Unplanned re-admission risk prediction method and system

Abstract

The invention provides an unplanned re-admission disease risk prediction method and system. The method comprises the following steps: constructing a dynamic classifier selection framework; setting binary variables representing medical diagnosis as a rule set, and extracting a plurality of data subsets from the data set by taking the rule set as an extraction rule; training a plurality of basic classifiers based on the data subset to obtain a classifier pool; defining a capability region and standards, namely meta-features, for measuring the capability of each basic classifier in the classifier pool in the capability region; extracting meta-features in the data set in the capability region by using a basic classifier; using the meta-features as a training set of a meta-classifier in a training dynamic classifier selection framework to train the meta-classifier; calculating the capability score of each basic classifier in the classifier pool by using the meta classifier; selecting the basic classifier with the highest ability score as a classifier integration pool; and predicting a to-be-predicted sample in the trained dynamic classifier selection framework. The method has high prediction accuracy.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a method and system for predicting the risk of unplanned hospital readmission. Background Art

[0002] Cirrhosis is a serious disease and the final common path for most chronic liver diseases, including chronic hepatitis B, hepatitis C, alcohol-related liver disease, and non-alcoholic fatty liver disease. Compared with other chronic diseases, chronic liver disease leads to higher hospitalization rates, longer hospital stays, and more readmissions. Studies have reported that the readmission rate for patients with cirrhosis within 7 to 90 days after discharge ranges from 10% to 50%, most of which are unplanned readmissions. A large proportion of unplanned readmissions are avoidable. Early identification of high-risk patients can allow for prospective interventions to improve medical outcomes.

[0003] In recent years, electronic health records, or EHRs, have been widely used in various clinical and research fields, and machine learning algorithms have achieved great success in mining EHR data. Various clinical indicators captured in EHRs, such as laboratory results and vital signs, provide data support for the development of cutting-edge prediction models, which can provide useful guidance for clinical decision-making. For example, Kenneth Berman et al. conducted a study using data from inpatient hepatology departments at two large academic medical centers for multivariate analysis to identify variables associated with 30-day readmission and investigate their association with 90-day mortality in patients with advanced liver disease. Chang Hu, Vikram Amjur et al. used logistic regression (LR), kernel support vector machine (SVM), and random forest classifier (RF) to analyze the 30-day and 90-day readmissions of a group of inpatients with cirrhosis from the North American Consortium for the Study of End-Stage Liver Disease (NACSELD). Various other machine learning methods, including Bayesian conditional probability, deep learning, gradient boosting, and novel ensemble methods, have also been used to develop risk prediction models for hospital readmission.

[0004] Although electronic health records provide rich information, the diversity of patients and data-related issues such as high dimensionality, incompleteness, sparsity, and temporal dynamics still bring many challenges that have not yet been fully addressed. A single machine learning model using the same feature set is usually not sufficient to effectively address the above challenges, so a multi-classifier system (MCS) is proposed.

[0005] Multi-classifier systems typically perform better than a single model by combining the outputs of multiple base classifiers. The principle of this technique is that not every classifier is an expert on all unknown samples. Instead, each base classifier is an expert in a different local area in the feature space. The development of a multi-classifier system usually includes three key steps: classifier pool generation, classifier selection, and classifier aggregation. The goal of the classifier pool generation step is to create a diverse set of classifiers that can complement each other to reduce errors in prediction. In order to explore diversity, a common strategy is to train classifiers using a subset of instances or a subset of the feature space. Once the classifier pool is generated, the next key step is to select appropriate base classifiers for a specific task.

[0006] Dynamic ensemble selection (DES) is a class of ensemble learning techniques that perform on-the-fly selection of base classifiers during inference. Dynamic ensemble selection works by estimating the capabilities of classifiers in a pool of classifiers in a local region of feature space. The local region is usually defined by the neighborhood of the query sample.

