A cognitive function evaluation method based on skeleton data signals and symmetric stability features

Abstract

The present application relates to a kind of cognitive function evaluation methods based on skeletal data signal and symmetry stability feature, comprising: the three-dimensional skeletal data of subject is collected, three-dimensional skeletal data is preprocessed, obtains the three-dimensional skeletal data after pre-processing respectively calculates relative position feature, relative direction feature and joint angle symmetry feature, and carries out feature extraction, obtains symmetry stability feature;Symmetry stability feature is carried out feature screening, and obtains complete data set and non-complete data set;Through complete data set and non-complete data set, complete data model and non-complete data model are trained respectively;Obtain the three-dimensional skeletal data to be evaluated, the three-dimensional skeletal data to be evaluated is input into complete data model or non-complete data model, and the cognitive function evaluation result is output.The present application is based on skeletal symmetry stability feature, realizes stable discrimination by constructing multiple sub-task model and carrying out integrated inference without removing missing test sample.

Description

Technical Field

[0001] This invention relates to the fields of cognitive impairment assessment and smart healthcare, and in particular to a cognitive function assessment method based on skeletal data signals and symmetry stability characteristics. Background Technology

[0002] Studies have shown that cognitive and motor functions share common pathways in the cortical control mechanisms, and early cognitive decline is often accompanied by subtle motor dysfunction. Therefore, objectively assessing the motor function of older adults has become an important starting point for identifying mild cognitive impairment (MCI).

[0003] In the field of skeletal data-based MCI recognition, current technologies face two major challenges: firstly, the lack of in-depth generalization and systematization of kinematic principles in manual feature extraction mechanisms; and secondly, the lack of robustness and correction capabilities of models when faced with non-random missing clinical data.

[0004] In terms of feature representation, existing technologies largely rely on intuitive kinematic parameters, but rarely delve into the underlying computational principles or generalize them to whole-body motion analysis. Taking gait analysis as an example, traditional methods typically determine the key feature of "stride length" based on the maximum difference in the anterior-posterior direction between the joints of both feet. However, this effective logic of measuring movement amplitude based on "relative displacement" is often limited to the lower limbs or specific gait cycles in existing studies, failing to extend to other key parts of the body. As a coordinated whole, the various relationships between the upper limbs and trunk also carry important balance control information. Existing methods ignore the possibility of extending these relative position, direction, and angle changes between any two points to the whole-body joint system and are limited to time-domain statistics, resulting in the inability to capture the extremely subtle systemic compensatory patterns and symmetry disruptions in MCI patients maintaining balance.

[0005] From a data modeling perspective, existing technologies have shortcomings in handling biased, missing data. Firstly, they lack complementary tasks involving tests of varying difficulty. Secondly, even when features are extracted from multiple tasks, a crude approach of directly mixing features is used, failing to intuitively assess the detection capabilities of different tests. Finally, when performing motor function assessments with varying levels of difficulty, older adults with cognitive impairment or physical decline often cannot complete the more challenging sub-tasks, leading to non-random missing data. Existing modeling strategies typically remove these incomplete samples, but this approach not only results in the loss of crucial information but also systematically ignores those high-risk groups with the most obvious symptoms and the greatest need for identification, severely weakening the model's sensitivity and generalization ability in actual clinical screening.

[0006] Therefore, how to construct a systematic feature extraction framework and develop an efficient ensemble model to improve the accuracy and robustness of balance capability assessment is a technical challenge that needs to be addressed. Summary of the Invention

[0007] The purpose of this invention is to provide a cognitive function assessment method based on skeletal data signals and symmetry stability features. Based on skeletal symmetry stability features, a multi-subtask model is constructed and integrated inference is performed to achieve stable discrimination without removing missing test samples.

