Dynamic balance capability assessment method based on inertial sensing signal and integrated elastic network

Abstract

The invention relates to a dynamic balance capability evaluation method based on an inertial sensing signal and an integrated elastic net, which comprises the following steps: collecting the inertial sensing signal, and carrying out feature extraction on the inertial sensing signal to obtain feature data; processing the feature data, and inputting the processed feature data into a dynamic balance capability evaluation model to obtain a dynamic balance capability evaluation result; the dynamic balance capability evaluation model is obtained by training an integrated elastic network model by using a training set; the training set comprises original feature data and labels thereof; in the training process of the training set, the integrated elastic network model is trained through self-service sampling training data and a dynamic weight adjustment strategy. According to the method, an integrated elastic network (E-E-N) model integrating multi-dimensional feature extraction and a self-service sampling technology is provided, and the prediction efficiency of high-risk groups is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of balance ability assessment, and in particular to a dynamic balance ability assessment method based on inertial sensor signals and an integrated elastic net. Background Art

[0002] Balance impairment in the elderly directly contributes to serious events such as falls, which in turn impacts quality of life. Therefore, effective assessment of balance is crucial. Traditional methods for assessing balance rely primarily on standardized scales and tools, such as the Berg Balance Scale (BBS), the Tinetti Motor Performance Assessment (POMA), and the Balance Evaluation System Test (BESTest). These tools, administered by physical therapists or other professionals, have been validated for their high reliability. However, these traditional methods suffer from limitations such as high cost and significant limitations on elderly individuals' activities. Therefore, there is an urgent need for new, cost-effective, and user-friendly methods that can provide objective and efficient assessments, enabling widespread adoption in the community and healthcare settings. Against this backdrop, sensor-based balance assessment technologies, particularly inertial measurement units (IMUs), have gained increasing attention. Due to their compactness, low cost, and ease of use, IMUs not only provide medical professionals with accurate motion signature data but also offer innovative avenues for developing personalized rehabilitation and treatment plans.

[0003] Existing IMU-based balance assessment processes typically involve three key steps: first, the detection of gait events and phases; second, the determination of relevant parameters; and finally, the construction of an assessment model. Action-based posture classification is often the first step in building a balance assessment model. Taking the classic TUG test as an example, its subtasks correspond to different balance and body control requirements. The sit-to-stand phase requires the subject to lift their hips from the seat, achieving maximum ankle dorsiflexion while simultaneously extending the trunk and hips. This series of movements requires coordinated momentum transfer and stability across the body. Initially, the body's forward momentum is converted into angular momentum. To prevent falls, both "lifting" and "lowering" strategies are required to mitigate potential tripping risks. Turning requires simultaneous translational and rotational movements while changing direction, while maintaining dynamic balance. Finally, the stand-to-sit phase requires coordinated trunk and knee flexion, precisely controlling the anterior-posterior and vertical displacement of the body's center of gravity.

[0004] Feature extraction of sensor data is a key step for identifying high-risk individuals. According to different experimental purposes, researchers have developed a variety of feature indicators related to balance ability and gait stability, mainly including demographic characteristics, descriptive statistical characteristics, spatiotemporal characteristics, and spectral characteristics. Among them, the spatial characteristics include step length, stride, and step width, which are based on inertial signal estimation parameters; the temporal characteristics include step speed, standing time, and swing time, which reflect the timing characteristics of gait; in the frequency domain, the commonly used is the harmonic ratio of acceleration in each direction, which is used to evaluate the coordination of body movement. In addition to these classic features, introducing system local stability and signal complexity into balance and gait analysis has great potential. However, there are two main limitations in current technology: first, the non-linear feature technology is mostly focused on single-task scenarios (such as simple walking), and its applicability under multi-task conditions needs to be verified; second, the feature extraction lacks a systematic framework, and various features are in a scattered state. Therefore, establishing a comprehensive feature extraction system and evaluating the contribution of each feature to balance ability is an important direction for future technology.

