ICU sepsis two-stage early prediction method and device and storage medium

Abstract

The invention discloses an ICU sepsis two-stage early prediction method and device and a storage medium, and the method comprises the steps: continuously collecting a clinical data set after a patient enters an ICU, and carrying out the feature extraction, and obtaining a prediction feature vector; based on the predicted feature vector corresponding to the clinical data set at the current observation moment, determining whether the patient is a non-sepsis sample by using a preset classification model; and if the patient is not determined as a non-sepsis sample by the classification model, predicting whether the patient has a sepsis infection risk in a future preset time period or not by using a preset prediction model according to the current observation moment and the prediction feature vector corresponding to the clinical data set in the previous set time. According to the application, a two-stage mode is adopted, patients are subjected to layered screening, and samples which do not obviously belong to sepsis are excluded; and for the samples which are not excluded, more accurate sepsis risk prediction is carried out by using the prediction model in combination with the prediction feature vectors corresponding to the clinical data of the plurality of time periods, so that the prediction accuracy is improved.

Description

Technical Field

[0001] The present application belongs to the field of medical data mining technology, and specifically relates to a two-stage early prediction method, device and storage medium for ICU sepsis. Background Art

[0002] Sepsis is a life-threatening organ dysfunction syndrome caused by infection, characterized by high morbidity and mortality, and is a leading cause of death and substantial medical costs for patients hospitalized in the ICU. Sepsis progresses rapidly, and it is estimated that for every hour of delay in diagnosis and treatment, the patient's survival rate decreases by approximately 7.6%. Therefore, in clinical practice, early warning diagnosis of sepsis plays a key role in patient treatment and prognosis.

[0003] However, there is currently no gold standard diagnostic method for sepsis in clinical practice, and diagnosis mainly relies on the patient's clinical manifestations and laboratory test results. However, these methods often have the following limitations.

[0004] 1. Lack of specificity in symptoms: The symptoms of early sepsis are similar to those of common diseases (such as fever and hypotension). Traditional diagnostic indicators (such as body temperature, heart rate, and blood pressure) are easily interfered with by other factors, leading to missed or misdiagnosis.

[0005] 2. Delayed laboratory test results: Although laboratory tests such as blood culture have high specificity, the results often take a long time to obtain and cannot meet the needs of early diagnosis.

[0006] 3. Limited predictive ability: Existing diagnostic and predictive methods can only be used after patients develop obvious symptoms and cannot be predicted in advance. This limits clinicians' ability to take timely intervention measures, thereby affecting treatment effectiveness and patient prognosis.

[0007] Data-driven artificial intelligence algorithms are gaining increasing attention in sepsis prediction research. However, data imbalance is a common challenge in time series prediction and disease diagnosis. In particular, in sepsis early warning, a significant imbalance exists between the proportion of minority patient samples (those with sepsis) and healthy samples. Unfortunately, this issue remains unaddressed in current sepsis early warning research. Summary of the Invention

[0008] The technical purpose of this application is to provide a two-stage early prediction method and device for ICU sepsis in order to address the technical problem that the current ICU sepsis prediction has a small number of sepsis patient samples, which is significantly unbalanced compared with healthy samples, resulting in low sepsis prediction accuracy.

[0009] In order to achieve the above technical objectives, this application adopts the following technical solutions.

[0010] In a first aspect, the present application provides a two-stage early prediction method for ICU sepsis, comprising:

[0011] Continuously collect clinical data sets after patients enter the ICU;

[0012] Perform feature extraction based on the clinical data set to obtain a prediction feature vector;

[0013] Based on the prediction feature vector corresponding to the clinical data set at the current observation moment, using a preset classification model to determine whether the patient is a non-sepsis sample;

[0014] If the patient is not determined as a non-sepsis sample by the classification model, the preset prediction model is used to predict whether the patient has a risk of sepsis infection within a future preset time period based on the prediction feature vector corresponding to the clinical data set at the current observation time and the previously set time.

[0015] Furthermore, the clinical data set includes vital sign parameters, auxiliary examination data, drug intervention data and demographic data.

