Malnutrition risk prediction system for critically ill patients based on XGBoost

Abstract

The present application discloses a malnutrition risk prediction system for critically ill patients based on XGBoost, which relates to the technical field of patient nutrition management, and includes an information acquisition module, a data processing module and a risk prediction module; the information acquisition module is used to obtain the target physiological characteristics of the target patient; the risk prediction module also includes a model construction submodule and a risk output submodule; the model construction submodule is used to obtain a training data set and a target parameter value; a risk prediction model is constructed and trained according to the target parameter value and the XGBoost model; during training, the difference between the minimum value of the objective function of the decision tree before updating and the minimum value of the objective function of the decision tree to be updated is calculated, a decision tree is generated according to the maximum difference, and the leaf node weight is obtained according to the minimum objective function value of the generated decision tree; the risk output submodule is used to obtain the target risk according to the output of the risk prediction model based on the target physiological characteristics, so as to realize automated feature engineering and reduce the difficulty of model training.

Description

Technical Field

[0001] The present application relates to the technical field of patient nutrition management, and specifically to a malnutrition risk prediction system for critically ill patients based on XGBoost. Background Art

[0002] Critically ill patients are highly susceptible to malnutrition due to their severe condition, metabolic disorders, and prolonged stress. Malnutrition not only compromises the patient's immune function and increases the risk of infectious complications, but can also lead to delayed wound healing, prolonged hospitalization, increased medical expenses, and even adversely impact the patient's prognosis.

[0003] Commonly used clinical methods for assessing malnutrition risk in critically ill patients are primarily based on subjective assessments by healthcare professionals, such as the Subjective Global Assessment (SGA) and the Nutritional Risk Screening 2002 (NRS2002). The accuracy of these assessments not only depends on the healthcare professional's expertise and clinical experience, but is also significantly influenced by subjective factors. Furthermore, fixed assessment indicators are difficult to adapt to patients with varying physiological and / or disease characteristics. This is particularly true when assessing malnutrition risk in patients with specific diseases and / or specific patients. Traditional assessment methods are difficult to implement and offer low accuracy.

[0004] Currently, a prediction model for malnutrition in nasopharyngeal carcinoma patients, application number CN202410365426.0, uses logistic regression analysis to construct a risk prediction model to assess the risk of malnutrition in nasopharyngeal carcinoma patients. However, differences in technical levels and equipment conditions among different medical institutions lead to significant differences in the collection of characteristic data of critically ill patients. Before training the risk prediction model, it is necessary to manually screen the model independent variables from the many potential influencing factors that the medical institution can collect. Feature engineering operations are difficult. Not only do the input features need to be manually adjusted and optimized based on the specific circumstances of the medical institution, increasing labor costs, but the data analysis skills of professionals are also required to be high. In particular, manual screening and adjustment are particularly difficult in the absence of data science professionals. Summary of the Invention

[0005] The purpose of the present invention is to solve the technical problem of difficult feature engineering operations when constructing a malnutrition risk prediction model for critically ill patients. A malnutrition risk prediction system for critically ill patients based on XGBoost is provided, which realizes automated feature engineering through the built-in feature importance evaluation of the risk prediction model.

[0006] The present invention seeks to protect a malnutrition risk prediction system for critically ill patients based on XGBoost, comprising an information acquisition module and a risk prediction module;

[0007] The information acquisition module is used to obtain target physiological characteristics of the target patient;

[0008] The risk prediction module also includes a model construction submodule and a risk output submodule; the model construction submodule also includes a parameter acquisition submodule, a model generation submodule and a model training submodule; the parameter acquisition submodule is used to obtain the training data set and target parameter values ​​of the risk prediction model; the training data set includes the physiological characteristics of all patients to be screened; the model generation submodule is used to build a risk prediction model based on the target parameter values ​​and the XGBoost model; the risk prediction model includes several decision trees; the model training submodule is used to generate each decision tree based on the training data set and the target parameter values, using the strategy of maximizing information gain; the information gain is obtained based on the difference between the minimum value of the objective function of the decision tree before updating and the minimum value of the objective function of the decision tree to be updated, and the weight of each leaf node is obtained according to the minimum value of the objective function of the generated decision tree; the risk output submodule is used to obtain the trained risk prediction model, and obtain the target risk of the target patient according to the output of the risk prediction model based on the target physiological characteristics.

