A perioperative patient data dimensionality reduction device and sample dataset acquisition system

Abstract

This invention provides a perioperative patient data dimensionality reduction device and a sample dataset acquisition system. The dimensionality reduction device includes: an input module for acquiring the patient's original perioperative feature data containing multidimensional features, and the corresponding classification labels; a primary dimensionality reduction module for performing dimensionality reduction on the original perioperative feature data using principal component analysis (PCA) to obtain first perioperative feature data; a secondary dimensionality reduction module for performing dimensionality reduction on the first perioperative feature data using a genetic algorithm to obtain more perioperative feature data; and an output module for outputting the perioperative feature data. By combining PCA and genetic algorithms for data dimensionality reduction, lower-dimensional perioperative feature data that is more beneficial for subsequent classification processing is selected, accelerating the efficiency of subsequent classification processing while maintaining high classification performance.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a device for reducing the dimensionality of perioperative patient data and a system for acquiring sample datasets. Background Technology

[0002] The perioperative period is the entire process surrounding surgery, from the moment the patient decides to undergo surgical treatment until basic recovery. It includes the period before, during, and after surgery. Specifically, it refers to the time from when surgical treatment is determined until the treatment related to this surgery is basically completed, which is about 5-7 days before surgery to 7-12 days after surgery.

[0003] According to the World Health Organization's (WHO) "World Health Statistics 2021" report, global life expectancy has increased to 73.3 years, and the number of elderly people worldwide is projected to exceed 1.5 billion by 2050. The ever-growing elderly population worldwide has been identified as a major target group for the surgical market, and the prediction of risk events in elderly patients has become a popular research area. Predicting postoperative risks for elderly surgical patients helps doctors develop treatment plans, allocate resources rationally, and reduce the probability of postoperative risk events. Currently, some diagnostic tools can help hospitals provide comprehensive and reliable treatment for high-risk patients. For example, Chinese patents with publication numbers CN111009322A and CN114038565A disclose perioperative risk assessment using predictive models based on perioperative patient datasets. However, many perioperative patient datasets suffer from high dimensionality, which directly affects the efficiency of perioperative prediction models. Blindly reducing dimensionality, on the other hand, reduces the predictive effectiveness of perioperative prediction models. Summary of the Invention

[0004] This invention aims to solve the technical problems existing in the prior art and provides a perioperative patient data dimensionality reduction device and a sample dataset acquisition system.

[0005] To achieve the above-mentioned objectives of the present invention, according to a first aspect of the present invention, the present invention provides a perioperative patient data dimensionality reduction device, comprising: an input module for acquiring original perioperative feature data of patients containing multidimensional features, and classification labels corresponding to the original perioperative feature data; a primary dimensionality reduction module for performing dimensionality reduction processing on the original perioperative feature data based on principal component analysis algorithm to obtain first perioperative feature data; a secondary dimensionality reduction module for performing dimensionality reduction processing on the first perioperative feature data based on genetic algorithm to obtain perioperative feature data; and an output module for outputting the perioperative feature data.

[0006] The above technical solution combines principal component analysis (PCA) and genetic algorithms for data dimensionality reduction. First, PCA is used for initial dimensionality reduction to obtain first-dimensional perioperative feature data that effectively represents the original perioperative feature data. Then, this first-dimensional feature data is used as the input and heuristic dataset for the genetic algorithm. This results in an initial population for the genetic algorithm that is a lower-dimensional and more optimal combination compared to the original perioperative feature data. This creates a smaller search range for further feature selection, improving operational efficiency and filtering out lower-dimensional perioperative feature data that is more beneficial for subsequent classification processing. This accelerates subsequent classification processing while maintaining high classification performance.

[0007] To achieve the above-mentioned objectives of the present invention, according to a second aspect of the present invention, the present invention provides a perioperative patient sample dataset acquisition system, comprising: a data acquisition module for acquiring original perioperative feature data and case records of multiple patients; a classification label set acquisition module for acquiring a set of classification labels based on multiple case records, wherein the classification labels characterize perioperative patient risk events; a classification label association module for associating the original perioperative feature data of patients with at least one classification label in the classification label set; and the perioperative patient data dimensionality reduction device described in the first aspect of the present invention for performing dimensionality reduction processing on the original perioperative feature data of all patients to obtain corresponding perioperative feature data; and a sample dataset acquisition module for using the perioperative feature data of patients as samples, associating the samples with the classification label set corresponding to the original perioperative feature data, and obtaining a perioperative patient sample dataset.

[0008] The above technical solution constructs a multi-class labeled sample dataset of perioperative patients. The features in this dataset have low feature dimensionality, which can accelerate the efficiency of subsequent classification processing and model training. At the same time, the features in the samples are all features that have a significant impact on subsequent classification, which can enable the subsequent classification to have better performance. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the perioperative patient data dimensionality reduction device in Embodiment 1 of the present invention;

[0010] Figure 2 This is a schematic diagram of the perioperative patient sample dataset acquisition system in Embodiment 2 of the present invention;

[0011] Figure 3 This is a schematic diagram of the sample dataset balancing method in Embodiment 3 of the present invention;

[0012] Figure 4 This is a schematic diagram of the sample dataset balancing device in Embodiment 4 of the present invention;

[0013] Figure 5This is a schematic diagram of the sample dataset acquisition system in Embodiment 5 of the present invention;

[0014] Figure 6 This is a flowchart illustrating the multi-label classification method for perioperative patient data in Embodiment 6 of the present invention;

[0015] Figure 7 This is a schematic diagram of the classification model in Example 6;

[0016] Figure 8 This is a preferred flowchart of the multi-label classification method for perioperative patient data in Example 6;

[0017] Figure 9 This is a schematic diagram of the multi-label classification device for perioperative patient data in Embodiment 7 of the present invention;

[0018] Figure 10 This is a schematic diagram of the perioperative patient risk event prediction system in Embodiment 8 of the present invention. Detailed Implementation

[0019] 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.

[0020] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship 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 limitations on this invention.

[0021] In the description of this invention, unless otherwise specified and limited, it should be noted that the terms "installation", "connection" and "linking" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two components. They can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0022] Example 1

[0023] This embodiment discloses a perioperative patient data dimensionality reduction device, such as... Figure 1 As shown, the device includes:

[0024] The input module obtains the patient's original perioperative feature data containing multidimensional features, as well as the classification labels corresponding to the original perioperative feature data;

[0025] The initial dimensionality reduction module uses principal component analysis algorithm to reduce the dimensionality of the original perioperative feature data to obtain the first perioperative feature data.