[0016] Furthermore, feature extraction is performed based on the clinical data set to obtain a prediction feature vector, including:

[0017] Determining a minimum value, a maximum value, a mean value, a standard deviation, a skewness, and / or a peak value of each of the life characteristic parameters within each sliding window to obtain a statistical indicator of each life characteristic, wherein the sliding window includes at least one observation moment;

[0018] Determining a characteristic difference between the vital characteristic statistical indicator at each observation moment and the corresponding vital characteristic statistical indicator at the previous observation moment;

[0019] For each sliding window, all vital sign statistical indicators, all feature differences, auxiliary examination data, drug intervention data, and demographic data are combined into an optimized feature vector. Variance analysis is performed based on the optimized feature vector, and the contribution of each feature in the optimized feature vector to the early prediction of sepsis is evaluated by calculating the F value. The top several features are selected from the largest to the smallest according to their contribution to form the prediction feature vector.

[0020] Furthermore, based on the predicted feature vector corresponding to the clinical data set at the current observation moment, a preset classification model is used to determine whether the patient is a non-sepsis sample, including:

[0021] Based on the predicted feature vector corresponding to the current observation moment, a pre-trained classification model is used to obtain the probability that the patient is infected with sepsis; if the probability is less than a preset threshold, the patient is determined to be a non-sepsis sample;

[0022] If the probability is greater than or equal to a preset threshold, the patient is determined to be a fuzzy sample.

[0023] Furthermore, the prediction model adopts LSTM or LightGBM architecture.

[0024] Furthermore, the method further comprises:

[0025] The Shapley feature importance analysis method is used to evaluate the contribution of each feature in the feature vector to the early prediction results of sepsis;

[0026] Present the contribution levels of each item in a chart or waterfall diagram.

[0027] Furthermore, the method further includes performing data preprocessing on the clinical data set, wherein the data preprocessing includes:

[0028] For the vital sign parameters, duplicate record rows in the clinical data set were deleted; outliers in the vital sign parameters were filtered; if there were missing values in the vital sign parameters, the corresponding missing values were initially filled with the most recent measurement value to achieve positive filling; for missing values that were still not filled after positive filling, the missing values were filled with the mean of the corresponding variables of all ICU sepsis patients; all continuous variables were standardized to the range of [0,1] using normalization; for categorical variables, one-hot encoding was used to convert them into a model-recognizable format.

[0029] In a second aspect, an embodiment of the present application provides a two-stage early prediction device for ICU sepsis, comprising:

[0030] Data collection module, used to continuously collect clinical data sets of patients after they enter the ICU;

[0031] A feature extraction module, configured to extract features based on the clinical data set to obtain a prediction feature vector;

[0032] The first stage module is configured to determine whether the patient is a non-sepsis sample using a preset classification model based on the prediction feature vector corresponding to the clinical data set at the current observation moment;

[0033] The second stage module is used to predict whether the patient has a risk of sepsis infection within a preset time period in the future based on the predicted feature vector corresponding to the clinical data set at the current observation time and the previous set time if the patient is not determined as a non-sepsis sample by the classification model.

[0034] Furthermore, the device also includes an analysis and display module, which is used to use the Shapley feature importance analysis method to evaluate the contribution of each feature in the prediction feature vector to the early prediction result of sepsis; and display each contribution in the form of a chart or waterfall chart.

[0035] In a third aspect, an embodiment of the present application further provides a computer-readable storage medium having program instructions stored thereon, which, when executed, implements a two-stage early prediction method for ICU sepsis as provided in any possible implementation scheme of the first aspect.