[0009] In one embodiment of the present application, the information acquisition module divides the target physiological characteristics into historical observation characteristics and target observation characteristics according to whether the acquisition time is a preset observation time; the system also includes a feature prediction module, which is used to predict the predicted physiological characteristics of the preset observation time based on the historical observation characteristics of the target patient; the risk output submodule also includes a risk calculation module and a risk verification module; the risk calculation module is used to obtain the predicted risk based on the predicted physiological characteristics before the preset observation time; the risk verification module is used to calculate the feature similarity between the predicted physiological characteristics and the target observation characteristics at the preset observation time. If the feature similarity is not less than the similarity threshold, the predicted risk is used as the target risk.

[0010] In one embodiment of the present application, if the feature similarity is less than the similarity threshold, the risk verification module is further configured to input the target observation feature into the risk calculation module and output the target risk.

[0011] In one embodiment of the present application, the risk verification module also includes screening target observation feature components from the target observation features according to the corresponding feature items of the intermediate nodes of all decision trees in the trained risk prediction model, and screening predicted physiological feature components from the predicted physiological features; and calculating the feature similarity based on the similarity between the predicted physiological feature components and the target observation feature components.

[0012] In one embodiment of the present application, the feature items corresponding to all intermediate nodes in the trained risk prediction model are recorded as a feature set, and the method for calculating feature similarity includes:

[0013] ;

[0014] ;

[0015] Where m represents the number of all feature items in the feature set; α i represents the feature weight of the i-th feature item in the feature set, F 1，i represents the value of the i-th feature item of the target observation feature component in the feature set, F 2，i represents the value of the predicted physiological feature component in the feature set, N i It represents the difference between the maximum value and the minimum value of the i-th feature item in the feature set in the historical medical records.

[0016] In one embodiment of the present application, the risk verification module also includes obtaining a corresponding feature weight based on the proportion of the number of intermediate nodes corresponding to each feature item in the feature set to the total number of intermediate nodes in the corresponding risk prediction model.

[0017] In one embodiment of the present application, the proportion of randomly selected samples when training each decision tree is less than 1. Before training each decision tree, the model training module obtains the current hit probability of each sample data in the training data set, and increases the hit probability of the sample data that was incorrectly predicted by the previous decision tree.

[0018] This application has the following beneficial effects:

[0019] 1. During the training process of the risk prediction model, the information gain of the greedy algorithm is used to automatically determine whether the leaf nodes in the current decision tree should continue to split and the specific splitting method to obtain the non-leaf nodes (i.e., intermediate nodes) of the decision tree, thereby realizing automated feature engineering. This system not only reduces labor costs and the difficulty of model training when training the risk prediction model, but also makes the trained risk prediction model more in line with the current characteristics of medical institutions.

[0020] 2. The system predicts the physiological characteristics of the target patient at the preset observation time in advance through the feature prediction module, and inputs the predicted physiological characteristics output by the feature prediction module into the risk prediction model to obtain the predicted risk in advance. When the patient's actual physiological characteristics at the preset observation time meet the requirements and the predicted physiological characteristics, the corresponding predicted risk is considered valid, so as to reduce the amount of calculation of the system's concurrent response at the preset observation time, especially in large-scale deployment and use, which reduces the server load and lowers the minimum requirements for system operation, which helps to optimize the system operation cost.

[0021] 3. The risk verification module reduces the misjudgment of the malnutrition risk of target patients due to the incorrect prediction results of the feature prediction model.

[0022] 4. The feature similarity calculated based on the similarity between the predicted physiological feature components and the target observed feature components not only reduces the computational complexity of the risk verification module, further reduces the computational complexity and system burden, but also reduces the prediction accuracy requirements for the feature prediction module.

[0023] 5. The difference between the maximum value and the minimum value of each feature item in the feature set in the historical medical records is used to obtain the gap between the target observed feature component and the predicted physiological feature component in the corresponding feature item, so as to reduce the impact of different dimensions between different feature items on the prediction validity.