[0008] To achieve the above objectives, the present invention provides the following solution: A cognitive function assessment method based on skeletal data signals and symmetry stability features, comprising: Collect three-dimensional skeletal data of the subject, preprocess the three-dimensional skeletal data, and obtain preprocessed three-dimensional skeletal data; The relative position features, relative direction features, and joint angle symmetry features of the preprocessed 3D skeleton data are calculated respectively, and feature extraction is performed to obtain symmetry stability features; Feature filtering is performed on the symmetry stability features to obtain complete and incomplete datasets; The complete data model and the incomplete data model are trained using the complete dataset and the incomplete dataset, respectively. The complete data model is constructed using a random forest model, and the incomplete data model is constructed using several random forest base models. Obtain the three-dimensional skeletal data to be evaluated. Depending on whether the three-dimensional skeletal data to be evaluated has missing sub-tasks, input the three-dimensional skeletal data to be evaluated into a complete data model or an incomplete data model, and output the cognitive function evaluation result.

[0009] Optionally, preprocessing the three-dimensional skeletal data includes: After Gaussian filtering and denoising the three-dimensional skeleton data, the continuous data stream is divided into independent data segments corresponding to different sub-tasks according to the timestamp, and the data at the beginning and end with a preset proportion is removed to eliminate interference.

[0010] Optionally, calculating the relative position feature includes: Based on the preprocessed 3D skeleton data, the relative distance between any two joints in 3D space is obtained, and the relative displacement sequence is obtained. Based on the relative displacement sequence, the relative displacement of adjacent interval frames is differentially divided by the interval time to obtain the change velocity sequence; Based on the changing velocity sequence, the velocity of adjacent interval frames is differentiated and divided by the interval time to obtain the acceleration sequence; The relative displacement sequence, the velocity change sequence, and the acceleration sequence are used as the relative position features.

[0011] Optionally, calculating the relative orientation feature includes: Based on the preprocessed 3D skeleton data, the connection between any two joint points is calculated using trigonometric functions. Calculate the angle between the line connecting any two joints and any axis of the coordinate system, and obtain the sequence of angles in the coordinate system. The angle between any axis of the coordinate system in adjacent interval frames is differentially divided by the interval time to calculate the angular velocity and obtain the angular velocity sequence. The angular velocity of adjacent frames is differentiated and divided by the interval time to calculate the angular acceleration, thus obtaining an angular acceleration sequence; The coordinate system's included angle sequence, the angular velocity sequence, and the angular acceleration sequence are used as the relative direction features.

[0012] Optionally, calculating the joint angle symmetry features includes: Based on the preprocessed 3D skeleton data, the joint angles are calculated using vector operations. The difference in joint angles, the rate of change of joint angles over time, and the acceleration of joint angles are calculated based on the joint angles. The difference sequence, the rate of change of joint angles, and the acceleration sequence of joint angles are obtained as the symmetry features of the joint angles.

[0013] Optionally, feature filtering is performed on the symmetry stability features to obtain complete and incomplete datasets, including: The symmetry stability features are divided into an initial complete dataset and an initial incomplete dataset. A hybrid feature selection method combining statistical tests and machine learning is adopted. First, features with statistically significant differences are selected through normality tests, homogeneity of variance tests, and inter-group difference tests. Then, highly correlated features are eliminated through correlation analysis. Finally, key features are selected based on feature importance assessment to obtain the complete dataset and the incomplete dataset. Specifically, for the initial complete dataset, all sub-test features are merged for feature selection, while for the initial incomplete dataset, each sub-test feature is selected individually.

[0014] Optionally, training the complete data model and the incomplete data model using the complete dataset and the incomplete dataset respectively includes: Outliers in the complete dataset are processed using the capping method. The complete dataset after outlier processing is divided into a training set and a test set. The complete data model is trained by performing 10-fold cross-validation on the training set. The random forest base model is constructed for each sub-task, and the prediction results of each random forest base model are integrated through a weighted soft voting mechanism to train the incomplete data model.