[0026] The secondary dimensionality reduction module uses a genetic algorithm to reduce the dimensionality of the first perioperative feature data to obtain the perioperative feature data.

[0027] The output module outputs perioperative characteristic data.

[0028] In this embodiment, to better reflect the patient's perioperative status, improve the accuracy of subsequent classification and processing, and avoid the problem of difficulty in collecting and managing postoperative patient data, preferably, the original perioperative characteristic data includes the patient's preoperative and intraoperative indicators, such as preoperative blood pressure, heart rate, and blood lipids, and intraoperative heart rate, blood pressure, blood loss, and operation time. Unlike some existing classification prediction models that only include the patient's preoperative baseline condition without considering the specific circumstances during the operation, numerous studies have confirmed that intraoperative indicators such as heart rate, blood pressure, blood loss, and operation time are all related to the patient's postoperative condition. Therefore, the original perioperative characteristic data provided in this embodiment can improve the accuracy of subsequent model predictions of postoperative events and does not depend on postoperative patient indicator data.

[0029] In this embodiment, the classification label is used to characterize perioperative patient risk events, which preferably include, but are not limited to, unplanned readmissions and death.

[0030] In this embodiment, to improve the richness of the data, the indicator data includes categorical data and numerical data. Categorical data represents the indicator data by categories, such as intraoperative bleeding volume, which can be represented as high, medium, or low. Numerical data represents the indicator data by numerical values, such as blood pressure values.

[0031] In this embodiment, the original perioperative feature data can be data from patients with known perioperative risk events. Therefore, known perioperative risk events can be used as corresponding classification labels associated with the original perioperative feature data. Alternatively, the original perioperative feature data can be data from patients with unknown perioperative risk events, with experts assigning corresponding classification labels to the original perioperative feature data. The original perioperative feature data can have one, two, or more classification labels.

[0032] In this embodiment, after processing by the principal component analysis algorithm, the feature dimension of the first perioperative feature data is smaller than the feature dimension of the original perioperative feature data, and the initial population of the genetic algorithm is constructed based on the first perioperative feature data.

[0033] In this embodiment, to further reduce the dimensionality of the first perioperative feature data using a genetic algorithm, preferably, the secondary dimensionality reduction module includes:

[0034] The initial population setting unit is set up based on the first perioperative period characteristic data. The number of genes of an individual is less than or equal to the total number of features in the first perioperative period characteristic data. Multiple individuals form the initial population. The genes of an individual are the features in the first perioperative period characteristic data. Under the condition that the number of genes of an individual is less than or equal to the total number of features in the first perioperative period characteristic data, the number of genes of each individual can be randomly set.

[0035] The evolutionary iteration unit repeats the following process until the termination condition is met, and outputs the individual with the highest fitness when the termination condition is met: obtain the fitness of each individual in the current generation; select a portion of individuals from the current generation based on their fitness as individuals for the next generation; perform crossover and mutation operations on the individuals in the next generation.

[0036] In this embodiment, the termination condition is preferably, but not limited to, reaching a preset maximum number of evolutionary iterations, or the maximum fitness value of an individual no longer increases during the evolutionary iterations, or the increase in the maximum fitness value of an individual during the evolutionary iterations is less than the increase threshold. In each iteration, the fitness of individuals in the current generation is sorted from high to low, and the top-ranked individuals are selected as individuals for the next generation. Crossover mainly involves exchanging common gene loci between paired parents to obtain offspring, which are then used as individuals for the next generation.

[0037] In this embodiment, to ensure that the dimensionality-reduced perioperative feature data performs better in subsequent classification processing and improves classification accuracy, the preferred process for obtaining individual fitness is as follows: Obtain the original perioperative feature data and corresponding classification labels of multiple patients; perform dimensionality reduction processing on the multiple original perioperative feature data according to the individual's feature information to obtain multiple dimensionality-reduced samples consistent with the individual's features; divide the multiple dimensionality-reduced samples into a dimensionality-reduced training set and a dimensionality-reduced test set; construct a dimensionality-reduced multilayer perceptron; train the constructed dimensionality-reduced multilayer perceptron using the dimensionality-reduced training set to obtain a dimensionality-reduced classification prediction model; test the dimensionality-reduced classification prediction model using the dimensionality-reduced test set to obtain the model's accuracy, and use this accuracy as the individual's fitness.

[0038] Example 2

[0039] This embodiment discloses a system for acquiring perioperative patient sample datasets, such as... Figure 2 As shown, the perioperative patient sample dataset acquisition system includes:

[0040] The data acquisition module is used to acquire raw perioperative characteristic data and case records of multiple patients; case data is generally text data, including doctor's diagnosis, past medical history, postoperative follow-up records, etc.

[0041] The classification label set acquisition module obtains a set of classification labels based on multiple cases. The classification labels represent perioperative patient risk events.

[0042] The classification label association module is used to associate the patient’s original perioperative feature data with at least one classification label in the classification label set. Therefore, the original perioperative feature data corresponds to a classification label set, and the classification label set includes at least one classification label.

[0043] The perioperative patient data dimensionality reduction device provided in Example 1 performs dimensionality reduction processing on the original perioperative feature data of all patients to obtain the corresponding perioperative feature data.

[0044] The sample dataset acquisition module uses the patient's perioperative feature data as samples, associates the samples with the corresponding classification label set of the original perioperative feature data, and obtains the sample dataset of perioperative patients.

[0045] In this embodiment, preferably, the classification tag set acquisition module specifically performs the following: word segmentation processing on the patient's case to obtain at least one postoperative event result (postoperative event result is the perioperative patient risk event); similar word analogy is performed on the postoperative event results of multiple patients using a trained CBOW model to obtain multiple sets of similar postoperative event results; the sets of similar postoperative event results are matched with the event dictionary; classification tags that match the sets of similar postoperative event results are searched in the event dictionary; and multiple classification tags constitute a classification tag set.