[0036] Compared with the existing technology, the beneficial technical effects achieved by the two-stage early prediction method and device for ICU sepsis provided by this application include at least the following: a two-stage approach is adopted, in which a classification model is first used to preliminarily determine whether the patient is a non-sepsis sample, and the patients are stratified and screened to exclude samples that are obviously not sepsis; for samples that are not excluded, the prediction feature vectors corresponding to clinical data of multiple time periods are combined, and a prediction model is used to make a more accurate sepsis risk prediction, which can reduce the impact of class imbalance on the two-stage early prediction performance of ICU sepsis and improve the accuracy of the prediction. The data used is derived from the patient's real clinical condition, so it has high feasibility and practicality in the actual ICU clinical environment and can effectively provide effective support for clinical diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The drawings described herein are for illustrative purposes only and are not intended to limit the scope of the present application in any way. In addition, the shapes and proportional dimensions of the components in the drawings are only schematic and are used to help understand the present application. They do not specifically limit the shapes and proportional dimensions of the components of the present application. Those skilled in the art can select various possible shapes and proportional dimensions to implement the present application according to the specific circumstances under the guidance of the present application. In the drawings:

[0038] Figure 1 A schematic flow chart of a two-stage early prediction method for ICU sepsis provided in an embodiment;

[0039] Figure 2 A schematic diagram illustrating the principle of predicting sepsis using the two-stage early prediction method for ICU sepsis provided in the embodiment;

[0040] Figure 3 This is a schematic diagram of the clinical onset time of sepsis according to the Third Sepsis International Consensus (Sepsis-3) criteria;

[0041] Figure 4 Schematic diagram of the inclusion process of the study population in the examples;

[0042] Figure 5Schematic diagram of the structure of a two-stage early prediction device for ICU sepsis provided in an embodiment. DETAILED DESCRIPTION

[0043] In order to enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of this application.

[0044] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of technical features indicated. Therefore, features specified as "first" or "second" may explicitly or implicitly include one or more features.

[0045] In response to the technical problem that the current ICU sepsis prediction has a small number of sepsis patient samples, which is significantly unbalanced compared with healthy samples, resulting in low sepsis prediction accuracy, the embodiments of the present application provide a two-stage early prediction method, device and storage medium for ICU sepsis, which can achieve relatively accurate early prediction of ICU sepsis.

[0046] The embodiment provides a two-stage early prediction method for ICU sepsis, such as Figure 1 Shown, including:

[0047] Step 1: Continuously collect clinical data sets after patients enter the ICU;

[0048] Step 2: Extract features based on the clinical dataset to obtain the prediction feature vector;

[0049] Step 3: Based on the predicted feature vector corresponding to the clinical dataset at the current observation moment, the preset classification model is used to determine whether the patient is a non-sepsis sample;

[0050] Step 4: If the patient is not determined as a non-sepsis sample by the classification model, the preset prediction model is used to predict whether the patient has a risk of sepsis infection within a preset time period in the future based on the prediction feature vector corresponding to the current observation time and the clinical data set within the previously set time.

[0051] Within healthcare settings, the ICU, due to its unique environment and requirements, has become a key area for big data applications. ICU wards utilize a wide range of advanced medical equipment, such as ventilators, ECG monitors, and various precision testing instruments, which collect patients' vital signs in real time, providing a rich data resource for disease prediction.

[0052] In a specific embodiment, the clinical data set includes the patient's vital sign parameters, auxiliary examination data, drug intervention data, and demographic data. Vital sign parameters may include body temperature, heart rate, blood pressure, respiratory rate, etc.; auxiliary examination data may include laboratory indicators (such as blood routine, biochemical indicators, coagulation function, etc.) and imaging examination reports. Drug intervention data includes information such as drug name, medication time, and medication dosage. Demographic data may include basic information of the patient, such as age, gender, race, underlying diseases, etc. All data are de-identified to ensure patient privacy and information security.

[0053] To ensure data accuracy and completeness, the examples systematically collected 79 features, including vital sign parameters, auxiliary examination data, medication intervention data, and demographic data. These feature data were collected at one-hour observation intervals. For multiple measurements within the same hour, the average was taken as the feature value at that time point.

[0054] In a specific embodiment, the duration of the observation time can be flexibly set according to the research purpose, the characteristics of the observed object, and the actual situation. For example, for monitoring the vital signs of critically ill ICU patients, an observation time interval may be set every 5-15 minutes to capture changes in the condition in a timely manner.

