Device, method, equipment, medium and program product for predicting 30-day death rate of cerebral hemorrhage patient

Abstract

The invention provides a 30-day death rate prediction device and method for cerebral hemorrhage patients, equipment, a medium and a program product, relates to the technical field of medical biology, and analyzes data in MIMIC and eICU-CRD databases. And performing feature selection by using Lasso regularization, mutual information and recursive feature elimination (RFE). And training to obtain a prediction model, and performing internal verification and external verification by applying comprehensive performance indexes. The feature importance is interpreted using an SHAP value. Determining key predictors includes acute physiological score III (APSIII), age, leukocyte (WBC), and albumin level. Compared with traditional scoring systems such as GCS and APSIII, the model shows higher prediction accuracy in the aspect of predicting the death rate in 30 days, and has more balanced sensitivity and specificity. The machine learning method provided by the embodiment of the invention shows a promising prognosis value in the aspect of predicting the 30-day death risk of ICH patients received by ICU. It needs to be noted that the models can only serve as supplementary judgments and cannot replace clinical judgments.

Description

TECHNICAL FIELD

[0001] The embodiment of the present application relates to the field of medical biotechnology, in particular to a device, method, equipment, medium and program product for predicting 30-day mortality of a patient with cerebral hemorrhage. BACKGROUND

[0002] Cerebral hemorrhage (ICH) is a severe subtype of stroke characterized by spontaneous bleeding within the brain parenchyma. ICH accounts for approximately 10-15% of all stroke cases, affecting nearly 2 million people worldwide each year. Within 30 days after the occurrence of ICH, the mortality rate is surprisingly more than 40%, and a large number of survivors suffer from severe disability. ICH has a significant impact on the world, especially on patients in intensive care units (ICU), which raises an urgent need for mortality prediction of ICH patients in order to quickly assess the risk of death and promote timely intervention. SUMMARY

[0003] The embodiment of the present application provides a device, method, equipment, medium and program product for predicting 30-day mortality of a patient with cerebral hemorrhage, aiming to fill the existing knowledge gap by predicting the mortality rate at 14 days and 30 days after ICH in young ICU patients, and providing a powerful auxiliary tool for clinicians to affect intervention measures, guide clinical judgment, and ultimately improve patient survival rate.

[0004] The first aspect of the embodiment of the present application provides a device for predicting 30-day mortality of a patient with cerebral hemorrhage, the device comprising:

[0005] an acquisition module configured to acquire mortality key prediction indicators of an ICH patient entering an ICU, the mortality key prediction indicators comprising at least: acute physiology score III (APS III), age, white blood cell level, and albumin level;

[0006] a prediction module configured to predict 30-day mortality of the patient based on the mortality key prediction indicators and a 30-day mortality prediction model, the 30-day mortality prediction model being trained based on sample mortality key prediction indicator data of ICH patients who died within 30 days after entering the ICU;

[0007] wherein APS III and age are positively correlated with 30-day mortality, albumin level is negatively correlated with 30-day mortality, and white blood cell level has a V-shaped relationship with 30-day mortality, and excessively high and low white blood cell levels both increase the risk of 30-day mortality.

[0008] Optionally, the device further comprises a training module, the training module comprising:

[0009] a sample acquisition submodule configured to acquire a plurality of medical indicator data of ICH patients who died within 30 days after entering the ICU from a plurality of public databases;

[0010] a feature determination sub-module, configured to determine, based on a feature selection method, key features related to the mortality rate within 30 days after entering the ICU from the plurality of medical indicators respectively; the key features include: age, systolic blood pressure SBP, APS III, mean blood pressure MBP, Glasgow coma scale GCS, body temperature, oxygen saturation SpO2, blood glucose level BGL, lymphocyte count, neutrophil count, albumin, chlorine, platelet count, and white blood cell count WBC;

[0011] a training sub-module, configured to train a preset model based on the medical indicator data corresponding to the key features, and to fine-tune the hyperparameters of the preset model by combining the Bayesian optimization of the Gaussian process in the training process to obtain a prediction model;

[0012] an analysis sub-module, configured to analyze the SHAP value of the prediction model, analyze the influence and correlation of each key feature in the preset model based on the SHAP value, and determine key prediction indicators related to the mortality rate within 30 days.

[0013] Optionally, the preset model is a plurality of preset models, and the training sub-module is specifically configured to train the plurality of preset models based on the medical indicator data corresponding to the key features respectively, fine-tune the hyperparameters of each preset model by combining the Bayesian optimization of the Gaussian process in the training process to obtain a plurality of prediction models, and the device further comprises:

[0014] a first evaluation module, configured to evaluate the plurality of prediction models based on accuracy, sensitivity, specificity, F1 score, and AUC respectively, determine a target prediction model based on the evaluation results, and take the preset model corresponding to the target prediction model as a target model;

[0015] a second training module, configured to train the target model based on the key prediction indicator data of the ICH patients who died within 30 days after entering the ICU to obtain a 30-day mortality rate prediction model.