[0046] In this embodiment, the Word2Vec CBOW Multi-Word Context Model is trained on a large medical corpus. The PKUSEG word segmentation tool (PKUSEG can segment words across multiple domains, including a separate model for the medical field) is used to segment the text information corresponding to the case set in this embodiment to obtain multiple postoperative event results. The event dictionary is preferably, but not limited to, the Chinese version of the ICD-11 event dictionary published by the World Health Organization, which contains many classification labels. Whether the set of similar postoperative event results matches the event dictionary is preferably, but not limited to, determined by semantic similarity. If the semantic similarity is greater than a preset similarity threshold, the two are considered to match; otherwise, they are not considered to match.

[0047] In this embodiment, preferably, to impute missing values ​​in the data and improve data quality, a missing value imputation device is also included. This device imputes missing values ​​in the patient's original perioperative characteristic data and inputs the imputed original perioperative characteristic data into a perioperative patient data dimensionality reduction device for dimensionality reduction. The missing value imputation device preferably, but not limited to, uses existing RandomForestRegressor imputation, MissForest imputation, Mean imputation, or Median imputation methods for imputation.

[0048] In this embodiment, more preferably, the missing data filling device performs missing data filling processing on the original perioperative feature data based on a Bayesian Gaussian process latent variable model.

[0049] In this embodiment, data imputation for missing values ​​inevitably introduces uncertainty into the original perioperative feature dataset. This embodiment uses a Bayesian Gaussian process latent variable model (BGPLVM) to impute missing values ​​for numerical features, specifically including:

[0050] First, the observed test data vector y is approximately calculated. * ∈R N×M The probability density p(y) * |Y)(where N is the total number of patient samples and M is the total number of features), and the observed value y * The variational distribution of the relevant latent variables is q(x) * Once the model parameters and latent variables are learned, BGPLVM can be used to estimate missing values. in It is a vector y * The values ​​that can be observed in These are the missing values ​​that need to be predicted. Given partially observed points y * This embodiment aims to reconstruct the lost parts. Missing data are filled by learning low-dimensional embeddings of observable variables on a small, complete dataset. BGPLVM is trained on the complete dataset D, introducing latent variables X and new test latent variables x. * As mentioned above A row vector representing the measurements of a single patient. Represents known observations. To represent missing values, y is obtained by maximizing the following probability density. * The corresponding hidden variable x * The Gaussian probability distribution.

[0051]

[0052] Next, by maximizing in To optimize the variational distribution q(x) using the variational lower bound. * ), keep apart from q(x) * All optimizations except for those for predicting missing values ​​remain unchanged. This invention employs the standard Gaussian process prediction method, while simultaneously taking the input x * Uncertainties were also taken into account, because x * There exists a distribution q(x) * Similar to the GP prediction format, in order to predict... This invention first predicts That is, with y * Corresponding implicit function value

[0053]

[0054] For x * Marginalization will produce a multivariate density that is not fully Gaussian dependent, but based on the squared exponential kernel, It can be analyzed and processed; in this invention, it is used... The mean and covariance are used; the mean provides an estimate of missing values, while the variance quantifies the uncertainty associated with the mean estimate. Using the BGPLVM model, the latent space and model hyperparameters learned on the training set are distributed to obtain the average estimate for each feature containing missing values.

[0055] In this embodiment, to facilitate data processing, it is further preferably included an encoding device for encoding the original perioperative characteristic data, and inputting the encoded data into the missing data filling device. The encoding device preferably, but is not limited to, using existing one-hot encoding rules.

[0056] In this embodiment, to facilitate data processing, it is further preferably included a normalization device for normalizing the original perioperative characteristic data after encoding processing, and inputting the normalized data into a missing data filling device. The normalization device preferably, but is not limited to, using the standard deviation normalization method for normalization processing.

[0057] Example 3

[0058] This embodiment provides a method for balancing sample datasets of perioperative patients, such as... Figure 3 As shown, the sample dataset balancing method includes:

[0059] Step S1: Oversample the minority class label samples in the perioperative patient sample dataset to obtain synthetic samples, and generate a corresponding synthetic label set for the synthetic samples. The sample dataset includes multiple samples and a classification label set corresponding to the samples. Each sample represents a patient's perioperative feature dataset, which can be the original perioperative feature data or the perioperative feature data obtained after dimensionality reduction of the original perioperative feature data in Example 1. The process of associating the classification labels of the samples has been described in detail in Example 1 and will not be repeated here.

[0060] Step S2: Add the synthetic samples and synthetic label set to the sample dataset to obtain a temporary sample dataset;

[0061] Step S3: Clean the samples in the temporary sample dataset to obtain a balanced sample dataset.

[0062] In this embodiment, SMOTE, SVM SMOTE, BorderlineSMOTE, K-Means SMOTE, or SMOTE-NC can be used to oversample minority class label samples in the perioperative patient sample dataset to obtain synthetic samples and generate corresponding synthetic label sets for the synthetic samples. Preferably, to improve the balance effect, the MLSMOTE algorithm is used to oversample minority class label samples in the perioperative patient sample dataset to obtain synthetic samples and generate corresponding synthetic label sets for the synthetic samples. The MLSMOTE algorithm, or Multi-label Synthetic Minority Over-sampling Technique (MLSMOTE), is commonly used to handle data imbalance problems in multi-label classification tasks. Its generation process includes: selecting minority class labels using the Imbalance Rate (IR); nearest neighbor search: once a sample belonging to the minority label is selected as a seed sample, its nearest neighbors are searched; feature set generation: after selecting a neighborhood, synthetic samples are obtained through interpolation; generation of synthetic label sets: a synthetic label set is needed for the generated synthetic samples.

[0063] In this embodiment, since oversampling algorithms such as MLSMOTE generate some noisy samples during the synthesis of minority class labeled samples, cleaning these noisy samples is essential. Therefore, step S3 is set to improve the quality of the sample dataset.

[0064] In this embodiment, preferably, in order to quickly determine the minority class label in the sample dataset, the ratio of the number of samples corresponding to each classification label to the total number of samples in the sample dataset is calculated. Classification labels with a ratio less than the ratio threshold are regarded as minority class classification labels, and classification labels with a ratio greater than or equal to the ratio threshold are regarded as majority class classification labels. The ratio threshold is preferably, but not limited to, less than 0.2.

