A multivariate time series data enhancement method for wound microenvironment data

Abstract

The present invention relates to a multivariate time series data enhancement method for wound microenvironment data, which belongs to the medical field and the field of machine learning technology. Based on the dynamic time warping algorithm, similarity measurement between samples is performed, by randomly selecting a sample as a benchmark sample, searching for K samples of the same category that are the nearest neighbors of the benchmark sample, performing dynamic time warping on all selected samples relative to the benchmark sample, and performing weighted averaging on the warped sample set, and taking the mean sequence data as a new synthetic sample. By increasing data diversity by random weights and adding noise, the problems of model overfitting and poor generalization ability caused by insufficient data volume are compensated, and the prediction effects of various classifiers on wound microenvironment data sets are effectively improved.

Description

Technical Field

[0001] The present invention relates to the fields of medical care and machine learning technology, and in particular to a multivariate time series data enhancement method for analyzing wound microenvironment data using a machine learning method. Background Art

[0002] In recent years, technologies such as data mining and artificial intelligence have rapidly advanced, achieving breakthroughs in various fields of applied science and engineering. Their accuracy and efficiency have earned them widespread recognition among experts across diverse fields. Existing mainstream data analysis methods, particularly supervised learning models based on deep neural networks, are based on the theory of asymptotic regression as the amount of data approaches infinity. Therefore, a large number of data samples is essential for developing a model with high predictive accuracy. However, in the medical field, due to the confidentiality and incompleteness of clinical data, the rarity of some cases, and the need to protect patient privacy, it is often difficult to obtain a large number of clinical diagnostic data samples suitable for analysis. Trauma is a common clinical condition, and the changing trends of wound microenvironment parameters are crucial to wound healing. Rapidly assessing wound status based on wound microenvironment data to guide physicians' diagnosis and treatment is of great practical significance. However, due to the difficulty of collecting large amounts of standardized wound microenvironment data, it is not easy to obtain. The question of how to perform intelligent diagnosis based on small amounts of wound microenvironment data is a hot topic among experts in related fields. A mainstream approach to addressing this data scarcity problem is to augment existing datasets to obtain richer data. This not only allows the model to be fully trained, but also, since data augmentation techniques usually increase the diversity of data, the generalization ability of the model can be further improved.

[0003] Multivariate time series (MTS) data is a common form of data that is prevalent in various fields of daily life, such as stock price trends in the economic field and historical records of tumor growth in the medical field. Wound microenvironment data is also in the form of MTS data. At each time point, various data such as exudate pH, wound temperature, humidity, and oxygen partial pressure can be collected. Discovering potential patterns from the multivariate time series data of the wound microenvironment and mining the value of the data can effectively guide doctors' clinical decisions and reduce misdiagnosis rates. Data augmentation methods have a good effect on strengthening and supplementing the fitting and generalization capabilities of machine learning models and deep learning models. Therefore, for wound microenvironment data with scarce data, it is very valuable to provide an effective multivariate time series data enhancement method. At present, most research on multivariate time series data focuses on the extraction of data features and the measurement of distances between samples. There are relatively few studies on data enhancement related methods. Existing typical solutions mainly include:

[0004] (1) Data enhancement methods based on time domain and frequency domain transformation

[0005] Data transformation in the time domain is the most common method for time series data enhancement, usually performed by injecting noise, extracting continuous slices from the original sequence, etc. At the same time, a few studies have also used perturbations of the frequency domain spectrum and amplitude to achieve this.

[0006] (2) Data enhancement methods based on statistical generation models

[0007] This type of method mainly models the original time series data through statistical models such as the autoregressive moving average model (ARMA) and the mixed autoregressive model (MAR). These models describe the conditional distribution of the time series by fitting the dependence of the value at time t on the value of the previous time step, and then generate a new time series sample by perturbing the initial value.

[0008] (3) Learning-based data enhancement methods

[0009] Currently, this type of method primarily generates new samples through learning embeddings or generative adversarial networks. Learning embeddings involves mapping data into a new representational space and enhancing the feature vectors within that representational space. Generative adversarial learning frameworks typically include a generative network and a discriminative network, which are jointly trained to optimize model parameters, ultimately yielding an efficient classifier.

