A plantar pressure-based diabetic foot assessment method and system

Abstract

This invention discloses a method and system for assessing diabetic foot based on plantar pressure, belonging to the field of gait and plantar pressure data analysis. The method includes: collecting plantar pressure data from subjects; analyzing and processing the collected plantar pressure data to extract original features related to foot lesions; enhancing and filtering the extracted original features to obtain feature data with high correlation to labels; classifying the obtained feature data into similar subdomain spaces using a multi-subdomain classification algorithm model; the subdomain spaces are multiple and are obtained by partitioning a single domain of a typical plantar pressure feature dataset feature space using a multi-subdomain partitioning method; and performing CPT / VPT regression prediction on the subject's feature data in the subdomain spaces to obtain assessment results. This invention, based on plantar pressure signals, can accurately and effectively assess diabetic foot and provide better decision support for clinicians, thereby improving the rehabilitation and treatment outcomes of diabetic patients.

Description

Technical Field

[0001] This invention relates to the field of gait and plantar pressure data analysis, and more particularly to a method and system for assessing diabetic foot based on plantar pressure. Background Technology

[0002] In recent years, while people's quality of life has continuously improved, unhealthy lifestyles and an increasingly aging population, among other factors, have led to a rapid increase in the number of people with diabetes worldwide. As of 2023, more than 500 million people aged 20 to 80 had diabetes, accounting for one-tenth of the global adult population. The WHO predicts that by 2030, the number of people with diabetes worldwide will exceed 600 million. China has the largest number of diabetes patients, with more than 140 million in 2023.

[0003] Diabetic foot is a serious chronic complication of diabetes, characterized by high incidence and difficulty in treatment and care. Diabetic foot causes excessive stress on the feet during daily activities, leading to musculoskeletal dysfunction in the feet or lower limbs, ultimately resulting in tissue necrosis and foot ulcers.

[0004] By using wearable footwear to monitor plantar pressure data in diabetic patients and analyzing dynamic changes in the foot, the risk of foot lesions can be screened in advance, which is of great significance for the prevention and treatment of foot ulcers. However, the accuracy of current assessments of diabetic foot lesions needs to be improved. Summary of the Invention

[0005] In order to at least partially solve one of the technical problems existing in the prior art, the purpose of this invention is to provide a method and system for assessing diabetic foot based on plantar pressure.

[0006] The technical solution adopted in this invention is:

[0007] A method for assessing diabetic foot based on plantar pressure includes the following steps:

[0008] Collect plantar pressure data from the subjects;

[0009] The collected plantar pressure data were analyzed and processed to extract the original features related to foot lesions;

[0010] The extracted raw features are enhanced and filtered to obtain feature data with high relevance to the labels;

[0011] The obtained feature data is classified into similar subdomain spaces using a multi-subdomain classification algorithm model. There are multiple subdomain spaces, which are obtained by partitioning a single domain of the feature space of a typical foot pressure feature dataset into multiple subdomain spaces.

[0012] In the subdomain space, regression prediction of CPT / VPT is performed on the subject's feature data to obtain the evaluation results.

[0013] Preferably, the feature extraction module performs preprocessing on the plantar pressure data before extracting features; the preprocessing involves cutting the plantar pressure data of the same subject into samples with a 30-second cycle.

[0014] Preferably, the features extracted by the feature extraction module include foot pressure peak value, left and right foot pressure peak value ratio, foot pressure peak value gradient, foot pressure time integral, left and right foot symmetry coefficient, gait phase, and foot pressure center trajectory;

[0015] Preferably, the multi-subdomain partitioning method divides the typical foot pressure feature dataset using the leave-one-out method, selecting a single data point as the test subset of the training set, and using the remaining data as the training subset of the training set; the training subset data is used to train a set of simple models; after the set of simple algorithm models has been trained, it is used to predict the test subset data.

[0016] The set of simple algorithm models includes multiple simple algorithms, each of which obtains a predicted label, thereby obtaining a set of predicted labels for the test subset data. Based on the results of the predicted label set, the test subset data is divided into a certain subdomain space. The leave-one-out method is repeated to initially divide the single domain of the feature space of the typical foot pressure feature dataset into three subdomain spaces: primary clear subdomain I, primary clear subdomain II, and primary fuzzy boundary subdomain.