[0065] In this embodiment, the number of samples required to generate for each minority class classification label is the oversampling rate of that label. To better determine the oversampling rate for each minority class classification label and ensure that the obtained balanced sample dataset performs better when applied to subsequent classification, preferably, in step S1, an oversampling rate is set for each minority class label based on a genetic algorithm, specifically including:

[0066] Step S11: Suppose the sample dataset includes W minority class labels. The oversampling rate of the samples with the W minority class labels is used as the W genes of an individual, where W is a positive integer. Each gene represents the oversampling rate of a minority class label. An initial population is constructed using multiple individuals. The initial population includes multiple initial individuals. The W gene values ​​of each initial individual are obtained by random selection. Preferably, a numerical range can be set for the oversampling rate of each minority class label. When constructing the initial population, values ​​within this numerical range are randomly selected as gene values. The numerical range can be set as needed.

[0067] Step S12, repeat the following evolutionary iteration process until the termination condition is met: obtain the fitness of each individual in the current generation population; select some individuals from the current generation population based on their fitness as individuals for the next generation population; perform crossover and mutation operations on the individuals in the next generation population.

[0068] Step S13: Output the individual with the highest fitness when the termination condition is met.

[0069] In this embodiment, the termination condition is preferably, but not limited to, reaching a preset maximum number of evolutionary iterations, or the maximum fitness value of an individual no longer increases during the evolutionary iterations, or the increase in the maximum fitness value of an individual during the evolutionary iterations is less than the increase threshold. In each iteration, the fitness of individuals in the current generation is sorted from high to low, and the top-ranked individuals are selected as individuals for the next generation of the population.

[0070] In this embodiment, to ensure that the obtained balanced sample dataset performs better when applied to subsequent classification, preferably, the process of obtaining the fitness of an individual is as follows:

[0071] The oversampling rate combination for minority class labels is obtained based on individual genetic information; the oversampling rate combination includes the oversampling rates for all minority class labels.

[0072] Based on the combination of minority class label oversampling rates, the minority class label samples in the sample dataset of perioperative patients are oversampled to obtain synthetic samples and synthetic label sets of synthetic samples. The synthetic samples and synthetic label sets are added to the sample dataset to obtain a balanced sample set. The balanced sample set is divided into a balanced training sample set and a balanced test sample set.

[0073] A balanced multilayer perceptron is constructed. The balanced multilayer perceptron is trained using a balanced training sample set to obtain a balanced prediction classification model. The balanced prediction classification model is tested using a balanced test sample set to obtain the accuracy of the balanced prediction classification model. This accuracy is used as the fitness of the individual.

[0074] In this embodiment, to effectively remove noisy samples and improve the quality of the sample set, preferably, step S3 involves cleaning each sample in the temporary sample dataset. The cleaning process includes:

[0075] Step S31: Select seed samples from the temporary sample dataset, select the k nearest neighbor samples of the seed samples, and form a nearest neighbor classification label set with the classification labels of the k nearest neighbor samples, where k is a positive integer; each sample in the temporary sample dataset can be selected as a seed sample in turn.

[0076] Step S32: Based on the nearest neighbor classification label set, predict the classification label set of the seed sample using Bayesian conditional probability to obtain the predicted classification label set of the seed sample.

[0077] Step S33: Determine whether the predicted classification label set of the seed sample is the same as its classification label set in the temporary sample dataset. If they are the same, retain the seed sample; if they are different, delete the seed sample and consider it a noise sample.

[0078] The above cleaning process directly predicts the classification label set of the seed sample based on the nearest neighbor classification label set of the seed sample using Bayesian conditional probability. The obtained predicted classification label set is then compared with the real classification label set of the seed sample in the temporary sample dataset. This process does not rely on the classifier's judgment but only on the data itself, reducing computational load and improving judgment efficiency and accuracy.

[0079] In this embodiment, more preferably, the specific process of selecting the k nearest neighbor samples of the seed sample in step S31 includes:

[0080] Obtain the Heterogeneous Value Difference Metric (HVDM) between the seed sample and all or part of the samples in the temporary sample dataset.

[0081] The out-of-class difference (HVDM) metric is corrected by adjusting the global imbalance weights of samples in a temporary sample dataset to obtain a corrected out-of-class difference metric.

[0082] The modified out-of-class difference metric is used to sort all samples in the temporary sample dataset with the seed sample, and the top k samples with the largest modified out-of-class difference metric values ​​are selected as the k nearest neighbors of the seed sample. Preferably, the modified out-of-class difference metric values ​​can be sorted from highest to lowest, and the top k samples with the largest modified out-of-class difference metric values ​​can be selected as the k nearest neighbors of the seed sample.

[0083] The above process of selecting the k nearest neighbors of the seed sample employs a weighted kNN (WkNN) method to improve the quality of the synthesized samples. If the true minority class label samples in the dataset are highly dispersed (spatially sparse), the synthesized minority class samples generated during the execution of algorithms such as MLSMOTE will still be scattered and sparse, lacking local balance. Directly using kNN for cleaning would likely discard sparse minority class samples and newly synthesized minority class samples from MLSMOTE, failing to establish appropriate classification boundaries. Therefore, a distance-weighted approach is needed to coordinate kNN cleaning. This means that when faced with sparsely distributed samples, instead of blindly deleting them, the local spatial density (i.e., the Hierarchical Value Difference (HVDM) metric and the global imbalance weights of the samples) is taken into account, preserving as many small samples as possible. kNN primarily relies on the label set of nearest neighbors for cleaning, making the calculation of nearest neighbor distances especially important when the data distribution is sparse. This is the main reason for adding distance weighting (i.e., using the globally imbalanced weights of samples in a temporary sample dataset to correct the out-of-class difference metric HVDM). WkNN cleans noisy samples by changing the calculation of the distance between nearest neighbors (correcting the out-of-class difference metric), which means it considers the influence of local density and uses the out-of-class difference metric to represent the distance between samples.

