A visual and interactive intelligent early warning method for sepsis

Abstract

The present invention discloses a visual interactive early intelligent warning method for sepsis, and the specific steps are as follows: the first step: patient database construction, through database technology, complete the collection and storage of patient data information; the second step: interact with database technology, complete the acquisition of patient clinical electronic medical record data, and construct a patient tabular time series with a step length of hours; the third step: data preprocessing and feature extraction, extract the features reflecting the measurement frequency, measurement time interval and other information in the patient information collection; the fourth step: complete the training and deployment of the model based on algorithms such as LightGBM, and use the Bayesian hyperparameter optimization algorithm during the model training process; the fifth step: clinically interpretable, visualized, and interactive interface construction. The present invention effectively improves the early warning ability of patient risks, and provides ICU doctors with early intervention before patients develop sepsis. At the same time, the interpretability, visualization and interactivity provide a strong guarantee for auxiliary decision-making for ICU patients.

Description

Technical Field

[0001] The present invention relates to the field of medical technology, and in particular to a visual interactive early intelligent warning method for sepsis. Background Art

[0002] Sepsis is one of the most common critical conditions in intensive care units and occurs when the body's response to infection loses control. It has been a focus of clinical and basic research in critical care medicine due to its severe morbidity, mortality and medical costs. Sepsis is a serious disease with a high mortality rate, with approximately 14,000 people dying from its complications every day worldwide. Early intervention with antibiotics can improve the survival rate of patients with sepsis.

[0003] As intensive care unit (ICU) clinicians are overwhelmed by big data, machine learning and big data will become more important in clinical practice. There is great value in developing and implementing machine learning algorithms to predict sepsis and evaluate its impact on clinical practice and patients. Therefore, it is feasible to use machine learning methods to process EHR big data based on the electronic health record (EHR) data of ICU patients to obtain an early warning model for sepsis. However, due to the lack of interpretability, the developed models based on machine learning show poor practicality in intensive care unit decision support. Therefore, it is of great value to develop a visual and interactive early intelligent warning method for sepsis. Summary of the invention

[0004] In view of the above defects or improvement needs of the prior art, the present invention proposes a visual interactive early intelligent warning method for sepsis, which has a high practical significance for assisting clinicians to intervene in the occurrence of sepsis in patients in advance. The present invention first builds a patient database, then obtains the patient's clinical electronic medical record data, preprocesses and extracts features of the data, and then completes the training and deployment of the model based on the LightGBM algorithm, uses the SHAP and LIME methods to perform interpretable analysis on the model, and uses the Flask framework and Vue framework to build the front-end and back-end systems to achieve clinically interpretable, visual, and interactive interface construction.

[0005] To achieve the above object, the present invention provides a visual interactive early intelligent warning method for sepsis, the steps of which are as follows:

[0006] S1: Patient database construction, collecting patient information through ventilators, electrocardiographs, blood glucose meters, blood gas analyzers, medical pendants, infusion pumps and other instruments in the ICU ward, and storing patient data using MySQL database technology;

[0007] S2: Acquisition of clinical electronic medical record data, extraction of patient information using SQL language, construction of patient tabular time series with hourly steps, and interaction between Flask framework and patient database. The definition of sepsis is based on Sepsis-3.0 standard;

[0008] S3: Data preprocessing and feature extraction, extracting features reflecting information such as measurement frequency and measurement time interval in patient information collection, and extracting clinical experience features;

[0009] S4: Complete the training and deployment of the model based on the LightGBM algorithm. First, use the LightGBM algorithm to train the model based on the Bayesian hyperparameter optimization method using the patient's electronic medical record information. Deploy the trained model and use the model to predict sepsis in patients.

[0010] S5: Build clinically interpretable, visualized, and interactive interfaces, using the Flask framework and Vue framework to build front-end and back-end systems to complete the interaction. Build a clinical visualization interactive interface based on interpretable information to assist clinical decision-making, and establish a new artificial intelligence decision-making assistance system from raw data and visualized warning information display to real-time risk tracking for doctors, and then to real-time intervention by doctors.

