A method and system for predicting malnutrition in hemodialysis patients based on AI technology

Abstract

The present invention provides a method and system for predicting malnutrition in hemodialysis patients based on AI technology. The system includes a data collection module, a data preprocessing module, a data standardization module, a dimensionality reduction module, and a nutritional status prediction module. The method includes the following steps: S1: Collect relevant attributes of hemodialysis patients to obtain a hemodialysis patient dataset; S2: Preprocess the hemodialysis patient dataset; S3: Standardize the hemodialysis patient dataset; S4: Feature selection and model construction methods; S5: Model training and generation. The method of the present invention can comprehensively analyze patients by using AI technology combined with multi-dimensional data, can more comprehensively predict the nutritional status of patients, and the evaluation results are comprehensive, avoiding the problems of predicting the nutritional status of hemodialysis patients based on only a few indicators in the prior art, the indicators being too single, and the prediction results being incomplete.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular, to a method and system for predicting malnutrition in hemodialysis patients based on AI technology. Background Art

[0002] Hemodialysis is a common method for treating chronic kidney diseases. However, long-term hemodialysis treatment may lead to malnutrition in patients, which in turn affects their quality of life and prognosis. Therefore, it is very necessary to monitor and predict the nutritional status of hemodialysis patients.

[0003] In the prior art, the nutritional status of hemodialysis patients is usually predicted based on a few indicators. The indicators are too single to comprehensively predict the nutritional status of patients, and the evaluation results lack comprehensiveness. At the same time, some indicators may be easily affected by other factors, resulting in inaccurate prediction of the nutritional status of patients. Summary of the Invention

[0004] The purpose of the present invention is to solve the deficiencies in the prior art and propose a method and system for predicting malnutrition in hemodialysis patients based on AI technology.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A system for predicting malnutrition in hemodialysis patients based on AI technology, comprising a data collection module, a data preprocessing module, a data standardization module, a dimensionality reduction module, and a nutritional status prediction module;

[0007] The data collection module is used to collect relevant data of hemodialysis patients;

[0008] The data preprocessing module is used to preprocess the data in the hemodialysis patient dataset;

[0009] The data standardization module is used to standardize the data in the hemodialysis patient dataset;

[0010] The dimensionality reduction module is used to perform clustering analysis on the data in the hemodialysis patient dataset, remove features with less relevance to malnutrition, and perform iterative clustering grouping to obtain features highly relevant to malnutrition;

[0011] The nutritional status prediction module is used to establish a prediction model and predict whether the patient is in a state of malnutrition.

[0012] A method for predicting malnutrition in hemodialysis patients based on AI technology, comprising the following steps:

[0013] S1: Collect relevant attributes of hemodialysis patients to obtain a hemodialysis patient dataset;

[0014] The data collection module collects relevant attributes of hemodialysis patients according to actual needs;

[0015] Specifically, the data collection module collects relevant attributes of hemodialysis patients from general demographic information, hemodialysis data, laboratory test indicators, and diagnosis results according to actual needs;

[0016] The general demographic information includes gender, age, body mass index (BMI), ethnicity, marital status, educational level, whether combined with hypertension, whether combined with diabetes, type of primary disease, post-dialysis systolic blood pressure, post-dialysis diastolic blood pressure, presence or absence of hematoma after dialysis, etc.;

[0017] The hemodialysis data includes information such as hemodialysis time, 24-hour urine volume, blood flow rate, ultrafiltration volume, type of vascular access, whether the access is infected, etc.; The types of vascular access include arteriovenous fistula, arteriovenous valve, central venous catheter, peritoneal dialysis catheter, etc.;

[0018] The laboratory test indicators include hemoglobin, total protein, prealbumin, blood glucose 2 hours after meal, serum creatinine, urea nitrogen, uric acid, etc.;

[0019] The diagnosis result of the hemodialysis patient is a specific disease;

[0020] After collecting the relevant attributes of the hemodialysis patient, a hemodialysis patient dataset is obtained. The hemodialysis patient dataset contains features and corresponding feature values; The features are various attributes of the collected patients, and the feature values are the specific values corresponding to various attributes.

[0021] S2: Preprocess the hemodialysis patient dataset;

[0022] The data preprocessing module preprocesses the data in the hemodialysis patient dataset by handling missing data and outliers, including the following sub-steps:

[0023] S21: Handle missing data;

[0024] Judge whether there is missing data in the hemodialysis patient dataset. If there is missing data, handle the missing data; if not, go to step S22;

[0025] S211: Judge whether there is missing data;

[0026] Judge whether there is missing data in the hemodialysis patient dataset;

[0027] Specific judgment methods include counting missing values, visualizing data distribution, etc.;

[0028] Statistical missing values: Use the isnull or isna function to identify missing data in the dataset;

[0029] Visualize data distribution: Present the data distribution in the dataset by plotting charts to identify missing data;

[0030] After the judgment, use the sum function to summarize the missing data to obtain the total number of missing data in the dataset;

[0031] After the judgment, use the sum function to obtain the number of missing data;

[0032] If there is no missing data in the hemodialysis patient dataset, go to step S22; if there is missing data, go to step S212;

[0033] S212: Evaluate the impact of missing data;

[0034] Evaluate the impact of missing data on the overall data distribution. Specific evaluation methods include comparing statistical indicators, calculating correlations, and comparing data distributions, etc.;

[0035] Specifically, comparing statistical indicators means calculating the statistical indicators of the hemodialysis patient dataset containing the characteristics corresponding to the missing data and the statistical indicators of the hemodialysis patient dataset after removing the characteristics corresponding to the missing data according to the missing data detected in step S211. The statistical indicators include mean, median, standard deviation, etc.; Compare the relative difference or percentage difference of the calculated statistical indicators. If the difference is small, it means that the missing data has little impact on the overall data distribution; otherwise, it means that the missing data has a large impact on the overall data distribution;

[0036] Calculating correlations means calculating the Pearson correlation coefficients between various characteristics in the hemodialysis patient dataset through the corr function to obtain a correlation matrix, and extracting the correlation coefficients between the characteristics corresponding to the missing data and other characteristics in the correlation matrix. If the correlation coefficient is small, it means that the missing data has little impact on the overall data distribution; otherwise, it means that the missing data has a large impact on the overall data distribution;

[0037] Comparing data distributions means plotting the data distributions of the hemodialysis patient dataset containing the characteristics corresponding to the missing data and the hemodialysis patient dataset after removing the characteristics corresponding to the missing data through histograms, and comparing the data distribution situations in the above two cases through box plots. If the data distributions are similar, the missing data has little impact on the overall data distribution; otherwise, the missing data has a large impact on the overall data distribution;

[0038] S213: Process the missing data;

[0039] Process the missing data according to the evaluation result of step S212. When the number of missing data is small and the missing data has little impact on the overall data distribution, the filling method or deletion method is adopted;

