Ai-based chemical index detection system for septicopyemia

Abstract

The invention discloses a septicopyemia death risk prediction method based on artificial intelligence, and the method comprises the following steps: data collection and preprocessing: collecting clinical data and inspection indexes of a septicopyemia patient, and carrying out the cleaning and standardization processing of the data; feature selection and engineering: selecting features related to septicopyemia death risks based on clinical knowledge and statistical analysis, and performing feature engineering; model training and verification: training the preprocessed data by using a machine learning algorithm, and verifying the model by using a training set, a verification set and a test set; model evaluation and optimization: evaluating model performance by using indexes such as AUC, accuracy rate, recall rate, F1 score and the like, and optimizing model hyper-parameters; and model explanation and application: explaining a model prediction result by using an explanatory tool, and applying the model in a clinical environment to assist death risk assessment. The invention provides a method for improving the early diagnosis and prognosis prediction of septicopyemia, thereby reducing the death rate.

Description

Technical Field

[0001] The present application belongs to the field of computer medicine technology, and in particular to a chemical indicator detection system for sepsis based on Ai. Background Art

[0002] Sepsis refers to life-threatening organ dysfunction caused by a disordered host response to infection. It is a systemic infection response and a response pattern of the immune system to infectious pathogens. Blood culture is currently the gold standard for diagnosing sepsis, with high specificity, but low sensitivity and long culture cycle are not conducive to early diagnosis and treatment guidance.

[0003] At present, there are many studies on the prognosis prediction of sepsis, such as procalcitonin, C-reactive protein, leukocytes, platelets, DD dimer, lactate, cystatin C, etc. However, most studies are limited in clinical prediction. Since sepsis is a complex clinical syndrome with strong individual differences and involves multiple aspects such as imbalance of inflammatory response, immune function and coagulation dysfunction, there are certain difficulties in diagnosis and prognosis. A single clinical indicator can no longer well reflect the nature of the rapid development of sepsis patients, and it also has certain limitations in the early diagnosis and prognosis of sepsis. At present, there is still a lack of effective biomarker combinations to distinguish between sepsis patients and non-sepsis patients. Although SOFA scores and qSOFA scores are also used clinically as screening tools for patients who may have sepsis, it is reported that qSOFA has high specificity in early detection of sepsis and hospital mortality, but qSOFA has low sensitivity in identifying sepsis and hospital mortality. Therefore, it is necessary to find a more sensitive and specific method to improve the early diagnosis and prognosis prediction of sepsis, thereby reducing mortality and improving the quality of life of patients. Therefore, predicting the occurrence of sepsis has become a popular research trend. Predictions are made for different populations and corresponding prediction models are established to play a role in prevention and guiding medication. Summary of the invention

[0004] In view of the technical problems raised by the background technology, the present invention proposes a sepsis chemical index detection system based on AI. In the first aspect, the present invention provides a method for predicting the risk of death from sepsis based on artificial intelligence, the method comprising the following steps: data collection and preprocessing, collecting clinical data and test indicators of sepsis patients, and cleaning and standardizing the data;

[0005] Feature selection and engineering: features associated with the risk of death from sepsis are selected based on clinical knowledge and statistical analysis, and feature engineering is performed;

[0006] Model training and validation: Use machine learning algorithms to train preprocessed data, and use training sets, validation sets, and test sets to validate the model;

[0007] Model evaluation and optimization, using indicators such as AUC, precision, recall, F1 score, etc. to evaluate model performance and optimize model hyperparameters;

[0008] Model interpretation and application: use explanatory tools to interpret model predictions and apply the model in clinical settings to assist in mortality risk assessment.

[0009] Preferably, the machine learning model includes one or more of XGB, LR, RF, SVM, and MLP.

[0010] Preferably, the first test is used as training data, and baseline data of age and gender are added, and whether antibiotics are used to complete this training. The training data set is split into training / validation / testing in a ratio of 3:1:1, and four models, including extreme gradient boosting tree (XGB), logistic regression model (LR), random forest model (RF), and support vector machine (SVM), are used for training;

[0011] The initial screening requirements were that the feature density of the positive group was ≥0.7 and that of the control group was ≥0.5. After the initial screening, a total of 103 features entered the formal model training, among which blood sodium, urea, lactate dehydrogenase, D-dimer, and prothrombin time were positively correlated with the mortality of hospitalized patients. Finally, the RF model was used to predict the mortality risk of hospitalized patients with sepsis. The results showed that RF had good predictive performance. The newly constructed model of the calibration curve can be used for the in-hospital mortality rate of elderly patients with sepsis. Blood procalcitonin, lactate dehydrogenase, and D-dimer were selected as predictors of patient examination and test indicators, and machine learning was used to establish a predictive model for assessing the mortality risk of patients with sepsis.