[0007] Heterogeneous classifiers can be selected and fused for unknown test instances at inference time. In particular, the META-DES framework assumes that the dynamic classifier selection problem can be viewed as a meta-problem. After generating a pool of classifiers, meta-features associated with different base classifiers are extracted using the training and dynamic selection datasets. These meta-features are then used to train a meta-classifier to estimate whether the base classifiers are able to correctly classify samples. When a new test instance is given, meta-features are extracted from this test instance and input to the trained meta-classifier to estimate the capability level of each base classifier. Based on these estimated capabilities, the most capable classifiers are selected and integrated for the final decision process.

[0008] It is worth noting that applying existing dynamic ensemble selection algorithms is not necessarily superior to state-of-the-art ensemble models in tabular data classification, and developing data-driven methods for tasks such as readmission risk prediction is challenging from multiple perspectives. First, due to the complexity and diversity of hospitalization and discharge processes, commonly used machine learning models have difficulty achieving high accuracy in predicting readmission risk within a certain time frame. Second, real-world electronic health record data have inherent limitations and quality issues, including incompleteness, sparsity, and temporal dynamics. Existing technologies cannot overcome these problems. Summary of the invention

[0009] In order to overcome the defects in the above-mentioned prior art, the purpose of the present invention is to provide a disease risk prediction method and system for unplanned readmission.

[0010] In order to achieve the above-mentioned object of the present invention, the present invention provides a disease risk prediction method for unplanned readmission, comprising the following steps:

[0011] Obtain disease datasets;

[0012] A dynamic selection classifier is constructed based on the DES framework. The dynamic selection classifier includes several basic classifiers and a meta-classifier. The dynamic selection classifier is trained based on the data set:

[0013] Setting a number of interpretable binary variables representing medical diagnoses as a rule set, and extracting the rule set from the data set into a number of data subsets as extraction rules;

[0014] Training the basic classifier based on the data subset to obtain a classifier pool;

[0015] defining a capability region and a criterion for measuring the capability of each basic classifier in the classifier pool within the capability region, one of the criteria including the binary variable, and using the criterion as a meta-feature;

[0016] Extracting meta-features from a data set using the base classifier within the capability region;

[0017] Training the meta-classifier using the meta-features as a training set for training the meta-classifier;

[0018] Use the meta-classifier to calculate the ability score of each base classifier in the classifier pool when predicting each set of data;

[0019] The base classifier with the highest ability score is selected as the best classifier to predict the corresponding data;

[0020] The prediction result of the best classifier is taken as the final prediction result.

[0021] This method has high prediction accuracy and also allows the generation of classifier pools and dynamic selection of classifiers in an interpretable way. In the dynamic classifier selection stage, the method selects the classifier with the highest ability score for prediction, and selects an optimal classifier for each individual to predict the risk of readmission. When the selected base classifier itself is understandable and interpretable, its prediction can be directly explained and understood. For example, models such as logistic regression LR, decision tree DT, K nearest neighbor KNN, etc. are interpretable as base classifiers. If the base classifier is an uninterpretable model, SHAP can be used to explain the entire prediction result.

[0022] Optionally, part of the data in the data set is used as a training set, and a set of rules representing binary variables of medical diagnosis is extracted from each training sample, and all the rules form the rule set.

[0023] In an optional solution of the disease risk prediction method for unplanned readmission, the capability area includes a first capability area and a second capability area;

[0024] The first capability region is a neighborhood surrounding the predicted data sample, the data samples in the neighborhood are similar samples of the predicted data sample, and the first component feature is extracted in the first capability region;

[0025] The second capability region is data containing the binary variable based on medical diagnosis, and a second component feature is extracted within the second capability region.

[0026] In an optional solution of the disease risk prediction method for unplanned readmission, the first component features include a standard for measuring local accuracy within the first capability area, a degree of consensus within the first capability area, an overall accuracy within the first capability area, accuracy in the decision space, and a degree of confidence of the input sample;

[0027] The second set of meta-characteristics includes local accuracy within a second capability region.