[0011] Furthermore, vital signs, laboratory tests, arterial blood gas values, laboratory observations, comorbidities, and demographic data were obtained from the electronic medical record data. Vital signs included: heart rate, temperature, systolic blood pressure, diastolic blood pressure, mean arterial pressure, respiratory rate, and oxygen saturation collected in real time. Laboratory test information included: base excess, urea nitrogen, calcium ions, chloride ions, creatinine, blood glucose, lactate, potassium ions, bilirubin, hematocrit, hemoglobin, thrombin time, white blood cell count, platelet count, anion gap, albumin, alanine aminotransferase, alkaline phosphatase, aspartate aminotransferase, international normalized ratio, and neutrophils; arterial blood gas values ​​included: pH, bicarbonate, blood carbon dioxide partial pressure, and blood oxygen partial pressure. Laboratory observations included: Glasgow Coma Index and oxygen uptake fraction. Comorbidities included: myocardial infarction, congestive heart failure, kidney disease, liver disease, diabetes, and malignant tumors. Demographic data include: age, gender, type of first ICU admission, height, weight, BMI, time difference from hospital admission to ICU admission, and time of ICU admission. A tabular time series integrating multiple physiological parameter information of critically ill patients is constructed with a step length of hours to reflect the physiological status of patients in real time. The definition of sepsis refers to the Sepsis-3.0 standard.

[0012] Furthermore, the data were preprocessed and feature extracted as follows. The preprocessing included removing features with missing values ​​exceeding 99.5%, using backward filling and forward filling methods for missing values, and partially using the median filling method; feature extraction calculated BMI based on clinical prior knowledge, as well as missing value marker sequences, the difference between the current parameter sampling and the last record, the maximum, minimum, median, standard deviation, and difference standard deviation of vital signs in the past 24 hours, and scoring features obtained by evaluating heart rate, temperature, respiratory rate, creatinine, mean blood pressure, systolic blood pressure, platelet count, and bilirubin features, and used One-hot encoding method for feature transformation.

[0013] Furthermore, the LightGBM algorithm is used to complete the training and deployment of the model, and the Bayesian hyperparameter optimization algorithm is used in the model training process. Bayesian optimization uses Bayesian technology to first assume the prior distribution model of the objective function, then obtain relevant information through samples, continuously optimize the model, and finally obtain the posterior distribution model of the objective function. Its core consists of two parts: one is Gaussian process regression, which calculates the mean and variance of the function value at each point. The other is to construct an acquisition function based on the mean and variance to determine at which point to sample during this iteration. The Bayesian optimization algorithm first assumes a prior distribution model for the objective function f(x). The commonly used assumption is that it satisfies the Gaussian process. After obtaining the prior distribution function, the model is corrected by sampling sample points. A collection function is defined in Bayesian optimization to determine the next sampling point. After the next sampling point is determined by the acquisition function, an experiment or observation can be conducted (i.e., (f(x)) is obtained through x. In the scenario of hyperparameter tuning, a parameter combination is tried to obtain an evaluation result. If the result of the sampling point has met the task requirements, the algorithm terminates; if it does not meet the task requirements, the sampling point is added to the currently observed sample point set, and then the current Gaussian process model is updated.

[0014] Furthermore, the local interpretable model-agnostic explanation method (LIME) and Shapley additive explanation method (SHAP) are used to perform interpretability operations on the model. The LIME method mainly uses a linear model as a local proxy model to explain the black box classification model. LIME emphasizes two model interpretation criteria in the process of local interpretation: one is that the explanation itself must be interpretable, and the other is local fidelity. LIME balances the above two criteria by minimizing the fidelity function and simplifying the interpretable proxy model. Its objective function is:

[0015] ξ(x)=arg ming∈G Γ(f,g,Π x )+Ω(g)

[0016] Where G represents the set of interpretable models; Ω(g) represents the complexity of model g; f represents the model to be explained; Π x Denotes the neighborhood metric of x to define the locality around x.

[0017] The SHAP method follows the description in LIME and uses the additive feature attribution method. The additive feature attribution method is expressed as follows, where z'∈{0,1} N , M is the number of features, φ i ∈R, φ0 represents the model output without simplified input, and g(z') represents the interpretable model output.