[0040] Furthermore, the filling methods include mean filling, median filling, mode filling, etc. Select a suitable filling method according to the data type and data distribution characteristics;

[0041] The deletion method refers to deleting the rows or columns containing missing data in the hemodialysis patient dataset;

[0042] When the number of missing data is large and the missing data has a great impact on the overall data distribution, deep learning methods, multiple imputation methods, etc. are used for data filling;

[0043] When there is missing data in continuous data, interpolation filling is performed using interpolation methods such as linear interpolation or polynomial interpolation;

[0044] S22: Process outliers;

[0045] Judge whether there are outliers in the hemodialysis patient dataset. If there are outliers, process the outliers; if not, go to step S23;

[0046] S221: Identify outliers;

[0047] Identify whether there are outliers in the hemodialysis patient dataset. The specific identification methods include the box plot method, Z-score comparison detection method, etc.;

[0048] Specifically, the box plot method means drawing a box plot, which contains the median, upper quartile, lower quartile, upper limit, and lower limit of the data in the hemodialysis patient dataset; if there are data points in the box plot above the upper limit or below the lower limit, the data corresponding to the data points is an outlier, otherwise it is a non-outlier;

[0049] The Z-score comparison detection method means calculating the mean and standard deviation of the data in the hemodialysis patient dataset, and calculating the standard score Z-score of each data point through the Z-score normalization formula according to the calculated mean, standard deviation and the data values in the hemodialysis patient dataset. Set a threshold in advance. If the standard score Z-score of the data point is greater than the threshold, it is an outlier, otherwise it is a non-outlier; The Z-score normalization formula is as follows:

[0050] ;

[0051] where, Z-score is the standard score, x is the data value in the hemodialysis patient dataset, where \(\bar{x}\) is the mean value and \(s\) is the standard deviation;

[0052] S222: Process the outliers;

[0053] Process the outliers by deleting or replacing them;

[0054] Specifically, replacing the outliers means replacing the outliers with specific values in the hemodialysis patient dataset, such as the median, mean, etc.;

[0055] S3: Standardize the hemodialysis patient dataset;

[0056] The data standardization module standardizes the data in the hemodialysis patient dataset through data standardization;

[0057] Standardize the diagnosis results of hemodialysis patients into mild malnutrition, moderate malnutrition, severe malnutrition, and normal nutrition status;

[0058] Furthermore, the hemodialysis patient dataset contains features of two types of tests: qualitative tests and quantitative tests;

[0059] For the features of qualitative tests, represent the test results, i.e., the feature values, numerically. Represent positive with the numerical value 1 and negative with the numerical value 0;

[0060] For the features of quantitative tests, determine the normal range of the test results of this feature in advance, compare the actual test results with the normal range, and calculate the difference value for each feature;

[0061] The normal range of the test results includes a single value and an interval;

[0062] If the normal range of the test results is a single value, the difference value = |actual test result - normal range of the test result|;

[0063] If the test result is an interval, calculate the central value of the interval. The central value of the interval is the average of the upper and lower limits of the interval, and the difference value = |actual test result - central value of the interval|;

[0064] Obtain the mean value and standard deviation of the difference values based on the calculated difference values of the features, and perform data standardization processing through the Z-score standardization formula;

[0065] At this time, in the Z-score standardization formula, Z-score is the standard score, \(x\) is the difference value, \(\bar{x}\) is the mean value of the difference values, and \(s\) is the standard deviation of the difference values of this feature.

[0066] S4: Feature selection and model construction method;

[0067] The dimensionality reduction module combines data algorithms to determine the features related to malnutrition in the hemodialysis patient dataset;

[0068] S41: Initially remove features with relatively small differences;

[0069] Perform clustering analysis on the data in the hemodialysis patient dataset through clustering algorithms;

[0070] The clustering algorithms include k-means clustering, hierarchical clustering, and DBSCAN clustering;

[0071] Record the data of each patient in the hemodialysis patient dataset as a sample, and run the above three clustering algorithms on each sample respectively. Each clustering algorithm assigns each sample to the corresponding cluster;

[0072] The method for setting the number of clusters for each clustering algorithm includes that the number of clusters for k-means clustering and hierarchical clustering is determined by the silhouette coefficient to find the optimal number of clusters. DBSCAN clustering does not require pre-setting the number of clusters, but requires setting the eps and min_samples parameters, and these parameters are determined by observing the data point distance graph;

[0073] Calculate the distribution of each feature in each cluster, and measure the distribution difference of each feature in different clusters through variance;

[0074] Calculate the variance of each feature included in each cluster according to the variance formula, sort the calculated variances of each feature in descending order, and draw a curve after variance sorting. In the curve, the horizontal axis is the feature name and the vertical axis is the variance of the corresponding feature;

[0075] Use the kneedle algorithm to obtain the slope change of the curve, detect the inflection point of the curve according to the slope change. The features before the first inflection point of the detected curve indicate that they show greater differences between different clusters and can better distinguish different clusters;

[0076] Remove the features in the dataset that do not show significant differences between different clusters, that is, the features after the first inflection point. Take the intersection of the features with greater differences obtained by the three clustering algorithms as the features after preliminary dimensionality reduction. The entire experimental process is iterative;

[0077] The conditions for ending the iteration can be determined according to the following three points:

[0078] (1) The curve of feature variance sorting is stable: When the variance sorting curve calculated by the kneedle algorithm does not change significantly in consecutive iterations at the position of the first inflection point, it indicates that the features that can distinguish different clusters have stabilized, and the iteration stops;

[0079] (2)Convergence of the feature intersection of the three clustering algorithms: Compare the significantly different feature sets obtained by K-means clustering, hierarchical clustering, and DBSCAN. If the feature intersection gradually decreases after each iteration and the feature set of the final intersection no longer changes significantly, it indicates that a stable preliminary dimension-reduced feature has been found and the iteration can be stopped;

[0080] (3)The clustering performance metrics reach the expected values: Evaluate the clustering effect after each iteration. The evaluation metrics include the silhouette coefficient, the purity of the nutritional status within the cluster, and the accuracy, etc. If these metrics do not improve significantly after multiple iterations, it indicates that the current feature set is already effective enough and there is no need to further remove features;

[0081] If at least two of the above conditions are met, it can be considered that the iteration process of S41 has reached the state of optimal feature screening;

[0082] S42: Generate a model through data clustering calculation;

[0083] After S41 preliminarily removes features with less difference to obtain new features, retain the new features for each cluster divided by each clustering algorithm in S41. For each sample, vote according to the clustering results of K-means, hierarchical clustering, and DBSCAN, and select the most frequent cluster as the final result; Use the cluster labels of the three algorithms as new features and add them to the dataset to further improve the clustering effect;

[0084] Assign a nutritional status label to each patient in each cluster; The nutritional status includes normal nutrition and malnutrition; Set the malnutrition status label to 1, and the malnutrition status includes mild, moderate, and severe malnutrition; Set the normal nutrition status label to 0;