[0017] More preferably, the basis for initially dividing the test subset data into a certain subdomain space is to determine whether the labels predicted by all simple models in the simple algorithm model for the test subset data are approximately equal;

[0018] Determine if all predicted labels are approximately equal. If they are approximately equal and the predicted labels are also approximately equal to the true labels, then the test data is assigned to the primary clear subdomain I. If all predicted labels are approximately equal but the predicted labels are not approximately equal to the true labels, then the test data is assigned to the primary clear subdomain II. If all predicted labels are not approximately equal, then the test data is assigned to the primary fuzzy subdomain.

[0019] Preferably, the simple model set consists of AdaBoost, MLP, GradientBoost, and RandomForest, and the parameters of each algorithm model are default or common parameters.

[0020] Preferably, the primary confusing domain is divided into two subdomains. Random forest algorithm models trained in primary clear subdomain I and primary clear subdomain II are used to predict the labels of the data in the primary fuzzy boundary subdomain. If the two predicted labels are approximately equal, the data is assigned to the secondary fuzzy subdomain; otherwise, it is assigned to the secondary clear subdomain.

[0021] For data belonging to the secondary easy-to-classify domain, further determine whether its predicted label is approximately equal to the true label; if they are approximately equal, it means that the data is correctly classified and belongs to the secondary clear subdomain I; otherwise, it is classified into the secondary clear subdomain II.

[0022] Thus, the single domain space of the typical foot pressure feature dataset is divided into 5 subdomain spaces: primary clear subdomain I, primary clear subdomain II, secondary clear subdomain I, secondary clear subdomain II, and secondary fuzzy boundary subdomain.

[0023] Preferably, the multi-subdomain classification algorithm model uses five subdomain spaces to determine which subdomain space the model input data is more similar to;

[0024] More preferably, the subdomain classification algorithm I uses the entire training set feature data to train a set of simple algorithm models, thereby enabling the determination of whether the model input data has a higher similarity to the primary clear subdomain or the primary fuzzy boundary subdomain.

[0025] If the input data is more similar to the primary clear subdomain, the domain similarity model is used to further determine which primary clear subdomain I or primary clear subdomain II is more similar to. The domain similarity model makes this determination by calculating and comparing the maximum mean difference between the two subdomains and the model input data.

[0026] If the input data is more similar to the primary fuzzy boundary subdomain, then the random forest model is trained using the primary clear subdomain I and primary clear subdomain II respectively through subdomain classification algorithm II to determine which subdomain or secondary clear subdomain is more similar to the data.

[0027] If the input data is more similar to the secondary clear subdomain, the domain similarity model is used to further determine whether the data is more similar to the secondary easily classified correct domain or the secondary easily classified incorrect domain.

[0028] The domain similarity model measures which subdomain the model input data is more similar to by calculating and comparing the maximum mean difference between two subdomains and the domain of the model input data.

[0029] Another technical solution adopted in this invention is:

[0030] A plantar pressure-based diabetic foot assessment system includes:

[0031] The data acquisition module is used to collect plantar pressure data from the subjects.

[0032] The feature extraction module is used to analyze and process the collected plantar pressure data and extract the original features related to foot lesions.

[0033] The feature enhancement and filtering module is used to enhance and filter the extracted raw features to obtain feature data with high relevance to the labels.

[0034] A multi-subdomain classification algorithm model is used to classify the obtained feature data into subdomain spaces that are similar to it; wherein, there are multiple subdomain spaces, which are obtained by using a multi-subdomain partitioning method from a single domain of the feature space of a typical foot pressure feature dataset;

[0035] The CPT / VPT regression prediction module is used to perform CPT / VPT regression prediction on the subject's feature data in the subdomain space to obtain the evaluation results.

