Ankle lateral collateral ligament injury assessment method

Abstract

The invention discloses an ankle lateral collateral ligament injury assessment method, and relates to the technical field of ligament injury detection. The ankle lateral collateral ligament injury assessment method specifically comprises the following steps: S1, experiment preparation; s2, ankle joint stability evaluation; s3, evaluating an ankle joint ligament relaxation index; s4, conclusion summarization; through combination of the leg shaking device and the inertial sensor, accurate monitoring of dynamic motion of the ankle joint is realized, and accuracy and real-time performance of data acquisition are improved. And secondly, data processing and analysis are carried out by using a deep learning algorithm, key features can be automatically extracted, data dimensions can be optimized, and the accuracy and efficiency of evaluation are effectively improved. In addition, according to the method, the stability of the ankle joint is evaluated, ligament relaxation index evaluation is introduced, and the functional state of the ankle joint is reflected more comprehensively. And finally, by introducing time-frequency analysis, a wavelet transform technology and ROC curve and AUC value evaluation, the reliability and scientificity of diagnosis are further enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of ligament injury detection, and in particular to a method for evaluating lateral collateral ligament injury of ankle. Background Art

[0002] Ankle instability is a sequela of sprained ankle. The incidence rate is very high, but it is difficult to make a clear diagnosis with existing technology. Doctors can only make a judgment based on medical history and physical examination. After ankle sprain, it is difficult for the ligaments to heal in position and restore the joint limiting function. Without surgical treatment or long-term standardized rehabilitation, it is very easy to have symptoms of ankle instability. This will cause abnormal wear and tear when the joint moves in a non-aligned state.

[0003] The current examination method is to use a "bracket" to apply lateral force to the ankle joint. If the patient has ankle instability, that is, the ligament function is not good, the X-ray taken at this time will show poor alignment of the bone structure. The above-mentioned examination method mainly focuses on the alignment of the bone structure and ignores key indicators such as ligament laxity, resulting in a single evaluation dimension that cannot fully reflect the functional status of the ankle joint; and insufficient data processing and analysis capabilities: the lack of efficient data processing and analysis means makes it difficult for doctors to interpret the examination results. Existing examination methods cannot automatically extract key features and optimize data dimensions, resulting in limited evaluation efficiency and accuracy. Summary of the invention

[0004] In view of the deficiencies of the prior art, the present invention provides a method for evaluating lateral ankle collateral ligament injury, which solves the problems of the prior art.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method for assessing lateral collateral ligament injury of the ankle, specifically comprising the following steps:

[0006] S1. Experimental preparation: Screen patients, conduct medical history inquiry and physical examination, obtain informed consent, and then fix the lower leg in the leg shaking device;

[0007] S2. Ankle stability assessment:

[0008] S2.1, record the data of inertial sensors during leg shaking and pre-process them using deep learning algorithms;

[0009] S2.2. Use deep learning networks to automatically extract key features, and combine PCA dimensionality reduction and Lasso regression feature selection to optimize data dimensions;

[0010] S2.3, train the SVM support vector machine classification model, and perform kernel function optimization and parameter tuning;

[0011] S2.4. Analyze the accuracy, recall and F1 score indicators and determine the evaluation method;

[0012] S3. Assessment of ankle ligament laxity index:

[0013] S3.1. Use the same leg shaking device and inertial sensor to record the leg shaking data;

[0014] S3.2, introduce time-frequency analysis and wavelet transform technology, and combine deep learning model to perform feature fusion and relaxation index prediction;

[0015] S3.3. The t-test method was used to compare the difference in relaxation index between the healthy group and the diseased group, and the ROC curve and AUC value were introduced to evaluate the diagnostic efficacy;

[0016] S4. Conclusion summary: Based on the evaluation results of S2 and S3, the evaluation method, data segment length, deep learning model parameters, and the calculation formula and threshold of the relaxation index are determined.

[0017] Preferably, when the leg shaking device is fixed, a special strap is used to fix the inertial sensor at the first metatarsophalangeal joint of the patient's dorsum of the foot or a designated position on the inner side of the foot, the parameters of the leg shaking device are adjusted, and an automatic calibration cycle is set; the leg shaking device drives the calf to shake back and forth through an electric drive structure, and uses the inertial sensor to collect data for evaluation in subsequent steps S2 and S3.