[0024] 6. Assigning higher weights to features with more corresponding intermediate nodes can not only be used to describe the importance of the corresponding feature items, but also when the value of the feature item with a larger feature weight changes, that is, when the value of the target observation feature component in the corresponding feature item is different from the value of the predicted physiological feature component in the corresponding feature item, it is more likely to affect the output result of the risk prediction model. The corresponding predicted risk should be deemed invalid as a priority, so as to achieve the goal of reducing the computational complexity of the system's concurrent response while ensuring the accuracy of the output target risk.

[0025] 7. As the number of decision trees generated increases, the decision trees to be trained gradually pay more attention to the sample data of special conditions and / or special patients, thereby improving the prediction accuracy of special conditions and / or special patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0027] Figure 1 This is a structural diagram of an implementable method of a critically ill patient malnutrition risk prediction system based on XGBoost involved in an embodiment of the present application;

[0028] Figure 2 This is a structural diagram of another possible implementation of the XGBoost-based critically ill patient malnutrition risk prediction system involved in an embodiment of the present application;

[0029] Figure 3 Schematic diagram of the target risk calculation process involved in the embodiment of this application;

[0030] Figure 4 This is a schematic diagram of the structure of an electronic device involved in an embodiment of the present application. DETAILED DESCRIPTION

[0031] To make the above-mentioned objects, features and advantages of the present application more clearly understood, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. The components of the embodiments of the present application generally described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, reference to the terms "one embodiment", "some embodiments", "implementation method", "embodiment", "illustrative embodiment", "example", "specific example" or "some examples", etc., the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed application, but merely indicates that the specific features, structures or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. Moreover, the specific features, structures or characteristics described can be combined in any one or more embodiments or examples in an appropriate manner. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making any creative work are within the scope of protection of the present application.

[0032] It should be noted that similar reference numerals and letters represent similar items in the following figures. Therefore, once an item is defined in one figure, it does not need to be further defined or explained in subsequent figures. At the same time, in the description of this application, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations.

[0033] The present invention claims protection for a malnutrition risk prediction system for critically ill patients based on XGBoost, with reference to the attached Figure 1 As shown, it includes an information acquisition module, a data processing module and a risk prediction module.

[0034] It should be noted that the information acquisition module is used to obtain the physiological characteristics of the target patient, namely, the target physiological characteristics. These physiological characteristics include, but are not limited to, basic information, vital signs, laboratory test indicators, food intake information, and disease diagnosis information. Basic information includes, but is not limited to, age, gender, height, weight, past medical history, and family medical history. Vital signs include, but are not limited to, heart rate, blood pressure, respiration, and body temperature. Laboratory test indicators include, but are not limited to, blood count, biochemical indicators, and electrolytes. Food intake information includes, but is not limited to, food type, time of intake, amount, and food processing method. Disease diagnosis information includes, but is not limited to, the disease name, symptoms during treatment, and historical medication information. Symptom information includes, but is not limited to, symptom content, duration, frequency, and severity. Symptom content includes, but is not limited to, whether the headache type is stabbing, distending, or other pain type, and the location of the headache. Historical medication information includes, but is not limited to, the time of medication use during treatment, the name, dosage, and method of medication for each medication use. The administration methods include oral administration and intravenous injection, but are certainly not limited thereto.

[0035] It should be noted that the data processing module is used to pre-process the target physiological characteristics. The pre-processing includes, but is not limited to, data conversion. The data conversion includes, but is not limited to, normalization and encoding. Encoding processing, for example, encodes disease names.

[0036] In this embodiment, the risk prediction module is used to input the pre-processed target physiological characteristics into a pre-trained risk prediction model. The risk prediction module includes a risk output submodule and a model construction submodule, but is not limited thereto. The model construction submodule is used to construct a risk prediction model, and the risk output submodule is used to obtain the trained risk prediction model, input the target physiological characteristics into the risk prediction model, and obtain the target risk of malnutrition in the target patient.