[0012] The model predicts the risk of death in patients with sepsis based on their clinical data and test indicators.

[0013] In a second aspect, the present invention provides a clinical decision support system, comprising the sepsis mortality risk prediction model described in the first aspect, and capable of assisting doctors in diagnosing and treating sepsis.

[0014] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0015] (1) More sensitive and specific methods can be used to improve the early diagnosis and prognosis prediction of sepsis, thereby reducing mortality and improving the quality of life of patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0017] Figure 1 is a sample processing flow chart provided in an embodiment of the present application;

[0018] Figure 2 : This is the XGB model change diagram provided in the embodiment of the present application, (a) training set change diagram; (b) validation set change diagram; (c) test set change diagram;

[0019] Figure 3 is a feature weight graph of the XGB model provided in an embodiment of the present application;

[0020] Figure 4 : These are the LR model training change diagrams provided in the embodiments of the present application, (a) LR training set change diagram, (b) validation set change diagram, and (c) test set change diagram;

[0021] Figure 5 is a feature weight graph of the LR model provided in an embodiment of the present application;

[0022] Figure 6 These are the first round RF model change diagrams provided in the embodiments of the present application, (a) training set change diagram, (b) validation set change diagram, and (c) test set change diagram;

[0023] Figure 7 is a feature weight diagram of the first round RF model provided in the embodiment of the present application,

[0024] Figure 8 : These are the SVM model training change graphs provided in the embodiments of the present application, (a) SVM training set change graph, (b) SVM validation set change graph, and (c) SVM test set change graph;

[0025] Fig. 9 is a feature weight graph of the first round of SVM model provided in the embodiment of the present application,

[0026] Fig.10 : These are the MLP model training change diagrams provided in the embodiments of the present application, (a) MLP training set change diagram, (b) MLP validation set change diagram, and (c) MLP test set change diagram;

[0027] Fig.11 It is a feature weight diagram of the first round MLP model provided in the embodiment of the present application. DETAILED DESCRIPTION

[0028] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0029] Example

[0030] The discharge method was used as the outcome label (death was the positive group, and discharge on doctor's orders was the negative group). The concept of enrollment time and time window was not set. The first test was used as training data, and age and gender baseline data, as well as "whether antimicrobial drugs were used" were added to complete this training. The training data set was split into training / validation / testing in a ratio of 3:1:1, and five models, including XGB, LR, RF, SVM, and MLP, were used for training.

[0031] The subjects of this example are patients who were admitted to Anning First People's Hospital and diagnosed with sepsis from January 1, 2019 to September 1, 2023. The diagnostic criteria for sepsis are: (1) infection at admission; (2) meeting the diagnostic criteria of sepsis 3.0. Exclusion criteria are: (1) age < 18 years old; (2) lack of complete clinical data or follow-up information; (3) patients whose family members gave up treatment or were discharged without medical advice.

[0032] This study included the test data of 1759 patients. The data were screened according to the patient name + patient ID, and 274 duplicate data were removed. 254 data with outcomes other than "discharge on doctor's orders and death" were removed. There were 573 test data without test data. After feature density screening, 4 data were removed. The final sample size was 654, which were divided into a positive group with death as the outcome and a negative group with discharge on doctor's orders according to the outcome. The feature density of the positive group was greater than or equal to 0.7, and the feature density of the control group was greater than or equal to 0.5 as the initial screening requirements. After the initial screening, a total of 103 features entered the formal model training.

[0033] The discharge method was used as the outcome label (death was the positive group, and discharge on doctor's orders was the negative group). The concept of enrollment time and time window was not set. The first test was used as training data, and age, gender baseline data, and "whether antimicrobial drugs were used" were added to complete this training. The training data set was split into training / validation / testing in a ratio of 3:1:1. Four models were used for training: extreme gradient boosting tree (XGB), logistic regression model (LR), random forest model (RF), and support vector machine (SVM).

[0034] Determine the sample screening process for this study, by identifying unique patients, screening patient test samples, and randomly assigning patients to test sets, validation sets, and training sets. See ( Figure 1 The initial screening requirement was that the density of the positive group features was ≥ 0.7 and the density of the control group features was ≥ 0.5. After the initial screening, a total of 103 features entered the formal model training. The detailed initial screening model of the included features is shown in (Table 1).

[0035] Among the 654 patients, the average age of the patients was between 70.45±15.5. 382 patients belonged to the training set, including 130 females and 252 males. 124 patients were assigned to the validation set, including 49 females and 75 males. 148 patients were assigned to the test set, including 46 females and 102 males. Specific sample baseline data statistics are shown in (Table 1).