[0018] Preferably, the preprocessing in S2.1 includes denoising, smoothing and feature extraction; and step S2.2 specifically includes:

[0019] S2.2.1. Feature extraction: Use deep learning networks to automatically extract key features and reduce human intervention;

[0020] S2.2.2, PCA dimensionality reduction and feature selection: Combine the PCA algorithm and feature importance evaluation to perform dimensionality reduction and feature selection on the data. The feature importance evaluation uses the Lasso regression algorithm. The formulas of the PCA algorithm and the Lasso regression algorithm are expressed as follows:

[0021] Y=XW PCA and feature selection = Lasso(X,λ);

[0022] Where X is the original data matrix, W PCA is the principal component loading matrix, Y is the data matrix after dimensionality reduction, and λ is the regularization parameter of Lasso regression;

[0023] S2.2.3. SVM model training and improvement: Use the SVM support vector machine classification model to train data from different time periods, and introduce kernel function optimization and parameter tuning to improve classification accuracy.

[0024] Preferably, the SVM model training in S2.2.3 requires preparation of a data set, including feature vectors and corresponding labels; extracting feature vectors from the preprocessed data, including angular velocity and acceleration, and assigning labels to each data point according to the experimental design;

[0025] The formula for training the model using the SVM algorithm is:

[0026]

[0027] subject to you i (w·φ(x i )+b)≥1-ξ i ,ξ i ≥0, i=1, 2, …, n;

[0028] Among them, w is the weight vector, b is the bias term, φ(x i ) is the feature vector x i The kernel function mapped to the high-dimensional space, C is the regularization parameter, ξ i is the slack variable, n is the number of data points;

[0029] Select RBF kernel as the kernel function, the formula is:

[0030] K(x p , x q )=exp(-γ║x p -x q ║ 2 );

[0031] Among them, γ is the parameter of the RBF kernel, which is determined by experiments. By adjusting the value of γ, the degree of fit of the model to the data can be changed; K(x p , x q ) represents two sample points x in the input space p and x q The inner product after being mapped to the high-dimensional feature space by the kernel function.

[0032] Preferably, the following improvements are made to improve the performance of the SVM model:

[0033] Parameter tuning: Use grid search or random search methods to tune the parameters of SVM to find the best parameter combination;

[0034] Feature selection: Select the features that have the greatest impact on the classification results through feature importance evaluation to reduce the number of features and improve the generalization ability of the model;

[0035] Model fusion: Use the fusion of multiple SVM models to improve the stability and accuracy of the model.

[0036] Preferably, the best SVM model is evaluated using the accuracy, recall and F1 score indicators in S2.4 to verify the classification effect; when the accuracy is high and the recall and F1 scores are also good, it means that the model has good classification performance;

[0037] The accuracy formula is expressed as:

[0038] The recall formula is expressed as:

[0039] The F1 score formula is expressed as:

[0040] Among them, Accuracy represents accuracy, Recall represents recall, TP represents the number of correctly classified positive examples, TN represents the number of correctly classified negative examples, FP represents the number of incorrectly classified positive examples, and FN represents the number of incorrectly classified negative examples.

[0041] Precision refers to the proportion of samples that are actually positive examples among those judged as positive examples by the classifier. The calculation formula is:

[0042] Recall refers to the proportion of samples that are actually positive examples that are correctly judged as positive examples by the classifier. The calculation formula is:

[0043] By comparing the performance of the SVM model under different parameter combinations, the optimal method for evaluating ankle stability is determined, including the optimal leg shaking frequency, data segment length, feature selection results, and parameter configuration of the SVM model.

[0044] Preferably, in the step S3.2, according to the correlation analysis results and weight adjustment, the formula for feature fusion and relaxation index prediction combined with the deep learning model is:

[0045] Relaxation index = A×w1+B×w2+C×w3;

[0046] Where A is the frequency domain sub-frequency of the X-axis angular velocity of the machine system, B is the 95th percentile of the wavelet reconstruction of the X-axis angular velocity of the machine system, C is the wavelet reconstruction variance of the X-axis acceleration of the machine system, and w1, w2, and w3 are weight coefficients obtained through correlation analysis and weight adjustment.

[0047] Preferably, in S3.3, the t-test method is used to compare the difference in relaxation index between the healthy group and the diseased group, and the ROC curve and AUC value are introduced to evaluate the diagnostic efficacy.

[0048] The t-test is used to compare whether there is a significant difference between the means of two independent samples. The formula is:

[0049]

[0050] in, and is the mean of the two samples, n1 and n2 are the sample sizes, and spooled is the pooled standard deviation.