[0055] In some embodiments, data preprocessing is also performed on clinical datasets, performing preliminary processing on collected sample data to ensure data quality and improve model performance. This processing includes operations such as data cleaning, missing value filling, and feature standardization.

[0056] As an example, demographic data preprocessing includes the following steps:

[0057] ① Outlier processing: For extreme outliers of variables such as weight, experts set a reasonable range and eliminate records that are obviously inconsistent with reality.

[0058] ② Filling missing values: Missing values of continuous variable data are filled with the group mean, while missing values of categorical variable data are filled with the value that appears most frequently in the group.

[0059] Vital signs and laboratory indicators are continuous variables. The specific data preprocessing steps are as follows:

[0060] ①Duplicate record removal: Delete duplicate record rows in the data.

[0061] ② Outlier filtering: For vital sign indicators, outliers were filtered according to the upper and lower limits of the variables in the literature (Rafiei et al., 2021). For laboratory indicators, the box-plot method was used to determine the range of outliers.

[0062] ③ Missing value handling: The "forward filling + mean filling" method takes into account both the local characteristics of the data and the statistical properties of the overall data, allowing for more reasonable handling of various missing value situations. As an example, subsequent missing values are first filled with the most recent measurement value (forward filling); if a missing value for a variable is not filled after admission to the ICU, it is filled with the mean value of the corresponding variable for all patients. In addition, laboratory indicators with a missing ratio ≥40% after the initial forward filling are deleted.

[0063] Afterwards, all continuous variables are normalized to the range of [0, 1]. For categorical variables, one-hot encoding is used to convert them into a format that can be recognized by the model.

[0064] In this embodiment, step 2 constructs and extracts more predictive features from the preprocessed clinical dataset to enhance the model's learning capabilities. This may include manually designing features based on domain knowledge and data analysis results, such as calculating new mean statistics and temperature fluctuation amplitudes, and selecting feature importance based on variance analysis.

[0065] As an example, feature extraction is performed based on a clinical dataset to obtain a prediction feature vector, including:

[0066] ① Assuming each sliding window is w hours, determine the minimum, maximum, mean, standard deviation, skewness, and / or peak values of each vital characteristic parameter within the past w hours. These statistics can reflect the fluctuation of the patient's physiological state and its distribution characteristics within the specified time window, thereby obtaining the statistical indicators of each vital characteristic. The sliding window includes at least one observation moment;

[0067] ② Determine the characteristic difference between the vital characteristic statistical indicator at each observation moment and the corresponding vital characteristic statistical indicator at the previous observation moment; that is, for each vital characteristic statistical indicator feature f, calculate the characteristic difference between the current moment t and the previous moment t-1. This difference is used to capture the trend and amplitude of the characteristic change over time, revealing the potential dynamic changes in the patient's condition.

[0068] ③ For each sliding window, all vital sign statistics, all feature differences, auxiliary examination data, drug intervention data, and demographic data were combined into an optimized feature vector. Analysis of variance (ANOVA) was performed on the optimized feature vector, and the F-value was calculated to evaluate the contribution of each feature in the optimized feature vector to the early prediction of sepsis. The top features were selected from the highest contribution to the lowest contribution to form the prediction feature vector. By selecting the features that contributed most to the prediction task, important variables that helped improve model performance were retained.

[0069] like Figure 2As shown, after extracting the prediction feature vector from the original unbalanced clinical data set, the prediction feature vector corresponding to the clinical data set at the current observation time (time t) is input into the preset classification model to determine whether the patient is a non-sepsis sample.

[0070] At the observation time (time t), vital sign parameters, auxiliary examination data, drug intervention data, and demographic data are integrated to obtain a prediction feature vector, and a high-efficiency classification model is used for rapid initial screening. The classification model can be a random forest (RF).

[0071] The pre-trained classification model is used to obtain the probability of a patient being infected with sepsis. If the probability is less than a preset threshold (δ = 0.1), the patient is judged as a non-sepsis sample; if the probability is greater than or equal to the preset threshold, the patient is judged as a fuzzy sample.