[0037] In this embodiment, the model construction submodule further includes a parameter acquisition module, a model generation module, and a model training module, but it is not limited thereto. The parameter acquisition module is used to obtain a training data set and a target parameter value of a risk prediction model. The model generation module is used to construct the risk prediction model based on the target parameter value and the XGBoost model. The model training module is used to train the risk prediction model based on the training data set and the target parameter value.

[0038] It should be noted that the training dataset includes the physiological characteristics of all patients to be screened. The corresponding sample information can be obtained by obtaining the physiological characteristics of each critically ill patient and whether they are malnourished from the medical institution's historical medical records. Alternatively, the physiological characteristics of sample patients can be obtained based on the medical institution's existing technology and equipment. All of this sample information constitutes the training dataset.

[0039] It should be noted that the target parameter value includes the value of the hyperparameter preset when training the risk prediction model. The target parameter value can be obtained in a preset manner or according to user input, for example, obtaining the target parameter value input by the user. If there is a target parameter value that has not been input, the default value is obtained. The target parameters include learning task parameters, learning rate, maximum depth of each tree, proportion of samples randomly extracted when training each tree, proportion of features randomly extracted when training each tree, minimum sample weight and required for node splitting, minimum loss reduction value for node splitting, penalty parameter of model complexity and performance evaluation index of the model, of course, it may not be limited to this.

[0040] In this embodiment, the proportion of randomly selected samples during training of each decision tree is less than 1. Before training each decision tree, the model training module obtains the current hit probability of each sample data in the training dataset, thereby increasing the hit probability of sample data that were incorrectly predicted by the previous decision tree. As the number of decision trees generated increases, the decision tree to be trained gradually increases its focus on sample data for specific conditions and / or patients, thereby improving the accuracy of predictions for specific conditions and / or patients.

[0041] It should be noted that the learning task parameters include loss functions, but are not limited to these. For example, when the learning task is a regression task, optional loss functions include root mean square error, mean absolute error, and mean absolute percentage error; when the learning task is a classification task, optional loss functions include cross entropy loss function and logarithmic loss.

[0042] In this embodiment, the learning task of the risk prediction model is a three-classification task, namely, a high-risk category, a medium-risk category, and a low-risk category. The high-risk category is used to describe the patient's high risk of malnutrition, and the medium-risk category is used to describe the patient's malnutrition. The softmax function is selected for probability distribution calculation, and the cross-entropy loss function is selected as the loss function of the decision tree. The risk prediction model includes several decision trees, and the output of the risk prediction model is obtained based on the outputs of all decision trees, which can be expressed as:

[0043] ;

[0044] It should be noted that Represents sample x i The output of the risk prediction model, f k (x i ) represents the sample x i At the output of the kth decision tree, K represents the number of decision trees in the risk prediction model. The risk prediction model uses the Boosting algorithm to train each decision tree in turn, that is, the first decision tree f1 (x i ) is used to output the prediction result of sample xi, and the second decision tree f2 (x i ) is used to output sample x i In the first decision tree f1(x i )’s prediction result and the actual result, the third decision tree f3 (x i ) is used to output sample xi in the second decision tree f2 (x i ) is the difference between the predicted error and the true error, and so on. Calculate the sum of the output results of all decision trees to get the sample x i Therefore, when training the t-th decision tree of the risk prediction model, the output of the risk prediction model is Equal to the output of the trained t-1 decision trees And the sum of the outputs of the t-th decision tree to be trained, that is:

[0045] .

[0046] It should be noted that The value of is preset to 0, which means that when training the first decision tree of the risk prediction model, The value of is 0.

[0047] It should be noted that the decision tree objective function O(t) of the t-th decision tree includes:

[0048] ;

[0049] Where n represents the total number of samples in the training data set; L represents the loss function; y i Represents sample x i The true result is whether malnutrition occurs; Ω represents the complexity, which can be obtained according to the regularization term, namely:

[0050] ;

[0051] Among them, γ represents the penalty parameter of the number of decision tree leaf nodes, T represents the number of decision tree leaf nodes, the more decision tree leaf nodes there are, the greater the corresponding penalty value; λ represents the L2 regularization coefficient, ω jRepresents the weight of the j-th leaf node in the decision tree.