[0051] Preferably, in S3.3, in order to evaluate the performance of the SVM model in the classification task, a ROC curve is drawn and the AUC value is calculated; the ROC curve represents the model performance by drawing a graph of the true positive rate TPR versus the false positive rate FPR; the ROC curve is a graph drawn with FPR as the horizontal axis and TPR as the vertical axis, showing the performance of the classifier at different thresholds, and the closer the ROC curve is to the upper left corner, the better the performance of the classifier, that is, the higher the true positive rate and the lower the false positive rate;

[0052] The true positive rate TPR formula is:

[0053] The false positive rate FPR formula is:

[0054] The AUC value represents the area under the ROC curve and is used to measure the performance of the classifier. The closer the AUC value is to 1, the better the classifier performance is. Conversely, the closer the AUC value is to 0, the worse the classifier performance is. The closer the AUC value is to 1, the better the model performance is. The AUC value is calculated by numerical integration or trapezoidal method.

[0055] The present invention provides a method for assessing lateral collateral ligament injury of the ankle. Compared with the prior art, the method has the following beneficial effects:

[0056] 1. This method for assessing lateral collateral ligament injury of the ankle realizes accurate monitoring of dynamic movement of the ankle joint by combining a leg shaking device with an inertial sensor, thereby improving the accuracy and real-time performance of data collection. Secondly, the use of deep learning algorithms for data processing and analysis can automatically extract key features and optimize data dimensions, effectively improving the accuracy and efficiency of the assessment. In addition, this method not only evaluates the stability of the ankle joint, but also introduces the evaluation of ligament laxity indicators, which more comprehensively reflects the functional status of the ankle joint. Finally, by introducing time-frequency analysis, wavelet transform technology, ROC curve and AUC value evaluation, the reliability and scientificity of the diagnosis are further enhanced. Overall, this method has the advantages of simple operation, comprehensive evaluation, high accuracy, and strong diagnostic efficiency, providing strong technical support for the clinical diagnosis and treatment of ankle instability and lateral collateral ligament injury of the ankle.

[0057] 2. This method for assessing lateral collateral ligament injuries of the ankle significantly improves the accuracy and efficiency of ankle stability assessment through processes such as automatic extraction of key features by deep learning networks, PCA dimension reduction and Lasso regression feature selection, and SVM model training and improvement. Compared with existing inspection methods, this method not only reduces human intervention, but also improves the classification performance of the SVM model by optimizing kernel functions and parameter tuning. At the same time, indicators such as accuracy, recall rate, and F1 score are introduced to evaluate the model to ensure the reliability of the classification effect. In addition, this method also determines the optimal evaluation method by comparing the model performance under different parameter combinations, providing strong support for the accurate assessment of ankle stability.

[0058] 3. This method for assessing lateral collateral ligament injury of the ankle not only achieves accurate assessment of ankle ligament laxity, but also greatly improves the efficiency and accuracy of the assessment. Through feature extraction, correlation analysis and weight adjustment, a laxity index prediction model was constructed, which provides strong support for clinical diagnosis and treatment. At the same time, statistical methods such as t-test, ROC curve and AUC value were introduced to comprehensively evaluate the diagnostic efficacy of the model and ensure the reliability and scientificity of the assessment results. This method has the advantages of simple operation, accurate results and wide applicability, and provides new ideas and technical means for the assessment of ankle ligament laxity. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a schematic diagram of the process of the present invention;

[0060] Figure 2 is a schematic diagram of the leg shaking device of the present invention;

[0061] Figure 3 The ligament laxity index diagram of the healthy person and the ligament injury patient of the present invention;

[0062] Figure 4 Schematic diagram of the signal acquisition, data processing and analysis program of the present invention.

[0063] In the figure: frame 1, calf swing arm 2, calf fixing structure 3, crank slider reciprocating motion mechanism 4, inertial sensor 5. DETAILED DESCRIPTION

[0064] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0065] The present invention provides three technical solutions:

[0066] Figure 1-Figure 2 A first embodiment is shown: a method for assessing lateral collateral ligament injury of the ankle, specifically comprising the following steps:

[0067] S1. Experimental preparation: Screen patients who meet the experimental conditions, and conduct detailed medical history inquiries and physical examinations. The same experienced chief physician performed ankle shaking and anterior drawer tests and judged the patients to have normal ankle joints, mild ankle instability, and severe ankle instability, respectively. Explain the experimental process, purpose, and potential risks to the patient, obtain their informed consent, and fix the calf to the leg shaking device; when the leg shaking device is fixed, use a special strap to fix the inertial sensor 5 at the first metatarsophalangeal joint of the patient's dorsum or a designated position on the inner side of the foot, adjust the parameters of the leg shaking device, such as the shaking frequency (the ankle joint stability assessment is set to 2.0Hz±0.1Hz, and the ligament laxity index assessment is also set to 2.0Hz±0.1Hz to maintain consistency), and set an automatic calibration cycle; the leg shaking device drives the calf to shake back and forth through an electric drive structure, and uses the inertial sensor 5 to collect data for the evaluation of subsequent steps S2 and S3;

