Clinical auxiliary pre-judgment system and method for lower respiratory tract infection of COPD (chronic obstructive pulmonary disease) patient

Abstract

The invention belongs to the technical field of clinical and biomedicine, and particularly relates to a clinical auxiliary prejudgment system and prejudgment method for lower respiratory tract infection of a critical patient with chronic obstructive pulmonary disease (COPD). The clinical auxiliary pre-judgment system for the lower respiratory tract infection of the COPD patient comprises a constructed multi-label clinical prediction model and an online interactive auxiliary pre-judgment model, synchronous and rapid pre-judgment of four types of lower respiratory tract fungi, GNB, GPB and MDRO infection of the COPD critical patient can be achieved, and overuse of antibacterial drugs is effectively controlled.

Description

Technical Field

[0001] The present invention belongs to the field of clinical and biomedical technology, and specifically relates to a clinical auxiliary prediction system and prediction method for lower respiratory tract infection in patients with severe chronic obstructive pulmonary disease. Background Art

[0002] Chronic obstructive pulmonary disease (COPD) is a common chronic respiratory disease with a high prevalence and heavy disease burden. Acute exacerbation of COPD is a common acute event in the development of COPD, which is closely related to the decline of lung function, the decline of quality of life, the increase of medical expenses, and the increase of death risk in COPD patients. Therefore, how to reduce the occurrence of acute exacerbation and prevent the impact of subsequent events is included as one of the important treatment goals of COPD. Clinically, multiple lower respiratory tract infections (LRTI) are the main causes of acute exacerbation of COPD, among which bacterial infections are very common and have a huge impact. Therefore, antibiotics play a vital role in the treatment of acute exacerbation. Long-term use of antibiotics and glucocorticoids can also easily cause fungal infections. Therefore, fungi are also one of the common microbial types of LRTI in COPD patients.

[0003] At present, the primary task of diagnosing and treating LRTI in patients in clinical work is to identify the type of pathogenic microorganism. Microbial culture of cytological samples is one of the routine clinical detection methods and is often used as the primary detection method for detecting various bacterial and fungal infections. Except for rare species, most bacterial and fungal infections can be accurately diagnosed by cytological specimen testing. However, the growth rate of various bacteria and fungi ranges from 1-3 days to 2-3 weeks, which greatly affects the turnaround time (TAT) of microbial culture. For the clinical treatment of critically ill patients, it is usually necessary to make a quick decision on the initial treatment of LRTI before the pathogen is clearly identified. Therefore, in actual work, the initial treatment of LRTI mainly relies on the empirical judgment of clinicians.

[0004] However, this empirical judgment is often accompanied by improper use and / or overuse of antibiotics, which increases the risk of infection and transmission of multidrug-resistant bacteria. Not only will the treatment effect of COPD patients be significantly negatively affected, but it will also pose a general threat to the health of other populations, causing serious impacts in multiple aspects and levels. For pathogens, improper use and / or overuse of antibiotics will increase the natural selection pressure of pathogens, leading to the development and continuous increase of pathogen resistance, and then the emergence of multidrug-resistant bacteria commonly known as superbugs. For treatment, the continuous development of pathogen resistance has led to the gradual ineffectiveness of many first-line drugs, the challenges faced by anti-infection treatment are increasing, and the development of new drugs is progressing slowly. At the individual patient level, improper use and / or overuse of antibiotics are often accompanied by the occurrence of adverse drug events, including allergic reactions involving organ systems such as the skin and subcutaneous tissue, and Clostridium difficile infection with intestinal symptoms as the main symptoms after the normal flora is disturbed.

[0005] In order to effectively control the above problems and achieve rapid prediction of LRTI, many academic studies are usually divided into two directions. The first direction is to summarize the diagnosis and treatment experience of LRTI through a large amount of clinical practice and research, continuously form new authoritative consensus in the field, and improve and update the professional knowledge system. Clinical decisions based on evidence-based medical knowledge have high accuracy, but the generation of clinical experience often requires a lot of practice and research accumulation. Relevant knowledge and consensus often need to be updated through semi-structured interviews, Delphi surveys and other methods. The process requires experts and scholars to spend a lot of time and energy, and they cannot respond to the rapid development of pathogen resistance in time. With the continuous accumulation of medical and health big data, biomedical prediction technology with machine learning as the core has become increasingly mature, and the second major direction of using infection prediction models to assist medical staff in achieving rapid clinical diagnosis has emerged.

[0006] However, it is worth noting that in clinical work, empirical treatment of LRTI (especially for critically ill patients) mostly uses broad-spectrum antimicrobial drugs covering multiple infection categories. This further increases the risk of increasing pathogen resistance and gradual failure of first-line anti-infective drugs, and there is an urgent need for auxiliary tools that can predict multiple infection categories. However, there are currently few prediction models that can achieve simultaneous and rapid prediction of multiple LRTI categories in critically ill COPD patients.

[0007] At present, infection-related prediction models usually have the following characteristics: (1) Personalized application scenarios: The infection model developed by Narayana et al. is only applicable to the scenario of predicting infection in patients with recurrent nephrotic syndrome, the model developed by Li et al. is only applicable to the scenario of predicting urinary tract infection in children under 3 years old, and the model developed by Zhang et al. is only applicable to the scenario of predicting secondary infection in patients after cardiovascular surgery; (2) Focus on specific infection categories: Eickelberg et al. developed a prediction model for bacterial infection in critically ill patients, Wei et al. developed a prediction model for multidrug-resistant bacterial infection in hospitalized patients in the respiratory department, and Han et al. developed a prediction model for fungal infection in patients with acute exacerbation of COPD; (3) Focus on specific species of pathogens: The model developed by Shahin et al. is used to predict whether critically ill patients are infected with Candida fungi, and the research of Kim et al. focuses on predicting Escherichia coli and Klebsiella that produce extended-spectrum β-lactamase (ESBL).

[0008] Therefore, there is an urgent need to build an interactive decision support system based on multi-label machine learning in this field to assist clinical medical staff in the simultaneous and rapid diagnosis of multiple types of lower respiratory tract infections in patients with severe COPD. Summary of the invention

[0009] To this end, the technical problem to be solved by the present invention is to provide a clinical auxiliary prediction system for lower respiratory tract infection in patients with severe chronic obstructive pulmonary disease;

[0010] The second technical problem to be solved by the present invention is to provide a clinical auxiliary prediction method for lower respiratory tract infection in patients with severe chronic obstructive pulmonary disease.