[0015] Optionally, integrating the prediction results of each random forest base model through a weighted soft voting mechanism includes: weighting and summing the class probabilities output by each random forest base model, and determining the final discrimination result based on the fused prediction probabilities.

[0016] The beneficial effects of this invention are as follows: (1) Multi-dimensional skeletal feature modeling framework: A multi-task feature analysis framework for three-dimensional skeletal data is constructed, which extracts symmetric stability features from multiple levels such as time domain, frequency domain and time frequency domain, and realizes multi-dimensional characterization of human balance and gait control state. This framework can capture the movement differences of various parts of the body under different test tasks, and provide a discriminative feature basis for subsequent modeling.

[0017] (2) Construction of multi-subtask ensemble model: Independent subtask classification base models are constructed for different test tasks, and ensemble inference is performed based on the predicted probabilities output by each subtask. When some test tasks are missing in the sample, the prediction results of the completed subtasks are used for fusion, thereby realizing unified modeling and discrimination of incomplete data samples and improving the robustness of the model in actual application scenarios.

[0018] (3) Inductive interpretation of movement differences based on feature analysis: Based on feature selection and modeling, volcano plots are used to jointly analyze the statistical significance and difference magnitude of key features in different test tasks, and to summarize and compare features with similar performance, thereby revealing the differences in the participation mode of body parts and movement control mode under different test conditions, and providing intuitive and interpretable movement level analysis basis for motor function assessment results. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of a cognitive function assessment method based on skeletal data signals and symmetry stability features according to an embodiment of the present invention; Figure 2 This is a flowchart of the data preprocessing process according to an embodiment of the present invention; Figure 3 This is the preprocessed signal according to an embodiment of the present invention; Figure 4This is a schematic diagram illustrating the distribution of 24 joints and the construction of symmetric stability features in an embodiment of the present invention; Figure 5 This is a model diagram of the SPPB test site according to an embodiment of the present invention; Figure 6 This is a volcano map of all SSF features of the upper limb based on STD, according to an embodiment of the present invention. Figure 7 This is a comparison chart of the performance and ROC curves of the complete data model and the incomplete data model on the training set in an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] To address the issues in existing technologies, such as the lack of systematic approach to skeletal motion feature extraction and the difficulty of handling incomplete data with non-random missing information, such as... Figure 1 As shown, this embodiment proposes a cognitive function assessment method based on skeletal data signals and symmetry stability features, including: Collect three-dimensional skeletal data of the subject, preprocess the three-dimensional skeletal data, and obtain preprocessed three-dimensional skeletal data; The relative position features, relative direction features, and joint angle symmetry features of the preprocessed 3D skeleton data are calculated respectively, and feature extraction is performed to obtain symmetry stability features; Feature filtering is performed on the symmetry stability features to obtain complete and incomplete datasets; The complete data model and the incomplete data model are trained using the complete dataset and the incomplete dataset, respectively. The complete data model is constructed using a random forest model, and the incomplete data model is constructed using several random forest base models. Obtain the three-dimensional skeletal data to be evaluated. Depending on whether the three-dimensional skeletal data to be evaluated has missing sub-tasks, input the three-dimensional skeletal data to be evaluated into a complete data model or an incomplete data model, and output the cognitive function evaluation result.

[0024] Further, the preprocessing of the three-dimensional skeletal data includes: After Gaussian filtering and denoising the three-dimensional skeleton data, the continuous data stream is divided into independent data segments corresponding to different sub-tasks according to the timestamp, and the data at the beginning and end with a preset proportion is removed to eliminate interference.

[0025] Specifically, data feature extraction includes: in preprocessing, using a depth camera such as Kinect V2 to acquire 3D skeletal data of the subject during SPPB, such as... Figure 2 As shown. After the original signal is smoothed and denoised by Gaussian filtering, the continuous data stream is precisely divided into independent data segments corresponding to different sub-tasks (such as standing side by side, standing in series, walking, standing up from a chair, etc.) according to the timestamp, and the first and last 10% of the data are removed to eliminate interference, such as... Figure 3 As shown. In feature extraction, features are constructed individually for each test segment to capture the micro-compensation patterns of MCI patients during movement.