[0068] like Figure 2 The leg shaking device shown includes a frame 1, a calf swing arm 2, a calf fixing structure 3, a crank slider reciprocating motion mechanism 4 and an inertial sensor 5. The top of the calf swing arm 2 is rotatably connected to the frame 1. The crank slider reciprocating motion mechanism 4 is driven by a closed-loop control 57 stepper motor with a torque of 3.6Nm. The calf swing arm 2 is driven to swing back and forth through reciprocating motion, thereby driving the calf to shake. The torque is amplified to 10.8Nm through a 1:3 planetary reducer. After testing, it can drive an adult male calf to do leg shaking exercise at an actual exercise frequency of 1.5Hz and 2Hz.

[0069] S2. Ankle stability assessment:

[0070] S2.1, recording the data of the inertial sensor 5 during the leg shaking process, and preprocessing it using a deep learning algorithm;

[0071] S2.2. Use deep learning networks to automatically extract key features, and combine PCA dimensionality reduction and Lasso regression feature selection to optimize data dimensions;

[0072] S2.3, train the SVM support vector machine classification model, and perform kernel function optimization and parameter tuning to improve classification accuracy;

[0073] S2.4. Analyze the accuracy, recall and F1 score indicators and determine the evaluation method;

[0074] S3. Assessment of ankle ligament laxity index:

[0075] S3.1, using the same leg shaking device and inertial sensor 5, recording leg shaking data;

[0076] S3.2, introduce time-frequency analysis and wavelet transform technology, and combine deep learning model to perform feature fusion and relaxation index prediction;

[0077] S3.3. The t-test method was used to compare the difference in relaxation index between the healthy group and the diseased group, and the ROC curve and AUC value were introduced to evaluate the diagnostic efficacy;

[0078] S4. Conclusion Summary: Based on the evaluation results of S2 and S3, the evaluation method, data segment length, deep learning model parameters, and the calculation formula and threshold of the relaxation index were determined for subsequent diagnosis and treatment.

[0079] By combining the leg shaking device with the inertial sensor 5, accurate monitoring of the dynamic movement of the ankle joint is achieved, and the accuracy and real-time performance of data collection are improved. Secondly, the use of deep learning algorithms for data processing and analysis can automatically extract key features and optimize data dimensions, effectively improving the accuracy and efficiency of the evaluation. In addition, this method not only evaluates the stability of the ankle joint, but also introduces the evaluation of ligament laxity indicators, which more comprehensively reflects the functional status of the ankle joint. Finally, by introducing time-frequency analysis, wavelet transform technology, ROC curve and AUC value evaluation, the reliability and scientificity of the diagnosis are further enhanced. Overall, this method has the advantages of simple operation, comprehensive evaluation, high accuracy, and strong diagnostic efficiency, providing strong technical support for the clinical diagnosis and treatment of ankle instability and lateral ankle collateral ligament injury.

[0080] The second implementation mode is mainly different from the first implementation mode in that the preprocessing in S2.1 includes denoising, smoothing and feature extraction; and step S2.2 specifically includes:

[0081] S2.2.1. Feature extraction: Use deep learning networks (such as convolutional neural networks (CNNs)) to automatically extract key features and reduce human intervention;

[0082] S2.2.2, PCA dimensionality reduction and feature selection: Combine the PCA algorithm and feature importance evaluation (such as Lasso regression) to reduce the dimension of the data and select features. The feature importance evaluation uses the Lasso regression algorithm. The formulas of the PCA algorithm and the Lasso regression algorithm are expressed as follows:

[0083] Y=XW PCA and feature selection = Lasso(X,λ);

[0084] Where X is the original data matrix, W PCA is the principal component loading matrix, Y is the data matrix after dimensionality reduction, and λ is the regularization parameter of Lasso regression;

[0085] S2.2.3. SVM model training and improvement: Use the SVM support vector machine classification model to train data from different time periods, and introduce kernel function optimization (such as RBF kernel) and parameter tuning (such as grid search) to improve classification accuracy.