[0010] However, the above data enhancement methods are not fully applicable to wound microenvironment indicator data. The main reasons are as follows:

[0011] (1) Wound microenvironment indicator data are discrete time series data. Due to the short duration of wounds and the difficulty in collecting data, the time step of such data is usually short. Extracting shorter data slices for data enhancement will largely lose some data features.

[0012] (2) Injecting noise into the original data, perturbing the frequency domain spectrum and amplitude, and generating methods based on statistical models all enhance the time series data of a feature dimension within that dimension. However, the wound microenvironment indicator data is a multivariate discrete time series data, and the correlation between its feature dimensions is crucial. Therefore, this type of method is not applicable to the wound microenvironment indicator dataset;

[0013] Learning-based data augmentation methods, depending on the network model they employ, can simultaneously account for inter-dimensional dependencies and temporal dependencies within each dimension. However, due to the numerous parameters in the network structure, training requires a large amount of raw data. For wound microenvironment indicator data, collecting a large number of usable samples is particularly difficult. Therefore, this paper proposes a multivariate time series data augmentation method, different from the above methods, suitable for wound microenvironment indicator datasets. Summary of the Invention

[0014] Technical problems to be solved

[0015] To improve the predictive performance of machine learning models and deep neural network models on data-scarce wound microenvironment data in supervised classification tasks, this paper proposes a multivariate time series data augmentation method based on dynamic time warping (DTW). This method synthesizes new data samples based on the input multivariate time series dataset, while increasing data diversity through random weighting and adding noise to compensate for the problems of model overfitting and poor generalization ability caused by insufficient data, effectively improving the predictive performance of various classifiers on wound microenvironment datasets.

[0016] Technical Solution

[0017] A multivariate time series data enhancement method for wound microenvironment data mainly includes measuring the distance between multivariate time series data samples based on DTW, aligning two multivariate time series data samples based on DTW, calculating the average sequence of multiple data samples, and introducing random weights and Gaussian noise. The overall steps for generating a new sample are as follows:

[0018] Step 1: Input the wound microenvironment dataset X to be enhanced, the label set Y corresponding to the dataset, and configurable parameters: the number of nearest neighbors to be found K and the benchmark sample weight parameter ω;

[0019] Step 2: Randomly select a sample from the dataset X as the benchmark sample, denoted as X0, and its corresponding category label is Y0;

[0020] Step 3: Calculate the DTW-warped distances between all data samples of category Y0 in dataset X and X0, find the nearest K samples, calculate the DTW-warped sequences of these K selected samples relative to X0, and form the dataset AvgTS to be averaged with X0, whose size is K+1;

[0021] Step 4: Generate a random weight set W. Weight set W = {ω, ω1,…,ω K}, ω is the benchmark sample weight value, which is a configurable input parameter, {ω1,…,ωK} is the weight of the K nearest neighbor samples, which is generated by pseudo-random number generation and normalized into K weight values ​​whose sum is 1-ω;

[0022] Step 5: Calculate the weighted average of the data at each time step of the dataset AvgTS according to the weights in W to form the mean time series X new ;

[0023] Step 6: Add noise. new Add Gaussian noise as a new data sample, and its label is consistent with the benchmark sample X0.

[0024] The wound microenvironment dataset X in step 1 is {X1, X2, ..., X N}, where X i ={X i,1 ,X i,2 ,…,X i,T} is a multivariate time series, where X i,j ={x1,x2,…,x n}, N is the number of sample sets, T is the time step of each sample, and n is the number of features in each time step. The label set Y corresponding to the sample set is {Y1, Y2, ..., Y N}, where Y i Represents sample X i The category label of .

[0025] In step 3, calculate a sample X in the sample set c The distance relative to X0 after normalization based on the dynamic time warping algorithm refers to the distance after X0 is normalized by the dynamic time warping algorithm. c Relative to X0, the time axis is distorted, that is, stretched or contracted, and the optimal regularization scheme is sought. The optimal here means that the sum of the distances of each time step after regularization is the smallest. The data of each time step in X0 after regularization corresponds to X c One or more time step data in the sequence is used to calculate the sample distance of each time step and the distance of the entire sequence.