[0028] Optionally, the interpretable binary variable representing a medical diagnosis is the presence or absence of comorbidities and / or complications.

[0029] Local accuracy within the second capability area is interpreted based on the presence or absence of complications and / or comorbidities.

[0030] In an optional scheme of the disease risk prediction method for unplanned readmission, if there are multiple base classifiers with the same ability score, the prediction results of such base classifiers are aggregated using the majority voting rule.

[0031] In an optional solution of the disease risk prediction method for unplanned readmission, before training, the data set is preprocessed, including calculating the P value of each variable in the data set;

[0032] When setting a binary variable representing a medical diagnosis, a variable with a P value less than the set value is selected.

[0033] This option can reduce the indicators that have a negative impact on the prediction results, thereby further improving the prediction accuracy.

[0034] The present application also proposes a disease risk prediction system for unplanned readmission, including a data acquisition unit for acquiring a disease data set, a processing unit, and a storage unit;

[0035] The data acquisition unit is communicatively connected to the processing unit and sends the data set to the processing unit. The processing unit is communicatively connected to the storage unit, and the storage unit is used to store at least one executable instruction. The executable instruction enables the processing unit to perform operations corresponding to the above-mentioned disease risk prediction method for unplanned readmission to predict the disease risk of unplanned readmission.

[0036] The beneficial effects of the present invention are:

[0037] The present invention combines the diversity of disease data with the dynamic selection of basic classifiers, and the proposed dynamic selection classifier improves the performance of predicting the risk of readmission. The rule-based generation of data subsets (patient subgroups) allows for interpretable training of diversified classifiers to support patient subgroup-specific decision making, that is, different patients use different models for prediction, for example, predicting whether they will be readmitted within 14 days and 30 days. Because different patients have different comorbidities and complications and the degree of occurrence, the dynamic selection of the most appropriate model for each patient is enhanced by using the local accuracy within the second capability region extracted by the second capability region, so that more accurate information acquisition can be achieved for different patients; different patients may use different classifiers or classifier combinations, and it is easier to achieve personalized predictions for patients, that is, select different classifiers according to different patient conditions, so as to obtain more accurate clinical prediction results. By addressing the challenges associated with real-world data, the invention contributes to the continued efforts of medical research, using the potential of real-world data to improve patient care and management, and is particularly suitable for identifying high-risk hospital readmission individuals in patients with cirrhosis.

[0038] Among them, by using different data subsets to train the base classifiers, the characteristics of the diverse feature selection of different base classifiers are utilized to solve the problems of incompleteness and sparsity of electronic health record data. By using different data subsets and training each base classifier, the classifier pool can capture the broad patterns in the electronic health record data, thereby enhancing the overall prediction performance.

[0039] In addition, by generating a pool of classifiers to obtain the most capable classifier, and using the most capable classifier to improve performance, the training and dynamic selection of each base classifier is made more interpretable and demonstrates its potential value in clinical settings. For example, interpretable rules used to generate data subsets for training base classifiers can accurately describe the clinical characteristics of data subsets (patient subgroups), that is, whether comorbidities and / or complications occur during hospitalization. These rules provide insights into specific features or conditions that affect predictions, which can support clinicians in making informed decisions and tailor interventions for different patient subgroups, that is, clinicians can develop strategies based on the different situations of different patients, such as some patients have complications, some patients have comorbidities, some patients have both, and some patients have none. In addition, there is information such as the severity of different patients, that is, "tailor-made". If multiple base classifiers are selected for prediction, analyzing the intersection of elements used by multiple rules can provide insights into characterizing data subsets (patient subgroups).