[0005] In the evaluation of balance ability, machine learning methods can more accurately predict balance ability due to their strong non-linear modeling capabilities. Support vector machine (SVM) is one of the most commonly used algorithms in balance ability evaluation in the past decade. In addition to SVM, other machine learning models, such as Naive Bayes (NB), K-Nearest Neighbors (KNN), and Random Forest (RF), are also widely used in balance ability evaluation due to their excellent performance. However, these technologies often ignore the problem of data imbalance, leading to a decrease in the prediction accuracy of the minority class.

[0006] In summary, in the field of balance ability evaluation, the current technology faces two major challenges: one is the lack of a systematic and multi-dimensional feature engineering framework, and the other is the widespread imbalance in population distribution in experimental data. In terms of feature extraction, existing methods are mostly limited to single-dimensional analysis of time-domain statistics or frequency-domain energy features, and a unified paradigm that integrates time-frequency domain joint analysis, nonlinear dynamics characteristics, and multi-modal data collaborative perception has not been established. It is particularly important to note that traditional feature extraction strategies rarely construct gait analysis models from the perspective of dynamic system theory, resulting in insufficient representation of the non-stationary characteristics, time-varying coupling relationships, and chaotic characteristics of gait signals. For example, in the processing of inertial sensor signals, most technologies only use statistical quantities such as mean and variance, while ignoring the analytical value of nonlinear dynamics indicators such as phase space reconstruction and Lyapunov exponent for balance control mechanisms.

[0007] On the data level, due to the low proportion of high-risk groups (such as patients with vestibular dysfunction and the elderly with high risk of falling) in the overall sample, the supervised learning model is susceptible to class distribution skewness, which is manifested in the following aspects: 1) the model is prone to overfitting to the majority class sample, and the sensitivity of the specific motion pattern recognition of high-risk individuals is insufficient; 2) the traditional resampling method is prone to introducing noise samples when processing high-dimensional biomechanical features, which aggravates the model generalization error. This phenomenon is particularly pronounced in long-term monitoring scenarios based on wearable devices, and its essence lies in the individual heterogeneity and dynamic variability of the biomechanical system. Population imbalance further affects the prediction performance of the model.

[0008] Therefore, how to construct a systematic feature extraction framework and develop an efficient model to improve the accuracy of balanced ability evaluation is a technical difficulty to be solved. SUMMARY

[0009] In order to solve the problems existing in the prior art, the purpose of the present application is to provide a dynamic balance ability evaluation method based on inertial sensing signals and ensemble elastic net, and to propose an ensemble elastic network (E-E-N) model integrating multi-dimensional feature extraction and bootstrap sampling technology for balance ability evaluation of the elderly population. The model uses a regularized elastic network as a base classifier, and optimizes the ensemble learning framework by introducing an adaptive weighting mechanism, significantly improving the prediction performance of high-risk populations.

[0010] To achieve the above purpose, the present application provides the following scheme:

[0011] A dynamic balance ability evaluation method based on inertial sensing signals and ensemble elastic net, comprising:

[0012] Collecting inertial sensing signals, extracting features from the inertial sensing signals, and obtaining feature data;

[0013] Processing the feature data, inputting the processed feature data into a dynamic balance ability evaluation model, and obtaining a dynamic balance ability evaluation result; the dynamic balance ability evaluation model is obtained by training an ensemble elastic net model using a training set; the training set includes: original feature data and its label;

[0014] In the training process of the training set, the ensemble elastic net model is trained by self-sampling training data and a dynamic weight adjustment strategy.

[0015] Optionally, the inertial sensing signals include three-axis acceleration and three-dimensional angular velocity data.

[0016] Optionally, the feature data includes demographic characteristics, descriptive statistical features, nonlinear features, spatiotemporal features, and spectral features.

[0017] Optionally, processing the feature data comprises normalizing the feature data.

[0018] Optionally, constructing the training set comprises:

[0019] performing feature space interpolation on the original feature data, constructing synthetic samples, locating k-nearest neighbors of minority class samples in the synthetic samples, generating new samples along sample-nearest neighbor lines, and obtaining the training set.