[0026] Furthermore, calculating the relative position features includes: Based on the preprocessed 3D skeleton data, the relative distance between any two joints in 3D space is obtained, and the relative displacement sequence is obtained. Based on the relative displacement sequence, the relative displacement of adjacent interval frames is differentially divided by the interval time to obtain the change velocity sequence; Based on the changing velocity sequence, the velocity of adjacent interval frames is differentiated and divided by the interval time to obtain the acceleration sequence; The relative displacement sequence, the velocity change sequence, and the acceleration sequence are used as the relative position features.

[0027] Furthermore, calculating the relative orientation features includes: Based on the preprocessed 3D skeleton data, the connection between any two joint points is calculated using trigonometric functions. Calculate the angle between the line connecting any two joints and any axis of the coordinate system, and obtain the sequence of angles in the coordinate system. The angle between any axis of the coordinate system in adjacent interval frames is differentially divided by the interval time to calculate the angular velocity and obtain the angular velocity sequence. The angular velocity of adjacent frames is differentiated and divided by the interval time to calculate the angular acceleration, thus obtaining an angular acceleration sequence; The coordinate system's included angle sequence, the angular velocity sequence, and the angular acceleration sequence are used as the relative direction features.

[0028] Furthermore, calculating the symmetry features of the joint angles includes: Based on the preprocessed 3D skeleton data, the joint angles are calculated using vector operations. The difference in joint angles, the rate of change of joint angles over time, and the acceleration of joint angles are calculated based on the joint angles. The difference sequence, the rate of change of joint angles, and the acceleration sequence of joint angles are obtained as the symmetry features of the joint angles.

[0029] Specifically, such as Figure 4 As shown, the selected symmetrical joint pairs include the shoulder (4-8), elbow (5-9), hand (6-10), hip (12-16), knee (13-17), and foot (14-18). The following three types of core kinematic features are extracted and combined with time-frequency domain analysis: ① Relative Position Features (RPF): Given a joint point in a certain frame Its coordinates are Now, calculate the relative distance between any two joints in three-dimensional space, assuming... and The first Any two joints along a certain axis during a frame The coordinates can be obtained by difference in this frame. When calculating all frames, a relative displacement sequence can be constructed. Furthermore, it is deduced that when adjacent interval frames... and The relative displacement can be obtained by taking the difference and dividing by the time interval. It is a sequence of changing rates. Similarly, perform the same calculation on the speed. Acceleration sequence can be obtained This feature extends the traditional concept of stride length to the dynamic relative displacement of all joints in the body, quantifying the coordinated control ability between limbs. ② Relative Orientation Features (ROF): Calculates the angle between the joint line and the spatial coordinate axes and its variation. Trigonometric functions are used to calculate the angle between any two joints. and connection With any axis of the coordinate system The included angle And calculate the included angle of all frames. Obtain the angle sequence and according to Derive its angular velocity sequence and Derive its angular acceleration sequence This feature reflects the body's tilting tendency and directional adjustment strategies during movement. ③ Joint Angle Symmetry Features (JASF): This feature calculates the angle between joints based on vector operations. It typically uses three joint points to find the angle between the midpoint and the two endpoints of a line. Then calculate the difference in angles. and the rate of change of joint angle over time With joint angle acceleration The timing representation of all the above frames is as follows: The joint angle rate is represented in the time sequence of all frames as follows: The timing representation of joint angle acceleration across all frames is as follows: This feature directly reflects the geometric symmetry between joints during the execution of movements.