[0011] In order to solve the above technical problems, the present invention provides a clinical auxiliary prediction system for lower respiratory tract infection in COPD patients, the system comprising a multi-label clinical prediction model;

[0012] The multi-label clinical prediction model is a multi-label clinical prediction model constructed based on the multi-model multi-label feature screening (MMFS) open algorithm framework and the composite multi-label imbalanced data processing (CMIDP) algorithm framework.

[0013] Specifically, the clinical auxiliary prediction system for lower respiratory tract infection in COPD patients, the lower respiratory tract infection includes Gram-negative bacteria (GNB) infection, Gram-positive bacteria (GPB) infection, multidrug-resistant bacteria (MDRO) infection and / or fungal infection of the lower respiratory tract.

[0014] Specifically, the clinical auxiliary prediction system for lower respiratory tract infection in COPD patients, the prediction indicators of the multi-label clinical prediction model include age, albumin (ALB), BT, aspartate aminotransferase to alanine aminotransferase ratio (AST / ALT), eosinophils (EOS), neutrophils (NEU), alkaline phosphatase (ALP), white blood cells (WBC), aspartate aminotransferase (AST), hematocrit (Hct), lactate dehydrogenase (LDH), urea nitrogen (BUN), blood glucose (GLU), prothrombin time (PT), blood chloride (Cl - ), blood calcium (Ca 2+ ), SB, PaO2, PaCO2, pH, TCO2, oxygen saturation (SaO2), body mass index (BMI), C-reactive protein (CRP), and lymphocytes (LYM).

[0015] Specifically, in the clinical auxiliary prediction system for lower respiratory tract infection in COPD patients, the prediction effect evaluation index of the multi-label clinical prediction model includes two types of indicators: example-based and label-based; wherein,

[0016] Example-based evaluation metrics include Hamming Loss, Zero-one Loss, Samples-precision, Samples-recall, and Samples-F1 score;

[0017] Label-based evaluation indicators include Micro-AUC, Macro-AUC, Micro-AP, and Macro-AP.

[0018] Specifically, the clinical auxiliary prediction system for lower respiratory tract infection in COPD patients, the prediction results of the system include simultaneous prediction of LRTI categories, and display the positive prediction probability of the corresponding category;

[0019] The LRTI categories include Gram-negative bacterial (GNB) infections, Gram-positive bacterial (GPB) infections, multidrug-resistant organisms (MDRO) infections, and / or fungal infections of the lower respiratory tract.

[0020] The present invention also discloses a method for constructing a clinical auxiliary prediction system for lower respiratory tract infection in COPD patients, comprising the following steps:

[0021] (1) Selecting infection categories that the system focuses on and performing partial correlation network analysis between different infection categories;

[0022] (2) Collect clinical information of patients and divide them into development samples and validation samples for subsequent data processing and modeling analysis;

[0023] (3) All patient data were analyzed for missing information and missing values ​​were filled, and the multi-model multi-label feature screening (MMFS) open algorithm framework was used to screen model features (i.e., predictor variables) and determine the indicators used for prediction;

[0024] (4) Based on the development samples, the composite multi-label imbalanced data processing (CMIDP) algorithm framework is used to process multi-label category imbalanced data to reduce data imbalance and obtain an optimized model;

[0025] (5) Model construction is performed based on the optimization model and combined with SHAP individual interpretability technology.

[0026] Specifically, in the method for constructing a clinical auxiliary prediction system for lower respiratory tract infection in COPD patients, in step (2), the patient's clinical information includes general condition, clinical diagnosis, physical examination, laboratory examination and / or microbial culture information.

[0027] Specifically, the method for constructing the clinical auxiliary prediction system for lower respiratory tract infection in COPD patients, in the step (3):

[0028] The missing analysis step uses Little's MCAR and partial correlation analysis test models;

[0029] The multi-model multi-label feature screening (MMFS) open algorithm framework steps are built based on scikit-learn using python, and based on recursive feature elimination (RFE) and permutation feature importance (PFI) algorithms, the personalized feature importance values ​​of all features are calculated for each label, and the average of the weighted personalized feature importance values ​​of all labels is used as the common feature importance value.

[0030] Specifically, in the method for constructing a clinical auxiliary prediction system for lower respiratory tract infection in COPD patients, in the step (4), the multi-label category imbalance processing step uses the three basic algorithms of REMEDIAL, MLSMOTE, and MLTL to propose a CMIDP algorithm framework, and forms four extended algorithms of REMLSMOTL, REMLSMOTE, REMLTL, and MLSMOTL.

[0031] The present invention also discloses a method for clinical auxiliary prediction of lower respiratory tract infection in COPD patients, comprising the steps of collecting clinical information of the patient to be tested, and performing prediction using the clinical auxiliary prediction system for lower respiratory tract infection in COPD patients.

[0032] The clinical auxiliary prediction system for lower respiratory tract infection in COPD patients described in the present invention uses the MIMIC-IV database, includes COPD severe hospitalized patients who meet the inclusion and exclusion criteria, and collects clinical diagnosis and treatment related data of the research subjects during suspected hospitalization from 2008 to 2019. In the study, a multi-model multi-label feature selection (MMFS) open algorithm framework and a compound multi-label imbalanced data processing (CMIDP) algorithm framework were proposed to provide technical support for solving feature selection and category imbalance problems in multi-label scenarios. Subsequently, three multi-label frameworks, Binary Relevance (BR), Classifier Chain (CC), and Label Powerset (LP), were combined with support vector machine (SVM), naive Bayes (NB), decision tree (DT), random forest (RF), LightGBM, XGBoost, and multi-layer perceptron (MLP), respectively, and MLkNN and HARAM were used to construct a total of 23 prediction models, which can predict four LRTIs, including fungi (Fungi), gram-negative bacteria (GNB), gram-positive bacteria (GPB), and multidrug-resistant bacteria (MDRO). The prediction model was developed in the development sample, and the model effect was verified in the verification sample. Finally, the model with the highest comprehensive performance was selected in combination with SHAP technology to assist in the construction of the prediction system. The multi-label decision support system constructed by the present invention can better predict Fungi infection, GNB infection, GPB infection, and MDRO infection in the lower respiratory tract of COPD patients with severe disease, and can control the excessive use of antimicrobial drugs to a certain extent. The system's high interactivity and interpretability can assist clinical medical staff in real-time simultaneous diagnosis of multiple LRTI categories.