[0020] Optionally, training the ensemble elastic net model by bootstrap sampling training data and a dynamic weight adjustment strategy comprises:

[0021] Step 1, constructing a first set by bootstrap sampling the training data to select a plurality of samples from the training data, and the remainder constitutes a second set;

[0022] Step 2, constructing a base classifier, and calculating the weight of each base classifier on the first set and the second set according to an evaluation index;

[0023] Step 3, setting a first target threshold, and determining whether the weight is less than the target threshold, when the weight is less than the target threshold, it indicates that the base classifier is an inefficient base classifier, and the base classifier is discarded;

[0024] Step 4, repeating step 3 until a set number of base classifiers is reached, and the remaining base classifiers are integrated into an ensemble classifier according to their performance ability.

[0025] Optionally, calculating the weight of each base classifier on the set comprises:

[0026]

[0027] wherein YI(D) is the Youden index of the first set D, F(D) is the F value of the set D, YI(R) is the Youden index of the second set R, and F(R) is the F value of the set R.

[0028] Optionally, the evaluation index comprises:

[0029] YI = sensitivity + specificity - 1

[0030]

[0031] wherein YI is the Youden index, F is the F value, sensitivity is the sensitivity, specificity is the specificity, and precision is the precision.

[0032] Optionally, the acquiring the dynamic balance ability evaluation result comprises:

[0033] inputting the processed feature data into a dynamic balance ability evaluation model, an integrated classifier in the dynamic balance ability evaluation model aggregating output results of base classifiers through a weighted voting strategy to acquire a plurality of classification results;

[0034] setting a second target threshold, selecting a final classification result from the plurality of classification results by using the second target threshold, and taking the final classification result as the dynamic balance ability evaluation result.

[0035] The present application has the following advantages:

[0036] In view of the limitations of the prior art, the present application innovatively proposes an Ensemble Elastic Net (E-E-N) model integrating multi-dimensional feature extraction and bootstrap sampling technology for balance ability evaluation of the elderly. The model takes regularized elastic network as a base classifier, optimizes the integrated learning framework by introducing an adaptive weighting mechanism, and significantly improves the prediction performance of high-risk groups. The advantages of the technology mainly lie in the following three aspects:

[0037] 1) Multi-level feature extraction framework: The present application constructs a multi-task feature extraction framework for wearable sensor data, integrating multi-disciplinary methods such as spectral domain analysis, dynamic system local stability evaluation and signal self-similarity detection, and realizing multi-dimensional analysis of human motion state. This scheme not only enhances the accuracy of risk prediction, but also reveals the cross-dimensional correlation mechanism of balance ability through multi-source feature fusion.

[0038] 2) E-E-N model construction: Elastic network regularization effectively solves the problems of feature multicollinearity and model overfitting, and integrates bootstrap sampling technology and dynamic weight adjustment strategy to significantly improve the recognition sensitivity of high-risk individuals while alleviating small sample data bias.

[0039] 3) Balance ability evaluation: The feature importance in the E-E-N model of the present application can provide ranking and provide a quantitative analysis method to understand the role of features in balance ability evaluation. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0041] Figure 1This is a flow chart of a method for evaluating dynamic balance capability based on inertial sensor signals and an integrated elastic net according to an embodiment of the present invention;

[0042] Figure 2 Schematic diagram of IMU wearing and signal cutting according to an embodiment of the present invention; (a) is a schematic diagram of IMU wearing, and (b) is a schematic diagram of signal cutting;

[0043] Figure 3 Schematic diagram of EEN feature importance estimation according to an embodiment of the present invention. DETAILED DESCRIPTION

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0045] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0046] like Figure 1 As shown, this embodiment discloses a dynamic balance ability assessment method based on inertial sensor signals and an integrated elastic net, including: collecting inertial sensor signals, extracting features from the inertial sensor signals, and obtaining feature data; processing the feature data, inputting the processed feature data into a dynamic balance ability assessment model, and obtaining a dynamic balance ability assessment result; the dynamic balance ability assessment model is obtained by training an integrated elastic net model using a training set; the training set includes: original feature data and its labels; during the training process of the training set, the integrated elastic net model is trained using self-sampling training data and a dynamic weight adjustment strategy.