[0086] SVM model training in S2.2.3 requires the preparation of a data set, including feature vectors and corresponding labels; extracting feature vectors from the preprocessed data, including angular velocity and acceleration, and assigning labels to each data point (such as high stability, medium stability, low stability, etc.) according to the experimental design;

[0087] The formula for training the model using the SVM algorithm is:

[0088]

[0089] subject to you i (w·φ(x i )+b)≥1-ξ i ,ξ i ≥0, i=1, 2, …, n;

[0090] Among them, w is the weight vector, b is the bias term, φ(x i ) is the feature vector x i The kernel function mapped to the high-dimensional space, C is the regularization parameter, ξ i is the slack variable, n is the number of data points;

[0091] Common kernel functions include linear kernel, polynomial kernel, radial basis function (RBF) kernel, etc. In this experiment, RBF kernel is selected as the kernel function, and the formula is:

[0092] K(x p , x q )=exp(-γ║x p -x q ║ 2 );

[0093] Among them, γ is the parameter of the RBF kernel, which is determined by experiments. By adjusting the value of γ, the degree of fit of the model to the data can be changed; K(x p , x q ) represents two sample points x in the input space p and x q The inner product after the kernel function is mapped to the high-dimensional feature space. In SVM, this inner product is used to calculate the distance or similarity between sample points, thereby helping the model to classify.

[0094] To improve the performance of the SVM model, the following improvements are made:

[0095] Parameter tuning: Use grid search or random search methods to tune the parameters of SVM (such as C, γ) to find the best parameter combination;

[0096] Feature selection: Select the features that have the greatest impact on the classification results through feature importance evaluation (such as Lasso regression) to reduce the number of features and improve the generalization ability of the model;

[0097] Model fusion: Use the fusion of multiple SVM models (such as Bagging, Boosting, etc.) to improve the stability and accuracy of the model.

[0098] In S2.4, the best SVM model is evaluated using the accuracy, recall, and F1 score indicators to verify the classification effect; when the accuracy is high and the recall and F1 scores are also good, it means that the model has good classification performance;

[0099] The accuracy formula is expressed as:

[0100] The recall formula is expressed as:

[0101] The F1 score formula is expressed as:

[0102] Among them, Accuracy represents accuracy, Recall represents recall, TP represents the number of correctly classified positive examples, TN represents the number of correctly classified negative examples, FP represents the number of incorrectly classified positive examples, and FN represents the number of incorrectly classified negative examples.

[0103] Precision refers to the proportion of samples that are actually positive examples among those judged as positive examples by the classifier. The calculation formula is:

[0104] Recall refers to the proportion of samples that are actually positive examples that are correctly judged as positive examples by the classifier. The calculation formula is:

[0105] By comparing the performance of the SVM model under different parameter combinations, the optimal method for evaluating ankle stability is determined, including the optimal leg shaking frequency, data segment length, feature selection results, and parameter configuration of the SVM model.

[0106] This example verifies the effect of data segment length on the analysis results. The experimental process is as follows:

[0107] Experimental methods and equipment: In the data processing stage, the motion data of each patient was processed in different time segments. Each segment of raw data was segmented into 5 seconds, 10 seconds, 15 seconds, or no segment, and a SVM support vector machine classification model was trained. The maximum accuracy of the model can reflect the impact of segment length on the analysis results.

[0108] Experimental results:

[0109] The svm classifier was trained using data that was segmented every 5, 10, 15, and 20 seconds and unsegmented data, respectively. Each segment number had data recorded at 1.5 Hz and 2.0 Hz. PCA was used to reduce the data features from 288 dimensions to 36 dimensions, and 20% of the data was randomly selected for testing.

[0110] Among them, 1.5 / 2.0 represent the frequency of the leg shaking device used when collecting the data of this segment, and 5000, 10000, 20000, and unseged represent the length of the segment. The first three represent 5 / 10 / 20 seconds segmentation, and unseged represents no segmentation. From the confusion matrix and accuracy of the classification results, the 2.0Hz vibration frequency is better than 1.5Hz in all segment lengths. In terms of segment length, too short or too long segment lengths have a negative impact on classification accuracy. The effect of no segmentation is the worst. In terms of specific performance, the 5-second segmentation is slightly better than the 20-second segmentation, and among the several selected segment length options, the 10-second segmentation seems to be the best.

[0111] Therefore, a vibration frequency that is close to the resonance frequency of the human foot and can make the vibration amplitude of the patient's foot in the leg shaking experiment more significantly different from that of the healthy foot can better make the experimental data of the patient and the healthy tester have a sufficiently large feature difference, which is convenient for the SVM model to classify. From the experimental results, 2.0Hz is obviously closer to the resonance frequency of the human foot than 1.5Hz, and can provide sufficient feature specificity for the acceleration data.