[0084] In this embodiment, more preferably, the formula for calculating the out-of-class difference (HVDM) metric between the seed sample and the samples in the temporary sample dataset is:

[0085]

[0086] Where f1 represents the feature vector of the seed sample; f2 represents the feature vector of any sample in the temporary sample dataset other than the seed sample; HVDM(f1,f2) represents the difference between the outliers of feature vectors f1 and f2; D(f1,f2) represents the distance between feature vectors f1 and f2; n represents the feature dimension of the samples in the temporary sample dataset; x represents the feature index; d x (f1, f2) represents the distance between eigenvectors f1 and f2 on feature x, d x (f1, f2) can be obtained using the following formula: C represents the number of categories for feature x when feature x is a categorical feature, and c represents the category index of feature x. This represents the number of samples in the temporary sample dataset where feature x belongs to feature vector f1 and the category feature of feature x is c; This represents the number of samples in the temporary sample dataset where feature x belongs to feature vector f2 and the category feature of feature x is c; This represents the number of samples in the temporary sample dataset whose feature x belongs to feature vector f1; σ represents the number of samples in the temporary sample dataset whose feature x belongs to feature vector f2; |f1-f2| represents the absolute value of the difference between feature vectors f1 and f2; x This represents the standard deviation of feature x in the temporary sample dataset.

[0087] In this embodiment, more preferably, the formula for calculating the corrected outlier difference metric between the seed sample and the samples in the temporary sample dataset is:

[0088]

[0089] Where f1 represents the feature vector of the seed sample; f2 represents the feature vector of any sample in the temporary sample dataset other than the seed sample; HVDM(f1,f2) represents the difference between the outlier values ​​of feature vectors f1 and f2; D W (f1,f2) represents the corrected outlier difference metric between feature vectors f1 and f2; n represents the feature dimension of the samples in the temporary sample dataset; IW represents the global imbalance weight of the sample with feature vector f2, IW = IR. nn / (IR + +IR - ),IR + IR represents the total imbalance rate of all minority class labels in the temporary sample dataset. - IR represents the total imbalance rate of all majority class labels in a temporary sample dataset. nn Let f be the total imbalance rate of all class labels in the class label set of the sample with feature vector f2.

[0090] In the above process of removing noisy samples, WkNN uses the Heterogeneous Value Difference Metric (HVDM) to calculate the distance, and corrects the HVDM using the globally imbalanced weights (IW) of the samples as weight coefficients. For temporary sample datasets, the more minority class labels the classification label set contains, the higher the IR (Integer Interference Rate). nn The larger the value, the larger the IW will be; for temporary sample datasets with sparse minority class label sample distribution and high imbalance rate, incorporating IW into the HVDM distance can improve the minority class sample density.

[0091] From the formula As can be seen, the weighting coefficients The value can be scaled by HVDM(f1,f2), and the more minority class labels in the nearest neighbor sample classification label set, the higher the weighting coefficient. The smaller the weighting coefficient, the higher the weighting coefficient of the nearest neighbor sample set. This is because a larger IW (indicating a higher percentage of minority class labels in the nearest neighbor sample label set) corresponds to a larger weighting coefficient of the nearest neighbor sample. The smaller the value, the more monotonically decreasing it becomes, which maintains the principle that, with a fixed feature dimension, the weighting coefficients of nearest neighbor samples remain constant. The scaling factor will vary depending on the majority and minority class labels contained in its label set; as the feature dimension increases, that is, as the sample distribution gradually becomes sparser, the scaling factor will also decrease.

[0092] It can be seen that WkNN can help filter nearest neighbor samples for samples with a large number of minority class labels in the label set by taking into account the distribution of labels in the nearest neighbor sample label set. This allows samples with more minority class labels in the label set to move closer to the seed sample, increasing the local minority class label density while reducing the majority class label density. The overall process is as follows: First, MLSMOTE is used to upsample the minority class labels, forming a relatively balanced temporary new sample set with the original samples. On this new sample set, the WkNN process is performed on each sample, which is to sort the k nearest neighbor samples based on weighted HVDM. Then, the label set of the seed sample is predicted based on the nearest neighbor samples. If the predicted label set is the same as the seed label set, the sample is retained; otherwise, it is deleted.

[0093] Example 4

[0094] This embodiment discloses a sample dataset balancing device for perioperative patients, such as... Figure 4 As shown, the sample dataset equalization device includes:

[0095] The sample synthesis module oversamples minority class label samples in the sample dataset of perioperative patients to obtain synthetic samples and generates corresponding synthetic label sets for the synthetic samples. The sample dataset includes multiple samples and the corresponding classification label sets of the samples.

[0096] The temporary sample dataset acquisition module adds the synthetic samples and synthetic label set to the sample dataset to obtain a temporary sample dataset.

[0097] The cleaning module cleans the samples in the temporary sample dataset to obtain a balanced sample dataset.

[0098] In this embodiment, preferably, the cleaning module includes:

[0099] The nearest neighbor sample acquisition unit selects a seed sample from the temporary sample dataset, selects k nearest neighbor samples of the seed sample, and the classification labels of the k nearest neighbor samples form a nearest neighbor classification label set, where k is a positive integer;

[0100] The predictive classification label set acquisition unit predicts the classification label set of the seed sample based on the nearest neighbor classification label set using Bayesian conditional probability, thereby obtaining the predicted classification label set of the seed sample.

[0101] The cleaning unit determines whether the predicted classification label set of the seed sample is the same as its classification label set in the temporary sample dataset. If they are the same, the seed sample is retained; otherwise, the seed sample is deleted.

[0102] In this embodiment, and more preferably, the specific process by which the nearest neighbor sample acquisition unit selects the k nearest neighbor samples of the seed sample includes:

[0103] Obtain the out-of-class difference (HVDM) measure between the seed sample and all or part of the samples in the temporary sample dataset;

[0104] The out-of-class difference (HVDM) metric is corrected by adjusting the global imbalance weights of samples in a temporary sample dataset to obtain a corrected out-of-class difference metric.

[0105] Sort all samples in the temporary sample dataset by the modified out-of-class difference metric with the seed sample, and select the top k samples with the largest modified out-of-class difference metric as the k nearest neighbors of the seed sample.

[0106] The equalization effect of the sample dataset equalization device provided in this embodiment was experimentally verified, and the results are as follows:

[0107]

[0108]

[0109] IR represents the Imbalance Rate of the sample set. The larger the IR, the more unbalanced the sample set. As can be seen from the experimental results in the table above, the maximum IR and average IR of the equalization device provided in this embodiment are the smallest, and the interval between the maximum value and the mean value of IR is narrowed, indicating that the balance of the sample set is better.