[0052] In this embodiment, since the first t-1 decision trees have completed training, the term related to the first t-1 decision trees in the objective function is a constant. The decision tree objective function is optimized by second-order Taylor expansion, regularization expansion and coefficient merging. , that is, the objective function of the decision tree after the optimization of the t-th decision tree is obtained:

[0053] ;

[0054] ;

[0055] ;

[0056] ;

[0057] ;

[0058] Among them, G j Represents the cumulative sum of the first-order partial derivatives of the samples contained in the j-th leaf node, H j represents the cumulative sum of the second-order partial derivatives of the samples contained in the j-th leaf node, g i represents the first-order partial derivative of the i-th sample in the corresponding leaf node, h i Represents the second-order partial derivative of the i-th sample in the corresponding leaf node. The optimal t-th decision tree of the risk prediction model is trained by minimizing the value of this objective function. Therefore, the number of leaf nodes of the t-th decision tree with a known structure is fixed, and the decision tree needs to be optimized by optimizing the weights corresponding to all leaf nodes of the decision tree. The leaf node objective function f(ω) of the j-th leaf node j )include:

[0059] ;

[0060] Therefore, the leaf node objective function of the jth leaf node is Take the minimum value at When the weights of all current leaf nodes make the corresponding leaf node objective function f take the minimum value, the t-th decision tree with the current known structure is in the decision tree objective function Take the minimum value.

[0061] In this embodiment, during the training of the risk prediction model, when the current depth of the kth decision tree is not greater than the preset depth threshold, a greedy algorithm is used to determine whether each leaf node needs to be split further and how to split it. When other leaf nodes remain unchanged, the decision tree objective function of the jth leaf node of the current tth decision tree before splitting is recorded as O non (t), the objective function of the decision tree after the leaf node is split at the split point of the dth feature dimension is recorded as , then the information gain corresponding to the jth leaf node is The calculation method is:

[0062] ;

[0063] Among them, G represents the sum of the first-order partial derivatives of the samples contained in the leaf node before splitting, G L It represents the sum of the first-order partial derivatives of the samples contained in the left node after the leaf node is split at the corresponding split point of the d-th feature dimension, G R It represents the sum of the first-order partial derivatives of the samples contained in the right node after the leaf node is split at the corresponding splitting point of the d-th feature dimension. H represents the sum of the second-order partial derivatives of the samples contained in the leaf node before the split. H L It represents the sum of the second-order partial derivatives of the samples contained in the left node after the leaf node is split at the corresponding split point of the d-th feature dimension, H R Represents the sum of the second-order partial derivatives of the samples contained in the right node after the leaf node is split at the corresponding splitting point of the d-th feature dimension.

[0064] It should be noted that if When the value of is greater than the preset splitting threshold, the corresponding splitting point is included in the candidate splitting point set; if When the value of is not greater than the preset splitting threshold, the corresponding splitting point is not included in the candidate splitting point set. The candidate splitting point set is obtained by screening all the splitting points of the j-th node of the current t-th decision tree in all feature dimensions. If the candidate splitting point set is an empty set, the leaf node is not split. If the candidate splitting point set is not an empty set, the leaf node is split and the splitting point corresponding to the maximum information gain in the candidate splitting point set is taken as the splitting method for the leaf node.

[0065] In this embodiment, the preset depth threshold and the preset splitting threshold can be obtained by presetting.

[0066] In this embodiment, the physiological characteristics of the target patient can be acquired in real time to monitor the risk of malnutrition in the target patient in real time. Nutritional support examinations and treatments can be reasonably arranged based on the risk of malnutrition in the target patient, thereby improving the effective utilization of medical resources. During the training process of the risk prediction model, the information gain of the greedy algorithm is used to automatically determine whether the leaf nodes in the current decision tree should continue to split and the specific splitting method to obtain the non-leaf nodes (i.e., intermediate nodes) of the decision tree, thereby realizing automated feature engineering. This system not only reduces labor costs and the difficulty of model training when training the risk prediction model, but also makes the trained risk prediction model more in line with the current characteristics of the medical institution.