[0033] In order to realize the prediction of multi-label outcomes, the clinical auxiliary prediction system for lower respiratory tract infection in COPD patients described in the present invention develops two algorithm frameworks, MMFS and CMIDP, in the feature screening stage and data imbalance processing stage respectively during the construction of the prediction model for multi-label task scenarios, thereby effectively realizing the construction of the prediction model.

[0034] The clinical auxiliary prediction system for lower respiratory tract infection in COPD patients described in the present invention is based on the optimal multi-label clinical prediction model, combined with the SHAP interpretable model and interactive web front-end technology. The prediction result of the infection type is obtained through the multi-label clinical prediction model, and the clinical interpretable decision curve of the prediction result is obtained through the SHAP interpretable model, and the above content is further interactively visualized through the web front-end technology.

[0035] The clinical auxiliary prediction system for lower respiratory tract infection in COPD patients described in the present invention uses real clinical medical record data to train and verify multiple multi-label models, selects the optimal prediction model and combines SHAP interpretability technology to build a visual decision support system, providing technical and tool support for the simultaneous prediction of multiple LRTI categories in severe COPD patients, and assisting clinical medical personnel to effectively control the improper use and / or excessive use of antibiotics.

[0036] The clinical auxiliary prediction system for lower respiratory tract infection in COPD patients described in the present invention makes full use of the relationship between multiple LRTI categories, uses basic information, laboratory indicators, blood gas analysis and other data of COPD critically ill hospitalized patients to build a multi-label clinical prediction model for synchronously predicting multiple infection categories, and uses an additive global interpretation method based on Shapley value to interpret the results of the developed prediction model.

[0037] The clinical auxiliary prediction system for lower respiratory tract infection in COPD patients described in the present invention considers the common characteristics and individual characteristics of each infection category, and proposes a multi-model multi-label feature selection (MMFS) open algorithm framework based on the permutation importance algorithm to construct a multi-label prediction model to screen features that are closer to clinical medical scenarios.

[0038] The clinical auxiliary prediction system for lower respiratory tract infection in COPD patients described in the present invention takes into account the possible impact of the coupling of majority class labels and minority class labels on the multi-label imbalance processing process, and reduces the impact of class imbalance in multi-label tasks on the model prediction effect. Based on the REMEDIAL, MLSMOTE, and MLTL algorithms, a compound multi-label imbalanced data processing (CMIDP) algorithm framework is proposed, and four multi-label imbalanced data processing algorithms, namely REMLSMOTL, REMLSMOTE, REMLTL, and MLSMOTL, are constructed to provide more sufficient technical support for improving the model prediction effect.

[0039] The clinical auxiliary prediction system for lower respiratory tract infection in COPD patients described in the present invention uses streamlit to build an online interactive clinical auxiliary prediction system to facilitate the promotion of the prediction model in the clinical environment and meet the needs of timeliness and explainability of the LRTI category of the target patient in the auxiliary diagnosis of clinical diagnosis and treatment. It is expected that any mobile Internet terminal in the clinical scenario can realize real-time prediction of the infection category.

[0040] The present invention realizes the synchronous and rapid prediction of four categories of lower respiratory tract fungal, GNB, GPB, and MDRO infections in COPD patients with severe illness by constructing a multi-label decision support system, and effectively controls the excessive use of antibiotics. The CC-LightGBM-based clinical auxiliary prediction system for LRTI in patients with severe COPD not only has significant predictive performance, but also can keep the misdiagnosis rate of uninfected patients at a low level. The study also found that BT, CRP, ALB, and LDH are important predictive indicators of the four LRTI categories. Smaller BMI values ​​are helpful for the prediction of GNB and MDRO infections when predicting the infection category, and smaller ALB values ​​are helpful for the prediction of Fungi, GNB, and MDRO infections.

[0041] The clinical auxiliary prediction method for lower respiratory tract infection in COPD patients described in the present invention has the following advantages: (1) It makes full use of low-cost indicators such as clinical routine examinations and clinical examinations, and realizes the synchronous and rapid prediction of multiple LRTI categories in severe COPD patients under the conditions of limited medical conditions, high diagnosis and treatment pressure, and long microbial culture TAT; (2) It proposes an MMFS open algorithm framework to realize the unified analysis and mining of common indicators and individual indicators with greater prediction importance for each infection category in multi-label tasks using multiple prediction models; (3) It proposes a CMIDP algorithm framework, which forms a variety of multi-label applications. (1) The algorithm for processing category imbalance in the multi-label task takes into full account the situations in which label coupling and label imbalance may affect the prediction performance of the downstream model; (2) The various algorithm technologies under the two major frameworks in the multi-label field are examined one by one, and the importance and significance of each indicator of the optimal model are visualized and explained in combination with the SHAP interpretability technology; (3) In combination with the individual interpretability technology of SHAP, an interactive decision support system is constructed to improve the interpretability of personalized prediction results and the generalizability of the model in clinical application, and to use the existing clinical networked equipment as a carrier to assist medical staff in making relevant decisions in real time online. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below according to specific embodiments of the present invention in conjunction with the accompanying drawings, wherein:

[0043] Figure 1The research design and technical framework flow chart for this invention;

[0044] Figure 2 The overall situation of missing value indicators in the present invention;

[0045] Figure 3 This is a heat map and significance analysis of the missing value correlation between the indicators with missing values ​​in the present invention. The missing value correlation between the corresponding indicators without coefficients indicated in the figure has no statistical significance;

[0046] Figure 4 Screening of research subjects of the present invention;

[0047] Figure 5 It is the partial correlation network between infection categories of each sample of severe COPD patients;

[0048] Figure 6 The importance values ​​of commonality and individuality features;

[0049] Figure 7 The screening overlaps corresponding to each infection category in all predictor variables;

[0050] Figure 8 It is the prediction effect of the multi-label model after being processed by the CMIDP framework's multiple category imbalance processing algorithms;

[0051] Fig. 9 To develop the ROC curves and PR curves of BR-LightGBM, BR-XGBoost, CC-RF, CC-XGBoost, LP-RF, LP-LightGBM, LP-XGBoost, and MLkNN in the sample;

[0052] Fig.10 To verify the ROC curves and PR curves of CC-LightGBM, LP-RF, LP-LightGBM and HARAM in the sample;

[0053] Fig.11 The feature importance and ranking of each indicator of CC-LightGBM;

[0054] Fig.12 The interactive interface of the clinical auxiliary prediction system and the example patient conditions. DETAILED DESCRIPTION

[0055] In the following embodiments of the present invention, the research data are all from the fourth-generation Medical Information Database for Intensive Care Medicine (MIMIC-IV), which contains various clinical diagnosis and treatment related information of inpatients admitted to the intensive care unit or emergency department of a tertiary academic medical center in Boston, Massachusetts, USA from 2008 to 2019. This research team has obtained legal authorization to use this database.