[0110] Example 5

[0111] This embodiment also discloses a perioperative patient sample dataset acquisition system. Compared with Embodiment 2, this embodiment adds a sample dataset balancing device, which performs sample balancing processing on the dimensionality-reduced sample dataset obtained in Embodiment 2. A schematic diagram of the device is shown below. Figure 5 As shown, it includes:

[0112] The data acquisition module is used to acquire raw perioperative characteristic data and case records of multiple patients;

[0113] The classification label set acquisition module obtains a set of classification labels based on multiple cases. The classification labels represent perioperative patient risk events.

[0114] The classification label association module is used to associate the patient's original perioperative characteristic data with at least one classification label in the classification label set.

[0115] The perioperative patient data dimensionality reduction device performs dimensionality reduction processing on the original perioperative feature data of all patients to obtain the corresponding perioperative feature data.

[0116] The sample dataset acquisition module uses the perioperative feature data of patients as samples, associates the samples with the corresponding classification label set of the original perioperative feature data, and obtains the sample dataset of perioperative patients.

[0117] It also includes the perioperative patient sample dataset balancing device provided in Example 4, which is used to balance the sample dataset.

[0118] In this embodiment, preferably, it also includes a missing value filling device for filling missing values ​​in the patient's original perioperative feature data, and inputting the filled original perioperative feature data into the perioperative patient data dimensionality reduction device for dimensionality reduction.

[0119] Example 6

[0120] This embodiment 6 discloses a multi-label classification method for perioperative patient data, such as... Figure 6 As shown, this multi-label classification method includes:

[0121] Step A: Obtain the feature data of the patients to be classified. The feature data of the patients to be classified are the feature data of perioperative patients, which may include multidimensional features. To improve the processability, reduce dimensionality, and improve the quality of the feature data of the patients to be classified, the feature data of the patients to be classified can be encoded, normalized, and then dimensionality reduced according to the feature dimensions of the samples output by the perioperative patient data dimensionality reduction device provided in Example 1. The dimensionality-reduced feature data of the patients to be classified is then input into the trained classification model.

[0122] Step B involves inputting the patient feature data to be classified into the trained classification model. The model outputs a classification result, which includes one or more classification labels and the classification confidence score for each label. The classification confidence score represents the probability that the patient feature data belongs to that label. The classification model includes a stacking-based classification ensemble model, a label association rule acquisition module, and a fusion module. The fusion module combines the classification matrix output by the classification ensemble model and the association rule matrix output by the label association rule acquisition module to obtain the classification result. The preferred fusion method is, but is not limited to, multiplying the classification matrix and the association rule matrix.

[0123] In an embodiment, preferably, the structural diagram of the classification model is as follows: Figure 7 As shown, the classification ensemble model includes a first multi-classification model, a second multi-classification model, a third multi-classification model, and a logistic regression model. The first multi-classification model, the second multi-classification model, and the third multi-classification model perform multi-label classification processing on the feature data of patients to be classified to obtain the first primary classification result, the second primary classification result, and the third primary classification result, respectively. The logistic regression model processes the first primary classification result, the second primary classification result, and the third primary classification result to obtain the classification matrix.

[0124] In this embodiment, preferably, the first multi-class classification model, the second multi-class classification model, and the third multi-class classification model are the Ranking-SVM model, the classification multilayer perceptron neural network model, and the Binary Relevance model, respectively. The Ranking-SVM model and the Binary Relevance model are relatively common basic models in stacking ensembles, and their use here ensures high reliability for model ensemble. The classification multilayer perceptron neural network model adopts a multilayer perceptron neural network structure (i.e., an MLP network structure), which can avoid overfitting problems and has low complexity.

[0125] In this embodiment, preferably, the method further includes a step of constructing a sample dataset of perioperative patients, such as... Figure 8 As shown, the steps for constructing a sample dataset of perioperative patients are preferably, but not limited to, using the system of Example 2 or Example 5.

[0126] In an embodiment, such as Figure 7 As shown, the training process of the classification ensemble model is as follows:

[0127] Construct a sample dataset of perioperative patients, in which each sample is associated with more than one classification label. Divide the sample dataset into a classification training set and a classification test set. The association of classification labels can be done manually.

[0128] Construct a classification ensemble model, namely the Stacking-based ensemble model mentioned above, which includes a first multi-class classification model, a second multi-class classification model, a third multi-class classification model, and a logistic regression model;

[0129] The classification ensemble model was trained using a classification training set and tested and validated using a classification test set. For validation, RandomizedSearchCV and GridSearchCV were used for cross-validation on the training set, and hyperparameters were selected based on the F1_Micro score.

[0130] In this embodiment, as Figure 7 As shown, preferably, the association rule acquisition module performs the following steps:

[0131] Obtain a sample dataset of perioperative patients, in which each sample is associated with one or more classification labels; the sample dataset is preferably, but not limited to, the perioperative patient sample dataset obtained in Example 2 or Example 5, which is the standard patient dataset.

[0132] Association rule mining is performed on the classification labels in the sample dataset to obtain the association rule matrix. The association rule matrix includes the association confidence between any two classification labels.

[0133] In this embodiment, as Figure 7 As shown, more preferably, when the number of classification labels in the sample dataset is small, specifically when it is less than a certain threshold, the FP-growth algorithm is used directly to mine association rules for the classification labels in the sample dataset. First, establish as follows: Figure 7 The classification label matrix shown has the labels in the first row and the patient ID in the first column. Then, the FP-growth algorithm is used to perform association rule analysis on the classification label matrix, outputting the association confidence score between any two classification labels. The association confidence score ranges from 0 to 1. Based on these association confidence scores, a system is established as follows: Figure 7 The association rule matrix shown has the first row and first column as category labels, and each element represents the association confidence between the category labels of the row and column containing that element. Figure 7 In this context, A(N-1) represents the association confidence between category label N and category label 1.

[0134] In this embodiment, preferably, when the number of classification labels in the sample dataset is large, the correlation patterns between the classification labels will be different. Directly performing association analysis will lead to frequent and complex itemset search processes, affecting the accuracy of association analysis. Specifically, when the number of classification labels is greater than or equal to a number threshold, the number threshold is preferably, but not limited to, 3, 4, or 5. The step of performing association rule mining on the classification labels in the sample dataset to obtain the association rule matrix specifically includes:

[0135] The classification labels in the sample dataset are clustered to obtain one or more clusters; the K-means++ algorithm is preferred, but not limited to, for clustering. Association rule mining is performed on the classification labels in each cluster to obtain an association rule sub-matrix. During fusion, the classification matrix is ​​divided into one or more sub-classification matrices according to the clustering results, with one sub-classification matrix corresponding to one cluster. The sub-classification matrix is ​​multiplied by the association rule sub-matrix corresponding to that cluster to obtain the classification sub-result for that cluster. All classification sub-results constitute the classification result.