[0072] The initial screening process eliminated a large number of apparently healthy samples while maintaining high accuracy for non-sepsis samples (e.g., under the RF model, on average, only one septic sample was misclassified out of 300 samples predicted as non-sepsis).

[0073] In step 4, samples marked as "fuzzy" in the first stage are analyzed in depth using complex time series models (such as LSTM and LightGBM) combined with the patient's time series data. The predicted feature vectors corresponding to the clinical dataset at the current observation time (time t) and the previously set time (within W hours) are input into the prediction model in the second stage. That is, the predicted feature vectors corresponding to time t-w+1 to time t are input into the prediction model. Using a more precise artificial intelligence algorithm, the "fuzzy" samples are classified more finely.

[0074] In the second-stage prediction algorithm, the initial screening of non-sepsis samples by the first-stage algorithm effectively alleviates the label imbalance problem of the "fuzzy" dataset. The final two-stage model dynamically updates the predicted probability of sepsis infection risk for the patient within the next N hours at each time point t+1, providing clinicians with real-time risk warnings and supporting earlier intervention decisions.

[0075] In some embodiments, the method further includes outputting the prediction results to provide auxiliary support for clinical decision making in order to improve the interpretability of the prediction results.

[0076] To improve the interpretability of prediction results, some embodiments also use the Shapley feature importance analysis method to evaluate the contribution of each feature in the feature vector to the early prediction results of sepsis; the contribution levels are displayed through charts or waterfall charts, such as how a certain vital sign or laboratory indicator affects sepsis risk assessment, and the decision basis of the model prediction is intuitively displayed through feature contribution charts or waterfall charts, helping doctors quickly understand the key drivers of patient risk.

[0077] The two-stage early prediction method for ICU sepsis provided in this application supports the continuous input of patient clinical data and the dynamic update of model results. In actual application, doctors can input the patient's real-time dynamic data into the system, and the system will give the prediction results in a short time. If the prediction results show that the patient has a higher risk of sepsis after N (such as N = 6) hours, the doctor can take corresponding intervention measures in advance, such as adjusting the antibiotic use plan, strengthening vital signs monitoring, optimizing fluid resuscitation strategies, etc., to improve the patient's diagnosis and treatment and prognosis.

[0078] In some embodiments, it is necessary to pre-build and train classification models and prediction models. In an embodiment, the classification model and prediction model can be trained using the MIMIC-IV v2.2 database. The MIMIC-IV database contains the admission records and clinical data of 73,181 patients from 2008 to 2019. The data content covers vital signs (such as body temperature, heart rate, blood pressure, respiratory rate, etc.), laboratory test results (such as blood routine, biochemical indicators, coagulation function, etc.), imaging examination reports, drug prescriptions, and basic information of patients (such as age, gender, underlying diseases, etc.). All data are de-identified to ensure patient privacy and information security. In order to ensure the accuracy and completeness of the data, a total of 79 features of vital sign parameters, auxiliary examination data, drug intervention data and demographic data are systematically collected in the embodiment. These feature data are collected according to a 1-hour time window, and the average of multiple measurements within the same hour is taken as the feature value at that time point.

[0079] According to the third international consensus on sepsis (Sepsis-3) criteria, the clinical onset time of sepsis was determined (see Figure 3 ). Suspected infection time (t suspicion ) is defined as the earlier timestamp between the administration of antibiotics and the sampling of blood cultures within the specified time window: if blood culture sampling is performed first, antibiotics must be given within 72 hours; whereas if antibiotics are given first, blood culture sampling must be performed within 24 hours. suspicion -48h,t suspicion The SOFA score was assessed every hour within 24 h, and the moment when the SOFA score increased by 2 points or more was defined as tsofa Considering that the change of SOFA score lags behind the actual sepsis infection, that is, when the patient's SOFA score changes by 2 points or more, the patient may have been infected with sepsis, this application will onset Defined as t suspicion and t sofa An earlier timestamp.