[0047] Specifically, the embodiment discloses a dynamic balance ability evaluation method based on inertial sensing signals and an integrated elastic net, including: combining multiple feature types, including demographic characteristics, descriptive statistical characteristics, nonlinear characteristics, spatiotemporal characteristics and spectral characteristics, and first discussing the nonlinear characteristics related to balance ability in the multi-task scene from the perspective of dynamic system. In addition, an integrated elastic net (E-E-N) model is constructed to cope with the data imbalance problem and improve the interpretability of the model. The integrated elastic net (E-E-N) model takes the regularized elastic network as the base classifier, trains the data through bootstrapsampling, and adopts a weighted mechanism for classifier fusion to evaluate the dynamic balance ability. The application uses real-world data to perform a series of verification experiments and conducts comprehensive comparative analysis. The results show that the application has superior prediction performance in the balance ability classification task, provides a cost-effective tool for balance ability evaluation, and can reduce the burden of continuous health monitoring.

[0048] Further, the inertial sensing signals include three-axis acceleration and three-dimensional angular velocity data. The feature data includes demographic characteristics, descriptive statistical characteristics, nonlinear characteristics, spatiotemporal characteristics and spectral characteristics.

[0049] Specifically, data processing: the application divides the data from the IMU into actions. The signal is first processed by a sixth-order Butterworth low-pass filter. The device collects three-axis acceleration (Accx, Accy, Accz) and three-dimensional angular velocity (Eulerx, Eulery, Eulerz) data through the built-in accelerometer and gyroscope, respectively, and then identifies different actions such as standing up, walking and sitting down according to the threshold method.

[0050] Feature extraction: The present application adopts a multi-level feature extraction strategy to cope with the complexity and diversity of signal feature systems. The extracted features include 1) demographic characteristics including age, gender, and whether to use auxiliary devices, etc.; 2) descriptive statistics features such as root mean square (Rms) and coefficient of variation (CV), maximum (Max), minimum (Min), range (Range), peak value of acceleration rate (Jerkmax) and mean value (Jerkmean), median (Median), etc.; 3) nonlinear features including approximate entropy (ApEn), sample entropy (SaEn), permutation entropy (PeEn) and fuzzy entropy (FuEn), maximum Lyapunov exponent (MLE), fractal dimension (FD) extracted from the whole data; 4) spatiotemporal features including subtask time and proportion, step count, step length, etc.; 5) spectral features including FFT first quartile (1stFFT), even and odd harmonic ratio (REOH), dominant frequency (DF), dominant frequency (DF). Among them, the demographic characteristics are obtained by questionnaire, and the last four kinds of characteristics are calculated based on the time series (i.e. a string of time-varying numerical values) formed by the IMU signal.

[0051] Specifically, the complete sequence is divided according to different actions, and then features are extracted according to different actions. Since different actions contain different information, feature extraction is performed according to different stages.

[0052] For example, descriptive statistics are extracted for different stages, nonlinearity is extracted for the whole process, and spatiotemporal and spectral features are extracted for the walking stage. Specifically:

[0053] 1) Demographic characteristics: including age, gender, and whether to use auxiliary devices in daily life, etc.

[0054] 2) Descriptive statistics features: for different subtask stages (such as standing up, walking, turning around, sitting down) after action segmentation, respectively extracted from three-axis acceleration signals (Accx, Accy, Accz) and three-dimensional angular velocity signals (Eulerx, Eulery, Eulerz). The extracted features include: maximum (Max), minimum (Min), range (Range), root mean square (Rms), peak value of acceleration rate (Jerkmax), mean value of acceleration rate (Jerkmean), median (Median), coefficient of variation (CV), etc. For the walking stage, forward walking and backward walking are combined into one "walking" stage.