[0085] For the patients in each cluster, count according to their nutritional status and calculate the proportion of the number of people in each nutritional status in the cluster;

[0086] Assign a nutritional status label to each cluster, and this label is the nutritional status with the highest proportion in each cluster; This step defines a representative nutritional status for each cluster, and these labels can help understand the differences between clusters;

[0087] Finally, generate a clustering model. The key metrics for evaluating the clustering model include the silhouette coefficient, the purity of the nutritional status within the cluster, and the accuracy. If the silhouette coefficient is above 0.5, the purity value of the nutritional status within the cluster is above 80%, and the accuracy is above 85%, it indicates that this clustering model can be used to identify the nutritional status of new data;

[0088] S43: Iteratively cluster and group by randomly combining features;

[0089] S43 is an optimization of step S42, in which representative nutritional status labels have been assigned to each cluster and the nutritional status ratios of each cluster have been calculated; in order to further optimize the clustering model and feature selection, S43 performs multiple iterative clustering analyses by randomly combining features;

[0090] From the features preliminarily screened in step S41, randomly select different combinations of feature subsets, and use these subsets to re-cluster the data; in each clustering, re-apply the k-means clustering, hierarchical clustering, and DBSCAN clustering methods, repeat the previous cluster number determination and clustering steps, and perform result fusion; in each iteration, maintain the same process as in S42, that is, assign nutritional status labels to each cluster and calculate the nutritional status ratios in each cluster;

[0091] By performing iterative clustering through multiple random combinations of different features, the aim is to observe the influence of different feature combinations on the clustering results and find the feature combination that can best distinguish the nutritional status;

[0092] After each clustering, evaluate the effect of the clustering model to ensure that the model can effectively distinguish different nutritional statuses, calculate the silhouette coefficient of each cluster, the purity and accuracy of the nutritional status within the cluster, and ensure that the clustering results can effectively distinguish the nutritional status;

[0093] When similar feature combinations obtain similar clustering results in multiple iterations, and indicators such as the purity of the nutritional status within the cluster, the silhouette coefficient, and the accuracy reach the set thresholds, that is, the silhouette coefficient continuously stabilizes above 0.5, the purity value remains above 80%, the accuracy remains above 85%, and there is no obvious improvement in further iterations, the iteration can be stopped;

[0094] Through the iterative optimization of S43, the feature set can be further reduced, and finally the most discriminative feature combination can be selected to improve the recognition accuracy of the clustering model for new data;

[0095] During the entire analysis process, the features highly correlated with malnutrition found in existing research remain unchanged;

[0096] The dimensionality reduction method proposed in the present invention starts from the data structure, captures the relationship between features and specific subgroups (clusters), and helps to retain the key features when distinguishing these subgroups; traditional dimensionality reduction methods usually reduce the dimension by transforming the data or based on linear algebra methods, without involving the clustering process;

[0097] S5: Model training and generation;

[0098] Based on the feature selection and clustering analysis of S4, the nutritional status prediction module predicts the malnutrition status through the finally screened features and the clustering model, including the following sub-steps:

[0099] S51: Prepare training data;

[0100] Use the feature selection results in step S4 to determine the feature set. These features are the most effective features for distinguishing the malnutrition status that are retained after multiple clustering iterations; each sample has been labeled with a nutrition status label (0 indicates normal nutrition, 1 indicates malnutrition), that is, mild, moderate, and severe malnutrition are classified as 1, and normal nutrition is 0;

[0101] S52: Use the clustering model to predict whether a patient is in a malnutrition status;

[0102] Through the previous clustering models (S42, S43), patient samples have been assigned to different clusters, and a nutrition status label has been assigned to each cluster. Use this clustering structure to build a prediction model.

[0103] S53: Train using the clustering model;

[0104] Use the filtered feature set (the final feature combination in step S43) as the input data.

[0105] According to the features of each new sample, use the trained clustering model to assign the new sample to a cluster. Use the previously calculated nutrition status proportion for each cluster to assign a predicted nutrition status label to the new sample. Specifically, the new sample will be assigned to a certain cluster, and the nutrition status proportion in the cluster determines the label (0 or 1) of the sample. Evaluate the prediction accuracy using the known nutrition status labels (0 or 1 labeled in the training set). The performance of the model can be evaluated by criteria such as accuracy, precision, recall, and F1 score;

[0106] S54: Model evaluation and optimization;

[0107] The evaluation metrics include accuracy, precision, recall, and F1 score; if the initial training result is not ideal, adjust the parameters of the clustering algorithm (such as eps and min_samples in DBSCAN, or the number of clusters in K-means), or recombine the features during the feature selection process;

[0108] S55: Use the trained clustering model to predict new data;

[0109] For each new patient, after inputting their features, the clustering model assigns the patient to a certain cluster, and based on the nutrition status label of the cluster, predicts whether the patient is in a malnutrition status. Through the prediction results of the model, patients with malnutrition can be intervened in advance, and corresponding nutritional support or other treatments can be provided.

[0110] Compared with the prior art, the beneficial effects of the present invention are:

[0111] The method of the present invention uses artificial intelligence (AI) technology to comprehensively analyze patients in combination with multi-dimensional data, which can more comprehensively predict the nutritional status of patients. The evaluation results are comprehensive, avoiding the problems of predicting the nutritional status of hemodialysis patients based on only a few indicators in the prior art, with overly single indicators and incomplete prediction results.

[0112] The method of the present invention avoids the problem of limitations of a single algorithm. By compensating each other among multiple algorithms, comparing through multiple channels, and corroborating each other, the accuracy of obtaining influencing factors of malnutrition is improved. There may be differences or deviations in the results of multiple algorithms. By cross-validating and comparing the results of multiple algorithms, they can compensate each other to ensure the accuracy of obtaining influencing factors of malnutrition. BRIEF DESCRIPTION OF THE DRAWINGS

[0113] Figure 1 It is a flowchart of the steps of a method for predicting malnutrition in hemodialysis patients based on AI technology of the present invention;

[0114] Figure 2 It is a diagram for realizing the module functions of a method for predicting malnutrition in hemodialysis patients based on AI technology of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0115] To further understand the purpose, structure, features, and functions of the present invention, the following is a detailed description in conjunction with embodiments.

[0116] As Figure 1 and Figure 2 shown, a system for predicting malnutrition in hemodialysis patients based on AI technology includes a data collection module, a data preprocessing module, a data standardization module, a dimensionality reduction module, and a nutritional status prediction module;

[0117] The data collection module is used to collect relevant data of hemodialysis patients;

[0118] The data preprocessing module is used to preprocess the data in the hemodialysis patient dataset;

[0119] The data standardization module is used to standardize the data in the hemodialysis patient dataset;

[0120] The dimensionality reduction module is used to perform clustering analysis on the data in the hemodialysis patient dataset, remove features with less relevance to malnutrition, and perform iterative clustering grouping to obtain features highly relevant to malnutrition;

[0121] The nutritional status prediction module is used to establish a prediction model and predict whether the patient is in a state of malnutrition.