[0056] In the following embodiments of the present invention, the specific research design and technical roadmap are as shown in the attached Figure 1 shown.

[0057] In the following embodiments of the present invention, regarding statistical analysis and implementation software, in each research indicator, continuous variables are used

[0058] use For description, categorical variables were described using n(%). For the difference analysis of each indicator in the two samples, continuous variables were tested using t-test when variance homogeneity was met, otherwise t'-test was used, and categorical variables were tested using Pearson χ 2 Test or Fisher's exact probability test. The test level was set at 0.050.

[0059] In the following examples of the present invention, partial correlation network analysis was performed on all infection categories in each sample. For the difference analysis of each evaluation index of the model, one-way analysis of variance was used when normality and homogeneity of variance were met, and Kruskal-Wallis H test was used when they were not met. Bonferroni method was used to correct the test level for multiple comparisons.

[0060] In the following embodiments of the present invention, python 3.10, scikit-multilearn 0.2, scikit-learn1.1, lightgbm3.3, xgboost 1.7 and the like are used to implement model training and verification, and seaborn 0.12, shap 0.41, Gephi0.10 and the like are used to implement drawing of various result graphs of the model.

[0061] Example 1

[0062] This example screens data for the above research objects. For details, see the attached Figure 4 shown.

[0063] In this example, the data inclusion criteria include: (1) patients admitted to the hospital aged ≥18 years; (2) patients whose diagnosis includes COPD; and (3) patients suspected of LRTI.

[0064] In this embodiment, the data exclusion criteria include: (1) patients who are missing important clinical examination information, including laboratory examination information and lower respiratory tract microbial culture information; (2) patients who are missing 60% or more of the data overall.

[0065] After screening through the above procedures, a total of 14,050 patients with severe COPD met the inclusion criteria of this study. After excluding 10,874 patients who met the exclusion criteria, a total of 3,176 patients entered the analysis phase. The average age of all subjects was 72.20 years old, of which 1,756 (55.29%) were male patients. The sample size of the development sample was 2,540 (79.97%), and the validation sample was 636 (20.03%). The details are shown in Table 1. After testing, it was found that except for the admission season (χ 2 =7.950, p=0.047), blood glucose (GLU) (t=2.544, p=0.011), and there were no significant statistical differences in other indicators between the two samples, that is, p≥0.050.

[0066] Table 1 Basic information of the research subjects and differences in basic information between the two samples

[0067]

[0068]

[0069]

[0070] In this embodiment, the infection categories are defined as follows: the basic infection categories of concern include Gram-negative bacteria (GNB) infection, Gram-positive bacteria (GPB) infection, multidrug-resistant bacteria (MDRO) infection, and fungal infection of the lower respiratory tract. Multidrug-resistant bacteria (MDRO) refers to bacteria that are resistant to three or more types of antimicrobial drugs used in clinical treatment. The specific infection category is determined by testing the lower respiratory tract cytology specimens after microbial culture. Among them, the cytology samples include clinically collected patient sputum, bronchoalveolar lavage fluid, bronchoscope brushing or washing fluid, lung fine needle aspirate and pleural effusion.

[0071] In this example, the partial correlation network analysis results between different infection categories in each sample of COPD severe patients are shown in Figure 5 , where nodes represent infection categories, and lines between nodes represent partial correlations between infection categories, and their thickness is proportional to the absolute value of the partial correlation coefficient. Since the development sample and the validation sample are formed by random stratification of all samples, the partial correlation analysis results between different infection categories in each sample are the same. The test level adjusted by the Bonferroni method is 0.008. Figure 5It can be seen that there is a positive correlation between each infection category, and the partial correlation coefficient between GNB and MDRO infection is the largest (r=0.437, 95%CI: 0.400-0.470, p<0.001), followed by the partial correlation coefficient between GNB and MDRO infection (r=0.161, 95%CI: 0.120-0.200, p<0.001), and the rest are less than 0.100 (p<0.001).

[0072] Example 2

[0073] The research data of this embodiment is based on the electronic health records of patients during hospitalization, and fully considers all aspects of information in the clinical diagnosis and treatment process. The data content mainly includes general conditions, clinical diagnosis, physical examination, laboratory tests, and microbial culture. General conditions include patient demographic characteristics, admission season, etc. Physical examination includes patient height, weight, etc. Laboratory tests include various blood tests, urine routine tests, liver and kidney function tests, etc. Clinical diagnosis includes various diseases such as COPD, coronary heart disease, hyperlipidemia, asthma, etc. All kinds of clinical examination results are collected from the first record results at admission.

[0074] The data preprocessing method of this embodiment is as follows: All patients are randomly divided into development samples and verification samples in a ratio of 8:2, and then subsequent data processing and modeling analysis tasks are performed. According to the research of Charlson et al., the comorbidity index (Charlson Comorbidity Index, CCI) is calculated in combination with clinical diagnostic information. Among them, the body mass index (BMI) is calculated based on the patient's height and weight. In the clinical examination data, for continuous variables with upper and lower limits of reference values, data exceeding 3 times the interquartile range are regarded as outliers and are directly deleted.

[0075] Subsequently, the missing data of all patients were analyzed, focusing on describing and analyzing all indicators with missing values. In this embodiment, Little's MCAR test was used to analyze whether the overall missing mechanism of the data was completely random missing (MCAR), and the test level was set to 0.001. Correlation heat map and significance test were used to describe and analyze the missing correlation between each indicator, and the test level was set to 0.000084 according to the Bonferroni method.

[0076] In this embodiment, referring to the research of Beaulieu-Jones et al., considering that the missing indicators are all continuous variables in physical examination and laboratory examination, in order to ensure that the model can make the most of the indicators in the data set without reducing the performance, this embodiment selects to delete the indicators with a missing ratio greater than or equal to 50%, and uses the chained equation multiple imputation with predicted mean matching (MICE-PMM) algorithm to fill in the indicators with other missing ratios. In order to ensure the robustness of the chained interpolation results, the mean of the data after 100 interpolations is used as the final filling result.