[0040] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0042] Figure 1 It is a schematic diagram of the framework for dynamically selecting a classifier;

[0043] Figure 2 This is the AUROC prediction curve for 14-day readmission;

[0044] Figure 3 This is the ROC prediction curve for 30-day readmission;

[0045] Figure 4 It is a SHAP summary diagram of the impact of input variables on the 14-day readmission rate;

[0046] Figure 5 The SHAP diagram shows the different predictions for 2 different patients, where A indicates that the patient has not been readmitted, or the probability of readmission is 0.41, and B indicates that the patient has been readmitted, and the probability of occurrence is 0.80. DETAILED DESCRIPTION

[0047] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0048] In the description of the present invention, unless otherwise specified and limited, it should be noted that the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a mechanical connection or an electrical connection, or it can be the internal connection between two components. It can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to the specific circumstances.

[0049] Embodiment 1

[0050] This embodiment takes liver cirrhosis as an example and provides a disease risk prediction method for unplanned readmission, including the following steps:

[0051] Get the disease dataset

[0052] In this example, cirrhosis data were collected from six tertiary hospitals in Chongqing, China based on electronic health records. The initial inclusion cohort included 13,476 patient records hospitalized between January 2011 and June 2020. These patients were diagnosed with cirrhosis according to ICD-9 or ICD-10. To refine the cohort, the following exclusion criteria were used: (1) patients hospitalized for cancer, malignant tumors, tuberculosis, and HIV (1906 cases); (2) patients lost to follow-up or died during hospitalization (151 cases); (3) patients without any readmission record (7150 cases); (4) patients with more than 15% missing data (1113 cases). The final cohort consisted of 3307 patients with at least one readmission record. Considering the complexity of factors affecting cirrhosis and the dynamics of the disease process, this example recorded the patient's demographic, etiology, medical history, and clinical examination information within 24 hours of the first visit.

[0053] The final dataset contained 3,307 patients who were readmitted in this example. The dataset included 2 demographic variables, 5 etiological variables, 35 comorbidity variables, 39 examination variables, 10 surgical variables, 7 medication variables, 1 comprehensive score variable, and 1 outcome variable. The dataset was preprocessed, including calculating the P value for each variable in the dataset. A complete list of indicators is shown in Table 1 in the Supplementary Materials. Of these patients, 1,121 (33.90%) were female, 2,186 (66.10%) were male, and the average age was 55 years. Among them, 423 (12.79%) were readmitted within 14 days and 879 (26.58%) were readmitted within 30 days. The association between readmission within 14 days and 30 days and a set of potential predictors was evaluated by univariate analysis. The Mann-Whitney U test was applied to quantitative predictors, while the χ2 test was used for binary predictors. 2 Test or Fisher's exact test. The indicators of the selected variables are those shown in Table 1.

[0054] Table 1: Characteristics of the 3307 patients readmitted

[0055]

[0056]

[0057]

[0058] Abbreviations: IQR: interquartile range. EMD, electrolyte metabolism disorder, refers to the imbalance of electrolytes (such as sodium, potassium, calcium, chloride, etc.) in the body, usually caused by electrolyte levels in the body exceeding the normal range. INR: International Normalization Ratio, INR = PTpatient / PTcontrol) ISI, PTpatient is the patient's prothrombin time, PTcontrol is the average normal prothrombin time, and ISI is the international sensitivity. The normal range of INR varies from hospital to hospital, laboratory to laboratory, and patient to patient, ranging from 0.8 to 1.2. MELD: Model for End-Stage Liver Disease, MELD = 9.57*ln(creatinine) + 3.78*ln(bilirubin) + 11.20*ln(INR) + 6.43. The unit of the patient's serum creatinine value is milligrams per deciliter (mg / dl), the unit of the patient's total bilirubin value is millimoles per liter (mg / l), and INR is the patient's international normalized ratio of coagulation. The higher the value, the higher the risk level.

[0059] Use part of the dataset as a training set and part as a test set.

[0060] A dynamic selection classifier is constructed based on the DES framework. The dynamic selection classifier includes several basic classifiers and a meta-classifier. In this embodiment, the dynamic selection classifier is developed based on the deslib library.

[0061] like Figure 1 As shown, the framework for dynamically selecting classifiers mainly focuses on two aspects: classifier pool generation and dynamic classifier selection.