[0018]

[0019] The SHAP value is mainly used to quantify the contribution of each feature to the model prediction. This method is derived from Shapley value in game theory. Its basic design idea is: first calculate the marginal contribution of a feature added to the model, then calculate the different marginal contributions of the feature in all feature sequences, and finally calculate the SHAP value of the feature, that is, the mean of all marginal contributions of the feature. Assume that the i-th sample is x i , the jth feature of the i-th sample is x ij , the marginal contribution of the feature is mc ij , the weight of the edge is w i , where f(x ij ) is x ij The SHAP value of, for example, the SHAP value of the first feature of the i-th sample is calculated as follows:

[0020] f(x i1 )=mc i1 w1+...+mc i1 w n

[0021] The model predicts the value of this sample to be y i , the baseline of the entire model (usually the mean of the target variable for all samples) is y base , then SHAPvalue obeys the following equation:

[0022] y i =y base +f(x i1 )+f(x i2)+...+f(x is )

[0023] f(x i1 ) is the final predicted value y of the first feature in the i-th sample i The SHAP value for each feature indicates the change in the model prediction when conditioned on that feature. For each feature, the SHAP value explains its contribution to illustrate the difference between the average model prediction and the actual prediction for the instance. i , 1)>0, indicating that the feature improves the prediction value, otherwise, it means that the feature reduces the contribution.

[0024] There are three important properties in the SHAP method: the local accuracy that indicates that the sum of feature attributions needs to be equal to the model output to be explained; the certainty that the attribution value of missing features must be equal to zero; and the consistency that if the model changes and causes the marginal contribution of a feature to increase or remain unchanged, the attribution value should also increase or remain unchanged.

[0025] This embodiment provides a visual interactive early intelligent warning method for sepsis, which specifically includes the following steps:

[0026] 1. Collect patient information through ventilators, electrocardiographs, blood glucose meters, blood gas analyzers, medical pendants, infusion pumps and other instruments in the ICU ward, and store patient data using MySQL database technology;

[0027] 2. Based on the Flask framework, the clinical electronic medical record data of patients are obtained from the MySQL database. The data obtained from the clinical electronic medical record data include vital signs, laboratory tests, arterial blood gas values, laboratory observation values, comorbidities, demographic data, and a tabular time series with a step length of hours that integrates multiple physiological parameter information of critically ill patients;

[0028] 3. Data preprocessing and feature extraction: extract the features reflecting the measurement frequency, measurement time interval and other information in the patient information collection, and extract clinical experience features;

[0029] 4. Complete model training and deployment based on algorithms such as LightGBM, and use the Bayesian hyperparameter optimization algorithm during model training;

[0030] 5. Use LIME and SHAP methods to perform interpretable analysis on the prediction results of the model. At the same time, the patient data is formatted and sent to the browser in JSON format. The front end parses the data and displays the patient's basic information analysis chart on the display interface. The front-end interpretable, visual, and interactive system mainly displays the changing trend of the patient's sepsis risk prediction, and also displays the patient's vital signs, laboratory tests, arterial blood gas values, laboratory observation values, comorbidities, demographic data, and other information to display the explanatory results. The present invention uses Echarts to visualize the patient's physiological data and uses Ajax technology to obtain the patient's background record information.

[0031] The above technical solution has the following effects:

[0032] 1. This method only selects patient vital signs, laboratory tests, arterial blood gas values, laboratory observation values, comorbidities, and demographic data that are easy to collect clinically. They are easy to obtain in various hospitals or institutions, which is conducive to improving the generalization ability of the model and providing favorable conditions for reducing the complexity of system deployment.

[0033] 2. This method performs feature processing based on clinical experience, which is conducive to mining the dynamic physiological and pathological information of patients that changes over time. The LightGBM model used in this method has good performance in tabular data prediction and shows good predictive ability in model training, which helps to improve the accuracy of early prediction of sepsis.

[0034] 3. This method uses the local interpretation method (LIME) and the Shapley additive interpretation method (SHAP) to provide doctors with interpretable analysis of the current model prediction results. It can solve the problem that the current "black box" model cannot explain to clinicians why the model outputs high risks, and has important practical application value for assisting clinicians in making decisions.