[0112] The first role of data segmentation is to perform data enhancement and expand the amount of input data. The second is to prevent individual outliers from having too much impact on the entire segment of data and to reduce the impact of overall baseline drift on feature calculations. In addition, data segmentation can also play a role in segment voting in individual tests, which can improve the overall judgment accuracy. From the experimental results, it can be seen that too short a segment length will reduce the accuracy of the model, which may be due to the fact that the shorter segment length contains fewer motion cycles. The poor accuracy of too long segments may be due to the small amount of training data and the overall offset caused by long-term error accumulation, which is reflected in the feature classification.

[0113] Since PCA has a limitation on the input data dimension, although this experiment shows that 2.0Hz, 10 seconds segmentation is the current best parameter selection, the second best 2.0Hz, 5 seconds segmentation is still selected in the end. PCA dimension reduction requires that the projected dimension is between 0 and the minimum value of the number of data and the number of features, so choosing 5 seconds segmentation can maintain the PCA dimension reduction at a reasonable level, that is, 12 dimensions. Too low a dimension reduction will also reduce the classification accuracy of the final model.

[0114] The accuracy and efficiency of ankle stability assessment were significantly improved through the processes of automatic extraction of key features by deep learning network, PCA dimension reduction and Lasso regression feature selection, SVM model training and improvement. Compared with the existing inspection methods, this method not only reduces human intervention, but also improves the classification performance of the SVM model by optimizing the kernel function and parameter tuning. At the same time, indicators such as accuracy, recall rate and F1 score were introduced to evaluate the model to ensure the reliability of the classification effect. In addition, this method also determined the optimal evaluation method by comparing the model performance under different parameter combinations, providing strong support for the accurate evaluation of ankle stability.

[0115] Figure 3-Figure 4 A third embodiment is shown, which differs from the first embodiment mainly in that the experimental purpose of step S3 is to find an indicator that can determine the laxity of the ankle joint ligament.

[0116] Experimental methods and equipment: leg shaking device, inertial sensor, straps. Fix the legs on the leg shaking device with straps, use the straps to fix the inertial sensor on the gap between the two thumbs on the inner side of the foot, and adjust the motor so that the frequency of the leg shaking device reaches 2Hz.

[0117] Experimental results and analysis: Experimental data recording and processing: Use software to record and save the original data of the inertial conductor shaking with the foot, obtain features through feature extraction, perform correlation analysis on each feature with the health and injury levels, select features with higher correlation, combine them and adjust the weights to obtain the relaxation index. At the same time, due to the small sample size, a t-test analysis is performed.

[0118] Among them, in step S3.2, according to the correlation analysis results and weight adjustment, the formula for feature fusion and relaxation index prediction combined with a deep learning model (such as a neural network) is:

[0119] Relaxation index = A×w1+B×w2+C×w3;

[0120] Where A is the frequency domain sub-frequency of the X-axis angular velocity of the machine system, B is the 95th percentile of the wavelet reconstruction of the X-axis angular velocity of the machine system, C is the wavelet reconstruction variance of the X-axis acceleration of the machine system, and w1, w2, and w3 are weight coefficients obtained through correlation analysis and weight adjustment.

[0121] S3.3. The t-test method was used to compare the difference in relaxation index between the healthy group and the diseased group, and the ROC curve and AUC value were introduced to evaluate the diagnostic efficacy;

[0122] The t-test is used to compare whether there is a significant difference between the means of two independent samples. The formula is:

[0123]

[0124] in, and is the mean of the two samples, n1 and n2 are the sample sizes, and spooled is the pooled standard deviation.

[0125] The model was applied to 22 subjects (15 patients and 7 healthy subjects) for testing. After t-test, the p-value of the data of healthy and patients was 1.22e-06, and the p-value of the data of patients with mild injury and severe injury was 0.0176, both of which were less than 0.05. It can be considered that there is a significant difference between the data of healthy and sick, and there is a significant difference between the data of mild injury and severe injury. However, due to the large difference in the number of samples of healthy and mild injury, a t-test was performed. Therefore, it can be considered that the laxity index calculated by this method can basically represent the laxity of the ankle joint.

[0126] In S3.3, in order to evaluate the performance of the SVM model in the classification task, the ROC (receiver operating characteristic) curve is drawn and the AUC (area under the curve) value is calculated; the ROC curve represents the model performance by plotting the true positive rate TPR against the false positive rate FPR; the ROC curve is a graph drawn with FPR as the horizontal axis and TPR as the vertical axis, showing the performance of the classifier at different thresholds. The closer the ROC curve is to the upper left corner, the better the performance of the classifier, that is, the higher the true positive rate and the lower the false positive rate;

[0127] The true positive rate TPR formula is:

[0128] The false positive rate FPR formula is:

[0129] The AUC value represents the area under the ROC curve and is used to measure the performance of the classifier. The closer the AUC value is to 1, the better the classifier performance is. Conversely, the closer the AUC value is to 0, the worse the classifier performance is. The closer the AUC value is to 1, the better the model performance is. The AUC value is calculated by numerical integration or trapezoidal method.