[0036] Table 1 Sample baseline data statistics

[0037]

[0038]

[0039] XGB Model

[0040] like Figures 2 to 3 As shown in the figure, the descriptive statistics and effect size estimates of sepsis biomarkers found after the application of XGB are summarized in detail. As can be seen from Table 1, the content of procalcitonin in patients with sepsis is negatively correlated with the mortality rate of patients during hospitalization. In addition, the eosinophil count and blood potassium of patients are positively correlated with the mortality rate of patients with sepsis. In addition, the specificity of the XG model is as high as 0.64 and the sensitivity is 0.77. This study is looking for possible markers of death in patients with sepsis, so the experimental results have a certain reference value.

[0041] Parameter name Weight Blood_Procalcitonin_Quantitative 0.329913256 Blood_Antithrombin III_Quantitative 0.129391671 Blood_eosinophil count_qualitative 0.105412261 Blood_Albumin_Qualitative 0.057133805 Blood_mean corpuscular volume_quantitative 0.055548578 Blood_Potassium_Qualitative 0.0531916 Blood_lymphocyte percentage_quantitative 0.042488301

[0042]

[0043] LR model training

[0044] The descriptive statistics and effect size estimates of sepsis biomarkers found after applying LR are summarized in detail. As can be seen in Table 1, the median age of patients with sepsis was significantly higher than that of non-septic patients. In addition, activated partial prothrombin time, lactate dehydrogenase, and blood chloride content were positively correlated with the mortality rate of patients with sepsis. However, blood triglycerides and activated partial thromboplastin time in the blood were negatively correlated.

[0045] Parameter name Weight age 0.449668535 Quantification of blood triglycerides 0.132801735 Fibrinogen quantification 0.122121561 Activated partial thromboplastin time 0.111289486 Blood lactate dehydrogenase quantification 0.081663038 Blood glucose [fasting] qualitative 0.079777422 Hemochloride 0.077366282

[0046]

[0047] RF model training

[0048] Training is performed with the following weight ratios:

[0049] Parameter name Weight Blood_Procalcitonin_Quantitative 0.039210549 Blood_Antithrombin III Quantification 0.018088358 Blood sodium quantitative 0.01494974 Blood_urea_qualitative 0.01197648 Blood_Lactate Dehydrogenase_Quantitative 0.00703109 Blood_D-dimer_quantitative 0.006940064 Blood_Prothrombin Time_Quantitative 0.004143118 Blood Chloride Quantification 0.002766368

[0050] We interpreted the sepsis prediction results of the RF model by studying the SHAP annotations. With the help of SHAP, we identified the levels of biomarkers that were significant in predicting hospital mortality in patients with sepsis. Blood sodium, urea, lactate dehydrogenase, D-dimer, prothrombin time, etc. were positively correlated with hospital mortality, but antithrombin III and procalcitonin were found to be negatively correlated with sepsis in the study.

[0051]

[0052] SVM Model

[0053] Parameter name Weight age 0.390792467 Qualitative blood sodium 0.18768486 Activated partial thromboplastin time 0.176695835 Fibrinogen quantification 0.090159528 Quantification of blood lymphocyte percentage 0.033984645 Mean corpuscular volume of blood 0.001261911

[0054] We explained the sepsis prediction results of the SVM model by studying SHAP annotations. The prediction results suggest that age is indeed an independent risk factor for death during hospitalization in patients with sepsis. In addition, hypernatremia and lymphocyte percentage are also positively correlated with death in patients with sepsis.

[0055]

[0056] The prognostic prediction model based on machine learning is used to predict the risk of death in patients with sepsis and provide a basis for clinical decision-making. Our model is very unique. It uses machine learning to sort common clinical laboratory tests according to the importance of variables, and selects 103 test features of patients' admission examinations to establish the final machine learning model, which is not available in traditional scoring systems. We used the RF model to predict the risk of death in hospitalized patients with sepsis, and the results showed that RF has good predictive performance. The calibration curve further confirmed that the newly constructed model can be used for the in-hospital mortality of elderly patients with sepsis. In this model, we selected blood procalcitonin, lactate dehydrogenase, and D-dimer as predictors of patient examination test indicators, which can be quickly and easily obtained in hospitals. Machine learning is used to establish a predictive model for assessing the risk of death in patients with sepsis. It provides a simpler and more effective prognostic tool for patients with sepsis.

[0057] Sepsis is characterized by high morbidity and mortality. Establishing a prediction model in the early stage of the disease is helpful for clinical decision support. The GNB classification method has been applied in many classification problem studies. Its algorithm is simple to calculate, the calculation speed is relatively fast, and the classification results obtained are relatively accurate. In addition, it can also show good results when the amount of data is large. LR is a widely used linear classification model with the characteristics of strong interpretability and good predictive performance.