[0122] A method for predicting malnutrition in hemodialysis patients based on AI technology, comprising the following steps:

[0123] S1: Collect relevant attributes of hemodialysis patients to obtain a hemodialysis patient dataset;

[0124] The data collection module collects relevant attributes of hemodialysis patients according to actual needs;

[0125] Specifically, the data collection module collects relevant attributes of hemodialysis patients from general demographic information, hemodialysis data, laboratory test indicators, and diagnosis results according to actual needs;

[0126] The general demographic information includes gender, age, body mass index (BMI), ethnicity, marital status, educational level, whether combined with hypertension, whether combined with diabetes, type of primary disease, post-dialysis systolic blood pressure, post-dialysis diastolic blood pressure, presence or absence of hematoma after dialysis, etc.;

[0127] The hemodialysis data includes information such as hemodialysis time, 24-hour urine volume, blood flow rate, ultrafiltration volume, type of vascular access, whether the access is infected, etc.; The type of vascular access includes arteriovenous fistula, arteriovenous valve, central venous catheter, peritoneal dialysis catheter, etc.;

[0128] The laboratory test indicators include hemoglobin, total protein, prealbumin, 2-hour postprandial blood glucose, serum creatinine, urea nitrogen, uric acid, etc.;

[0129] The diagnosis result of the hemodialysis patient is a specific disease;

[0130] After collecting relevant attributes of hemodialysis patients, a hemodialysis patient dataset is obtained. The hemodialysis patient dataset contains features and corresponding feature values; The features are various attributes of the collected patients, and the feature values are specific values corresponding to various attributes;

[0131] Furthermore, determine the inclusion and exclusion criteria;

[0132] The inclusion and exclusion criteria include inclusion criteria and exclusion criteria;

[0133] The inclusion criteria refer to the conditions that must be met in patient research or treatment, including the following:

[0134] (1) Age ≥ 18 years old;

[0135] (2) Diagnosed with end-stage renal disease;

[0136] (3) The patient undergoes regular dialysis (2 - 4 times / week) in this blood purification center for ≥ 3 months;

[0137] (4) Have complete dialysis data for not less than 1 month;

[0138] The exclusion criteria refer to the conditions that should be excluded in patient research or treatment, including the following:

[0139] (1) The patient has co-existing cardiovascular and cerebrovascular diseases, infections, malignant tumors, gastrointestinal and liver diseases, pulmonary tuberculosis, arrhythmias, etc.;

[0140] (2) The clinical data and experimental data are incomplete;

[0141] According to the inclusion and exclusion criteria, the patient data that meet the requirements in the hemodialysis patient dataset are screened out;

[0142] By screening out the patient data that meet the requirements through the inclusion and exclusion criteria, the accuracy and credibility of the data can be improved, unnecessary data processing and analysis can be reduced, and the efficiency of the research can be enhanced.

[0143] S2: Preprocess the hemodialysis patient dataset;

[0144] The data preprocessing module preprocesses the data in the hemodialysis patient dataset by handling missing data, handling outliers, and feature selection, including the following sub-steps:

[0145] S21: Handle missing data;

[0146] Judge whether there is missing data in the hemodialysis patient dataset. If there is missing data, handle the missing data; if not, go to step S22;

[0147] S211: Judge whether there is missing data;

[0148] Judge whether there is missing data in the hemodialysis patient dataset;

[0149] Specific judgment methods include counting missing values, visualizing data distribution, etc.;

[0150] Count missing values: Use the isnull or isna function to identify the missing data in the dataset;

[0151] Visualize data distribution: Present the data distribution in the dataset by drawing charts to identify missing data;

[0152] After the judgment is completed, summarize the missing data through the sum function to obtain the total number of missing data in the dataset;

[0153] After the judgment is completed, obtain the number of missing data through the summation function sum;

[0154] If there is no missing data in the hemodialysis patient dataset, go to step S22; if there is missing data, go to step S212;

[0155] S212: Evaluate the impact of missing data;

[0156] Evaluate the impact of missing data on the overall data distribution. Specific evaluation methods include comparing statistical indicators, calculating correlations, comparing data distributions, etc.;

[0157] Specifically, comparing statistical indicators means calculating the statistical indicators of the hemodialysis patient dataset containing the features corresponding to the missing data and the statistical indicators of the hemodialysis patient dataset after removing the features corresponding to the missing data based on the missing data detected in step S211. The statistical indicators include mean, median, standard deviation, etc.; compare the relative difference or percentage difference of the calculated statistical indicators. If the difference is small, it indicates that the missing data has little impact on the overall data distribution; otherwise, it indicates that the missing data has a large impact on the overall data distribution;

[0158] Calculating correlations means calculating the Pearson correlation coefficients between various features in the hemodialysis patient dataset through the corr function to obtain a correlation matrix, and extracting the correlation coefficients between the features corresponding to the missing data and other features in the correlation matrix. If the correlation coefficient is small, it indicates that the missing data has little impact on the overall data distribution; otherwise, it indicates that the missing data has a large impact on the overall data distribution;

[0159] Comparing data distributions means plotting the data distributions of the hemodialysis patient dataset containing the features corresponding to the missing data and the hemodialysis patient dataset after removing the features corresponding to the missing data through histograms, and comparing the data distribution situations in the above two cases through box plots. If the data distributions of the two are similar, it indicates that the missing data has little impact on the overall data distribution; otherwise, it indicates that the missing data has a large impact on the overall data distribution;

[0160] S213: Process the missing data;

[0161] When the number of missing data is small and the missing data has little impact on the overall data distribution, for example, when the proportion of the missing data in the total amount of data in the hemodialysis patient dataset is less than 10%, the filling method or deletion method is adopted;

[0162] Furthermore, the filling method includes mean filling, median filling, mode filling, etc., and a suitable filling method is selected according to the data type and data distribution characteristics;

[0163] The deletion method means deleting the rows or columns containing the missing data in the hemodialysis patient dataset;

[0164] When the number of missing data is large and the missing data has a large impact on the overall data distribution, deep learning methods, multiple imputation methods, etc. are used for data filling;

[0165] When there are missing data in continuous data, interpolation methods are used for interpolation filling, such as linear interpolation or polynomial interpolation, etc.;

[0166] By interpolating and filling the missing data, the continuity of the data can be ensured, which is convenient for subsequent data analysis and other applications;

[0167] S22: Process outliers;

[0168] Judge whether there are outliers in the hemodialysis patient dataset. If there are outliers, process the outliers; if not, go to step S23;

[0169] S221: Identify outliers;

[0170] Identify whether there are outliers in the hemodialysis patient dataset. The specific identification methods include the box plot method, the Z-score comparison detection method, etc.;