[0062] Training phase

[0063] In this embodiment, a number of interpretable binary variables representing medical diagnoses are set as rule sets, and the rule sets are used as extraction rules to extract a number of data subsets from the training set.

[0064] When setting a binary variable representing a medical diagnosis, a variable with a P value less than a set value in the data set is selected. In this embodiment, the set value is preferably, but not limited to, 0.001, that is, not all variables in Table 1, such as gender, whether peritonitis occurs, whether hepatic encephalopathy occurs, albumin count, red blood cell count, total bilirubin, etc., are used. Only variables with a P value < 0.001 are selected, and the binary variable is selected from these variables, while variables with a P value > 0.001 are not used in this embodiment.

[0065] In the training phase, a set of rules representing binary variables of medical diagnosis was extracted from each training sample in the training set using 5-fold cross validation. There can be overlaps between these rules, that is, some features can be shared between different rules. In this embodiment, a total of 80 rules were extracted from the training set to form a rule set, ensuring that each rule covers at least 500 instances, i.e., training samples. These rules are used as patterns to identify specific features or conditions in patient records. By using these rules, data subsets (SG1, SG2, ..., SG n ), which can also be called patient subgroups.

[0066] The interpretable binary variable representing the medical diagnosis is preferably, but not limited to, the presence or absence of comorbidities and / or complications of the disease. For example, a rule may be: [(electrolyte metabolism disorder = True) and (hepatic encephalopathy = False)], or the presence or absence of other comorbidities and / or complications, and these rules together constitute a rule set.

[0067] Based on the data subset, several basic classifiers are trained. The training process uses the PyCaret library to effectively train the basic classifiers and adjust the hyperparameters, and obtains a classifier pool C = {c1, c2, ..., c m},c1,c2,...,c m are the basic classifiers.

[0068] Basic classifiers include but are not limited to: Naive Bayes, K-nearest neighbor classifier, logistic regression, linear discriminant analysis, and quadratic discriminant analysis; and ensemble models: Random Forest Classifier, Extreme Random Trees Classifier, Adaptive Boosting Classifier, Gradient Boosting Classifier, Lightweight Gradient Boosting Machine, and Extreme Gradient Boosting.

[0069] By using different data subsets and training base classifiers, the classifier pool can capture broad patterns in electronic health record data, thereby enhancing the overall prediction performance.

[0070] The purpose of dynamically selecting a classifier is to find one or more basic classifiers from the classifier pool C to classify a given test sample. This is achieved in the present embodiment by the following steps:

[0071] Define capability regions and standards for measuring the capability of each basic classifier in the classifier pool within the capability region, one of the standards being the binary variable, and use all the standards as meta-features, that is, the binary variable is also one of the meta-features; use the basic classifier within the capability region to extract the meta-features in the training set.

[0072] Among them, the capability area includes the first capability area θ j and the second capability region θ j'. Define the second capability area θ j ' is to effectively incorporate complications and / or comorbidities to characterize disease patients (cirrhosis patients in this example) and to improve dynamic classifier selection.

[0073] First capability area θ j It is a neighborhood around the predicted data sample. The data samples in this neighborhood are similar samples of the predicted data sample. It aims to capture the similar sample features around this predicted data sample, so as to estimate the performance of the basic classifier in this specific area. Define the first ability region θ j The methods include but are not limited to: K nearest neighbor and clustering based on the entire feature space.

[0074] In the first capability region θ j The first set of meta-features is extracted from the first set of meta-features; the first set of meta-features includes a standard f1 for measuring local accuracy within the first capability area, a degree of consensus f2 within the first capability area, an overall accuracy f3 within the first capability area, accuracy f4 in the decision space, and a degree of confidence f5 of the input sample. The method for extracting the first set of meta-features is to directly use the meta-classifier.