[0030] In summary, this embodiment constructs a system to express the spatial components and dynamic changes of quantified motion at any two joint points. The scope is now narrowed to verify the effectiveness of the SSF feature set, and the calculation principle of "relative change" is generalized to six pairs of symmetrical joints (e.g., left and right shoulders, elbows, hands, hips, knees, and ankles). Based on this, multidimensional statistical indicators are extracted from the above three types of time-series signals, including: ① Time-domain features: Standard Deviation (STD), Impulse Factor (IF), and Shap Factor (SF), used to capture the instantaneous impact and fluctuation amplitude of the motion; ② Frequency-domain features: Ratio of Even-to-Odd Harmonics (REOH), Power Spectral Entropy (PSE), and Wavelet Entropy (WE), used to evaluate the spectral symmetry and complexity of the motion signal. Specific formulas are shown in Table 1. The signal's first One sample; The total number of signal samples; : No. Power spectral density of each frequency component; Normalized power spectral components; : No. The first decomposition level Wavelet coefficients; The total number of wavelet decomposition layers; : No. Normalized energy distribution of layer coefficients.

[0031] Table 1 Furthermore, feature filtering is performed on the symmetry stability features to obtain complete and incomplete datasets, including: The symmetry stability features are divided into an initial complete dataset and an initial incomplete dataset. A hybrid feature selection method combining statistical tests and machine learning is adopted. First, features with statistically significant differences are selected through normality tests, homogeneity of variance tests, and inter-group difference tests. Then, highly correlated features are eliminated through correlation analysis. Finally, key features are selected based on feature importance assessment to obtain the complete dataset and the incomplete dataset. Specifically, for the initial complete dataset, all sub-test features are merged for feature selection, while for the initial incomplete dataset, each sub-test feature is selected individually.

[0032] Specifically, feature selection: In this stage, the samples can be divided into complete datasets (containing only samples with all completed test data) and incomplete datasets (containing both samples with and without all completed test data). For the former, all sub-test features are combined for feature selection, while for the latter, each sub-test feature is selected individually. The selection steps are as follows: To address the issues of high dimensionality and redundancy in 3D skeleton data, a hybrid feature selection method combining statistical tests and machine learning is proposed. The Shapiro–Wilk test and Levene test are applied to evaluate the normality and homogeneity of variance of each feature. Based on the above test results, the independent samples t-test, Welch t-test, or Mann–Whitney U test is used to identify features with statistically significant inter-group differences (p-value < 0.05). Correlation analysis is then performed to eliminate highly correlated features; for two features with a correlation coefficient greater than 80%, the one with the lower p-value is retained. Finally, the average impurity reduction index derived from the random forest algorithm is used to evaluate feature importance, and the top 20 features with the highest importance scores are selected for subsequent analysis. Furthermore, based on volcano plots, inductive analysis is performed on the feature distributions across different test tasks to characterize the differences in action patterns reflected in each test, such as... Figure 6 As shown.

[0033] Furthermore, training the complete data model and the incomplete data model using the complete dataset and the incomplete dataset respectively includes: Outliers in the complete dataset are processed using the capping method. The complete dataset after outlier processing is divided into a training set and a test set. The complete data model is trained by performing 10-fold cross-validation on the training set. The random forest base model is constructed for each sub-task, and the prediction results of each random forest base model are integrated through a weighted soft voting mechanism to train the incomplete data model.

[0034] Furthermore, the integration of prediction results from various random forest base models through a weighted soft voting mechanism includes: weighting and summing the class probabilities output by each random forest base model, and determining the final discrimination result based on the fused prediction probabilities.