[0080] The SOFA score (Sequential Organ Failure Assessment) is an indicator that measures organ function and failure rate, and plays a vital role in clinically identifying patients with sepsis in the ICU. This application defines non-infectious sepsis patients in the control group according to the SOFA score standard, that is, the SOFA score should not change by more than 1 point during any consecutive 72 hours of ICU admission, and the SOFA score should remain below 2 points at any given time.

[0081] All patients were divided into a sepsis group (case group) and a non-sepsis group (control group). In order to accurately evaluate the early prediction ability, this application excluded patients with sepsis who developed the disease within 5 hours of ICU admission, and focused on cases in which the patient was in normal condition at admission but subsequently deteriorated. Data of patients aged 18 years and above who were admitted to the ICU for the first time, with an ICU length of stay (LoS) between 8 hours and 20 days, were selected; according to the clinical identification criteria of Sepsis-3, all patients were divided into a case group (sepsis patients) and a control group (non-sepsis patients). Once sepsis is confirmed, the patient's data records after the onset time will be eliminated. In the case group, the records of patients who contracted sepsis within 5 hours of ICU admission were further eliminated, and finally 2077 sample records of case group patients and 11,435 control group patients were obtained. The inclusion criteria for the study are shown in the table. Figure 4 , and collects hourly physiological characteristics and corresponding sepsis infection labels. These data contain missing values, which will be processed and supplemented by subsequent modules.

[0082] In the embodiment, data preprocessing is performed on the data set. For specific steps, please refer to the above data preprocessing steps.

[0083] Furthermore, to predict the onset of sepsis N hours in advance, the sepsis onset time of the case group was adjusted N hours in advance. After this adjustment, all infection labels N hours before onset were marked as 1, indicating that the patient was likely to develop sepsis within the next N hours. This method allows the model to learn the characteristic patterns of patients before onset, improving its early prediction capabilities.

[0084] Divide the prepared dataset into a training set, a validation set, and a test set. Typically, a 70% / 15% / 15% split is used. The training set is used for model training, the validation set is used for tuning model hyperparameters, and the test set is used to evaluate the performance of the final model.

[0085] For example, training a classification model (such as a random forest model) is performed using the training set data (feature vectors and corresponding labels, where the label indicates whether the patient is non-septic). The validation set data is used to evaluate the performance of the classification model under different hyperparameter settings. Metrics such as accuracy, recall, and F1 score can be used for evaluation. Cross-validation (such as GridSearchCV or RandomizedSearchCV) is used to adjust the hyperparameters of the random forest model for optimal performance.

[0086] The method for training the prediction model can adopt existing technology, which will not be described in detail in this application.

[0087] This application example utilizes vital sign parameters, auxiliary examination data, drug intervention data, and demographic data to construct a precise sepsis early warning model. This model comprises a first-stage classification model and a second-stage prediction model. While traditional methods typically focus on a single or a few indicators, this application fully exploits the vast data resources provided by modern medical information technology, significantly improving the accuracy and reliability of predictions.

[0088] In the real-time early warning task of sepsis, this application proposes a two-stage prediction framework to address the imbalance of sepsis labels (positive samples account for only 15.37%). The first stage quickly screens obvious non-sepsis samples and reduces the label imbalance in the training data through a simple and efficient algorithm. The second stage uses a more complex time series model to perform detailed classification on the remaining "fuzzy" samples. This two-stage prediction process fully utilizes the combination of simple and complex models to alleviate the sample imbalance problem in the prediction process, while improving the accuracy of early warning of sepsis. It also has good flexibility and scalability, and can optimize model performance through the combination of different algorithms.

[0089] This application's AI-based two-stage early prediction system for ICU sepsis can predict risk N hours before onset, providing doctors with valuable time for early intervention, early warning, and improved intervention efficiency. Doctors can take effective treatment measures before the disease has fully developed and arrange more targeted examinations and treatments based on the predicted results, reducing unnecessary diagnostic steps and waiting time, making the medical process more efficient and orderly.