[0077] After data processing, the overall situation of missing value indicators in the data of this implementation is shown in the attached Figure 2 As shown in the figure. It can be seen that the indicators included in the study with missing data are mainly continuous variables of laboratory tests and blood gas analysis, with a total of 35 items. Among them, there are 9 blood gas analysis variables, including carbon dioxide partial pressure (PaCO2), oxygen partial pressure (PaO2), plasma carbon dioxide content (TCO2), oxygen saturation (SaO2), pH, standard bicarbonate (SB), standard base excess (SBE), anion gap (AG) and body temperature (BT), and the remaining 26 items are laboratory test variables. From the perspective of variable missing, the missing proportion of each variable is different. Among them, the indicator with the largest missing proportion is BT (49.53%), and the smallest is red blood cells (RBC, 0.03%), hemoglobin (HB, 0.03%) and SB (0.03%). From the perspective of sample missing, there are also differences in the missing variables of each sample, among which the largest missing amount is 18 (51.43%) and the smallest is 0 (0%).

[0078] In the attached Figure 2 Based on the missing data shown in the figure, we further conducted a heat map of the missing data correlation and a significance analysis between the indicators with missing values. After Little's MCAR test, the χ2 value was 41912.710, p<0.001. Obviously, the test result rejected the hypothesis that the missing data mechanism was MCAR. The correlation coefficients between the indicators are shown in the attached figure. Figure 3 As shown in the figure, the correlation between the two indicators without coefficients is not significant (p≥0.000084). Figure 3 As can be seen from the results, there are correlations of varying sizes between the missingness of the indicators, so it can be considered that the missing mechanism of the data in this study is more inclined to random missing (MAR) as a whole.

[0079] Example 3

[0080] This embodiment adopts a comprehensive strategy to screen model features (ie, predictor variables).

[0081] The screening strategy is divided into two stages: (1) data-driven feature screening stage; (2) improvement stage combined with clinical expert opinions. In stage (1), the importance of predictive features is evaluated using the MMFS algorithm framework proposed in this invention, and features that meet the requirements are screened. In stage (2), based on the results of the previous stage and combined with the knowledge and experience of clinical experts, all screened indicators are reviewed and improved to form the final results of the prediction indicator screening of this study.

[0082] In this embodiment, the MMFS open algorithm framework is built using python based on scikit-learn, and the corresponding pseudo code is shown in Table 2. At the same time, in order to consider the common prediction features and individual prediction features of each label in the clinical medical scenario, MMFS calculates the individual feature importance values ​​of all features for each label based on the recursive feature elimination (RFE) and permutation feature importance (PFI) algorithms, and takes the average of the weighted individual feature importance values ​​of all labels as the common feature importance value.

[0083] Table 2 MMFS framework pseudo code

[0084]

[0085]

[0086] Furthermore, in order to obtain relatively robust feature importance results under relatively limited computing resources, this embodiment sets the base model combination of the MMFS framework to logistic regression, support vector machine and random forest, the number of cycles N is set to 100, and the weights of the personalized feature importance values ​​of each label (i.e., w1, w2, ..., w l ) are all set to 1. In this embodiment, the union of the top 10 features with the largest personalized importance values ​​corresponding to each label and the top 10 features with the largest common importance values ​​is selected as the prediction indicator set in the data-driven stage. It should be noted that this embodiment uses a Venn diagram to describe the MMFS feature screening.

[0087] This embodiment uses a comprehensive indicator screening strategy to determine a total of 25 prediction indicators. The importance values ​​of common and personalized features calculated in the data screening stage are as follows: Figure 6Through the data-driven screening stage, age, albumin (ALB), BT, aspartate aminotransferase to alanine aminotransferase ratio (AST / ALT), eosinophils (EOS), neutrophils (NEU), alkaline phosphatase (ALP), white blood cells (WBC), aspartate aminotransferase (AST), hematocrit (Hct), lactate dehydrogenase (LDH), urea nitrogen (BUN), blood glucose (GLU), prothrombin time (PT), blood chloride (Cl - ), blood calcium (Ca 2+ ), SB, PaO2, PaCO2, pH, TCO2 and oxygen saturation (SaO2), a total of 22 indicators.

[0088] Figure 7 The overlap of variables for subsequent modeling tasks corresponding to each infection category after the data-driven screening stage is shown. Among them, Fungi, GNB, GPB, and MDRO infections all contain ALB and BT, while the above four categories contain AST, Hct, Cl separately. - ,LDH,BUN,GLU,SB,PT,PaO2,Ca 2+ . During the clinical expert review and recommendation stage, three new indicators were included: body mass index (BMI), C-reactive protein (CRP), and lymphocytes (LYM). Therefore, under the comprehensive feature screening strategy, this study determined a total of 25 predictive variables and processed various downstream modeling tasks accordingly.

[0089] Example 4

[0090] This embodiment performs category imbalance processing based on the data processing results of the previous embodiment.

[0091] This embodiment uses the three basic algorithms of REMEDIAL, MLSMOTE, and MLTL to propose a CMIDP algorithm framework, and forms a total of four extended algorithms, namely REMLSMOTL, REMLSMOTE, REMLTL, and MLSMOTL.

[0092] Specifically, the above seven multi-label category imbalanced data processing algorithms were used in the development samples respectively, and CC-LightGBM, LP-RF, LP-LightGBM and HARAM multi-label prediction models were taken as examples to evaluate the effects of various imbalance processing algorithms in the verification samples.

[0093] This embodiment also uses Python to build a CMIDP algorithm framework based on scikit-learn, and the corresponding pseudo code is shown in Table 3 below.

[0094] Table 3 CMIDP framework pseudo code

[0095]

[0096]

[0097] Among them, in the extended algorithm, REMLSMOTL is composed of REMEDIAL, MLSMOTE and MLTL, REMLSMOTE is composed of REMEDIAL and MLSMOTE, REMLTL is composed of REMEDIAL and MLTL, and MLSMOTL is composed of MLSMOTE and MLTL. REMEDIAL decouples the majority class labels from the minority class labels to solve the problem that when the majority class labels and the minority class labels coexist in the same sample, the subsequent class imbalance processing and even the overall prediction effect of the model are affected. MLSMOTE and MLTL are two algorithms that use the K nearest neighbor (KNN) principle for oversampling and undersampling in multi-label tasks. Referring to the research suggestions for algorithm development, the number of nearest neighbors K in MLSMOTL is set to 5, K in MLTL is set to 1, and the threshold TH is set to:

[0098]

[0099] Among them, I is an indicator of the imbalance level of the quantitative data set with a value range of [0,1], which is specifically defined as:

[0100]

[0101] Among them, y is the label being analyzed, Y i is the label set of sample i, is the indicator function.