[0055] 3) Nonlinear features: extracted on the whole TUG test data (i.e. unsegmented), including approximate entropy (ApEn), sample entropy (SaEn), permutation entropy (PeEn), fuzzy entropy (FuEn), largest Lyapunov exponent (MLE), fractal dimension (FD), to characterize the complexity, irregularity and local dynamic stability of the movement signal.

[0056] 4) Spatiotemporal features: extracted based on the action segmentation results, including the duration of each subtask and its percentage of the whole TUG test time (POD), the number of steps, step length, step frequency, etc. in the walking phase, to reflect the individual movement rhythm and gait characteristics.

[0057] 5) Spectral features: mainly for the acceleration signal in the walking phase, extracted by Fourier transform (FFT), including the first quartile energy of FFT (1st FFT), the ratio of even and odd harmonics (REOH), dominant frequency (DF), spectral edge frequency (SEF), spectral median frequency (SMF), etc. to describe the frequency characteristics and rhythm in gait.

[0058] Further, the processing of the feature data comprises: standardizing the feature data.

[0059] Specifically, to reduce the influence of individual heterogeneity, the maximum-minimum normalization method is used to standardize the original features to 0-1. This preprocessing technique eliminates the dimensional differences of feature quantities, ensuring that each variable has comparable contribution in the subsequent modeling process.

[0060] Further, the construction of the training set comprises: performing feature space interpolation on the original feature data, constructing synthetic samples, locating the k-nearest neighbors of the minority class samples in the synthetic samples, randomly generating new samples along the sample-neighbor line to obtain the training set.

[0061] Specifically, to optimize the class distribution, the synthetic minority over-sampling technique (SMOTE) is used to construct synthetic samples through feature space interpolation: first, locate the k-nearest neighbors of the minority class samples, then generate new samples along the sample-neighbor line to expand the data volume;

[0062] Among them, the original samples contain two types of labels, one representing good health status and the other representing poor health status. Under normal circumstances, the number of samples with good health status is relatively large, and the number of samples with poor health status is small. Therefore, new samples with poor health status are synthesized by smote to balance the two types of samples.

[0063] Further, the training of the integrated elastic net model through self-sampling training data and a dynamic weight adjustment strategy comprises the following steps: step 1, selecting a plurality of samples from the training data to construct a first set through self-sampling training data, and the remaining data constitutes a second set; step 2, constructing a base classifier, and calculating the weight of each base classifier on the first set and the second set according to an evaluation index; step 3, setting a first target threshold, and determining whether the weight is less than the target threshold; when the weight is less than the target threshold, it indicates that the base classifier is an inefficient base classifier, and the base classifier is discarded; step 4, repeating step 3 until there is no inefficient base classifier, and the remaining base classifiers are integrated into an integrated classifier.

[0064] Further, the acquisition of the dynamic balance ability evaluation result comprises the following steps: inputting the processed feature data into a dynamic balance ability evaluation model, aggregating the output results of the base classifiers through a weighted voting strategy to obtain a plurality of classification results, setting a second target threshold, selecting a final classification result from the plurality of classification results by using the second target threshold, and taking the final classification result as the dynamic balance ability evaluation result.

[0065] Specifically, the model is constructed as follows:

[0066] For the original training set S: S i ={X i ,y i}, i = 1, 2, …, n and the test set T: T i ={X i ,y i}, i = 1, 2, …, m, wherein X and y respectively represent input features and corresponding labels. First, bootstrap sampling is used to sample the original training data to obtain a set D and a remaining data set R (step 4), and a base classifier f i is constructed on the set D, and the remaining data set is used for base classifier performance judgment.

[0067] The base classifier f i is constructed on the set D: an elastic net model is trained on the set D, and the base classifier f i is obtained. This classifier can classify data, and since f i is correct and wrong, f i is used to classify data on the two sets, and the classification results are obtained, which need to be compared with the real labels. In the comparison process, a plurality of commonly used evaluation indexes including sensitivity, specificity and precision are generated.