[0130] This embodiment not only achieves accurate evaluation of ankle ligament laxity, but also greatly improves evaluation efficiency and accuracy. Through feature extraction, correlation analysis and weight adjustment, a laxity index prediction model is constructed, which provides strong support for clinical diagnosis and treatment. At the same time, statistical methods such as t-test, ROC curve and AUC value are introduced to comprehensively evaluate the diagnostic efficacy of the model and ensure the reliability and scientificity of the evaluation results. This method has the advantages of simple operation, accurate results and wide applicability, and provides new ideas and technical means for the evaluation of ankle ligament laxity.

[0131] In the model using machine learning after SVM screening and classification, the data of 22 subjects were divided into training set and validation set, and the accuracy was 80%. The results of the training set and validation set are shown in Table 1-2 below:

[0132] Table 1 Training set results 2

[0133]

[0134] Table 2 Validation set results 2

[0135]

[0136] Figure 4 A schematic diagram of the signal acquisition, data processing and analysis program is shown, which implements the first two steps of feature extraction of data on the Matlab side. The main time-frequency features are: effective value, mean, peak, peak factor, kurtosis, etc., as well as centroid frequency, PSD, frequency standard deviation, etc.; the processing of outliers is optimized.

[0137] To facilitate code packaging, the code is ported to a .py file, and the feature extraction of wavelet transform is supplemented to facilitate the use of subsequent models.

[0138] Meanwhile, the contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0139] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0140] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for assessing lateral collateral ligament injury of an ankle, characterized in that: The specific steps include: S1. Experimental preparation: Screen patients, conduct medical history inquiry and physical examination, obtain informed consent, and then fix the lower leg in the leg shaking device; S2. Ankle stability assessment: S2.1, record the data of inertial sensors during leg shaking and pre-process them using deep learning algorithms; S2.

2. Use deep learning networks to automatically extract key features, and combine PCA dimensionality reduction and Lasso regression feature selection to optimize data dimensions; S2.3, train the SVM support vector machine classification model, and perform kernel function optimization and parameter tuning; S2.

4. Analyze the accuracy, recall and F1 score indicators and determine the evaluation method; S3. Assessment of ankle ligament laxity index: S3.

1. Use the same leg shaking device and inertial sensor to record the leg shaking data; S3.2, introduce time-frequency analysis and wavelet transform technology, and combine deep learning model to perform feature fusion and relaxation index prediction; S3.

3. The t-test method was used to compare the difference in relaxation index between the healthy group and the diseased group, and the ROC curve and AUC value were introduced to evaluate the diagnostic efficacy; S4. Conclusion summary: Based on the evaluation results of S2 and S3, the evaluation method, data segment length, deep learning model parameters, and the calculation formula and threshold of the relaxation index are determined.

2. A method for assessing lateral collateral ligament injury of ankle according to claim 1, characterized in that: When the leg shaking device is fixed, a special strap is used to fix the inertial sensor at the first metatarsophalangeal joint of the patient's dorsum of the foot or a designated position on the inner side of the foot, the parameters of the leg shaking device are adjusted, and an automatic calibration cycle is set; the leg shaking device drives the calf to shake back and forth through an electric drive structure, and uses the inertial sensor to collect data for evaluation in subsequent steps S2 and S3.

3. A method for assessing lateral collateral ligament injury of ankle according to claim 1, characterized in that: The preprocessing in S2.1 includes denoising, smoothing and feature extraction; Step S2.2 specifically includes: S2.2.

1. Feature extraction: Use deep learning networks to automatically extract key features and reduce human intervention; S2.2.2, PCA dimensionality reduction and feature selection: Combine the PCA algorithm and feature importance evaluation to perform dimensionality reduction and feature selection on the data. The feature importance evaluation uses the Lasso regression algorithm. The formulas of the PCA algorithm and the Lasso regression algorithm are expressed as follows: Y=XW PCA and feature selection = Lasso(X,λ); Where X is the original data matrix, W PCA is the principal component loading matrix, Y is the data matrix after dimensionality reduction, and λ is the regularization parameter of Lasso regression; S2.2.

3. SVM model training and improvement: Use the SVM support vector machine classification model to train data from different time periods, and introduce kernel function optimization and parameter tuning to improve classification accuracy.