[0171] Specifically, the box plot method refers to drawing a box plot, which contains the median, upper quartile, lower quartile, upper limit, and lower limit of the data in the hemodialysis patient dataset; if there are data points in the box plot that are higher than the upper limit or lower than the lower limit, the data corresponding to the data points is an outlier, otherwise it is a non-outlier;

[0172] The Z-score comparison detection method refers to calculating the mean and standard deviation of the data in the hemodialysis patient dataset, and calculating the standard score Z-score of each data point through the Z-score normalization formula according to the calculated mean, standard deviation, and the data values in the hemodialysis patient dataset. Set a threshold in advance. If the standard score Z-score of the data point is greater than the threshold, it is an outlier, otherwise it is a non-outlier; The Z-score normalization formula is as follows:

[0173] ;

[0174] Among them, Z-score is the standard score, x is the data value in the hemodialysis patient dataset, is the mean, and s is the standard deviation;

[0175] S222: Process outliers;

[0176] Process the outliers by deleting or replacing the outliers;

[0177] Specifically, replacing outliers means replacing the outliers with specific values in the hemodialysis patient dataset, such as the median, mean, etc.;

[0178] Furthermore, outliers can be processed by methods such as trimmed mean and median absolute deviation;

[0179] Processing outliers through methods such as trimmed mean and median absolute deviation can better resist the influence of outliers and make the description of the overall characteristics of the data more accurate;

[0180] S3: Standardize the hemodialysis patient dataset;

[0181] The data standardization module standardizes the data in the hemodialysis patient dataset through data standardization;

[0182] Standardize the diagnosis results of hemodialysis patients into mild malnutrition state, moderate malnutrition state, severe malnutrition state, and normal nutrition state;

[0183] Furthermore, the hemodialysis patient dataset contains features of two detection types: qualitative detection and quantitative detection;

[0184] For the features of qualitative detection, represent the detection results, i.e., the feature values, numerically. Use the numerical value 1 for positive and 0 for negative;

[0185] For the features of quantitative detection, determine the normal range of the detection results of this feature in advance, compare the actual detection results with the normal range, and calculate the difference value of each feature;

[0186] The normal range of the detection results includes a single value and an interval;

[0187] If the normal range of the detection result is a single value, the difference value = |actual detection result - normal range of the detection result|;

[0188] If the detection result is an interval, calculate the central value of the interval. The central value of the interval is the average of the upper and lower limits of the interval, and the difference value = |actual detection result - central value of the interval|;

[0189] Obtain the mean and standard deviation of the difference values based on the calculated difference values of the features, and perform data standardization processing through the Z-score standardization formula;

[0190] At this time, in the Z-score standardization formula, Z-score is the standard score, x is the difference value, is the mean of the difference values, and s is the standard deviation of the difference values of this feature.

[0191] S4: Feature selection and model construction method;

[0192] The dimensionality reduction module combines data algorithms to determine the features related to malnutrition in the hemodialysis patient dataset;

[0193] S41: Initially remove features with relatively small differences;

[0194] Perform clustering analysis on the data in the hemodialysis patient dataset through a clustering algorithm;

[0195] The clustering algorithm includes k-means clustering, hierarchical clustering, and DBSCAN clustering;

[0196] Record the data of each patient in the hemodialysis patient dataset as a sample, and run the above three clustering algorithms on each sample respectively. Each clustering algorithm assigns each sample to the corresponding cluster;

[0197] The method for setting the number of clusters for each clustering algorithm includes: for k-means clustering and hierarchical clustering, the silhouette coefficient is used to determine the optimal number of clusters. DBSCAN clustering does not require pre-setting the number of clusters, but requires setting the eps and min_samples parameters, and these parameters are determined by observing the data point distance graph;

[0198] Calculate the distribution of each feature in each cluster, and measure the distribution difference of each feature in different clusters through variance;

[0199] Calculate the variance of each feature included in each cluster according to the variance formula, sort the calculated variances of each feature in descending order, and draw a curve after variance sorting. In the curve, the horizontal axis is the feature name and the vertical axis is the variance of the corresponding feature;

[0200] Use the kneedle algorithm to obtain the slope change of the curve, detect the inflection point of the curve according to the slope change. The features before the first inflection point of the detected curve indicate that they show greater differences between different clusters and are more capable of distinguishing different clusters;

[0201] Remove the features in the dataset that do not show significant differences between different clusters, that is, the features after the first inflection point. Take the intersection of the features with greater differences obtained by the three clustering algorithms as the features after preliminary dimensionality reduction. The entire experimental process is iterative;

[0202] The conditions for ending the iteration can be determined according to the following three points:

[0203] (1) The curve of feature variance sorting is stable: When the variance sorting curve calculated by the kneedle algorithm does not change significantly in consecutive iterations at the position of the first inflection point, it indicates that the features capable of distinguishing different clusters have stabilized, and the iteration stops;

[0204] (2) The intersection of features of the three clustering algorithms converges: Compare the feature sets with greater differences obtained by k-means clustering, hierarchical clustering, and DBSCAN. If the intersection of features gradually decreases after each iteration and the feature set of the final intersection no longer changes significantly, it means that stable preliminary dimensionality reduction features have been found and the iteration can be stopped;

[0205] (3) The clustering performance metrics meet the expectations: Evaluate the clustering effect after each iteration. The evaluation metrics include the silhouette coefficient, the purity of the nutritional status within the cluster, and the accuracy rate, etc. If these metrics do not improve significantly in multiple iterations, it indicates that the current feature set is already sufficiently effective and there is no need to further remove features.

[0206] If at least two of the above conditions are met, it can be considered that the iterative process of S41 has reached the state of optimal feature screening.

[0207] S42: Generate a model through data clustering calculation.

[0208] After S41 preliminarily removes features with small differences, new features are obtained. Retain the new features for each cluster divided by each clustering algorithm in S41. For each sample, vote according to the clustering results of K-means, hierarchical clustering, and DBSCAN, and select the most frequent cluster as the final result. Use the cluster labels of the three algorithms as new features and add them to the dataset to further improve the clustering effect.

[0209] Assign a nutritional status label to each patient in each cluster. The nutritional status includes normal nutrition and malnutrition. Set the malnutrition status label to 1, and the malnutrition status includes mild, moderate, and severe malnutrition. Set the normal nutrition status label to 0.

[0210] For the patients in each cluster, conduct statistics according to their nutritional status and calculate the proportion of the number of people in each nutritional status in this cluster.

[0211] Assign a nutritional status label to each cluster. This label is the nutritional status with the highest proportion in each cluster. This step defines a representative nutritional status for each cluster, and these labels can help understand the differences between clusters.

