test set; Model training and optimization module: Construct a temporomandibular movement classification model, which includes a decision tree model, an ensemble learning model, and a support vector machine model. After optimizing and training the temporomandibular movement classification model using the training set and corresponding classification labels, the optimal model is obtained. Classification module: Uses the optimal model to predict the test set and obtains the classification result of normal or abnormal temporomandibular joint movement.

7. A computer storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the steps of the method as described in any one of claims 1 to 5.

Images