[0035] 4. This method proposes a visual and interactive early intelligent warning method for sepsis based on clinical patient information, which can not only assist clinicians in making early judgments on the occurrence of sepsis in patients, but also provide explainable analysis of the judgments, further visualize the patient's information, and make the clinician's diagnosis more intelligent. The explainable, visual, and interactive early intelligent warning system for sepsis proposed in this invention has played a good role in inspiring medical intelligence. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0037] Figure 1The figure is a flow chart of the visual interactive early intelligent warning method for sepsis of the present invention.

[0038] Figure 2 Schematic diagram of a tabular time series with a step size of hours.

[0039] Figure 3 Schematic diagram of the process of building a LightGBM model.

[0040] Figure 4 This is a flowchart of the Bayesian optimization algorithm.

[0041] Figure 5 Schematic diagram of using LIME method to explain the model.

[0042] Figure 6 Schematic diagram of using SHAP method to explain the model.

[0043] Figure 7 Schematic diagram of the explainable, visual, and interactive decision-making support system for the intensive care unit. DETAILED DESCRIPTION

[0044] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0045] Embodiment 1

[0046] This embodiment provides a visual interactive early intelligent warning method for sepsis, such as Figure 1 As shown, the steps are as follows:

[0047] Step 1: Patient database construction, through the ICU ward ventilator, electrocardiograph, blood glucose meter, blood gas analyzer, medical pendant, infusion pump and other instruments to complete the collection of patient information, using HDFS distributed file storage and SPARK framework, using MySQL database technology to complete the storage of patient data;

[0048] Step 2: Obtain clinical electronic medical record data, use SQL language to complete the extraction of patient information, and construct a patient tabular time series with hourly steps ( Figure 2 ) to complete the interaction between the Flask framework and the patient database. The definition of sepsis refers to the Sepsis-3.0 standard;

[0049] Step 3: Data preprocessing and feature extraction, extracting features reflecting measurement frequency, measurement time interval and other information in patient information collection, and extracting clinical experience features;

[0050] Step 4: Complete model training based on the LightGBM algorithm ( Figure 3 ), firstly using the LightGBM algorithm, based on the Bayesian hyperparameter optimization method ( Figure 4 ), using the patient's electronic medical record information to train the model. The trained model is deployed and used to make early predictions of sepsis in patients;

[0051] Step 5: Use LIME Figure 5 )、SHAP( Figure 6 ) method to analyze the predictive results of machine learning models for interpretability. Figure 5 As shown in the figure, the LIME method gives the top features for model prediction. The picture is divided into three parts. The leftmost part is the probability of the prediction result, the middle part gives the contribution value of the top-ranked features, and the right side is the value of the feature. Figure 6 As shown in the figure, the SHAP method gives the features with larger contribution values ​​in the model and their contribution values, where the contribution on the right is positive and the contribution on the left is negative.

[0052] Step 6: Build a clinical visualization, interactive, and explainable interface, and use the Flask framework and Vue framework to build the front-end and back-end systems. Build a clinical visualization interactive interface based on explainable information to assist clinical decision-making, and establish a new artificial intelligence decision-making assistance system from raw data and visual warning information display to real-time risk tracking for doctors, and then to real-time intervention by doctors ( Figure 7 ).

[0053] Experimental results:

[0054] The present invention obtains vital signs, laboratory tests, arterial blood gas values, laboratory observations, complications, and demographic data from electronic medical record data. A tabular time series integrating multiple physiological parameter information of critically ill patients is constructed with a step length of hours to reflect the patient's physiological state in real time; the definition standard of sepsis refers to the Sepsis-3.0 standard. The data is preprocessed and feature extracted as follows, and the preprocessing includes removing features with missing values ​​exceeding 99.5%, using the backward filling and forward filling methods for the missing values, and partially using the median filling method; the prediction results of the model are analyzed for interpretability using the LIME and SHAP methods. At the same time, the patient data is formatted and sent to the browser in JSON format, and the front end parses the data and displays the patient's basic information analysis chart on the display interface. The front-end visualization and interactive system mainly displays the patient's sepsis risk prediction change trend and the patient's vital signs, laboratory tests, arterial blood gas values, laboratory observations, complications, demographic data and other information, and displays the explanatory results at the same time. The present invention uses Echarts to visualize the patient's physiological data and uses Ajax technology to obtain the patient's background record information.