4. A method for assessing lateral ankle collateral ligament injury according to claim 3, characterized in that: The SVM model training in S2.2.3 requires the preparation of a data set, including feature vectors and corresponding labels; extracting feature vectors from the preprocessed data, including angular velocity and acceleration, and assigning labels to each data point according to the experimental design; The formula for training the model using the SVM algorithm is: subject to y i (w·φ(x i )+b)≥1-ξ i ,ξ i ≥0,i=1,2,…,n; Among them, w is the weight vector, b is the bias term, φ(x i ) is the feature vector x i The kernel function mapped to the high-dimensional space, C is the regularization parameter, ξ i is the slack variable, n is the number of data points; Select RBF kernel as the kernel function, the formula is: K(x p ,x q )=exp(-γ║x p -x q ║ 2 ); Among them, γ is the parameter of the RBF kernel, which is determined by experiments. By adjusting the value of γ, the degree of fit of the model to the data can be changed; K(x p , x q ) represents two sample points x in the input space p and x q The inner product after being mapped to the high-dimensional feature space by the kernel function.

5. A method for assessing lateral collateral ligament injury of ankle according to claim 4, characterized in that: To improve the performance of the SVM model, the following improvements are made: Parameter tuning: Use grid search or random search methods to tune the parameters of SVM to find the best parameter combination; Feature selection: Select the features that have the greatest impact on the classification results through feature importance evaluation to reduce the number of features and improve the generalization ability of the model; Model fusion: Use the fusion of multiple SVM models to improve the stability and accuracy of the model.

6. A method for assessing lateral collateral ligament injury of ankle according to claim 5, characterized in that: In S2.4, the best SVM model is evaluated using the accuracy, recall and F1 score indicators to verify the classification effect; when the accuracy is high and the recall and F1 scores are also good, it means that the model has good classification performance; The accuracy formula is expressed as: The recall formula is expressed as: The F1 score formula is expressed as: Among them, Accuracy represents accuracy, Recall represents recall, TP represents the number of correctly classified positive examples, TN represents the number of correctly classified negative examples, FP represents the number of incorrectly classified positive examples, and FN represents the number of incorrectly classified negative examples. Precision refers to the proportion of samples that are actually positive examples among those judged as positive examples by the classifier. The calculation formula is: Recall refers to the proportion of samples that are actually positive examples that are correctly judged as positive examples by the classifier. The calculation formula is: By comparing the performance of the SVM model under different parameter combinations, the optimal method for evaluating ankle stability is determined, including the optimal leg shaking frequency, data segment length, feature selection results, and parameter configuration of the SVM model.

7. A method for assessing lateral collateral ligament injury of ankle according to claim 1, characterized in that: In the step S3.2, according to the correlation analysis results and weight adjustment, the formula for feature fusion and relaxation index prediction combined with the deep learning model is: Relaxation index = A×w1+B×w2+C×w3; Where A is the frequency domain sub-frequency of the X-axis angular velocity of the machine system, B is the 95th percentile of the wavelet reconstruction of the X-axis angular velocity of the machine system, C is the wavelet reconstruction variance of the X-axis acceleration of the machine system, and w1, w2, and w3 are weight coefficients obtained through correlation analysis and weight adjustment.

8. The method for assessing lateral collateral ligament injury of ankle according to claim 1, characterized in that: S3.3: Use t-test method to compare the difference of relaxation index between healthy group and diseased group, and introduce ROC curve and AUC value to evaluate the diagnostic efficacy The t-test is used to compare whether there is a significant difference between the means of two independent samples. The formula is: in, and is the mean of the two samples, n1 and n2 are the sample sizes, and spooled is the pooled standard deviation.

9. A method for assessing lateral collateral ligament injury of ankle according to claim 6, characterized in that: In S3.3, in order to evaluate the performance of the SVM model in the classification task, the ROC curve is drawn and the AUC value is calculated; the ROC curve represents the model performance by drawing a graph of the true positive rate TPR versus the false positive rate FPR; the ROC curve is a graph drawn with FPR as the horizontal axis and TPR as the vertical axis, showing the performance of the classifier at different thresholds. The closer the ROC curve is to the upper left corner, the better the performance of the classifier, that is, the higher the true positive rate and the lower the false positive rate; The true positive rate TPR formula is: The false positive rate FPR formula is: The AUC value represents the area under the ROC curve and is used to measure the performance of the classifier. The closer the AUC value is to 1, the better the classifier performance is. Conversely, the closer the AUC value is to 0, the worse the classifier performance is. The closer the AUC value is to 1, the better the model performance is. The AUC value is calculated by numerical integration or trapezoidal method.

Images