[0036] The beneficial effects of this invention are as follows: By extracting and analyzing plantar pressure data features, and then constructing a regression algorithm model for diabetic foot using a multi-subdomain classification algorithm, this invention can accurately estimate the values ​​of the current perception threshold (CPT) and vibrating perception threshold (VPT), ​​which can be used to assist in assessing the degree of foot lesions. This provides better decision support for clinicians, thereby improving the rehabilitation and treatment outcomes for diabetic patients. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following description is provided with accompanying drawings of the relevant technical solutions in the embodiments of the present invention or the prior art. It should be understood that the accompanying drawings described below are only for the purpose of clearly illustrating some embodiments of the technical solutions of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is an overall block diagram of a diabetic foot assessment system based on plantar pressure according to an embodiment of the present invention;

[0039] Figure 2 This is a flowchart of feature enhancement and filtering in an embodiment of the present invention;

[0040] Figure 3 This is a flowchart of the multi-subdomain partitioning process of the training set in an embodiment of the present invention;

[0041] Figure 4This is a detailed diagram of the subdomain classification algorithm model in an embodiment of the present invention;

[0042] Figure 5 This is a flowchart of the multi-subdomain classification algorithm for the test set in this embodiment of the invention;

[0043] Figure 6 This is a detailed diagram of the subdomain classification algorithm model in an embodiment of the present invention;

[0044] Figure 7 This is a flowchart illustrating the steps of a diabetic foot assessment method based on plantar pressure, as described in an embodiment of the present invention. Detailed Implementation

[0045] The embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. The step numbers in the following embodiments are set only for ease of explanation, and there is no limitation on the order between the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0046] In the description of this invention, it should be understood that the orientation descriptions, such as up, down, front, back, left, right, etc., are based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0047] In the description of this invention, "several" means one or more, "multiple" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. If "first" or "second" is used, it is only for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features. Furthermore, "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0048] In the description of this invention, unless otherwise explicitly defined, terms such as "set up," "install," and "connect" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this invention in conjunction with the specific content of the technical solution.

[0049] Terminology Explanation:

[0050] MMD: Maximum Mean Difference, a measure of the difference between two distributions.

[0051] AdaBoost, or Adaptive Boosting, is an ensemble learning method based on the boosting strategy, designed to reduce bias. The core idea of ​​AdaBoost is to train different classifiers (weak classifiers) on the same training set, and then combine these weak classifiers to form a stronger final classifier (strong classifier).

[0052] MLP: Multilayer Perceptron.

[0053] GradientBoost: A gradient boosting method.

[0054] RandomForest is a classifier that consists of multiple decision trees, and the class of its output is determined by the mode of the classes output by the individual trees.

[0055] like Figure 7 As shown, this embodiment provides a method for assessing diabetic foot based on plantar pressure, including the following steps:

[0056] S1. Collect plantar pressure data from the subject;

[0057] S2. Analyze and process the collected plantar pressure data to extract the original features related to foot lesions;

[0058] S3. Enhance and filter the extracted original features to obtain feature data with high relevance to the labels;

[0059] S4. Using a multi-subdomain classification algorithm model, the obtained feature data is classified into subdomain spaces that are similar to it; there are multiple subdomain spaces, which are obtained by using a multi-subdomain partitioning method from a single domain of the feature space of a typical foot pressure feature dataset.

[0060] S5. Perform CPT / VPT regression prediction on the subject's feature data in the subdomain space to obtain the evaluation results.

[0061] This embodiment first collects foot pressure data using a foot pressure shoe, then extracts features such as peak foot pressure, foot pressure time-integral pressure, foot pressure stage, and foot pressure center trajectory. After a series of feature enhancements and filtering, features with high correlation to CPT / VPT regression labels are obtained. The multi-subdomain classification algorithm first divides the single domain space of the training set into multiple subdomain spaces that are easier to regress based on these features and labels. The test set data is then classified into the corresponding subdomain spaces using a domain classification algorithm. Finally, foot lesion assessment is performed on the test set data within the subdomain spaces. This embodiment's method, based on plantar pressure signals, can accurately and effectively identify and assess diabetic foot, providing better decision support for clinicians and thus improving the rehabilitation and treatment outcomes for diabetic patients.