[0026] The dynamic time warping algorithm in step 3 is an optimization problem, solved based on the idea of ​​dynamic programming. The dynamic time programming algorithm can recursively calculate the cost of optimal warping, that is, the distance between the two sequences after alignment, and its calculation formula is as follows:

[0027]

[0028] A and B are two multivariate time series. Time series A is defined as A={a1,a2,...,a T}, the time step is T. The vector a of each time stepi ={a i1 ,a i2 ,...,a in}, the feature vector of each time step has n features. i ,B j ) represents the distance between the first i time steps of time series A and the first j time steps of B, d(a i ,b j ) represents the vector a at time step i in sequence A i and vector b at time step j in sequence B j The distance between them. The data of each time step of the multivariate time series is a vector, and the distance between the vectors of the corresponding time steps is measured using the Euclidean distance, that is:

[0029]

[0030] The distance in step 3 is D(X) obtained by the above DTW algorithm. C,T ,X 0,T ), a matrix can be defined to memorize the distances that have been calculated to reduce the time consumption during the calculation process.

[0031] The DTW regularized sequence in step 3 refers to the alignment sequence of each sample relative to X0. When calculating the DTW distance between two sequences, the alignment scheme for the optimal alignment can be obtained by backtracking based on the distance matrix obtained in the process. The so-called alignment scheme refers to the alignment scheme in which a time step in X0 corresponds to X0. C One or more time steps. For example, X0 = {X 0,1 ,X 0,2 ,...,X 0,T}, X C The alignment scheme corresponding to X0 may be XC′={{X C,1},{X C,2 ,X C,3 ,X C,4},...,{X C,T}}, that is, the first time step of X0 after alignment corresponds to X C The first time step data of X0 corresponds to the second time step data of X C The second, third, and fourth time steps of data, and so on.

[0032] The random weight set W in step 4 is W={ω,ω1,…,ω K}, whose size is K + 1 and sums to 1, representing the weight of each sample in the sample set to be averaged. Here, the baseline sample is assigned a weight of ω, and the weights of the remaining K samples are generated using pseudo-random number generation, with the sum of their weights being 1-ω. ω is a configurable parameter. For the wound microenvironment dataset, its empirical value is 0.5.

[0033] The weighted average sequence of the data set AvgTS calculated in step 5 refers to the alignment sequence of the obtained K samples relative to the reference sample X0, and the weighted average value is calculated for each time step as the numerical vector of the synthetic sample at that time step. Taking the alignment scheme mentioned above as an example, X0 and X C The mean vector at the second time step is:

[0034]

[0035] Where W0 and W c are the weight values ​​corresponding to the two samples in the random weight set.

[0036] The Gaussian noise in step 6 is a type of noise whose probability density function follows a Gaussian distribution. There are two configurable parameters: the expected value and variance of the Gaussian distribution. Adding appropriate Gaussian noise to the data introduces a certain amount of error, making the data more valuable for experimental purposes and improving the model's generalization ability.

[0037] X in step 6 new The label of the synthesized sample is the same as the label of the randomly selected benchmark sample X0. Because the original sample subsets selected for generating new samples have the same label value during the data generation process, they are all sample data of the same category as the benchmark sample. Therefore, the label of the final synthesized sample is consistent with the label of the benchmark sample.

[0038] Beneficial effects

[0039] The present invention proposes a multivariate time series data enhancement method based on a dynamic time warping algorithm. The innovation lies in performing weighted averaging on the relatively warped sequence based on DTW to obtain a synthetic sample, and at the same time improving the diversity of the synthetic data by assigning random weights and Gaussian noise to each sample. The present invention measures the similarity between samples based on the dynamic time warping algorithm, randomly selects a sample as a benchmark sample, searches for K samples of the same category that are the nearest neighbors to the benchmark sample, performs dynamic time warping on all selected samples relative to the benchmark sample, and performs weighted averaging on the warped sample set, and uses the mean sequence data as a new synthetic sample. At the same time, the data is further enriched by introducing random weights and Gaussian noise. By combining a specific machine learning model or a deep learning model, the data enhancement method can effectively improve the prediction accuracy of the model on wound microenvironment data, improve the generalization ability of the model, and is also effective for a variety of other multivariate time series data sets. For deep neural network models, since the training samples are greatly enriched, the model convergence speed can be effectively improved and the occurrence of overfitting can be reduced. This method is targeted at wound microenvironment datasets, but can also be used for multivariate time series data in various other fields. It can effectively address the limitations of applying current mainstream machine learning models and deep learning models when sample data is insufficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The accompanying drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like parts throughout the drawings.