[0035] Specifically, the ensemble model is constructed as follows: two sets of models are built: a complete data model and an incomplete data model. The former is trained and tested using a complete dataset, while the latter is tested using an incomplete dataset. In terms of overall structure, the complete data model uses a random forest model, with each sample having 20 selected SSF features. The incomplete data model, on the other hand, builds a separate random forest base model for each subtest, with each base model having different 20 SSF features. Finally, a soft voting method is used to integrate the seven base models, achieving a complementary effect between tests. The modeling process is as follows: ① The complete data model is constructed using a conventional modeling process. First, using a threshold of... The first step is to use a capping method to handle outliers. Secondly, the dataset is divided into training and test sets in an 8:2 ratio. To ensure the robustness of the model evaluation and to fine-tune hyperparameters, 10-fold cross-validation is performed on the training set. Furthermore, to address the imbalance problem, the Borderline-SMOTE (Synthetic Minority Oversampling) algorithm is applied to the training set. Finally, after data preparation, a random forest algorithm is used for classification. ② Incomplete Data Model: Seven base models are constructed based on individual subtests. The modeling process for each base model is consistent with the above method, including feature extraction, feature selection, and classifier training. The standard SPPB test consists of seven subtasks, each corresponding to different motor function performances. For subjects missing some subtasks, the base models corresponding to completed subtasks are included in the integration process. During the integration phase, a weighted soft voting strategy is used to fuse the prediction results of each base model, i.e., the class probabilities output by each base model are weighted and summed to obtain the comprehensive predicted probability of the sample belonging to each class. The weights of each base model are allocated based on its classification performance on the training data to reflect the contribution of different sub-task models to the overall discrimination. Finally, the discrimination result of the sample is determined based on the fused prediction probability. The algorithm steps are shown in Table 2. The evaluation metrics are defined as follows: ; ; ; ; ; This represents the number of samples that were correctly classified as positive. This represents the number of samples that were misclassified as negative, even though they were originally positive. This represents the number of samples that were misclassified as positive but were actually negative. This represents the number of samples that were correctly classified as negative.

[0036] Table 2 This invention proposes a Symmetric Stability Features (SSF) extraction scheme, constructing a comprehensive symmetry quantification system covering "relative position, relative direction, and joint angle" based on any two joints. It combines multiple feature types, including descriptive statistical features, spatiotemporal features, and spectral features, and for the first time deeply explores the motor compensation patterns of MCI patients from the perspective of dynamic system stability. Secondly, this invention designs a multi-task ensemble learning strategy to resist incomplete data through multi-difficulty level motor function assessment tests, such as the Short Physical Performance Battery (SPPB). By constructing independent base classifiers for each sub-task and employing a weighted soft voting mechanism, it effectively solves the problem of missing data for high-difficulty tasks due to insufficient subject ability, achieving highly robust identification of MCI, and can be combined with Shap interpretability analysis to analyze the specific feature contributions. Results show that this framework exhibits robust and excellent predictive performance on real-world datasets, providing a low-cost and cost-effective tool for large-scale screening and personalized detection of MCI.

[0037] Example 1: A cognitive function assessment method based on skeletal data signals and symmetry stability features includes: Identifying motor function (MCI) in community-dwelling older adults requires them to complete the Short Physical Performance Battery (SPPB) test. Taking this test as an example, subjects must complete three parts sequentially: balance test (side-by-side standing, semi-tandem standing, tandem standing), gait test (two 3-meter walks), and chair sit-up test (single sit-up, five repetitions), named SPPB1-7. During the test, a depth camera is mounted on a tripod and placed directly in front of the subject, with a sampling frequency set to 20 Hz. Figure 5 As shown, the venue setup strictly followed the technical specifications, with the camera positioned 1.5 meters behind the finish line at a height of 1.4 meters to ensure complete capture of the full-body skeletal data.

[0038] The acquired raw 3D skeleton data underwent data cleaning and segmentation. First, a coordinate system with the camera position as the origin was defined. To eliminate high-frequency noise in the signal, a Gaussian filter was used to smooth the signal for each joint. After data preprocessing, the untrimmed continuous data was divided into seven sub-task segments based on timestamps. To prevent interference from equipment vibration, the first and last 10% of the data in each sub-task were removed. Then, the required features were extracted according to a strategy and input into the constructed model. We propose an extraction scheme based on Symmetric Stability Feature (SSF), which calculates the symmetry features of joint relative position, relative direction, and joint angle, and extracts time-domain (standard deviation, impulse factor, waveform factor) and frequency-domain (harmonic ratio, power spectral entropy, wavelet entropy) indices.