[0055] Test data: MIMIC-IV data, the most influential public database in the field of critical care, was used as test data. 85% of the MIMIC-IV data was used as the training set, and 15% of the MIMIC-IV data was used as the test set. The ROC curve with a prediction interval of 6 hours was drawn, and the AUROC value of the test set at this time was calculated to be 0.8156. At the same time, the interpretable step gave high-risk factors. These results all indicate that the proposed interpretable early warning method for sepsis has practical clinical application value.

[0056] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modifications, improvements, etc. made within the principles of the present invention should be included in the scope of protection of the present invention. Based on the technical solution of the present invention, various modifications or deformations that can be made by those skilled in the art without creative work are still within the scope of protection of the present invention.

Claims

1. A visual interactive early intelligent warning method for sepsis, characterized in that: The following steps are involved: S1: Patient database construction: patient information is collected through ventilators, electrocardiographs, blood glucose meters, blood gas analyzers, medical pendants, and infusion pumps in the ICU ward, and patient data is stored using MySQL database technology; S2: Acquisition of clinical electronic medical record data; extraction of patient information using SQL language, construction of a patient tabular time series with hourly steps, interaction between the Flask framework and the patient database, and reference to the Sepsis-3.0 standard for the definition of sepsis; S3: Data preprocessing and feature extraction: extract the features reflecting the measurement frequency and measurement time interval information in the patient information collection, and extract clinical experience features; S4: Complete the training and deployment of the model based on the LightGBM algorithm. First, use the LightGBM algorithm, based on the Bayesian hyperparameter optimization method, to train the model using the patient's electronic medical record information, deploy the trained model, and use the model to make early predictions of sepsis in patients. S5: Construction of clinically interpretable, visual, and interactive interfaces; using the Flask framework and the Vue framework to build front-end and back-end systems, building clinical visual interactive interfaces based on interpretable information to assist clinical decision-making, and establishing a new artificial intelligence decision-making assistance system from raw data and visual warning information display to real-time risk tracking for doctors, and then to real-time intervention by doctors; Clinically interpretable, visualized, and interactive interfaces are built, and local interpretation methods LIME and Shapley additive interpretation method SHAP are used to interpret the model. The LIME method uses a linear model as a local proxy model to interpret the black box classification model. LIME emphasizes two model interpretation criteria in the process of local interpretation: one is that the explanation itself must be interpretable, and the other is local fidelity. LIME balances the above two criteria by minimizing the fidelity function and simplifying the interpretable proxy model. Its objective function: ξ(x)=argmin g∈G C(f,g,P x )+Ω(g) Where G represents the set of interpretable models; Ω(g) represents the complexity of model g; f represents the model to be explained; Π x represents the neighborhood metric of x to define the locality around x; The SHAP method follows the description in LIME and uses the additive feature attribution method. The additive feature attribution method is expressed as follows, where z'∈{0,1} N , M is the number of features, φ i ∈R, φ0 represents the model output without simplified input, and g(z') represents the interpretable model output; The SHAP value is mainly used to quantify the contribution of each feature to model prediction. This method is derived from Shapley value in game theory. The basic design idea is: first calculate the marginal contribution of a feature added to the model, then calculate the different marginal contributions of the feature in all feature sequences, and finally calculate the SHAP value of the feature, that is, the mean of all marginal contributions of the feature; assuming that the i-th sample is x i , the jth feature of the i-th sample is x ij , the marginal contribution of the feature is mc ij , the weight of the edge is w i , where f(x ij ) is x ij The SHAP value of the first feature of the i-th sample is calculated as follows: f(x i1 )=mc i1 w1+...+mc i1 w n The model predicts the value of this sample to be y i , the baseline of the entire model is y base , then SHAPvalue obeys the following equation: y i =y base +f(x i1 )+f(x i2 )+...+f(x is ) f(x i1 ) is the final predicted value y of the first feature in the i-th sample i The SHAP value of each feature indicates the change in the model prediction when the feature is conditioned; for each feature, the SHAP value explains its contribution to illustrate the difference between the average model prediction and the actual prediction of the instance; when f(x i , 1)>0, indicating that the feature improves the prediction value, otherwise, it means that the feature reduces the contribution.