[0041] Figure 1 This is a flow chart of the multivariate time series data enhancement method for wound microenvironment data proposed by the present invention;

[0042] Figure 2 It is the overall classification framework when the present invention is combined with the machine learning model;

[0043] Figure 3 It is the improvement effect of different data sets under different models when applying the invention. DETAILED DESCRIPTION

[0044] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only intended to illustrate the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0045] A multivariate time series dataset refers to data with multiple consecutive time steps and multiple characteristic data values ​​in each time step. It is the data form of the wound microenvironment dataset and is also a common data form in the medical field, with important analytical significance. A large number of standard data samples is the basic prerequisite for obtaining a high-accuracy analysis model. Therefore, for wound microenvironment data that is difficult to collect, effective data enhancement technology is crucial. To this end, for the multivariate time series dataset of the wound microenvironment, the present invention proposes a data enhancement scheme based on a dynamic time warping algorithm.

[0046] Figure 1 The multivariate time series data enhancement scheme for wound microenvironment data proposed in the present invention mainly includes the following parts: DTW-based measurement of similarity between multivariate time series data samples, DTW-based time series data regularization, time series data set weighted averaging, and Gaussian noise addition. Figure 1 , the specific steps for actually amplifying multiple data samples are explained in detail.

[0047] Step 1: Determine whether the input data is in the form of time series data. The input data dimensions should be (N, time_length, feature_nums), where N is the number of samples, time_length is the time step, and feature_nums is the number of features in each time step. You also need to enter the corresponding label set data, whose size is N, as well as the amount of data to be amplified, the number of nearest neighbors to be found, K, and the baseline sample weight value, ω.

[0048] Step 2: Based on the dynamic time warping algorithm, calculate the DTW distance between samples and the relative warping path. To avoid subsequent repeated calculations, define two arrays to store the distance between related samples and the alignment path respectively. Define D(X i ,X j ) represents the sample X i and X j At the same time, in the process of calculating the DTW distance, the alignment scheme for the optimal alignment is obtained by backtracking based on the obtained distance matrix, which is defined as P BA , represents the regular sequence when sequence B is aligned relative to A, where P BA ={p1,p2,...,p L}, which means a in the regularized sequence A i The index sequence of the time step sequence of the corresponding sequence B is p i ={I1,I2,...,I m}, that is, a i The data corresponding to sequence B is

[0049] Step 3: Randomly select a sample X0 from the sample set as the benchmark sample, with its label Y0. Find the K nearest neighbors of the benchmark sample and the same category. Based on the stored alignment path, construct the normalized dataset AvgTS of these K nearest neighbors relative to X0.

[0050] Step 4: Generate a weight set W, where the weight value of the benchmark sample is the input configuration parameter ω, and the remaining K samples are randomly generated by generating random numbers, and the sum of their weights is 1-ω.

[0051] Step 5: Calculate the weighted average of the regularized vectors of the K+1 samples in AvgTS at each time step as the vector of the synthetic data at that time step.

[0052] Step 6: Add Gaussian noise to the synthesized data. There are two configurable parameters here: the expected value and variance of the Gaussian distribution. Noise is a method of data augmentation and essentially a form of regularization. Its purpose is to make the trained model more robust to the effects of noise, improving its robustness and preventing overfitting. Here, Gaussian noise is added to introduce a certain bias into the synthesized data, thereby increasing the richness of the augmented data.

[0053] Step 7: Repeat steps 3 to 6 until amount of synthetic data is generated.