[0062] Accordingly, such as Figure 1 As shown, this embodiment also provides a diabetic foot assessment system based on plantar pressure, comprising:

[0063] The data acquisition module is used to collect plantar pressure data, i.e., foot pressure data;

[0064] The feature extraction module is used to analyze and process plantar pressure data and extract original features related to foot lesions;

[0065] The feature enhancement and filtering module is used to enhance and filter the extracted raw features to obtain feature data with high relevance to the labels.

[0066] The multi-subdomain classification algorithm model classifies the feature data of each subject into a subdomain space similar to it. There are multiple subdomain spaces, which are obtained by using a multi-subdomain partitioning method from a single domain of the feature space of a typical foot pressure feature dataset.

[0067] The CPT / VPT regression prediction module is used to perform CPT / VPT regression prediction on the subject's feature data in the similar subdomain space, thereby providing data support for the assessment of their foot lesions.

[0068] In this embodiment, new subject foot pressure data are collected through a data acquisition module, and original features with high correlation to the labels are extracted through a feature extraction module. The original features are then enhanced and filtered. The enhanced and filtered feature data are then classified into similar subdomain spaces using a multi-subdomain classification algorithm model. Finally, CPT / VPT regression prediction is performed on the data assigned to the subdomain spaces. The subdomain spaces used in the subdomain classification algorithm model are divided based on the training set using a set of simple algorithm models.

[0069] In this embodiment, the feature extraction module performs preprocessing on the plantar pressure data before extracting features. The preprocessing involves dividing the plantar pressure data of the same subject into samples with a 30-second interval, and performing the process in subsequent feature extraction and model training in 30-second intervals.

[0070] All extracted features were statistical characteristics of approximately 30 foot pressure cycles in the 30-second sample data, namely, mean, variance, maximum, and minimum features. A total of 450 original features were extracted from 7 categories, including foot pressure peak value, left-right foot foot pressure peak value ratio, foot pressure peak value gradient, foot pressure time integral, left-right foot symmetry coefficient, gait phase, and foot pressure center trajectory.

[0071] like Figure 2 As shown, starting from the training set, feature enhancement and feature selection are performed, which is an alternating iterative process. First, features are enhanced, then selected, and then enhanced and selected again, specifically as follows:

[0072] For the left-foot CPT regression prediction task, new features are obtained from the seven main feature categories through pairwise comparisons to expand the information content of the original dataset. Then, an initial feature selection process is performed, removing features with a correlation coefficient less than 0.12. Next, the remaining features are multiplied pairwise to further enhance their effectiveness. Finally, a second feature selection process is conducted, removing features with a correlation coefficient less than 0.15, resulting in 797 features highly relevant to the left-foot CPT regression prediction task. For the right-foot CPT regression prediction task, the initial feature selection process removes features with a correlation coefficient less than 0.11. Next, the remaining features are multiplied pairwise to further enhance their effectiveness. Finally, a second feature selection process is conducted, removing features with a correlation coefficient less than 0.13, resulting in 725 features highly relevant to the regression task. For the left-foot VPT regression prediction task, the initial feature selection process removes features with a correlation coefficient less than 0.1. Next, the remaining features are multiplied pairwise to further enhance their effectiveness. Finally, a second feature selection process is conducted, removing features with a correlation coefficient less than 0.12, resulting in 611 features highly relevant to the regression task. For the right-foot VPT regression prediction task, the feature preprocessing method is exactly the same as that for the left-foot VPT regression task, resulting in 652 features that are highly correlated with the regression task.

[0073] First, let's systematically introduce the construction process of the multi-subdomain classification algorithm model in this embodiment. Firstly, all collected foot pressure datasets are divided into training and test sets through cross-validation. To build a better model, the training set is first divided into five subdomain spaces: primary clear subdomain I, primary clear subdomain II, secondary clear subdomain I, secondary clear subdomain II, and secondary fuzzy boundary subdomain. Then, the multi-subdomain classification algorithm is constructed using these five subdomain spaces, enabling it to predict which subdomain space is more similar to the test set data. The multi-subdomain partitioning of the training set is implemented as follows: Figure 3 , Figure 4 As shown, the implementation of the multi-subdomain classification algorithm model is as follows: Figure 5 , Figure 6 As shown below, each item will be introduced in detail.