[0136] In this embodiment, more preferably, the association rule submatrix is ​​obtained by mining association rules for the classification labels in each classification cluster using the FP-growth algorithm. The acquisition process is as follows: Figure 7 The process is consistent and has been explained in detail in the above preferred solutions, so it will not be repeated here.

[0137] Example 7

[0138] This embodiment discloses a multi-label classification device for perioperative patient data, such as... Figure 9 As shown, it includes:

[0139] The data acquisition module is used to acquire characteristic data of patients to be classified.

[0140] The classification module is used to input the feature data of patients to be classified into the trained classification model. The classification model outputs the classification result, which includes one or more classification labels and the classification confidence of each classification label. The classification model includes a stacking-based classification ensemble model, a label association rule acquisition module, and a fusion module. The fusion module is used to fuse the classification matrix output by the classification ensemble model and the association rule matrix output by the label association rule acquisition module to obtain the classification result.

[0141] In this embodiment, preferably, the classification ensemble model includes a first multi-classification model, a second multi-classification model, a third multi-classification model, and a logistic regression model; the first multi-classification model, the second multi-classification model, and the third multi-classification model respectively perform multi-label classification processing on the patient feature data to be classified to obtain a first primary classification result, a second primary classification result, and a third primary classification result; the logistic regression model processes the first primary classification result, the second primary classification result, and the third primary classification result to obtain a classification matrix.

[0142] In this embodiment, preferably, it further includes a classification ensemble model training module, which performs the following process:

[0143] A sample dataset of perioperative patients is constructed, in which each sample is associated with more than one classification label. The sample dataset is divided into a classification training set and a classification test set. Preferably, but not limited to, the sample dataset of perioperative patients is constructed using the system provided in Example 2 or Example 5.

[0144] Construct classification ensemble models; the classification ensemble models include first-class multi-class classification models, second-class multi-class classification models, third-class multi-class classification models, and logistic regression models;

[0145] The classification ensemble model is trained using a classification training set and tested and validated using a classification test set.

[0146] In this embodiment, the classification device constructs an ensemble classification model for multi-label postoperative events in the perioperative period, incorporating association rule analysis. Multiple postoperative risk events may occur. This study aims to predict these multiple postoperative events. A multi-label prediction model is constructed by integrating a Ranking-SVM model, a multilayer perceptron neural network model, and a Binary Relevance model. To further improve the model's stability and accuracy, association rules are incorporated into the prediction model for optimization.

[0147] Example 8

[0148] This embodiment discloses a perioperative patient risk event prediction system, such as... Figure 10 As shown, it includes:

[0149] The data acquisition module is used to acquire characteristic data of patients to be classified.

[0150] The classification module is used to input the feature data of the patients to be classified into the trained classification model. The classification model outputs the classification result, which includes one or more classification labels and the classification confidence of each classification label. Each classification label corresponds to a perioperative patient risk event.

[0151] The classification model includes a stacking-based classification ensemble model, a label association rule acquisition module, and a fusion module. The fusion module is used to fuse the classification matrix output by the classification ensemble model and the association rule matrix output by the label association rule acquisition module to obtain the classification result.

[0152] The conversion module converts the classification labels in the classification results into corresponding perioperative patient risk events to obtain risk prediction results.

[0153] In this embodiment, preferably, the classification ensemble model includes a first multi-classification model, a second multi-classification model, a third multi-classification model, and a logistic regression model; the first multi-classification model, the second multi-classification model, and the third multi-classification model respectively perform multi-label classification processing on the patient feature data to be classified to obtain a first primary classification result, a second primary classification result, and a third primary classification result; the logistic regression model processes the first primary classification result, the second primary classification result, and the third primary classification result to obtain a classification matrix.

[0154] In this embodiment, preferably, it further includes a classification ensemble model training module, which performs the following process:

[0155] A sample dataset of perioperative patients is constructed, in which each sample is associated with more than one classification label. The sample dataset is divided into a classification training set and a classification test set. Preferably, but not limited to, the sample dataset of perioperative patients is constructed using the system provided in Example 2 or Example 5.

[0156] Construct classification ensemble models; the classification ensemble models include first-class multi-class classification models, second-class multi-class classification models, third-class multi-class classification models, and logistic regression models;

[0157] The classification ensemble model is trained using a classification training set and tested and validated using a classification test set.

[0158] In this embodiment, during the acquisition of the system sample dataset provided in Embodiment 2 or Embodiment 5, prediction of perioperative risk events for patients (especially elderly surgical patients) is performed. Based on improvements to the missing and imbalanced datasets, association rule analysis is integrated to build a multi-label prediction model for postoperative events. Postoperative event labels are extracted based on patient case texts using the Word2Vec CBOW label extraction model. A large medical-related corpus is collected to train a medical word vector model, achieving the extraction of the postoperative event label set (i.e., the classification label set). Next, missing data is imputed using a Bayesian Gaussian process latent variable model, and imbalanced label data is processed using MLSMOTE, weighted kNN (WKNN), and genetic algorithms. Finally, a feature dimensionality reduction model is built by combining principal component analysis (PCA) and genetic algorithms to provide more relevant input for the classification ensemble model.

[0159] 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.