[0067] In one feasible embodiment, the target physiological characteristics include historical observation characteristics and target observation characteristics. The historical observation characteristics include the actual physiological characteristics of the target patient before the preset observation time during the current treatment process. The target observation characteristics include the actual physiological characteristics of the target patient at the preset observation time. The actual physiological characteristics include the values ​​of the patient's physiological characteristics that can be directly detected by the existing professional technology and / or detection equipment of the medical institution, of course, it is not limited to this.

[0068] In this embodiment, refer to the attached Figure 2 and attached Figure 3 As shown, the information acquisition module includes a target acquisition submodule and a reference acquisition submodule. The target acquisition submodule is used to obtain the target observation characteristics of the target patient. For example, if it is necessary to determine whether the target patient is at risk of malnutrition at a first time point, the target acquisition submodule obtains the physiological characteristics of the target patient at the preset observation point. The reference acquisition submodule is used to obtain the historical observation characteristics of the target patient.

[0069] It should be noted that the preset observation time can be obtained in a preset manner. For example, if the observation interval is preset to be observed every 2 hours, it can be calculated from the start of monitoring the nutritional support of the target patient. The first preset observation time is the time when monitoring starts, and the corresponding physiological characteristics collected are the target observation characteristics of the first observation, and there is no corresponding historical observation characteristic; the second preset observation time is 2 hours after the start of monitoring, and the corresponding physiological characteristics collected are the target observation characteristics of the second observation, and the target observation characteristics of the first observation are updated to the historical observation characteristics of the second observation.

[0070] In this embodiment, the system further includes a feature prediction module, which is used to predict the predicted physiological characteristics of the target patient at the preset observation time based on the actual physiological characteristics of the target patient before the preset observation time.

[0071] It should be noted that the neural network model can be trained by taking the physiological characteristics of historical patients at different time points in the historical medical records, and the actual physiological characteristics of the target patient before the preset observation time are input into the trained neural network model to obtain the predicted physiological characteristics.

[0072] In this embodiment, the risk output submodule also includes a risk calculation submodule and a risk verification submodule. The risk calculation submodule is used to input the predicted physiological characteristics into the risk prediction model before the preset observation time to obtain the predicted risk of the target patient. The risk verification submodule is used to calculate the feature similarity between the predicted physiological characteristics and the actual physiological characteristics collected by the target acquisition submodule at the predicted observation time; compare the feature similarity with the similarity threshold; if the feature similarity is not less than the similarity threshold, the corresponding predicted risk is deemed valid, and the predicted risk is used as the target risk to reduce the amount of calculation required for concurrent responses of the system at the preset observation time, especially in large-scale deployment and use, which reduces the server load, lowers the minimum requirements for system operation, and helps optimize system operation costs; if the feature similarity is less than the similarity threshold, the target observation characteristics are sent to the risk calculation submodule for input, and the target risk is output to reduce the misjudgment of the target patient's malnutrition risk due to the erroneous prediction results of the feature prediction model.

[0073] It should be noted that the feature similarity can be calculated based on cosine similarity, Euclidean distance, etc. The similarity threshold can be obtained by presetting. For example, the similarity threshold is pre-set to 80%.

[0074] In this embodiment, the system predicts the physiological characteristics of the target patient at a preset observation time in advance through the feature prediction module, and inputs the predicted physiological characteristics output by the feature prediction module into the risk prediction model to obtain the predicted risk in advance. When the actual physiological characteristics of the patient at the preset observation time meet the requirements and the predicted physiological characteristics, the corresponding predicted risk is deemed valid to reduce the amount of calculation of the system's concurrent response at the preset observation time, especially in large-scale deployment and use, which reduces the server load and lowers the minimum requirements for system operation, thereby helping to optimize the system's operating costs.

[0075] In a feasible embodiment, refer to the attached Figure 2As shown, the risk verification module also includes screening target observation feature components from the target observation features according to the feature items corresponding to the intermediate nodes of all decision trees in the currently trained risk prediction model, and screening predicted physiological feature components from the predicted physiological features, that is, the target observation feature components are the values ​​of the feature items corresponding to the intermediate nodes of all decision trees in the current risk prediction model of the target patient at the preset observation time, and the predicted physiological feature components are the values ​​of the feature items corresponding to the intermediate nodes of all decision trees in the current risk prediction model of the target patient output by the feature prediction module.