[0102] This embodiment performs MLTL undersampling while fully considering the existence of three categories of samples after MLSMOTE oversampling: minority class samples, majority class samples, and newly generated samples. The MLSMOTE and MLTL algorithms are fused and improved to form MLSMOTL. The corresponding pseudo code is shown in Table 4-5 below.

[0103] Table 4 Pseudo code of MLSMOTL algorithm

[0104]

[0105]

[0106] Table 5 EMLTL algorithm pseudo code

[0107]

[0108]

[0109] In this embodiment, the performance of each unbalanced data processing algorithm in different multi-label models is as follows: Figure 8 As shown in the figure, in order to be consistent with the direction of F1 score and AUC evaluation of superiority, the results of 1-Hamming Loss and 1-Zero-one Loss are shown on the corresponding axes. It can be found that most of the evaluation indicators of all the above models after MLSMOTE processing are large, and the individual prediction indicators of HARAM are very small after processing by the seven class imbalance processing algorithms. Based on this, the data processed by MLSMOTE is used for learning and training in the subsequent model development stage.

[0110] Example 5

[0111] In this embodiment, under-sampling is performed on majority class samples and newly generated samples, that is, the union of majBag and newBag is selected as the source of under-sampling samples.

[0112] Model training and validation

[0113] This embodiment uses two types of multi-label algorithm frameworks, algorithm adaptation and problem transformation, to construct a prediction model. Among them, the algorithm adaptation method includes two algorithm models, MLkNN and HARAM. MLkNN obtains the label set statistics of adjacent samples by identifying the K nearest neighbors of the target sample, and determines the label set of the target sample using the maximum a posteriori (MAP) principle. HARAM is based on the adaptive resonance theory (ART) neural network, adding an additional unsupervised fuzzy ART layer to achieve rapid classification of multi-label tasks. The problem transformation method includes three multi-label frameworks, Binary Relevance (BR), Classifier Chain (CC), and Label Powerset (LP). Among them, BR is a framework for converting multi-label tasks into multiple binary classification tasks, CC is a framework for converting multi-label tasks into a chain of classifiers that follow one another according to the Bayesian rule, and LP is a framework for converting multi-label tasks into multi-classification tasks according to the label combinations that appear in the data. The above problem transformation framework is combined with support vector machine (SVM), naive Bayes (NB), decision tree (DT), random forest (RF), LightGBM, XGBoost, multi-layer perceptron (MLP), and algorithm adaptive models MLkNN and HARAM to form 23 prediction models.

[0114] The development sample data was randomly divided into a training set and a test set in a ratio of 8:2. In the training set, the grid search algorithm under the five-fold cross-validation framework was used to learn and verify the hyperparameters of each model. The optimal parameter combination of each model was selected for performance evaluation in the test set, and the results were used as the model prediction effect of the development sample. At the same time, the performance of each model was evaluated in the validation sample. The 95% CI of all evaluation indicators was calculated using Bootstrap.

[0115] Prediction effect evaluation

[0116] This embodiment uses two types of indicators, example-based and label-based, to evaluate the model prediction effect. Example-based evaluation indicators include Hamming Loss, Zero-one Loss, Samples-precision, Samples-recall, and Samples-F1 score, and label-based evaluation indicators include Micro-AUC, Macro-AUC, Micro-AP, and Macro-AP. Among them, the larger the Hamming Loss and Zero-one Loss values, the lower the model effect, and the larger the values ​​of other indicators, the higher the model effect. The definitions of all evaluation indicators are introduced as follows:

[0117] (1) Hamming Loss

[0118]

[0119] is the predicted value of label j in sample i, y i,j is the corresponding true value, n is the total number of samples, and m is the total number of labels.

[0120] (2)Zero-one Loss

[0121]

[0122] is the predicted value of sample i, y i is the corresponding true value, and n is the total sample size.

[0123] (3)Samples-Precision

[0124]

[0125] is the set of predicted labels for sample i, Y i is the set of true labels of the corresponding samples, |·| is the total number of labels contained in the set, and n is the total number of samples.

[0126] (4)Samples-Recall

[0127]

[0128] is the set of predicted labels for sample i, Y i is the set of true labels of the corresponding samples, |·| is the total number of labels contained in the set, and n is the total number of samples.

[0129] (5)Samples-F1 score

[0130]

[0131] is the set of predicted labels for sample i, Y i is the set of true labels of the corresponding samples, |·| is the total number of labels contained in the set, and n is the total number of samples.

[0132] (6) Micro-AUC and Macro-AUC

[0133] The false positive rate (FPR) and true positive rate (TPR) are defined as FPR = FP / (FP+TN) and TPR = TP / (TP+FN), respectively, where TP, TN, FP, and FN are the number of true positive samples, true negative samples, false positive samples, and false negative samples, respectively. The horizontal axis of the receiver operating characteristic curve (ROC) is FPR, and the vertical axis is TPR. The area under the receiver operating characteristic curve (AUC) can be obtained by integrating the ROC corresponding to different thresholds x:

[0134]

[0135] All samples of all categories are converted into binary samples, and the overall TPR and FPR are calculated to obtain Micro-AUC. After calculating TPR and FPR of samples of each category, the average AUC of all categories is taken to obtain Macro-AUC.

[0136] (7) Micro-AP and Macro-AP

[0137] Average Precision (AP) is also called the area under the Precision-Recall curve, where Precision and Recall are defined as P = TP / (TP+FP) and R = TP / (TP+FN) respectively. The horizontal axis of the Precision-Recall curve (PRC) is Recall and the vertical axis is Precision, so the AP can be obtained by integrating the PRC corresponding to different thresholds x:

[0138]

[0139] Similarly, all samples of all categories are converted into binary samples, and the overall Precision and Recall are calculated to get Micro-AP. After calculating the Precision and Recall of each category of samples, the average of all category APs is taken to get Macro-AP.

[0140] The prediction effects and statistical tests of the 23 multi-label models constructed in this study in the development samples are shown in Table 6, and the test level correction for multiple comparisons is 0.002.