[0039] The model was trained using real data from 176 subjects (99 with normal cognition and 77 with mild cognitive impairment). Differences in all features between the two groups were tested and visualized to analyze the differences in data distribution across different difficulty levels and explain behavioral phenomena. Features were then processed and divided into two datasets, with different modeling strategies employed for each dataset. The primary model used was an incomplete data model, incorporating random forest base classifiers trained separately for each of the seven SPPB subtasks. In ensemble learning, a weighted soft voting strategy was used, where the weight of each base model was determined by its accuracy on the training set. During training, the dataset was divided into training and test sets in an 8:2 ratio, and 10-fold cross-validation was used for hyperparameter optimization on the training set. Data processing and model development primarily relied on Python (version 3.10.9) for statistical analysis and plotting.

[0040] After inputting the data into the model, the MCI prediction results can be obtained. In past test sets, this ensemble strategy model has shown greater robustness compared to a single model using only complete data, especially in data scenarios with missing subtasks, as shown in Table 3. Figure 7 As shown, the model achieved an Area Under the Curve (AUC) of 0.9781 and an accuracy of 0.8889 on the test set, demonstrating excellent screening performance that surpasses the performance of each base model and the complete data model, reflecting the complementary effect between subtasks. Furthermore, combined with SHAP (Shapley Additive exPlanations) analysis, the model further revealed the unique motor stability patterns of MCI patients under tasks of varying difficulty.

[0041] Table 3 This embodiment integrates multi-dimensional analysis of relative position, relative direction, and joint angle symmetry based on three-dimensional skeletal data, overcoming the limitations of existing studies that only focus on local limbs or single test task features. By fusing statistical features in the time and frequency domains, it can comprehensively reflect the differences in overall stability and compensatory mechanisms among the elderly population when performing SPPB sub-tasks of varying difficulty. In particular, it introduces feature difference analysis based on volcano maps, intuitively revealing specific movement patterns in patients with mild cognitive impairment at different balance difficulties, such as hand compensation at higher difficulty levels, further improving the specificity of feature selection and the interpretability of pathological mechanisms.

[0042] Multi-task ensemble model construction for incomplete data: By combining random forest base classifiers for each sub-task with a weighted soft voting mechanism, the comprehensive decision-making process for incomplete clinical data is optimized. When dealing with scenarios where some sub-tasks are missing due to physical limitations of elderly individuals, such as the inability to complete the tandem standing exercise, this model effectively overcomes the problems of traditional single models' strong dependence on data integrity and poor robustness, significantly improving the model's adaptability in real-world community screening scenarios and its accurate identification of MCI patients.

[0043] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A cognitive function assessment method based on skeletal data signals and symmetry stability characteristics, characterized in that, include: Collect three-dimensional skeletal data of the subject, preprocess the three-dimensional skeletal data, and obtain preprocessed three-dimensional skeletal data; The relative position features, relative direction features, and joint angle symmetry features of the preprocessed 3D skeleton data are calculated respectively, and feature extraction is performed to obtain symmetry stability features; Feature filtering is performed on the symmetry stability features to obtain complete and incomplete datasets; The complete data model and the incomplete data model are trained using the complete dataset and the incomplete dataset, respectively. The complete data model is constructed using a random forest model, and the incomplete data model is constructed using several random forest base models. Obtain the three-dimensional skeletal data to be evaluated. Depending on whether the three-dimensional skeletal data to be evaluated has missing sub-tasks, input the three-dimensional skeletal data to be evaluated into a complete data model or an incomplete data model, and output the cognitive function evaluation result.