[0090] The overall performance of the two-stage early prediction method for ICU sepsis provided in this application was compared with the common SMOTE technique and cost-sensitive loss function. The results showed that most prediction algorithms showed enhanced performance after applying the two-stage prediction framework compared to using SMOTE or cost-sensitive loss function alone.

[0091] Based on the same inventive concept as the two-stage early prediction method for ICU sepsis provided in the above embodiment, the present application embodiment also provides a two-stage early prediction device for ICU sepsis, such as Figure 5 As shown, it includes: a data acquisition module, a feature extraction module, a first-stage module, and a second-stage module. Among them, the data acquisition module is used to continuously collect clinical data sets of patients after they enter the ICU; the feature extraction module is used to extract features based on the clinical data sets to obtain prediction feature vectors; the first-stage module is used to determine whether the patient is a non-sepsis sample based on the prediction feature vector corresponding to the clinical data set at the current observation time, using a preset classification model; if the patient is not determined to be a non-sepsis sample by the classification model, the second-stage module is used to predict whether the patient has a risk of sepsis infection within a preset time period in the future based on the prediction feature vector corresponding to the clinical data set at the current observation time and the previously set time.

[0092] In some embodiments, the device further includes an analysis and display module (not shown in the figure), which is used to use the Shapley feature importance analysis method to evaluate the contribution of each feature in the prediction feature vector to the early prediction results of sepsis; and to display the contribution of each feature in a chart or waterfall chart.

[0093] The first-stage module and the second-stage module implement two-stage early prediction of ICU sepsis. For patient i who enters the ICU for the tth hour, the first-stage module first uses the first-stage sepsis prediction algorithm (such as random forest, kNN, support vector machine SVM, etc.) to perform initial screening based on the patient's simple original physiological indicators, with the aim of quickly identifying non-septic patients with obvious mild symptoms.

[0094] If the probability of sepsis infection predicted by the first-stage algorithm is lower than the predetermined threshold σ, the patient is predicted to have a non-septic infection within the next N hours. Then, the algorithm proceeds to time t+1 and collects new patient indicator data for real-time sepsis prediction updates.

[0095] Otherwise, for those "fuzzy" datasets where the first-stage algorithm predicts a high probability of sepsis infection (exceeding a threshold σ), the sample imbalance issue is alleviated after initial screening with negative samples. Subsequently, the second-stage module uses more complex patient time series metrics, combined with a weighted loss function, to train different types of machine learning and deep learning models to perform secondary classification predictions on the "fuzzy" datasets. If the secondary classification prediction indicates a high risk of sepsis infection, an alarm is issued to alert clinical staff. Otherwise, the system proceeds to time t+1 and continues to collect new patient metric data to update the two-stage sepsis prediction in real time.

[0096] The two-stage early prediction device for ICU sepsis provided in this application implements a two-stage sepsis early warning framework to optimize prediction performance. The two-stage framework gradually improves the model's performance on imbalanced datasets through a two-step process of initial screening and refined classification.

[0097] It is understood that the specific definition of the two-stage early prediction device for ICU sepsis provided in this application can be found in the definition of the two-stage early prediction method for ICU sepsis above, and will not be repeated here. Each module in the above-mentioned two-stage early prediction device for ICU sepsis can be implemented in whole or in part by software, hardware, and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to the above modules.

[0098] The embodiments of the present application also provide a computer-readable storage medium having program instructions stored thereon, which, when executed, implement the steps of the two-stage early prediction method for ICU sepsis provided in the above embodiments.

[0099] The above is a detailed introduction to the two-stage early prediction method, device and storage medium for ICU sepsis provided by this application. This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the concept of this application and should not be understood as limiting the scope of protection of this application.

Claims

1. A two-stage early prediction method for ICU sepsis, characterized by: include: Continuously collect clinical data sets after patients enter the ICU; Perform feature extraction based on the clinical data set to obtain a prediction feature vector; Based on the prediction feature vector corresponding to the clinical data set at the current observation moment, using a preset classification model to determine whether the patient is a non-sepsis sample; If the patient is not determined as a non-sepsis sample by the classification model, the preset prediction model is used to predict whether the patient has a risk of sepsis infection within a future preset time period based on the prediction feature vector corresponding to the clinical data set at the current observation time and the previously set time.