[0141] Table 6 Performance of various multi-label models in development samples,

[0142]

[0143]

[0144] The results show that all indicators are statistically different overall (p<0.001). In the results of multiple comparisons, for Hamming Loss, it was found that BR-XGBoost, CC-XGBoost, LP-LightGBM, and LP-XGBoost all had the highest performance, with no statistical difference (all p≥0.002). As for Zero-one Loss, CC-XGBoost, LP-LightGBM, and LP-XGBoost all had the highest performance, with no statistical difference (all p≥0.002). For Samples-Precision, the models CC-RF and LP-RF both reached the maximum mean of 0.999. For Samples-Recall, the model MLkNN reached the maximum mean of 0.944. For Samples-F1 score, the models BR-LightGBM, BR-XGBoost, CC-XGBoost, LP-LightGBM, and LP-XGBoost all had the highest performance, with no statistical difference (all p≥0.002).

[0145] In order to further compare the performance of BR-LightGBM, BR-XGBoost, CC-RF, CC-XGBoost, LP-RF, LP-LightGBM, LP-XGBoost and MLkNN, this example plots the ROC curves and PR curves of micro-average, macro-average and each infection category, and calculates the area under each curve. The results are shown in the attached figure. Fig. 9As shown. It can be seen that for Micro-AUC, the models LP-LightGBM and LP-XGBoost are both the best, with no statistical difference between the two [0.982 (95% CI: 0.982-0.983) vs 0.982 (95% CI: 0.981-0.982), p = 0.204]. For Macro-AUC, the models LP-LightGBM and LP-XGBoost are also the best, with no statistical difference between the two [0.980 (95% CI: 0.980-0.981) vs 0.980 (95% CI: 0.979-0.981), p = 0.368]. However, for Micro-AP, the LP-LightGBM model outperforms LP-XGBoost [0.963 (95% CI: 0.962-0.964) vs 0.962 (95% CI: 0.961-0.963), p = 0.045]. Similarly, for Macro-AP, the LP-LightGBM model outperforms LP-XGBoost [0.961 (95% CI: 0.959-0.962) vs 0.959 (95% CI: 0.958-0.960), p = 0.033]. In summary, LP-LightGBM performs best in the development sample.

[0146] In order to verify the analysis results in the development sample, the models were systematically compared in the validation sample. The results are shown in Table 7 below. The test level for multiple comparisons was also corrected to 0.002.

[0147] Table 7 Verifies the performance of each multi-label model in the sample.

[0148]

[0149]

[0150] It can be seen that for Hamming Loss, CC-LightGBM and LP-LightGBM were found to have the highest performance, with no statistical difference between the two (p ≥ 0.002). For Zero-one Loss, model LP-RF achieved the minimum mean of 0.370. For Samples-Precision, LP-RF was found to achieve the maximum mean of 0.997. For Samples-Recall, model HARAM achieved the maximum mean of 0.814. For Samples-F1 score, model CC-LightGBM achieved the maximum mean of 0.647.

[0151] In this example, various ROC curves and PR curves of CC-LightGBM, LP-RF, LP-LightGBM and HARAM were further plotted for further analysis. The results are shown in the attached Fig.10 . It can be seen that for Micro-AUC, the model CC-LightGBM is the best [0.733 (95% CI: 0.732 ~ 0.735)]. For Macro-AUC, the model CC-LightGBM is also the best [0.733 (95% CI: 0.731 ~ 0.734)]. For Micro-AP, the model CC-LightGBM is also the best [0.331 (95% CI: 0.329 ~ 0.334)]. Similarly, for Macro-AP, the model CC-LightGBM is the best [0.351 (95% CI: 0.348 ~ 0.353)]. In summary, although the best model LP-LightGBM in the development sample is not the best in the validation sample, it also has relatively good results. However, considering the universality of the model's prediction effect, the optimal model is finally determined to be CC-LightGBM, which performs best in the validation sample.

[0152] Example 6

[0153] In this example, in order to further analyze the importance and direction of the prediction indicators in the optimal model CC-LightGBM for the prediction of each infection category, the SHAP interpretability technology is combined and the violin global interpretation map is used for visualization. The results are shown in the attached Fig.11 .

[0154] The results showed that BT, CRP, ALB, and LDH were among the top 10 predictors of all infection categories. BT had a positive effect in all LRTI categories. CRP also had a positive effect, but was more important in predicting GNP, GPB, and MDRO infections, ranking 7th, 4th, 1st, and 1st in importance in the four categories. It is worth noting that ALB had a negative effect in Fungi, GNB, and MDRO infections (i.e., the smaller the ALB value, the greater the probability of predicting these three infections), while it had a positive effect in GPB infection. BMI was less important in Fungi and GPB infections, ranking 13th and 11th, respectively. In comparison, its overall value for predicting GNB and MDRO infections was slightly higher (9th and 5th). LYM was not important in the four infection categories, ranking 18th, 8th, 22nd, and 9th, respectively.

[0155] Example 7

[0156] This embodiment builds an interactive auxiliary prediction system based on the above research results.

[0157] This embodiment selects the model with the highest comprehensive performance to build an auxiliary prediction system. In order to solve the common black box problem in the interpretability of machine learning models, the SHAP technology is used to build a special interpretability framework for multi-label models. In order to facilitate the wide application of clinical practice, an interactive decision support system is specially built to achieve online synchronous prediction of various infection categories of target patients.

[0158] In this embodiment, the online interactive auxiliary prediction system can realize the rapid prediction of multiple LRTI categories of patients by inputting the clinical information of the target patients. Information can be input on the left panel of the system, and the right panel includes three parts: indicator description, prediction results, and explanation. Among them, the prediction results include simultaneous predictions of 4 LRTI categories, showing the positive prediction probability of the corresponding category. The explanation part uses the SHAP decision curve to explain the prediction results of each category, that is, from the perspective of game theory, it explains whether the prediction indicators of the corresponding IRTI category promote the increase or decrease of the final prediction probability of the category based on the baseline prediction probability.

[0159] In this embodiment, the clinical auxiliary prediction system interface and the situation of the example patient are shown in the attached Fig.12 After entering the relevant information of the patient, you can see that the system predicts the probability of Fungi, GNB, GPB, and MDRO infection to be 82.55%, 69.70%, 12.01%, and 19.53%, respectively. In the explanation section, you can see that BT, ALB, and EOS are the top three most important indicators for predicting Fungi infection in this patient, while BT, AST / ALT, and ALB are the top three most important indicators for predicting GNB infection in this patient.