[0212] Finally, generate a clustering model. The key metrics for evaluating the clustering model include the silhouette coefficient, the purity of the nutritional status within the cluster, and the accuracy rate. If the silhouette coefficient is above 0.5, the purity value of the nutritional status within the cluster is above 80%, and the accuracy rate is above 85%, it indicates that this clustering model can be used to identify the nutritional status of new data.

[0213] S43: Randomly combine features for iterative clustering grouping.

[0214] S43 is an optimization of step S42. In step S42, a representative nutritional status label has been assigned to each cluster, and the nutritional status proportions of each cluster have been calculated. To further optimize the clustering model and feature selection, S43 conducts multiple iterative clustering analyses by randomly combining features.

[0215] From the features preliminarily screened in step S41, randomly select different combinations of feature subsets, and use these subsets to re-cluster the data; in each clustering, re-apply the k-means clustering, hierarchical clustering, and DBSCAN clustering methods, repeat the previous cluster number determination and clustering steps, and perform result fusion; in each iteration, maintain the same process as in S42, that is, assign nutritional status labels to each cluster and calculate the proportion of nutritional status in each cluster.

[0216] Through iterative clustering by randomly combining different features multiple times, the purpose is to observe the impact of different feature combinations on the clustering results and find the feature combination that can best distinguish nutritional status.

[0217] After each clustering, evaluate the effect of the clustering model to ensure that the model can effectively distinguish different nutritional statuses, calculate the silhouette coefficient of each cluster, the purity and accuracy of the nutritional status within the cluster, and ensure that the clustering results can effectively distinguish nutritional status.

[0218] When similar feature combinations obtain similar clustering results in multiple iterations, and indicators such as the purity of nutritional status within the cluster, silhouette coefficient, and accuracy reach the set thresholds, that is, the silhouette coefficient remains stable above 0.5, the purity value remains above 80%, the accuracy remains above 85%, and there is no obvious improvement in further iterations, the iteration can be stopped.

[0219] Through the iterative optimization of S43, the feature set can be further reduced, and finally the most discriminative feature combination can be selected to improve the recognition accuracy of the clustering model for new data.

[0220] During the entire analysis process, the features highly correlated with malnutrition found in existing research remain unchanged.

[0221] The dimensionality reduction method proposed in the present invention starts from the data structure, captures the relationship between features and specific subgroups (clusters), and helps to retain the key features when distinguishing these subgroups; traditional dimensionality reduction methods usually reduce the dimension by transforming the data or based on linear algebra methods, without involving the clustering process.

[0222] S5: Model training and generation;

[0223] The nutritional status prediction module, based on the feature selection and clustering analysis in S4, predicts the malnutrition status through the finally screened features and the clustering model, including the following sub-steps:

[0224] S51: Prepare training data;

[0225] Use the feature selection results of step S4 to determine the feature set, which are the features that are retained after multiple clustering iterations and are most effective in distinguishing malnutrition status; each sample has been labeled with a nutrition status label (0 indicates normal nutrition, 1 indicates malnutrition), that is, mild, moderate, and severe malnutrition are classified as 1, and normal nutrition is 0;

[0226] S52: Use the clustering model to predict whether a patient is in a malnutrition state;

[0227] Through the previous clustering models (S42, S43), patient samples have been assigned to different clusters, and a nutrition status label has been assigned to each cluster. Use this clustering structure to build a prediction model.

[0228] S53: Train using the clustering model;

[0229] Use the filtered feature set (the final feature combination of step S43) as the input data.

[0230] According to the features of each new sample, use the trained clustering model to assign the new sample to a cluster. Using the nutrition status ratio calculated for each cluster before, assign a predicted nutrition status label to the new sample. Specifically, the new sample will be assigned to a certain cluster, and the nutrition status ratio in the cluster determines the label (0 or 1) of the sample. Use the known nutrition status labels (0 or 1 labeled in the training set) to evaluate the prediction accuracy. The performance of the model can be evaluated by criteria such as accuracy, precision, recall, and F1 score;

[0231] S54: Model evaluation and optimization;

[0232] The evaluation metrics include accuracy, precision, recall, and F1 score; if the initial training result is not satisfactory, adjust the parameters of the clustering algorithm (such as eps and min_samples in DBSCAN, or the number of clusters in K-means), or recombine features during the feature selection process;

[0233] S55: Use the trained clustering model to predict new data;

[0234] For each new patient, after inputting their features, the clustering model assigns the patient to a certain cluster, and based on the nutrition status label of the cluster, predicts whether the patient is in a malnutrition state. Through the prediction results of the model, patients with malnutrition can be intervened early, and corresponding nutritional support or other treatments can be provided.

[0235] The method of the present invention preprocesses the data in the hemodialysis patient dataset by dealing with missing data, outliers, and feature selection. By calculating the Pearson correlation coefficient between each feature in the dataset, the influence of the corresponding data on the overall data distribution is judged, avoiding the problem that certain indicators may be easily affected by other factors, resulting in inaccurate prediction of the patient's nutritional status.

[0236] The present invention has been described by the above related embodiments. However, the above embodiments are only examples for implementing the present invention. It must be pointed out that the disclosed embodiments do not limit the scope of the present invention. On the contrary, modifications and refinements made without departing from the spirit and scope of the present invention fall within the scope of the patent protection of the present invention.