Claims

1. A system for acquiring perioperative patient sample datasets, characterized in that, include: The data acquisition module is used to acquire raw perioperative characteristic data and case records of multiple patients; The classification label set acquisition module obtains a set of classification labels based on multiple cases. The classification labels represent perioperative patient risk events. The classification label association module is used to associate the patient's original perioperative characteristic data with at least one classification label in the classification label set. The perioperative patient data dimensionality reduction device performs dimensionality reduction processing on the original perioperative feature data of all patients to obtain the corresponding perioperative feature data. The sample dataset acquisition module uses the perioperative feature data of patients as samples, associates the samples with the corresponding classification label set of the original perioperative feature data, and obtains the sample dataset of perioperative patients. It also includes a perioperative patient sample dataset balancing device for balancing the sample dataset; wherein the perioperative patient sample dataset balancing device includes: The sample synthesis module uses the MLSMOTE algorithm to oversample the minority class label samples in the sample dataset of perioperative patients to obtain synthetic samples, and generates a corresponding synthetic label set for the synthetic samples. The sample dataset includes multiple samples and the corresponding classification label set of the samples. The temporary sample dataset acquisition module adds the synthetic samples and synthetic label set to the sample dataset to obtain a temporary sample dataset. The cleaning module cleans the samples in the temporary sample dataset to obtain a balanced sample dataset. The cleaning process involves cleaning each sample in the temporary sample dataset, including: Step S31: Select seed samples from the temporary sample dataset, select the k nearest neighbor samples of the seed samples, and form a nearest neighbor classification label set with the classification labels of the k nearest neighbor samples, where k is a positive integer; each sample in the temporary sample dataset can be selected as a seed sample in turn. Step S32: Based on the nearest neighbor classification label set, predict the classification label set of the seed sample using Bayesian conditional probability to obtain the predicted classification label set of the seed sample. Step S33: Determine whether the predicted classification label set of the seed sample is the same as its classification label set in the temporary sample dataset. If they are the same, keep the seed sample; if they are different, delete the seed sample and consider it a noise sample. In step S31, the specific process of selecting the k nearest neighbor samples of the seed sample includes: Obtain the out-of-class difference (HVDM) measure between the seed sample and all or part of the samples in the temporary sample dataset; The corrected out-of-class difference (HVDM) is obtained by adjusting the global imbalance weights of the samples in the temporary sample dataset. The formula for calculating the corrected HVDM between the seed sample and the temporary sample dataset is as follows: D W (f1,f2)=e (1-IW)n ·HVDM(f1,f2); Where f1 represents the feature vector of the seed sample; f2 represents the feature vector of any sample in the temporary sample dataset other than the seed sample; HVDM(f1,f2) represents the difference between the outlier values ​​of feature vectors f1 and f2; D W (f1,f2) represents the corrected outlier difference metric between feature vectors f1 and f2; n represents the feature dimension of the samples in the temporary sample dataset; IW represents the global imbalance weight of the sample with feature vector f2, IW = IR. nn / (IR + +IR - ),IR + IR represents the total imbalance rate of all minority class labels in the temporary sample dataset. - IR represents the total imbalance rate of all majority class labels in a temporary sample dataset. nn Let f be the total imbalance rate of all class labels in the class label set of the sample with feature vector f2; Sort all samples in the temporary sample dataset by the modified out-of-class difference metric with the seed sample, and select the top k samples with the largest modified out-of-class difference metric as the k nearest neighbors of the seed sample.

2. The perioperative patient sample dataset acquisition system as described in claim 1, characterized in that, The perioperative patient data dimensionality reduction device includes: The input module obtains the patient's original perioperative feature data containing multidimensional features, as well as the classification labels corresponding to the original perioperative feature data; The initial dimensionality reduction module uses principal component analysis algorithm to reduce the dimensionality of the original perioperative feature data to obtain the first perioperative feature data. The secondary dimensionality reduction module uses a genetic algorithm to reduce the dimensionality of the first perioperative feature data to obtain the perioperative feature data. The output module outputs perioperative characteristic data.

3. The perioperative patient sample dataset acquisition system as described in claim 2, characterized in that, The original perioperative characteristic data includes patient index data before and during surgery.

4. The perioperative patient sample dataset acquisition system as described in claim 2 or 3, characterized in that, The secondary dimensionality reduction module includes: The initial population setting unit is set up based on the first perioperative period characteristic data. The number of genes in each individual is less than or equal to the total number of features in the first perioperative period characteristic data. Multiple individuals form the initial population. The evolutionary iteration unit repeats the following process until a termination condition is met, and outputs the individual with the highest fitness when the termination condition is met: Obtain the fitness of each individual in the current generation population; select a portion of individuals from the current generation population based on their fitness as individuals for the next generation population; perform crossover and mutation operations on the individuals in the next generation population.

5. The perioperative patient sample dataset acquisition system as described in claim 4, characterized in that, The process of obtaining an individual's fitness: The original perioperative feature data and corresponding classification labels of multiple patients were obtained. The original perioperative feature data were then dimensionality-reduced according to the individual's feature information to obtain multiple dimensionality-reduced samples consistent with the individual's features. Multiple dimensionality reduction samples are divided into a dimensionality reduction training set and a dimensionality reduction test set; Construct a dimension-reduced multilayer perceptron neural network; The constructed dimensionality-reduced multilayer perceptron is trained using a dimensionality-reduced training set to obtain a dimensionality-reduced classification prediction model; the dimensionality-reduced classification prediction model is tested using a dimensionality-reduced test set to obtain the accuracy of the dimensionality-reduced classification prediction model, and the accuracy is used as the fitness of an individual.

6. The perioperative patient sample dataset acquisition system as described in claim 1, characterized in that, The specific execution of the category tag set acquisition module is as follows: The patient's case is segmented to obtain at least one postoperative event result. The postoperative event results of multiple patients are compared with similar words using a trained CBOW model to obtain multiple sets of similar postoperative event results. The sets of similar postoperative event results are matched with the event dictionary. The event dictionary is searched for classification labels that match the sets of similar postoperative event results. Multiple classification labels constitute a classification label set.

7. The perioperative patient sample dataset acquisition system as described in claim 1 or 6, characterized in that, It also includes a missing value filling device, which is used to fill in the missing values ​​in the patient's original perioperative feature data, and input the filled original perioperative feature data into the perioperative patient data dimensionality reduction device for dimensionality reduction processing.

8. The perioperative patient sample dataset acquisition system as described in claim 7, characterized in that, The missing data filling device performs missing data filling on the original perioperative feature data based on a Bayesian Gaussian process latent variable model.

9. The perioperative patient sample dataset acquisition system as described in claim 8, characterized in that, Also includes: The encoding device is used to encode the original perioperative characteristic data.

10. The perioperative patient sample dataset acquisition system as described in claim 9, characterized in that, It also includes: The normalization device is used to normalize the original perioperative characteristic data.

Images