[0076] In this embodiment, the feature similarity is calculated based on the similarity between the predicted physiological feature component and the target observed feature component.

[0077] It should be noted that the intermediate nodes of the decision tree in the risk prediction model serve as the basis for the corresponding decision tree's judgment. The structure of all decision trees in the trained risk prediction model is known, meaning that the basis for risk judgment in the current risk prediction model is known. When the differences between the predicted physiological characteristics and the target observed characteristics are more concentrated in the feature items corresponding to all intermediate nodes, the probability of a discrepancy between the predicted risk and the actual risk is greater. When the differences between the predicted physiological characteristics and the target observed characteristics are more concentrated in features other than the feature items corresponding to all intermediate nodes, the probability of a discrepancy between the predicted risk and the actual risk is smaller. Therefore, the feature similarity calculated based on the similarity between the predicted physiological characteristic components and the target observed characteristic components not only reduces the computational effort of the risk verification module, further reducing the computational effort and system burden, but also lowers the prediction accuracy requirements for the feature prediction module.

[0078] In this embodiment, the feature items corresponding to all intermediate nodes in the risk prediction model are combined into a feature set, which is recorded as F. The method for calculating the feature similarity includes:

[0079] ;

[0080] Where m represents the number of all feature items in the feature set; α i represents the feature weight of the i-th feature item, F 1，i represents the value of the i-th feature item of the target observation feature component in the feature set, F 2，i represents the value of the predicted physiological feature component in the feature set, N i It represents the difference between the maximum value and the minimum value of the i-th feature item in the feature set in the historical medical records.

[0081] It should be noted that the gap between the target observed feature component and the predicted physiological feature component in the corresponding feature item is obtained by taking the difference between the maximum value and the minimum value of each feature item in the feature set in the historical medical records, so as to reduce the influence of different feature items on the prediction validity due to different dimensions.

[0082] It should be noted that the feature weights can be pre-set according to the feature importance.

[0083] In this embodiment, the feature weight α i The calculation methods include:

[0084] ;

[0085] Among them, M represents the number of all intermediate nodes in the risk prediction model, M i Indicates the number of intermediate nodes corresponding to the i-th feature item in the feature set, that is, there are M in the risk prediction model. i The intermediate nodes are judged according to the i-th feature item of the feature set.

[0086] It should be noted that the trained risk prediction model has the characteristics of the medical institution itself. The intermediate nodes in the decision tree generated by the risk prediction model that serve as the basis for risk judgment are in line with the current professional and technical level and equipment characteristics of the medical institution. When a feature item appears in multiple intermediate nodes of multiple decision trees, it means that the feature item plays an important role in the decision-making process of the model. Giving higher weights to features with more corresponding intermediate nodes can not only be used to describe the importance of the corresponding feature items, but also when the values ​​of feature items with larger feature weights change, that is, the values ​​of the target observed feature components in the corresponding feature items are different from the values ​​of the predicted physiological feature components in the corresponding feature items, there is a greater probability that it will affect the output results of the risk prediction model. The corresponding predicted risk should be identified as invalid as a priority, so as to achieve the accuracy of the output target risk while reducing the computational complexity of the system's concurrent response.

[0087] Refer to the attached Figure 4 As shown, an embodiment of the present application provides an electronic device, including: a processor and a memory, the processor and the memory are interconnected and communicate with each other through a communication bus and / or other forms of connection mechanisms (not shown), the memory stores a computer program executable by the processor, and when the computing device is running, the processor executes the computer program to execute the system in any optional implementation mode of the above embodiment.

[0088] An embodiment of the present application provides a storage medium, wherein when the computer program is executed by a processor, the system of any optional implementation of the above embodiment is executed. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0089] In the embodiments provided in this application, it should be understood that the disclosed system can be implemented in other ways. The system embodiments described above are merely schematic. For example, the division of the modules is only a logical function division, and can be implemented in another way. For example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interface, indirect coupling or communication connection of the system or unit, which can be electrical, mechanical or other forms.