2. The cognitive function assessment method based on skeletal data signals and symmetry stability characteristics according to claim 1, characterized in that, Preprocessing of the three-dimensional skeletal data includes: After Gaussian filtering and denoising the three-dimensional skeleton data, the continuous data stream is divided into independent data segments corresponding to different sub-tasks according to the timestamp, and the data at the beginning and end with a preset proportion is removed to eliminate interference.

3. The cognitive function assessment method based on skeletal data signals and symmetry stability characteristics according to claim 1, characterized in that, Calculating the relative position features includes: Based on the preprocessed 3D skeleton data, the relative distance between any two joints in 3D space is obtained, and the relative displacement sequence is obtained. Based on the relative displacement sequence, the relative displacement of adjacent interval frames is differentially divided by the interval time to obtain the change velocity sequence; Based on the changing velocity sequence, the velocity of adjacent interval frames is differentiated and divided by the interval time to obtain the acceleration sequence; The relative displacement sequence, the velocity change sequence, and the acceleration sequence are used as the relative position features.

4. The cognitive function assessment method based on skeletal data signals and symmetry stability characteristics according to claim 1, characterized in that, Calculating the relative orientation features includes: Based on the preprocessed 3D skeleton data, the connection between any two joint points is calculated using trigonometric functions. Calculate the angle between the line connecting any two joints and any axis of the coordinate system, and obtain the sequence of angles in the coordinate system. The angular velocity is calculated by subtracting the angle between any axis of the coordinate system for adjacent frames and dividing by the interval time, thus obtaining the angular velocity sequence. The angular velocity of adjacent frames is differentiated and divided by the interval time to calculate the angular acceleration, thus obtaining an angular acceleration sequence; The coordinate system's included angle sequence, the angular velocity sequence, and the angular acceleration sequence are used as the relative direction features.

5. The cognitive function assessment method based on skeletal data signals and symmetry stability characteristics according to claim 1, characterized in that, Calculating the joint angle symmetry features includes: Based on the preprocessed 3D skeleton data, the joint angles are calculated using vector operations. The difference in joint angles, the rate of change of joint angles over time, and the acceleration of joint angles are calculated based on the joint angles. The difference sequence, the rate of change of joint angles, and the acceleration sequence of joint angles are obtained as the symmetry features of the joint angles.

6. The cognitive function assessment method based on skeletal data signals and symmetry stability characteristics according to claim 1, characterized in that, Feature filtering is performed on the symmetry stability features to obtain complete and incomplete datasets, including: The symmetry stability features are divided into an initial complete dataset and an initial incomplete dataset. A hybrid feature selection method combining statistical tests and machine learning is adopted. First, features with statistically significant differences are selected through normality tests, homogeneity of variance tests, and inter-group difference tests. Then, highly correlated features are eliminated through correlation analysis. Finally, key features are selected based on feature importance assessment to obtain the complete dataset and the incomplete dataset. Specifically, for the initial complete dataset, all sub-test features are merged for feature selection, while for the initial incomplete dataset, each sub-test feature is selected individually.

7. The cognitive function assessment method based on skeletal data signals and symmetry stability characteristics according to claim 1, characterized in that, Training the complete data model and the incomplete data model using the complete dataset and the incomplete dataset respectively includes: Outliers in the complete dataset are processed using the capping method. The complete dataset after outlier processing is divided into a training set and a test set. The complete data model is trained by performing 10-fold cross-validation on the training set. The random forest base model is constructed for each sub-task, and the prediction results of each random forest base model are integrated through a weighted soft voting mechanism to train the incomplete data model.

8. The cognitive function assessment method based on skeletal data signals and symmetry stability characteristics according to claim 7, characterized in that, Integrating the prediction results of each random forest base model through a weighted soft voting mechanism includes: weighting and summing the class probabilities output by each random forest base model, and determining the final discrimination result based on the fused prediction probabilities.

Images