[0054] Figure 2 It is the overall framework of the classification model combined with this data enhancement method. For the original data set, data preprocessing is first required, and corresponding methods are used to deal with missing values, outliers, etc. Then the data set is divided, usually into a training set and a test set. The training set is used to generate new samples and train the classifier, and the test set is used to finally verify the prediction effect of the classifier. We use the above-mentioned data enhancement method to amplify the training set data to generate a certain number of new samples. The new synthetic samples and the original training set samples are input as training data into the constructed classification model for training. After the training is completed, the model fitting effect is tested on the test set. The above-mentioned classification model can be a traditional machine learning model or a constructed neural network classifier.

[0055] Figure 3The results of the classification task test of the data augmentation method on the wound microenvironment dataset and several open source datasets are shown. For the classification model, we chose the benchmark model NN-DTW for the TSC task, which is the nearest neighbor method based on the DTW similarity metric, as well as support vector classification (SVC) and a neural network classifier based on CNN. The amount of amplified data for the test is 1 times and 2 times the amount of data in the initial training set. The number of neighbors configured during the test is 5, the weight ω is 0.5, and the Gaussian noise introduced is noise that obeys a Gaussian distribution with a mean of 0 and a standard deviation of 0.005. Since the purpose of this experiment is to verify the effectiveness of the data augmentation method, the SVC parameters and the structure of the neural network model are fixed for multiple experiments, and they are all optimal parameter configurations after adjustment when no data augmentation is performed. Reference Figure 3 The values ​​in the table are the prediction accuracy of the test set. It can be seen that the data enhancement method can effectively improve the prediction effect of the classifier on the wound microenvironment dataset, and also has certain improvements on other datasets.

[0056] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed in the present invention, and these modifications or replacements should all be included in the scope of protection of the present invention.

Claims

1. A multivariate time series data enhancement method for wound microenvironment data, characterized in that Here are the steps: Step 1: Input the wound microenvironment dataset to be enhanced The data includes exudate pH value, wound temperature, humidity, oxygen partial pressure data, and the label set corresponding to this data set is , and configurable parameters: the number of neighbors to find K, the benchmark sample weight parameter ; The wound microenvironment dataset ,in is a multivariate time series, where , N is the number of sample sets, T is the time step of each sample, n is the number of features in each time step; the label set corresponding to the sample set is ,in Representation sample Category label of Step 2: From the dataset A sample is randomly selected as the benchmark sample, denoted as , and its corresponding category label is ; Step 3: Calculate the dataset All categories in The data samples and Based on the distance between the two after the dynamic time warping algorithm is adjusted, find the nearest K samples and calculate the K selected samples relative to The dynamic time warping algorithm is used to normalize the sequence, and The data set AvgTS to be averaged is composed of a size of K+1; In step 3, a sample in the data set is calculated Relative to The distance after normalization based on the dynamic time warping algorithm refers to the distance after normalization based on the dynamic time warping algorithm. Relative to Distort the time axis, that is, stretch or shrink, and seek the optimal regularization scheme; the optimal here means that the sum of the distances of each time step after regularization is the smallest; after regularization The data of each time step corresponds to One or more time step data in the sequence, and then calculate the sample distance of each time step and the distance of the entire sequence; The regularized sequence in step 3 refers to each sample relative to When calculating the DTW distance between two sequences, the alignment scheme for the optimal alignment is obtained by backtracking based on the distance matrix obtained in the process. The so-called alignment scheme refers to A time step corresponds to One or more time steps; Step 4: Generate a random set of weights ; Weight set , is the benchmark sample weight value, which is a configurable input parameter. is the weight of K nearest neighbor samples, which is generated by pseudo-random number generation and normalized to the sum of K weight values ​​of ; The random weight set in step 4 , whose size is K+1 and whose sum is 1, represents the weight of each sample in the sample set to be averaged; here the benchmark sample is given The weight of the remaining K samples is generated by pseudo-random number generation, and the sum of their weights is ; It is a configurable parameter. For the wound microenvironment dataset, its empirical value is 0.5; Step 5: Calculate the weighted average of the data at each time step of the dataset AvgTS according to the weights in W to form the mean time series ; The weighted average sequence of the AvgTS data set in step 5 is the weighted average sequence of the K samples obtained relative to the benchmark sample. The weighted average of each time step is calculated as the numerical vector of the synthetic sample at that time step; Step 6: Add noise: Add Gaussian noise as a new data sample, and its label is the same as the baseline sample consistent.

Images