[0074] The following section introduces a method for dividing the training set into five subdomains. For example... Figure 3 As shown, the first step is to train subdomain classification algorithm I using training data and then use it to predict the spatial distribution characteristics of the test data. The initial partitioning results in three types: primary clear subdomain I, primary clear subdomain II, and primary fuzzy boundary subdomain. If the test data is classified as a primary fuzzy boundary subdomain, it indicates that its spatial distribution characteristics are not yet clear and further partitioning is needed. Then, subdomain classification algorithm II is trained using training data consisting of primary clear subdomain I and primary clear subdomain II to determine the spatial distribution characteristics of the data in the primary fuzzy boundary subdomain. The secondary partitioning results also have three types: secondary clear subdomain I, secondary clear subdomain II, and secondary fuzzy boundary subdomain. After two subdomain partitions, the spatial distribution of the training set is finally subdivided into five subdomain spaces: primary clear subdomain I, primary clear subdomain II, secondary clear subdomain I, secondary clear subdomain II, and secondary fuzzy boundary subdomain.

[0075] Figure 4This paper further demonstrates the structural details of Subdomain Classification Algorithm I and Subdomain Classification Algorithm II, providing an in-depth perspective on how to achieve accurate partitioning of the training set from a single spatial domain to five subdomains. Subdomain Classification Algorithm I is used to divide the training set into Primary Clear Subdomain I, Primary Clear Subdomain II, and Primary Fuzzy Subdomain. First, each simple model in a pre-defined set of simple models is trained using the training data. Then, these models perform regression prediction on the test data to obtain the corresponding regression prediction label set. The algorithm determines whether all predicted labels are approximately equal. If they are approximately equal and the predicted labels are also approximately equal to the true labels, the test data is assigned to Primary Clear Subdomain I. If all predicted labels are approximately equal but the predicted labels are not approximately equal to the true labels, the test data is assigned to Primary Clear Subdomain II. If all predicted labels are not approximately equal, the test data is assigned to Primary Fuzzy Subdomain. In this experiment, the set of simple models consists of AdaBoost, MLP, GradientBoost, and RandomForest, and the parameters of each algorithm model are default or common parameters.

[0076] Specifically, whether all predicted labels are approximately equal is determined by calculating the following formula:

[0077]

[0078] Among them, y i Let be the regression predicted label value obtained from the i-th simple model. The mean of all predicted labels is used, and the CV obtained by the equation measures the dispersion of the discrete sequence. When the CV is greater than a certain threshold, it is considered that all predicted labels are not approximately equal.

[0079] Subdomain classification algorithm II is used to divide a primary fuzzy subdomain into secondary sharp subdomain I, secondary sharp subdomain II, and secondary fuzzy subdomain. First, a random forest algorithm model is trained using training data. Since the training data consists of two parts, primary sharp subdomain I and primary sharp subdomain II, two random forest models can be trained separately. The models then perform regression prediction on the test data, yielding two label values, y1 and y2. Finally, it is determined whether the two label values ​​are approximately equal. If they are approximately equal and the predicted label is also approximately equal to the true label, the test data is classified into secondary sharp subdomain I; if the two label values ​​are approximately equal but the predicted label is not approximately equal to the true label, the test data is classified into primary sharp subdomain II; if the two label values ​​are not approximately equal, the test data is classified into the primary fuzzy subdomain.

[0080] Through a two-stage source domain partitioning strategy, this embodiment meticulously divides the training set into five distinct subdomains: primary sharp subdomain I, primary sharp subdomain II, secondary sharp subdomain I, secondary sharp subdomain II, and secondary blurred boundary subdomain. A significant characteristic of these subdomains is the strong correlation between labels and features within the same subdomain, while the similarity of feature space distribution varies considerably between different subdomains.