2. The two-stage early prediction method for ICU sepsis according to claim 1, characterized in that: The clinical data set includes vital sign parameters, auxiliary examination data, drug intervention data and demographic data.

3. The two-stage early prediction method for ICU sepsis according to claim 2, characterized in that: Feature extraction is performed based on the clinical data set to obtain a prediction feature vector, including: Determining a minimum value, a maximum value, a mean value, a standard deviation, a skewness, and / or a peak value of each of the life characteristic parameters within each sliding window to obtain a statistical indicator of each life characteristic, wherein the sliding window includes at least one observation moment; Determining a characteristic difference between the vital characteristic statistical indicator at each observation moment and the corresponding vital characteristic statistical indicator at the previous observation moment; For each sliding window, all vital sign statistical indicators, all feature differences, auxiliary examination data, drug intervention data, and demographic data are combined into an optimized feature vector; variance analysis is performed based on the optimized feature vector, and the contribution of each feature in the optimized feature vector to the early prediction of sepsis is evaluated by calculating the F value. The top several features are selected from the largest to the smallest according to their contribution to form the prediction feature vector.

4. The two-stage early prediction method for ICU sepsis according to claim 1, characterized in that: Determining whether the patient is a non-sepsis sample using a preset classification model based on the predicted feature vector corresponding to the clinical data set at the current observation moment includes: Based on the predicted feature vector corresponding to the current observation moment, a pre-trained classification model is used to obtain the probability that the patient is infected with sepsis; if the probability is less than a preset threshold, the patient is determined to be a non-sepsis sample; If the probability is greater than or equal to a preset threshold, the patient is determined to be a fuzzy sample.

5. The two-stage early prediction method for ICU sepsis according to claim 1, characterized in that: The prediction model adopts LSTM or LightGBM architecture.

6. The two-stage early prediction method for ICU sepsis according to claim 1, characterized in that: The method further comprises: The Shapley feature importance analysis method is used to evaluate the contribution of each feature in the feature vector to the early prediction results of sepsis; Present the contribution levels of each item in a chart or waterfall diagram.

7. The two-stage early prediction method for ICU sepsis according to claim 2, characterized in that: The method further includes performing data preprocessing on the clinical data set, wherein the data preprocessing includes: For the vital sign parameters, duplicate record rows in the clinical data set were deleted; outliers in the vital sign parameters were filtered; if there were missing values in the vital sign parameters, the corresponding missing values were initially filled with the most recent measurement value to achieve positive filling; for missing values that were still not filled after positive filling, the missing values were filled with the mean of the corresponding variables of all ICU sepsis patients; all continuous variables were standardized to the range of [0,1] using normalization; for categorical variables, one-hot encoding was used to convert them into a model-recognizable format.

8. A two-stage early prediction device for ICU sepsis, characterized by: The device comprises: Data collection module, used to continuously collect clinical data sets of patients after they enter the ICU; A feature extraction module, configured to extract features based on the clinical data set to obtain a prediction feature vector; The first stage module is configured to determine whether the patient is a non-sepsis sample using a preset classification model based on the prediction feature vector corresponding to the clinical data set at the current observation moment; The second stage module is used to predict whether the patient has a risk of sepsis infection within a preset time period in the future based on the predicted feature vector corresponding to the clinical data set at the current observation time and the previous set time if the patient is not determined as a non-sepsis sample by the classification model.

9. The two-stage early prediction device for ICU sepsis according to claim 8, characterized in that: The device also includes an analysis and display module, which is used to use the Shapley feature importance analysis method to evaluate the contribution of each feature in the prediction feature vector to the early prediction result of sepsis; and to display each contribution in a chart or waterfall chart.

10. A computer-readable storage medium, characterized in that Program instructions are stored thereon, and when the program instructions are executed, the steps of the two-stage early prediction method for ICU sepsis according to any one of claims 1 to 7 are implemented.

Images