[0068] The weight ω of each base classifier is calculated according to evaluation indexes such as Youden Index (YI) and F value (F) i , to reduce the impact of underperforming base classifiers:

[0069] YI = sensitivity + specificity - 1

[0070]

[0071] where the four evaluation indexes are defined as follows:

[0072] Sensitivity = TP / (TP + FN)

[0073] Specificity = TN / (TN + FP)

[0074] Precision = TP / (TP + FP)

[0075] Accuracy = (TP + TN) / (TP + FP + TN + FN)

[0076] where TP represents the number of individuals correctly identified as imbalanced, FN represents the number of individuals incorrectly identified as balanced, FP represents the number of individuals incorrectly identified as imbalanced, and TN represents the number of individuals correctly identified as balanced. For each base classifier, the YI and F value for D and R are denoted as YI(D), YI(R), F(D), F(R), respectively, and the weight is scaled to 0-1:

[0077]

[0078] When the weight is less than a determined threshold c, the base classifier is considered inefficient and is discarded in step 7. In the proposed ensemble classifier, the integration of base classifiers is accomplished through a weighted voting strategy.

[0079] The ensemble classifier aggregates the output results of base classifiers through a weighted voting strategy, which are represented as probabilities of classification results. In step 11, the probability output obtained by the ensemble is converted into real classification results, and in step 12, a threshold is applied to determine the final classification results:

[0080] More specifically, the ensemble framework under imbalanced data:

[0081] 1: Input: training set S = {S1, S2, …, S n}, where S i = {X i , y i}, test set T = {T1, T2, …, T n}, where T j = {X i ,y j}, X is the sample feature, y is the sample label, the number of base trainers N, the number of samples k in the sub-training set, the weight threshold c;

[0082] 2: Output: the prediction result P(T) of the test set T;

[0083] 3: For i from 1 to N do;

[0084] 4: Randomly select k samples from the training set S with replacement to form D, and the remaining as set R:

[0085] D = {D1, D2, …, D k}, R = {R1, R2, …, R n-k}

[0086] 5: Construct base classifier f i on D;

[0087] 6: Calculate the weight ω i of f i according to the Youden index and F-score and scale it to 0-1;

[0088] 7: If ω i < c, then ω i = 0;

[0089] 8: End if;

[0090] 9: End for;

[0091] 10: The total weight is

[0092] 11: Calculate the prediction probability of all f i on T and combine them

[0093] 12: Convert the prediction probability into the final regression or classification result

[0094] This embodiment is dedicated to evaluating the balance ability of the elderly population, and the elderly can complete fixed tests, for example, to complete a timed up and go test (TUG): stand up from a chair, walk 3 meters in a straight line, turn around, walk 3 meters back to the original position, and finally sit down in a standard armchair. The use of gait aids is prohibited during the test. The elderly wear a nine-degree-of-freedom IMU between the L3-L5 vertebrae, such as Figure 2(a)-(b) are shown. Then, the signal is processed by a six-order Butterworth low-pass filter with a cutoff frequency of 3 Hz. After data preprocessing, the TUG task is divided into five sub-stages: standing-up stage, going stage, turning stage, returning stage and sitting-down stage.

[0095] Then, the features required according to the strategy are extracted and input into the constructed model. In the training of the model of the present application, real data of 41 subjects are used for training, and the E-E-N model has N = 30 elastic nets for integrated construction. Each bootstrap operation selects k = 26 instances to form a set D. In ensemble learning, base classifiers with a weight lower than 0.7 will be removed. Monte Carlo cross-validation is used for 50 repeated experiments, and a random seed is set for each iteration to divide the data set. The test set is isolated throughout and only used for final performance evaluation. The average of the test set indicators of the 50 experiments is recorded as the final result. Data processing and model development mainly rely on R (version 4.2.1) and Python (version 3.10.9). In the implementation and result visualization of the machine learning model, Scikit-Learn (version 1.2.1) and Matplotlib (version 3.6.3) are selected as the main tools.

[0096] After inputting the model, the balance ability status can be obtained. In the past test set of the present application, the prediction result of E-E-N is better than that of the classic machine learning algorithm, especially in the identification of individuals with poor balance ability, as shown in Table 1.