[0081] After dividing the training set into 5 subdomains, these subdomains can be used to construct a multi-subdomain classification algorithm model. This model can determine which subdomain space the input foot pressure feature data is most similar to, i.e., which subdomain space it belongs to. For example... Figure 5 The subdomain classification algorithm for the test set, as demonstrated, involves three key components: Subdomain Classification Algorithm I, Subdomain Classification Algorithm II, and a domain similarity model. These components work together in the feature space of a single test set (target domain) to determine which subdomain among the following—Primary Clear Subdomain I, Primary Clear Subdomain II, Secondary Clear Subdomain I, Secondary Clear Subdomain II, and Secondary Blurred Boundary Subdomain—has higher similarity. Here, the single test set data refers to all sample feature data from a single individual subject.

[0082] Specifically, after training on the entire training set (source domain), Subdomain Classification Algorithm I is used to predict whether the original target domain feature distribution is closer to the primary sharp boundary subdomain or the primary fuzzy boundary subdomain. If the target domain feature distribution is similar to the primary fuzzy boundary subdomain, further classification is performed using Subdomain Classification Algorithm II. This algorithm is trained using the primary sharp boundary subdomain obtained by merging primary sharp subdomains I and II, and is used to predict the similarity between the target domain and the secondary sharp boundary subdomain or the secondary fuzzy boundary subdomain. If the target domain feature distribution is similar to the primary / secondary sharp boundary subdomain, the domain similarity model is used to obtain the next result. The domain similarity model calculates and compares the MMD (Mean Difference of Data) of the target domain and the primary / secondary sharp boundary subdomain distributions to determine the degree of similarity between the target domain and primary / secondary sharp boundary subdomains I or II. By evaluating the degree of similarity and difference between different domains, the domain similarity model effectively classifies the test set data into primary or secondary easily classifiable domains, thereby enhancing the accuracy of subdomain classification.

[0083] Figure 6This paper details the structure of Subdomain Classification Algorithm I and Subdomain Classification Algorithm II, providing an in-depth perspective on how test set data is efficiently and accurately classified into corresponding subdomains. Subdomain Classification Algorithm I is used to determine which of the initial test set single-sample data feature distributions is more similar to the primary clear subdomain and the primary fuzzy subdomain. Subdomain Classification Algorithm I has a similar structure to Subdomain Classification Algorithm II, obtaining the regression prediction label set through the same simple model set. Unlike Subdomain Classification Algorithm I, the actual labels of the test data cannot be used in this domain classification judgment logic; therefore, only the criterion of whether all predicted labels are approximately equal can be used. Specifically, if all predicted labels are approximately equal, the test data feature distribution is more similar to the primary clear boundary subdomain; otherwise, it is more similar to the primary fuzzy subdomain. Subdomain Classification Algorithm II is used to further determine which of the test set single-sample data feature distributions similar to the primary fuzzy subdomain is more similar to the secondary clear subdomain and the secondary fuzzy subdomain. Subdomain Classification Algorithm II has a similar structure to Subdomain Classification Algorithm II, obtaining two regression prediction labels, y1 and y2, through the same simple model. Similarly, since the true labels of the test data cannot be used in this domain classification judgment logic, only the criterion of whether the two regression predicted labels are approximately equal can be used. Specifically, if the two regression predicted labels are approximately equal, then the feature distribution of the test data is more similar to the secondary clear boundary subdomain; otherwise, it is more similar to the secondary fuzzy subdomain.

[0084] In summary, subdomain classification of the test set can determine which of the five subdomains in the training set has the highest similarity between the data feature spatial distribution of a single subject and that subdomain. The application of the subdomain classification algorithm allows the test set to be accurately classified into subdomains with highly correlated feature labels. Subsequently, the model is trained using these specific subdomains for label prediction, achieving higher prediction performance.

[0085] Finally, after accurately dividing the data of each subject in the test set into subdomains with significant label feature correlation, the random forest algorithm model trained in the corresponding subdomain is used to perform CPT / VPT regression prediction, thereby obtaining more accurate label results.