Claims

1. A method for predicting malnutrition in hemodialysis patients based on AI technology, characterized in that: It includes the following steps: S1: Collect relevant attributes of hemodialysis patients to obtain a hemodialysis patient dataset; The data collection module collects relevant attributes of hemodialysis patients according to actual needs; Specifically, the data collection module collects relevant attributes of hemodialysis patients from general demographic information, hemodialysis data, laboratory test indicators, and diagnostic results according to actual needs; After collecting relevant attributes of hemodialysis patients, a hemodialysis patient dataset is obtained, and the hemodialysis patient dataset contains features and corresponding feature values; The features are various attributes of the collected patients, and the feature values are the specific values corresponding to various attributes; S2: Preprocess the hemodialysis patient dataset; The data preprocessing module preprocesses the data in the hemodialysis patient dataset by processing missing data and outliers, including the following sub-steps: S21: Process missing data; Judge whether there is missing data in the hemodialysis patient dataset. If there is missing data, process the missing data; if not, go to step S22; S211: Judge whether there is missing data; Judge whether there is missing data in the hemodialysis patient dataset; Specific judgment methods include counting missing values and visualizing data distribution; After the judgment is completed, obtain the number of missing data through the sum function sum; If there is no missing data in the hemodialysis patient dataset, go to step S22; if there is missing data, go to step S212; S212: Evaluate the impact of missing data; Evaluate the impact of missing data on the overall data distribution. Specific evaluation methods include comparing statistical indicators, calculating correlation, and comparing data distribution; S213: Process the missing data; Process the missing data according to the evaluation result of step S212. When the number of missing data is small and the impact of missing data on the overall data distribution is small, use the filling method or deletion method; When the number of missing data is large and the impact of missing data on the overall data distribution is large, use deep learning methods and multiple imputation methods for data filling; When there is missing data in continuous data, use the interpolation method for interpolation filling, and the interpolation method includes linear interpolation or polynomial interpolation; S22: Process outliers; Judge whether there are outliers in the hemodialysis patient dataset. If there are outliers, process the outliers; if not, go to step S23; S221: Identify outliers; Identify whether there are outliers in the hemodialysis patient dataset. Specific identification methods include the box plot method and the Z-score comparison detection method; S222: Process the outliers; Process the outliers by deleting or replacing the outliers; Specifically, replacing outliers means replacing outliers with specific values in the hemodialysis patient dataset, including the median and the mean; S3: Standardize the hemodialysis patient dataset; The data standardization module standardizes the data in the hemodialysis patient dataset; Standardize the diagnostic results of hemodialysis patients into mild malnutrition, moderate malnutrition, severe malnutrition, and normal nutrition status; By calculating the difference values of each feature, obtain the mean and standard deviation of the difference values based on the calculated difference values of the features, and perform data standardization processing through the Z-score standardization formula; S4: Feature selection and model construction method; The dimensionality reduction module combines data algorithms to determine the features related to malnutrition in the hemodialysis patient dataset; Including the following sub-steps: S41: Initially remove features with relatively small differences; Perform cluster analysis on the data in the hemodialysis patient dataset through a clustering algorithm, and assign each sample to the corresponding cluster; calculate the distribution of each feature in each cluster, and measure the distribution difference of each feature in different clusters through variance; remove features with relatively small differences; Perform cluster analysis on the data in the hemodialysis patient dataset through a clustering algorithm; The clustering algorithm includes k-means clustering, hierarchical clustering, and DBSCAN clustering; Record the data of each patient in the hemodialysis patient dataset as a sample, and run the above three clustering algorithms on each sample respectively. Each clustering algorithm assigns each sample to the corresponding cluster; The method for setting the number of clusters for each clustering algorithm includes determining the optimal number of clusters using the silhouette coefficient for k-means clustering and hierarchical clustering. DBSCAN clustering does not require pre-setting the number of clusters, but requires setting the eps and min_samples parameters, and these parameters are determined by observing the data point distance graph; Calculate the distribution of each feature in each cluster, and measure the distribution difference of each feature in different clusters through variance; Calculate the variance of each feature included in each cluster according to the variance formula, sort the calculated variances of each feature in descending order, and draw a curve after variance sorting. The horizontal axis of the curve is the feature name, and the vertical axis is the variance of the corresponding feature; Use the kneedle algorithm to obtain the slope change of the curve, detect the inflection point of the curve according to the slope change, and the features before the first detected inflection point of the curve indicate greater differences between different clusters and distinguish different clusters; Remove the features in the dataset that do not show significant differences between different clusters, that is, the features after the first inflection point. Take the intersection of the features with large differences obtained by the three clustering algorithms as the features after preliminary dimensionality reduction. The entire experimental process is iterative; S42: Generate a model through data clustering calculation; Generate a clustering model. The key indicators for evaluating the clustering model include the silhouette coefficient, purity, and accuracy of the nutritional status within the cluster; S43: Randomly combine features for iterative clustering grouping; Randomly select different combinations of feature subsets from the features initially screened in step S41, and use these subsets to re-cluster the data and perform multiple iterative cluster analyses; In each clustering, re-apply the k-means clustering, hierarchical clustering, and DBSCAN clustering methods, repeat the previous cluster number determination and clustering steps, and perform result fusion; In each iteration, maintain the process consistent with S42, that is, assign nutritional status labels to each cluster and calculate the proportion of nutritional status in each cluster; After each clustering, evaluate the effect of the clustering model to ensure that the model can effectively distinguish different nutritional statuses, calculate the silhouette coefficient of each cluster, the purity and accuracy of the nutritional status within the cluster, and ensure that the clustering results can effectively distinguish the nutritional status; S5: Model training and generation; Based on the feature selection and clustering analysis in S4, use the finally selected features and the clustering model to predict the malnutrition status, including the following sub-steps: S51: Prepare training data; Determine the feature set through the feature selection results of step S4 as the training data; S52: Use the clustering model to predict whether the patient is in a malnourished state; S53: Use the clustering model for training; S54: Model evaluation and optimization; S55: Use the trained clustering model to predict new data.

2. A method for predicting malnutrition in hemodialysis patients based on AI technology as claimed in claim 1, wherein: In step S211, the method for determining whether there is missing data in the dataset is as follows: Statistical missing values: Use the isnull or isna function to identify the missing data in the dataset; Visualize the data distribution: Present the data distribution in the dataset by drawing a graph to identify the missing data; After the judgment, use the sum function to summarize the missing data to obtain the total number of missing data in the dataset.

3. A method for predicting malnutrition in hemodialysis patients based on AI technology as claimed in claim 1, wherein: In step S212, compare statistical indicators: According to the missing data detected in step S211, calculate the statistical indicators of the hemodialysis patient dataset containing the features corresponding to the missing data and the statistical indicators of the hemodialysis patient dataset after removing the features corresponding to the missing data. The statistical indicators include mean, median, and standard deviation; Compare the relative difference or percentage difference of the calculated statistical indicators. If the difference is small, it means that the missing data has little impact on the overall data distribution; otherwise, it means that the missing data has a large impact on the overall data distribution; Calculate the correlation: Calculate the Pearson correlation coefficient between each feature in the hemodialysis patient dataset through the corr function to obtain the correlation matrix, and extract the correlation coefficient between the feature corresponding to the missing data and other features in the correlation matrix. If the correlation coefficient is small, it means that the missing data has little impact on the overall data distribution; otherwise, it means that the missing data has a large impact on the overall data distribution; Compare the data distribution: Draw the data distribution of the hemodialysis patient dataset containing the features corresponding to the missing data and the data distribution of the hemodialysis patient dataset after removing the features corresponding to the missing data through a histogram, and compare the data distribution in the above two cases through a box plot. If the two data distributions are similar, the missing data has little impact on the overall data distribution; otherwise, the missing data has a large impact on the overall data distribution.

4. A method for predicting malnutrition in hemodialysis patients based on AI technology as claimed in claim 1, wherein: In step S221, the box plot method: by drawing a box plot, which includes the median, upper quartile, lower quartile, upper limit, and lower limit of the data in the hemodialysis patient dataset; If there are data points above the upper limit or below the lower limit in the box plot, the data corresponding to the data points is an outlier, otherwise it is a non-outlier; Z-score comparison detection method: Calculate the mean and standard deviation of the data in the hemodialysis patient dataset. According to the calculated mean, standard deviation, and the data values in the hemodialysis patient dataset, calculate the standard score Z-score of each data point through the Z-score normalization formula. Set a threshold in advance. If the standard score Z-score of a data point is greater than the threshold, it is an outlier point, otherwise it is a non-outlier point; The Z-score normalization formula is as follows: Among them, Z-score is the standard score, x is the data value in the dataset of hemodialysis patients, is the average value, and s is the standard deviation.