[0058] Our prediction model (LR: AUC 0.76, GNB: AUC 0.76) is more accurate than the previous prediction model based on SOFA score in the stratification of mortality risk in sepsis patients. Machine learning can obtain unknown dependencies from existing data sets and use the learned dependencies to analyze and process new data sets. Different detection indicators have different weights in predicting the risk of death in sepsis patients.

[0059] Procalcitonin is a calcitonin propeptide glycoprotein without hormone activity. It is a soluble protein released into the patient's circulatory system under conditions of severe systemic infection, especially bacterial infection. In the case of bacterial infection, proinflammatory cytokines (IL-1 and IL-6) activate the gene CALC-1 responsible for the production of procalcitonin in many cells throughout the body. Since most cells cannot convert procalcitonin into calcitonin, the concentration of procalcitonin increases after entering the blood circulation, while the level of calcitonin remains unchanged. Therefore, the level of procalcitonin generally reflects the severity of infection in the body. Lactate dehydrogenase catalyzes the reversible conversion of pyruvate into lactate in aerobic glycolysis. Previous studies have shown that an increase in the blood content of lactate dehydrogenase is significantly associated with an increase in the 28-day mortality rate of patients with sepsis. Secondly, the level of lactate dehydrogenase is positively correlated with the production of lactate and IL-1β inflammatory substances. In patients with sepsis, intracellular glucose metabolism reprogramming may be the cause of metabolic abnormalities in patients with sepsis, so the level of lactate dehydrogenase can indirectly reflect the level of glucose metabolism in patients with sepsis. According to the research results, the extracted blood procalcitonin, lactate dehydrogenase and D-dimer detection indicators of patients combined with machine learning models can better predict the mortality rate of patients during hospitalization. The advantages are as follows: First, the data of this study comes from the Medical Laboratory Department of Anning First People's Hospital, which can better reflect the real data of local hospitals compared with downloading data analysis in the database. The machine learning method is used to process missing values ​​and establish a death risk prediction model, which is more in line with objective facts. And through machine learning, the characteristic variables detected in sepsis patients are selected to further establish a prediction model, which fully considers the different weights of various examination indicators in sepsis patients and covers comprehensively.

[0060] The embodiments described above are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A method for predicting mortality risk of sepsis based on artificial intelligence, characterized in that: The method comprises the following steps: data collection and preprocessing, collecting clinical data and test indicators of sepsis patients, and cleaning and standardizing the data; Feature selection and engineering: features associated with the risk of death from sepsis are selected based on clinical knowledge and statistical analysis, and feature engineering is performed; Model training and validation: Use machine learning algorithms to train preprocessed data, and use training sets, validation sets, and test sets to validate the model; Model evaluation and optimization, using indicators such as AUC, precision, recall, F1 score, etc. to evaluate model performance and optimize model hyperparameters; Model interpretation and application: use explanatory tools to interpret model predictions and apply the model in clinical settings to assist in mortality risk assessment.

2. The method according to claim 1, characterized in that The machine learning model includes one or more of XGB, LR, RF, SVM, and MLP.

3. The method according to claim 1, characterized in that The first test was used as training data, and baseline data of age and gender were added, as well as whether antibiotics were used to complete this training. The training data set was split into training / validation / testing at a ratio of 3:1:

1. Four models were used for training: extreme gradient boosting tree (XGB), logistic regression model (LR), random forest model (RF), and support vector machine (SVM); The initial screening requirement was that the feature density of the positive group was ≥0.7 and the feature density of the control group was ≥0.

5. After the initial screening, a total of 103 features entered the formal model training. Among them, blood sodium, urea, lactate dehydrogenase, D-dimer, and prothrombin time were positively correlated with the mortality of hospitalized patients. Finally, the RF model was used to predict the mortality risk of hospitalized patients with sepsis. The results showed that RF had good predictive performance. The newly constructed model of the calibration curve can be used for the in-hospital mortality rate of elderly patients with sepsis. Blood procalcitonin, lactate dehydrogenase, and D-dimer were selected as predictive factors for patient examination and test indicators, and machine learning was used to establish a predictive model for assessing the mortality risk of patients with sepsis.

4. The sepsis mortality risk prediction model according to the method of claim 1, characterized in that: The model predicts the risk of death in patients with sepsis based on their clinical data and test indicators.

5. A clinical decision support system, characterized in that: The method comprises the sepsis mortality risk prediction model as described in claim 1 and can assist doctors in diagnosing and treating sepsis.

Images