[0075] Second capability area θ j ' is the data containing the binary variable based on medical diagnosis, in the second capability area θ j '. The second component feature includes the local accuracy f6 in the second capability region. In this embodiment, the local accuracy f6 in the second capability region is the above-mentioned binary variable, and the local accuracy f6 in the second capability region is explained according to the presence or absence of complications and / or comorbidities. Therefore, the second capability region θ based on medical diagnosis j 'Focusing on complications and / or comorbidities, treating them as binary variables indicating the absence or presence of these conditions, local accuracy f6 within the second capability area is used to focus on a subset of data (subgroup of patients) with specific complications and / or comorbidities.

[0076] The second capability area θ j 'Contains an equal number of positive and negative samples to give a balanced estimate of the performance of the base classifier.

[0077] The base classifier c i and data sample x j The extracted meta-feature is denoted as v ij ={f1,f2,f3,f4,f5,f6}.

[0078] Assign labels to each meta-feature: For each meta-feature, if the classifier c i The data sample is correctly classified, and its label is α ij =1; otherwise αij = 0. Then the meta-features are used as the training set for training the meta-classifier to train the meta-classifier. In this embodiment, the meta-classifier is preferably but not limited to a classifier such as META-DES.

[0079] The trained meta-classifier is used to calculate the capability scores of each base classifier in the classifier pool. The capability scores are usually expressed as the probability estimated by the meta-classifier. The classifier with the highest capability score is selected as the classifier ensemble pool C that predicts the corresponding data. * When there is only one basic classifier with the highest ability score, the classifier integration pool C * There is only one basic classifier in the pool, which is the best classifier and its prediction result is the final prediction result. When there are multiple basic classifiers with the highest ability score, that is, the ability scores of multiple basic classifiers are the same and the highest, the classifier integration pool C * If there are multiple base classifiers in , the prediction results of these base classifiers are aggregated using the majority voting rule as the final prediction result.

[0080] Testing Phase

[0081] Extract meta-features from the test set in the classifier pool C, and use the trained meta-classifier to calculate the ability scores of each basic classifier in the classifier pool when predicting the test sample, thus obtaining the classifier integration pool C * When the classifier ensemble pool C * There is only one basic classifier in the ensemble pool C, which is the best classifier and its prediction result is used as the final prediction result. * If there are multiple base classifiers in , the prediction results of these base classifiers are aggregated using the majority voting rule as the final prediction result.

[0082] Prediction results

[0083] Unplanned readmissions within 14 and 30 days were assessed separately. The area under the receiver operating characteristic curve (AUROC) was Figure 2 and Figure 3 As shown in Figure 2, the higher the AUROC value, the stronger the discrimination ability of the model. In order to avoid the model from simply predicting the majority class and obtaining high accuracy in imbalanced classification, a threshold selection process based on sensitivity was adopted. The threshold selection for calculating the accuracy ensured that the sensitivity was greater than 0.5, indicating a focus on correctly identifying positive cases. As shown in Table 2, the average score and the maximum and minimum ranges observed during the cross-validation process are presented. The proposed method outperforms the baseline model in terms of both Accuracy and AUROC.

[0084] Table 2 Comparison of results of different prediction models

[0085]

[0086] The prediction results can be explained by interpretability techniques such as SHapley Additive exPlanations (SHAP). For example, Figure 4 showed that higher IBIL, D_TBIL, PT-INR, ascitic fluid concentrate, and lower D_ALB were associated with higher predicted outcomes in patient subgroups. Figure 5 Individual power plots for (A) patients who were not readmitted and (B) patients who were readmitted are shown. The predicted probability values ​​f(x) of the model were 0.41 and 0.80 for Patient 1 and Patient 2, respectively. For the first patient, TBA, PLT, glutathione, AST, and peritonitis had a negative impact on the predicted outcome, while Ca and HGB had a positive impact on the predicted outcome. For the second patient, Ca, HGB, gender, EMD, autoimmune disease, D_TBIL, and D_Na had a positive impact on the predicted outcome, while glutathione had a negative impact on the predicted outcome.