5. The method for predicting malnutrition in hemodialysis patients based on AI technology according to claim 1, characterized in that: In step S3, the data standardization module performs standardization processing on the data in the hemodialysis patient dataset through data standardization; The hemodialysis patient dataset contains the characteristics of two detection types: qualitative detection and quantitative detection; For the characteristics of qualitative detection, represent the detection result, that is, the characteristic value, numerically. Represent positive with the value 1 and negative with the value 0; For the characteristics of quantitative detection, determine the normal range of the detection result of this characteristic in advance, compare the actual detection result with the normal range, and calculate the difference value of each characteristic; The normal range of the detection result includes a single value and an interval; If the normal range of the detection result is a single value, the difference value = |actual detection result - normal range of the detection result|; If the detection result is an interval, calculate the central value of the interval. The central value of the interval is the average of the upper limit and the lower limit of the interval, and the difference value = |actual detection result - central value of the interval|; Obtain the mean and standard deviation of the difference values of the characteristics according to the calculated difference values of the characteristics, and perform data standardization processing through the Z-score normalization formula; At this time, in the Z-score normalization formula, Z-score is the standard score, x is the difference value, is the mean of the difference values, and s is the standard deviation of the difference values of this feature.

6. The method for predicting malnutrition in hemodialysis patients based on AI technology according to claim 1, characterized in that: The specific content of step S4 is as follows: S41: Initially remove the features with relatively small differences; The conditions for ending the iteration are determined according to the following three points: (1) The curve of the feature variance ranking is stable: When the variance ranking curve calculated by the kneedle algorithm does not change significantly at the position of the first inflection point in consecutive iterations, it indicates that the features that can distinguish different clusters have stabilized, and the iteration stops; (2) The feature intersection of the three clustering algorithms converges: Compare the larger difference feature sets obtained by K-means clustering, hierarchical clustering, and DBSCAN. If the feature intersection gradually decreases after each iteration and the feature set of the final intersection no longer changes significantly, it means that stable preliminary dimensionality reduction features have been found, and the iteration stops; (3) The clustering performance index reaches the expectation: Evaluate the clustering effect after each iteration. The evaluation indicators are silhouette coefficient, purity of nutritional status within the cluster, accuracy, and other indicators; If these metrics do not show significant improvement over multiple iterations, it indicates that the current feature set is already sufficiently effective and no further feature removal is required; If at least two of the above conditions are met, it means that the iterative process in step S41 has reached the optimal feature screening state; S42: Generate a model through data clustering calculation; In S41, after initially removing features with relatively small differences to obtain new features, retain the new features for each cluster obtained by each clustering algorithm. For each sample, vote based on the clustering results of K-means, hierarchical clustering, and DBSCAN, and select the most frequent cluster as the final result; Use the cluster labels of the three algorithms as new features and add them to the dataset to further improve the clustering effect; Assign a nutritional status label to each patient in each cluster; the nutritional status includes normal nutrition and malnutrition status; Set all malnutrition status labels to 1, and the malnutrition status includes mild, moderate, and severe malnutrition; Set the normal nutrition status label to 0; For the patients within each cluster, conduct statistics based on their nutritional status and calculate the proportion of the number of people in each nutritional status in that cluster; Assign a nutritional status label to each cluster, and this label is the nutritional status with the highest proportion in each cluster; Finally, generate a clustering model. The key metrics for evaluating the clustering model include the silhouette coefficient, the purity and accuracy of the nutritional status within the cluster. If the silhouette coefficient is above 0.5, the purity value of the nutritional status within the cluster is above 80%, and the accuracy is above 85%, it indicates that this clustering model can be used to identify the nutritional status of new data; S43: Randomly combine features for iterative clustering grouping; When similar feature combinations obtain similar clustering results in multiple iterations, and metrics such as the purity of the nutritional status within the cluster, the silhouette coefficient, and the accuracy reach the set thresholds, that is, the silhouette coefficient continuously stabilizes above 0.5, the purity value remains above 80%, the accuracy remains above 85%, and there is no obvious improvement in further iterations, stop the iteration; During the entire analysis process, the features highly correlated with malnutrition found in existing studies remain unchanged.

7. A method for predicting malnutrition in hemodialysis patients based on AI technology as claimed in claim 1, characterized in that: The specific content of step S5 is as follows: The nutritional status prediction module, based on the feature selection and clustering analysis in S4, uses the finally screened features and the clustering model to predict the malnutrition status, including the following sub-steps: S51: Prepare training data; Use the feature selection results of step S4 to determine the feature set. These features are the most effective features for distinguishing the malnutrition status that are retained after multiple clustering iterations; each sample has already been labeled with a nutritional status label, and mild, moderate, and severe malnutrition are classified as 1, and normal nutrition is 0; S52: Use the clustering model to predict whether a patient is in a malnutrition status; Through the clustering model, patient samples have been assigned to different clusters, and a nutritional status label has been assigned to each cluster. Use this clustering structure to build a prediction model; S53: Use the clustering model for training; Use the screened feature set as the input data; Based on the characteristics of each new sample, the trained clustering model is used to assign the new sample to a cluster, and the predicted nutritional status label is given to the new sample by using the nutritional status ratios calculated for each cluster previously. Specifically, the new sample will be assigned to a certain cluster, and the nutritional status ratio in the cluster determines the label of the sample. The known nutritional status labels are used to evaluate the prediction accuracy, and the performance of the model is evaluated by criteria such as accuracy, precision, recall, and F1 score. S54: Model evaluation and optimization; The evaluation metrics include accuracy, precision, recall, and F1 score; if the initial training result is not satisfactory, adjust the parameters of the clustering algorithm or recombine the features during the feature selection process. S55: Use the trained clustering model to predict new data; For each new patient, after inputting their characteristics, the clustering model assigns the patient to a certain cluster, and based on the nutritional status label of the cluster, predicts whether the patient is in a malnourished state.

8. A system for predicting malnutrition in hemodialysis patients based on AI technology, which is used to implement the method for predicting malnutrition in hemodialysis patients based on AI technology according to any one of claims 1-7, characterized in that: It includes a data collection module, a data preprocessing module, a data normalization module, a dimensionality reduction module, and a nutritional status prediction module; The data collection module is used to collect relevant data of hemodialysis patients; The data preprocessing module is used to preprocess the data in the hemodialysis patient dataset; The data normalization module is used to normalize the data in the hemodialysis patient dataset; The dimensionality reduction module is used to perform clustering analysis on the data in the hemodialysis patient dataset, remove features with low relevance to malnutrition, and perform iterative clustering grouping to obtain features highly relevant to malnutrition; The nutritional status prediction module is used to establish a prediction model and predict whether the patient is in a malnourished state.

Images