[0086] In the multi-subdomain partitioning process and multi-subdomain classification algorithm model, the data used for model training and prediction are all data bound to the subjects as samples. During leave-one-out training and testing, the training data and test data will not contain samples belonging to the same subject at the same time, which can prevent data leakage during the algorithm model training process. Accordingly, the label predicted by each algorithm model is the majority vote result of multiple sample predictions, ensuring that the prediction is made with the subject as a single subject.

[0087] In practical applications, after collecting plantar pressure data from diabetic patients, features are extracted using a feature extraction module, and then processed using a feature enhancement and filtering module to obtain feature data with high correlation to the labels. A multi-subdomain classification algorithm then categorizes these label-correlated feature data into subdomain spaces similar to their own. Finally, a CPT / VPT regression prediction module predicts the label for these feature data within this subdomain space, thereby obtaining the subject's foot assessment. The subdomain space is obtained by dividing typical plantar pressure feature data (which performs the same function as the training set in the above embodiment) into subdomains.

[0088] In summary, the system of this invention, based on collected plantar pressure data, integrates multiple modules including a plantar pressure signal acquisition module, a feature extraction module, a feature enhancement and screening module, a multi-subdomain classification algorithm module, and a CPT / VPT regression prediction module. It can accurately and effectively assess diabetic foot and provide better decision support for clinicians, thereby improving the rehabilitation and treatment outcomes for diabetic patients. Specifically, compared with existing technologies, this invention has the following technical advantages in assessing diabetic foot:

[0089] (1) In the feature enhancement and filtering module, different filtering thresholds are provided for different regression tasks, and the same regression algorithm model framework is adopted for all of them. While ensuring the accuracy of the task, the application scope of the whole system is expanded.

[0090] (2) Dividing the mixed single domain of the training set feature space into 5 subdomains with distinct feature distributions provides more information and can provide new research content and directions.

[0091] (3) By using a multi-subdomain classification algorithm, the test set data or newly collected data is classified into a subdomain space with similar feature distribution, and the label is predicted, thereby improving the prediction effect.

[0092] In the foregoing description of this specification, references to terms such as "one embodiment," "another embodiment," or "some embodiments" indicate that a specific feature, structure, material, or characteristic described in connection with an embodiment or example is included in at least one embodiment or example of the present invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0093] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0094] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for assessing diabetic foot based on plantar pressure, characterized in that, Includes the following steps: Collect plantar pressure data from the subjects; The collected plantar pressure data are analyzed and processed to extract original features related to foot lesions; the original features include peak plantar pressure, ratio of peak plantar pressure between left and right feet, peak plantar pressure gradient, plantar pressure time integral, left and right foot symmetry coefficient, gait stage, and plantar pressure center trajectory; The extracted raw features are enhanced and filtered to obtain feature data that are highly correlated with the CPT and VPT of the left and right feet of the label; The obtained feature data is classified into similar subdomain spaces using a multi-subdomain classification algorithm model. There are multiple subdomain spaces, which are obtained by partitioning a single domain of the feature space of a typical foot pressure feature dataset into multiple subdomain spaces. In the subdomain space, the feature data of the subjects are used to perform regression prediction of the left and right foot CPT and VPT using the trained random forest algorithm model to obtain the evaluation results; The multi-subdomain classification algorithm model includes subdomain classification algorithm I, subdomain classification algorithm II, and a domain similarity model; The subdomain classification algorithm I and subdomain classification algorithm II are used to predict the feature distribution of the original target domain. If the feature distribution of the target domain is similar to that of the primary / secondary clear boundary subdomain I, the domain similarity model will be used to obtain the next result. The domain similarity model uses the method of calculating and comparing the MMD size of the target domain with that of the primary / secondary clear boundary subdomain I to determine the degree of similarity between the target domain and the primary / secondary clear boundary subdomain I. The specific subdomain classification algorithm I is as follows: First, each model in the preset model set is trained using training data. Then, these models are used to perform regression prediction on the test data to obtain the corresponding regression prediction label set. The model set includes AdaBoost, MLP, GradientBoost, and RandomForest. Determine whether all predicted labels are approximately equal. If they are approximately equal and the predicted labels are also approximately equal to the true labels, then the test data is assigned to the primary clear subdomain I. If all predicted labels are approximately equal but the predicted labels are not approximately equal to the true labels, then the test data is assigned to the primary clear subdomain II. If all predicted labels are not approximately equal, then the test data is assigned to the primary fuzzy subdomain. The specific subdomain classification algorithm II is as follows: First, two random forest algorithm models are trained using the data from two primary clear subdomains, I and II, respectively. Then, the random forest algorithm models are used to perform regression prediction on the test data to obtain two label values. Determine whether two label values ​​are approximately equal. If they are approximately equal and the predicted label is also approximately equal to the true label, then the test data is assigned to the secondary clear subdomain I. If the two label values ​​are approximately equal but the predicted label is not approximately equal to the true label, then the test data is assigned to the primary clear subdomain II. If the two label values ​​are not approximately equal, then the test data is assigned to the secondary fuzzy subdomain.