[0090] In addition, the units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0091] Furthermore, the functional modules in the various embodiments of the present application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0092] Flowcharts are used herein to illustrate the steps of the methods of the embodiments of the present disclosure. It should be understood that the preceding or following steps do not necessarily need to be performed in exact order. Instead, the various steps may be evaluated in reverse order or simultaneously. Furthermore, other operations may be added to these processes.

[0093] Unless otherwise defined, all terms used herein have the same meaning as commonly understood by those skilled in the art to which the present disclosure belongs. It should also be understood that terms such as those defined in common dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant art and should not be interpreted in an idealized or highly formal sense unless expressly defined as such herein.

[0094] The above is a detailed introduction to the critically ill patient malnutrition risk prediction system based on XGBoost. This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only an embodiment of this application. It is only used to help understand the critically ill patient malnutrition risk prediction system based on XGBoost of this application, and is not used to limit the scope of protection of this application. At the same time, for those skilled in the art, this application can have various changes and variations. Any modifications and equivalent substitutions made within the spirit and principles of this application should be included in the scope of protection of this application.

Claims

1. The malnutrition risk prediction system for critically ill patients based on XGBoost is characterized by: Including information acquisition module and risk prediction module; The information acquisition module is used to obtain target physiological characteristics of the target patient; The risk prediction module also includes a model construction submodule and a risk output submodule; the model construction submodule also includes a parameter acquisition submodule, a model generation submodule and a model training submodule; the parameter acquisition submodule is used to obtain the training data set and target parameter values ​​of the risk prediction model; The training dataset includes the physiological characteristics of all patients to be screened; the model generation module is used to build a risk prediction model based on the target parameter values ​​and the XGBoost model; the risk prediction model includes several decision trees; the model training module is used to generate each decision tree based on the training dataset and the target parameter values ​​using a strategy that maximizes information gain; The information gain is obtained based on the difference between the minimum value of the objective function of the decision tree before the update and the minimum value of the objective function of the decision tree to be updated. The weight of each leaf node is obtained based on the minimum value of the objective function of the generated decision tree. The risk output submodule is used to obtain the trained risk prediction model and obtain the target risk of the target patient based on the output of the risk prediction model based on the target physiological characteristics; The target parameters include the learning task parameters, learning rate, maximum depth of each tree, proportion of randomly selected samples when training each tree, proportion of randomly selected features when training each tree, minimum sample weight sum required for node splitting, minimum loss reduction value for node splitting, penalty parameter for model complexity, and performance evaluation index of the model; the learning task parameters include loss function; The information acquisition module divides the target physiological characteristics into historical observation characteristics and target observation characteristics according to whether the acquisition time is the preset observation time; the system also includes a feature prediction module, which is used to predict the predicted physiological characteristics of the preset observation time based on the historical observation characteristics of the target patient. The feature prediction module is constructed based on a neural network model trained based on the physiological characteristics values ​​of historical patients at different time points in the historical medical records; the risk output submodule also includes a risk calculation submodule and a risk verification submodule; The risk calculation module is used to obtain the predicted risk based on the predicted physiological characteristics before the preset observation time; The risk verification module is used to calculate the feature similarity between the predicted physiological features and the target observation features at a preset observation time. If the feature similarity is not less than the similarity threshold, the predicted risk is used as the target risk. If the feature similarity is less than the similarity threshold, the risk verification module is also used to input the target observation features into the risk calculation module and output the target risk.

2. The system according to claim 1, wherein: The risk verification module also includes screening target observation feature components from target observation features and screening predicted physiological feature components from predicted physiological features according to corresponding feature items of intermediate nodes of all decision trees in the trained risk prediction model; The feature similarity is calculated based on the similarity between the predicted physiological feature components and the target observed feature components.

3. The system according to claim 1 or 2, characterized in that When training each decision tree, the proportion of randomly selected samples is less than 1. Before training each decision tree, the model training module obtains the current hit probability of each sample data in the training data set and increases the hit probability of the sample data that was incorrectly predicted by the previous decision tree.

Images