2. A visual interactive early intelligent warning method for sepsis according to claim 1, characterized in that: Data from the electronic medical record were obtained, including vital signs, laboratory tests, arterial blood gas values, laboratory observations, comorbidities, and demographic information; Vital signs include: heart rate, temperature, systolic blood pressure, diastolic blood pressure, mean arterial pressure, respiratory rate, and blood oxygen saturation collected in real time; laboratory test information includes: base excess, urea nitrogen, calcium ions, chloride ions, creatinine, blood glucose, lactate, potassium ions, bilirubin, hematocrit, hemoglobin, thrombin time, white blood cell count, platelet count, anion gap, albumin, alanine aminotransferase, alkaline phosphatase, aspartate aminotransferase, international normalized ratio, and neutrophils; arterial blood gas values ​​include: pH, bicarbonate, and blood carbon dioxide concentration. Laboratory observation values ​​included Glasgow Coma Index and oxygen uptake fraction. Comorbidities included myocardial infarction, congestive heart failure, kidney disease, liver disease, diabetes, and malignant tumors. Demographic data included age, sex, race, type of first ICU admission, height, weight, BMI, time difference from hospital admission to ICU admission, and time of ICU admission. A tabular time series integrating multiple physiological parameter information of critically ill patients was constructed with a step length of hours to reflect the patient's physiological status in real time. The definition of sepsis referred to the Sepsis-3.0 standard.

3. A visual interactive early intelligent warning method for sepsis according to claim 1, characterized in that: The data were preprocessed and feature extracted as follows. The preprocessing included removing features with missing values ​​exceeding 99.5%, using the backward filling and forward filling methods for missing values, and partially using the median filling method; feature extraction calculated the body mass index based on clinical prior knowledge, as well as the missing value marker sequence, the difference between the current parameter sampling and the previous record, the maximum, minimum, median, standard deviation, and difference standard deviation of vital signs in the past 24 hours, and the scoring features obtained by evaluating the heart rate, temperature, respiratory rate, creatinine, mean blood pressure, systolic blood pressure, platelets, and bilirubin characteristics, and used the One-hot encoding method for feature transformation.

4. A visual interactive early intelligent warning method for sepsis according to claim 1, characterized in that: The LightGBM algorithm is used to complete the training and deployment of the model, and the Bayesian hyperparameter optimization algorithm is used in the model training process; Bayesian optimization uses Bayesian technology to first make assumptions about the prior distribution model of the objective function, and then obtain relevant information through samples, continuously optimize the model, and finally obtain the posterior distribution model of the objective function.

5. The visual interactive early intelligent warning method for sepsis according to claim 1, characterized in that: The specific steps include: 1) Collect patient information through ventilators, electrocardiographs, blood glucose meters, blood gas analyzers, medical pendants, and infusion pumps in the ICU ward, and store patient data using MySQL database technology; 2) Based on the Flask framework, the clinical electronic medical record data of patients were obtained from the MySQL database. The data obtained from the clinical electronic medical record data included vital signs, laboratory tests, arterial blood gas values, laboratory observation values, comorbidities, and demographic information. The tabular time series integrating the multi-physiological parameter information of critically ill patients was constructed with a step length of hours; 3) Data preprocessing and feature extraction: extracting features reflecting measurement frequency and measurement time interval information in patient information collection, and extracting clinical experience features; 4) Complete model training and deployment based on the LightGBM algorithm, and use the Bayesian hyperparameter optimization algorithm during model training; 5) Use LIME and SHAP methods to perform interpretable analysis on the prediction results of the model; at the same time, format the patient data and send it to the browser in JSON format. The front-end parses the data and displays the patient's basic information analysis chart on the display interface; the front-end interpretable, visual, and interactive system mainly displays the patient's sepsis risk prediction trend, while displaying the patient's vital signs, laboratory tests, arterial blood gas values, laboratory observation values, comorbidities, and demographic information, and displays explanatory results; use Echarts to visualize the patient's physiological data, and use Ajax technology to obtain the patient's background record information.

Images