2. The method for assessing diabetic foot based on plantar pressure according to claim 1, characterized in that, The analysis and processing of the collected plantar pressure data includes: The collected plantar pressure data is preprocessed; the preprocessing involves cutting the plantar pressure data of the same subject into samples according to a preset time period.

3. The method for assessing diabetic foot based on plantar pressure according to claim 1, characterized in that, The enhancement and filtering process for the extracted original features includes: Different screening thresholds are provided for different regression tasks, and the same regression algorithm model framework is used for all of them, so as to improve the accuracy of the task and expand the application scope of the entire system.

4. The method for assessing diabetic foot based on plantar pressure according to claim 1, characterized in that, The multi-subdomain partitioning method includes: The training data is used to train the subdomain classification algorithm I, and the trained subdomain classification algorithm I is used to predict the test data, determine its spatial distribution characteristics, and thus classify the test data into types. The initial classification results include three types: primary clear subdomain I, primary clear subdomain II, and primary fuzzy boundary subdomain. If the test data is identified as a primary fuzzy boundary subdomain, it indicates that its spatial distribution characteristics are not yet obvious and further classification is needed. Then, the subdomain classification algorithm II is trained using the training data consisting of primary clear subdomain I and primary clear subdomain II to determine the spatial distribution characteristics of the data in the primary fuzzy boundary subdomain. The secondary partitioning results include three types: secondary sharp subdomain I, secondary sharp subdomain II, and secondary fuzzy boundary subdomain. After two subdomain partitionings, the spatial distribution of the test data is finally subdivided into five subdomain spaces: primary sharp subdomain I, primary sharp subdomain II, secondary sharp subdomain I, secondary sharp subdomain II, and secondary fuzzy boundary subdomain.

5. The method for assessing diabetic foot based on plantar pressure according to claim 1, characterized in that, The following formula is used to determine whether all predicted labels are approximately equal: in, For the first The regression-predicted label values ​​obtained from the model The mean of all predicted labels; To measure the degree of dispersion of a discrete sequence, when If the value exceeds a preset threshold, it is determined that all predicted labels are not approximately equal.

6. A diabetic foot assessment system based on plantar pressure, applied to the method described in any one of claims 1-5, characterized in that, include: The data acquisition module is used to collect plantar pressure data from the subjects. The feature extraction module is used to analyze and process the collected plantar pressure data and extract the original features related to foot lesions. The original features include the peak plantar pressure, the ratio of the peak plantar pressure of the left and right feet, the peak plantar pressure gradient, the time integral of the plantar pressure, the symmetry coefficient of the left and right feet, the gait stage, and the center trajectory of the plantar pressure. The feature enhancement and filtering module is used to enhance and filter the extracted raw features to obtain feature data that is highly correlated with the left and right foot CPT and VPT of the label. A multi-subdomain classification algorithm model is used to classify the obtained feature data into subdomain spaces that are similar to it; wherein, there are multiple subdomain spaces, which are obtained by using a multi-subdomain partitioning method from a single domain of the feature space of a typical foot pressure feature dataset; The CPT and VPT regression prediction module is used to perform regression prediction of the left and right foot CPT and VPT on the subject's feature data in the subdomain space using a trained random forest algorithm model, and obtain evaluation results.

Images