[0097] Table 1 Comparison of E-E-N and four machine learning models

[0098]

[0099] In addition, the E-E-N model provides the importance ranking result, as shown in Table 2, which can provide the contribution ranking of the input model features, thereby providing an objective and quantifiable perspective for understanding the differences among the population. Figure 3

[0100] The above-described embodiments are only descriptions of the preferred modes of the present application and do not limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements to the technical solutions of the present application made by those skilled in the art shall fall within the protection scope determined by the claims of the present application.​

Claims

1. A dynamic balance ability evaluation method based on inertial sensor signals and integrated elastic net, characterized in that: include: Collecting inertial sensor signals, performing feature extraction on the inertial sensor signals, and obtaining feature data; Processing the characteristic data, inputting the processed characteristic data into a dynamic balance ability assessment model, and obtaining a dynamic balance ability assessment result; The dynamic balance ability assessment model is obtained by training an integrated elastic net model using a training set; the training set includes: original feature data and its labels; During the training process of the training set, the integrated elastic net model is trained through self-sampling training data and dynamic weight adjustment strategy.

2. The dynamic balance ability evaluation method based on inertial sensor signals and integrated elastic net according to claim 1, characterized in that: The inertial sensing signal includes: three-axis acceleration and three-dimensional angular velocity data.

3. The dynamic balance ability evaluation method based on inertial sensor signals and integrated elastic net according to claim 1, characterized in that: The characteristic data include: demographic characteristics, descriptive statistical characteristics, nonlinear characteristics, spatiotemporal characteristics and spectral characteristics.

4. The dynamic balance ability evaluation method based on inertial sensor signals and integrated elastic net according to claim 1, characterized in that: Processing the characteristic data includes: performing standardization processing on the characteristic data.

5. The dynamic balance ability evaluation method based on inertial sensor signals and integrated elastic net according to claim 1, characterized in that: Constructing the training set includes: Performing feature space interpolation on the original feature data to construct synthetic samples, locating the k nearest neighbors of the minority class samples in the synthetic samples, and randomly generating new samples along the sample-nearest neighbor connection to obtain the training set.

6. The dynamic balance ability evaluation method based on inertial sensor signals and integrated elastic net according to claim 1, characterized in that: Training the ensemble elastic net model using self-sampling training data and a dynamic weight adjustment strategy includes: Step 1: self-sampling the training data, selecting multiple samples from the training data to form a first set, and the remaining samples to form a second set; Step 2: construct a base classifier, and calculate the weight of each base classifier on the first set and the second set according to the evaluation index; Step 3: setting a first target threshold, and judging whether the weight is less than the target threshold. When the weight is less than the target threshold, it indicates that the base classifier is an inefficient base classifier, and the base classifier is discarded. Step 4: Repeat step 3 until the set number of base classifiers is reached. The retained base classifiers are weighted according to their performance capabilities and integrated into an integrated classifier.

7. The dynamic balance ability evaluation method based on inertial sensor signals and integrated elastic net according to claim 6, characterized in that: Calculating the weight of each base classifier on the set involves: Where YI(D) is the Youden index of the first set D, F(D) is the F value of set D, YI(R) is the Youden index of the second set R, and F(R) is the F value of set R.

8. The method for evaluating dynamic balance ability based on inertial sensor signals and integrated elastic net according to claim 6, characterized in that: The evaluation indicators include: YI=sensitivity+specificity-1 Among them, YI is Youden index, F is F value, sensitivity is sensitivity, specificity is specificity, and precision is precision.

9. The dynamic balance ability evaluation method based on inertial sensor signals and integrated elastic net according to claim 1, characterized in that: Obtaining the dynamic balance ability assessment result includes: Inputting the processed feature data into a dynamic balance ability evaluation model, wherein the integrated classifier in the dynamic balance ability evaluation model aggregates the output results of the base classifiers through a weighted voting strategy to obtain multiple classification results; A second target threshold is set, and a final classification result is selected from multiple classification results using the second target threshold, and the final classification result is used as the dynamic balance ability assessment result.

Images