[0160] It can be seen that the present invention realizes the synchronous and rapid prediction of four categories of lower respiratory tract fungal, GNB, GPB, and MDRO infections in COPD patients with severe illness by constructing a multi-label decision support system, and effectively controls the excessive use of antibiotics. The CC-LightGBM-based clinical auxiliary prediction system for LRTI in patients with severe COPD not only has significant predictive performance, but also can keep the misdiagnosis rate of uninfected patients at a low level. The study also found that BT, CRP, ALB and LDH are important predictive indicators of the four LRTI categories. Smaller BMI values ​​are helpful for the prediction of GNB and MDRO infections when predicting the infection category, and smaller ALB values ​​are helpful for the prediction of Fungi, GNB and MDRO infections.

[0161] Example 8

[0162] This embodiment performs clinical auxiliary prognosis of the lower respiratory tract based on the clinical patient condition.

[0163] For a 62-year-old with a BMI of 28.53 kg / m 2A severe COPD patient with Fungi infection and GNB infection in the lower respiratory tract. After clinical examination, it was determined that his ALB, BT, AST / ALT, EOS, NEU, ALP, WBC, AST, Hct, LDH, BUN, GLU, PT, Cl - , Ca 2+ , SB, PaO2, PaCO2, pH, TCO2, SaO2, CRP and LYM were 2.56 g / dL, 37.34°C, 3.4, 5%, 77%, 94 IU / L, 12.7 K / μL, 17 IU / L, 28.1%, 241.89 IU / L, 18 mg / dL, 104 mg / dL, 14.54 sec, 106 mEq / L, 7.9 mg / dL, 24 mEq / L, 69 mmHg, 75 mmHg, 7.09, 24 mEq / L, 75.61%, 100.82 mg / L and 1.29 K / μL, respectively.

[0164] After inputting the above data, the predicted probability of the clinical auxiliary prediction system for lower respiratory tract infection according to the present invention shows that the probability of Fungi infection, GNB infection, GPB infection and MDRO infection of the patient is 82.55%, 69.70%, 12.01% and 19.53%, respectively. The predicted results are Fungi infection and GNB infection, which are consistent with the actual situation of the patient. For details, see Fig.12 .

[0165] Obviously, the above embodiments are merely examples for the purpose of clear explanation, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the invention.

Claims

1. A clinical auxiliary prediction system for lower respiratory tract infection in COPD patients, characterized in that: The system includes a multi-label clinical prediction model; The multi-label clinical prediction model is a multi-label clinical prediction model constructed based on a multi-model multi-label feature screening open algorithm framework and a composite multi-label unbalanced data processing algorithm framework.

2. According to claim 1, the clinical auxiliary prediction system for lower respiratory tract infection in COPD patients is characterized by: The predictive indicators of the multi-label clinical prediction model include age, albumin, BT, aspartate aminotransferase to alanine aminotransferase ratio, eosinophils, neutrophils, alkaline phosphatase, white blood cells, aspartate aminotransferase, hematocrit, lactate dehydrogenase, urea nitrogen, blood glucose, prothrombin time, blood chloride, blood calcium, SB, PaO2, PaCO2, pH, TCO2, oxygen saturation, body mass index, C-reactive protein and / or lymphocytes.

3. The clinical auxiliary prediction system for lower respiratory tract infection in COPD patients according to claim 1 or 2, characterized in that: The prediction effect evaluation index of the multi-label clinical prediction model includes two types of indexes: example-based and label-based; wherein, Example-based evaluation metrics include Hamming Loss, Zero-one Loss, Samples-precision, Samples-recall, and Samples-F1 score; Label-based evaluation indicators include Micro-AUC, Macro-AUC, Micro-AP, and Macro-AP.

4. The clinical auxiliary prediction system for lower respiratory tract infection in COPD patients according to any one of claims 1 to 3, characterized in that: The prediction results of the system include simultaneous prediction of LRTI categories and display of positive prediction probabilities of corresponding categories; The LRTI categories include Gram-negative bacterial infections, Gram-positive bacterial infections, multidrug-resistant bacterial infections, and / or fungal infections of the lower respiratory tract.

5. The clinical auxiliary prediction system for lower respiratory tract infection in COPD patients according to any one of claims 1 to 4, characterized in that: The system also includes an online interactive auxiliary prediction system built based on the CC-LightGBM model combined with SHAP individual interpretability technology, which is used for the visualization output of prediction results and clinical interpretations.

6. A method for constructing a clinical auxiliary prediction system for lower respiratory tract infection in COPD patients as claimed in any one of claims 1 to 5, characterized in that: The steps include: (1) Selecting infection categories that the system focuses on and performing partial correlation network analysis between different infection categories; (2) Collect clinical information of patients and divide them into development samples and validation samples for subsequent data processing and modeling analysis; (3) Analyze missing data and fill missing values ​​for all patient data, and use a multi-model multi-label feature screening open algorithm framework to screen model features and determine the indicators used for prediction; (4) Based on the development samples, a composite multi-label imbalanced data processing algorithm framework is used to process multi-label category imbalanced data, reduce data imbalance and thus obtain an optimized model; (5) Model construction is performed based on the optimization model and combined with SHAP individual interpretability technology.

7. The method for constructing a clinical auxiliary prediction system for lower respiratory tract infection in COPD patients according to claim 6, characterized in that: In step (2), the patient's clinical information includes general condition, clinical diagnosis, physical examination, laboratory test and / or microbial culture information.

8. The method for constructing a clinical auxiliary prediction system for lower respiratory tract infection in COPD patients according to claim 6 or 7, characterized in that: In the step (3): The missingness analysis step uses Little's MCAR and partial correlation analysis test models; and / or, The multi-model multi-label feature screening open algorithm framework steps are built based on scikit-learn using python, and are based on recursive feature elimination and permutation feature importance algorithms. The personalized feature importance values ​​of all features are calculated for each label, and the average of the weighted personalized feature importance values ​​of all labels is used as the common feature importance value.

9. The method for constructing a clinical auxiliary prediction system for lower respiratory tract infection in COPD patients according to any one of claims 6 to 8, characterized in that: In the step (4), the multi-label category imbalance processing step uses the three basic algorithms of REMEDIAL, MLSMOTE, and MLTL to propose a CMIDP algorithm framework, and forms four extended algorithms of REMLSMOTL, REMLSMOTE, REMLTL, and MLSMOTL.

10. A clinical auxiliary prediction method for lower respiratory tract infection in COPD patients, characterized in that: The method comprises the steps of collecting clinical information of the patient to be tested, and performing prediction using the clinical auxiliary prediction system for lower respiratory tract infection in COPD patients as described in any one of claims 1 to 5.

Images