[0087] Embodiment 2

[0088] The present embodiment provides a disease risk prediction system for unplanned readmission, including a data acquisition unit for acquiring a disease data set, a processing unit and a storage unit; the data acquisition unit is communicatively connected to the processing unit and sends the data set to the processing unit, the processing unit is communicatively connected to the storage unit, and the storage unit is used to store at least one executable instruction, the executable instruction enables the processing unit to perform operations corresponding to the disease risk prediction method for unplanned readmission as described in Example 1, and predict the disease risk of unplanned readmission.

[0089] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0090] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A disease risk prediction method for unplanned readmission, characterized in that: The following steps are involved: Obtain disease datasets; A dynamic selection classifier is constructed based on the DES framework. The dynamic selection classifier includes several basic classifiers and a meta-classifier. The dynamic selection classifier is trained based on the data set: Setting a number of interpretable binary variables representing medical diagnoses as a rule set, and extracting the rule set from the data set into a number of data subsets as extraction rules; Training the basic classifier based on the data subset to obtain a classifier pool; defining a capability region and a criterion for measuring the capability of each basic classifier in the classifier pool within the capability region, one of the criteria including the binary variable, and using the criterion as a meta-feature; Extracting meta-features from a data set using the base classifier within the capability region; Training the meta-classifier using the meta-features as a training set for training the meta-classifier; Use the meta-classifier to calculate the ability score of each base classifier in the classifier pool when predicting each set of data; The base classifier with the highest ability score is selected as the best classifier to predict the corresponding data; The prediction result of the best classifier is taken as the final prediction result.

2. The disease risk prediction method for unplanned readmission according to claim 1, characterized in that: Part of the data in the data set is used as a training set, and a set of rules representing binary variables of medical diagnosis is extracted from each training sample, and all the rules form the rule set.

3. The disease risk prediction method for unplanned readmission according to claim 1, characterized in that: The capability area includes a first capability area and a second capability area; The first capability region is a neighborhood surrounding the predicted data sample, the data samples in the neighborhood are similar samples of the predicted data sample, and the first component feature is extracted in the first capability region; The second capability region is data containing the binary variable based on medical diagnosis, and a second component feature is extracted within the second capability region.

4. The disease risk prediction method for unplanned readmission according to claim 3, characterized in that: The first set of meta-features includes a standard for measuring local accuracy within the first capability region, a degree of consensus within the first capability region, overall accuracy within the first capability region, accuracy in the decision space, and a degree of confidence in the input sample; The second set of meta-characteristics includes local accuracy within a second capability region.

5. The disease risk prediction method for unplanned readmission according to claim 1, characterized in that: The interpretable binary variables representing medical diagnoses were the presence or absence of comorbidities and complications.

6. The disease risk prediction method for unplanned readmission according to claim 4, characterized in that: Local accuracy within the second capability area was interpreted based on the presence or absence of complications and comorbidities.

7. The disease risk prediction method for unplanned readmission according to claim 1, characterized in that: If there are multiple base classifiers with the same and highest ability scores, the prediction results of such base classifiers are aggregated using the majority voting rule as the final prediction result.

8. The disease risk prediction method for unplanned readmission according to claim 1, characterized in that: SHAP was used to interpret the prediction results of the dynamic selection classifier.

9. The disease risk prediction method for unplanned readmission according to claim 1, characterized in that: Before training, the data set is preprocessed, including calculating the P value of each variable in the data set; When setting a binary variable representing a medical diagnosis, a variable with a P value less than the set value is selected.

10. A disease risk prediction system for unplanned readmission, characterized in that: including a data acquisition unit, a processing unit and a storage unit for acquiring a disease data set; The data acquisition unit is communicatively connected to the processing unit and sends the data set to the processing unit. The processing unit is communicatively connected to a storage unit, and the storage unit is used to store at least one executable instruction. The executable instruction enables the processing unit to perform operations corresponding to the disease risk prediction method for unplanned readmission as described in any one of claims 1 to 9, and predict the disease risk of unplanned readmission.

Images