[0016] Optionally, the preset model is a plurality of preset models, and the training sub-module is specifically configured to train the plurality of preset models based on the medical indicator data corresponding to the key features respectively, fine-tune the hyperparameters of each preset model by combining the Bayesian optimization of the Gaussian process in the training process to obtain a plurality of prediction models, and the device further comprises:

[0017] a second evaluation module, configured to evaluate the plurality of prediction models based on accuracy, sensitivity, specificity, F1 score, and AUC respectively, and determine a target prediction model based on the evaluation results;

[0018] The third training module is configured to perform parameter updating on the target prediction model based on the key prediction index data of the ICH patients who died within 30 days after entering the ICU to obtain a 30-day mortality prediction model.

[0019] Optionally, the device further comprises a reminding module configured to output reminding information according to the 30-day mortality.

[0020] The second aspect of the embodiment of the present application provides a 30-day mortality prediction method for ICH patients, and the method comprises the following steps of:

[0021] The key prediction index of mortality of the ICH patients entering the ICU is obtained, and the key prediction index of mortality at least comprises an acute physiology score III (APS III), an age, a white blood cell level and an albumin level.

[0022] The 30-day mortality of the patient is predicted based on the 30-day mortality prediction model according to the key prediction index of mortality, and the 30-day mortality prediction model is obtained based on the sample key prediction index of mortality of the ICH patients who died within 30 days after entering the ICU.

[0023] The APS III and the age are positively correlated with the 30-day mortality, the albumin level is negatively correlated with the 30-day mortality, the white blood cell level has a V-shaped relationship with the 30-day mortality, and the excessively high and low white blood cell levels increase the risk of 30-day mortality.

[0024] The third aspect of the embodiment of the present application provides an electronic device, which comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor implements the 30-day mortality prediction method for ICH patients according to the second aspect when executing the computer program.

[0025] The fourth aspect of the embodiment of the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the 30-day mortality prediction method for ICH patients according to the second aspect.

[0026] The fifth aspect of the embodiment of the present application provides a computer program product, which comprises a computer program / instruction, and the computer program / instruction is executed by a processor to implement the steps in the 30-day mortality prediction method for ICH patients according to the second aspect.

[0027] In the embodiments of the present application, data in the MIMIC and eICU-CRD databases are analyzed, including patient demographics, clinical indicators, and survival outcomes. Lasso regularization, mutual information, and recursive feature elimination (RFE) are used for feature selection. Various ML models are evaluated, including logistic regression, multilayer perceptron (MLP), random forest, K-nearest neighbors (KNN), and XGBoost, and comprehensive performance metrics are applied for internal and external validation. SHAP values are used to explain feature importance. The random forest model shows higher discriminative ability, with an AUC of 0.877 for internal validation and an AUC of 0.838 for external validation. Key predictors include acute physiology score III (APS III), age, white blood cell (WBC), and albumin levels. Compared with traditional scoring systems such as GCS and APS III, the model shows higher prediction accuracy in predicting 30-day mortality, with more balanced sensitivity and specificity. The machine learning method proposed in the embodiments of the present application shows promising prognostic value in predicting the 30-day mortality risk of ICH patients admitted to the ICU. It should be noted that the 30-day mortality prediction model proposed in the embodiments of the present application can only be used as a supplementary judgment and cannot replace clinical judgment. BRIEF DESCRIPTION OF DRAWINGS

[0028] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the description of the embodiments of the present application. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0029] Figure 1 A schematic diagram showing the performance of the machine learning algorithm in internal and external validation sets by ROC curve is shown;

[0030] Figure 2 A SHAP summary plot of the random forest model in external validation is shown;

[0031] Figure 3 A SHAP dependence plot of the random forest model in external validation is shown;

[0032] Figure 4 A structure block diagram of the 30-day mortality prediction device for patients with cerebral hemorrhage provided by the embodiments of the present application is shown. DETAILED DESCRIPTION

[0033] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the following will further describe the present application in combination with the drawings and specific embodiments.

[0034] Advances in information technology have fueled the rapid expansion of medical data, making comprehensive analysis of electronic medical records (EMRs) and related administrative datasets possible. These analyses are critical in refining medical diagnoses, treatment regimens, prognostic assessments, and identifying risk factors to improve patient care. Machine learning (ML) has emerged as an important tool in this field, uncovering non-linear patterns, providing deep data insights, and improving prediction accuracy. Various ML techniques, including logistic regression (LR), random forest (RF), and support vector machines (SVM), have been explored for ICH prognosis and mortality estimation, with varying results.

[0035] Given the acute nature of intracerebral hemorrhage (ICH) and the complexity of medical data interpretation, embodiments of the present invention aim to develop a predictive model using machine learning techniques, focusing on ICH patients admitted to the ICU. By studying the nuances of model selection, fitting, and prediction, the goal is to provide valuable perspectives for medical professionals and researchers, aiding in model development and clinical protocol formulation, primarily predicting 30-day mortality risk in ICH patients.

[0036] Specifically, in embodiments of the present invention, the eICU Collaborative Research Database (eICU-CRD) and the Medical Information MIMIC-III (v1.4) and MIMIC-IV (v2.0) databases were used. These publicly available databases contain de-identified health data collected during routine clinical care in a regional medical center ICU (MIMIC) and a geographically distributed ICU (eICU-CRD).

[0037] In embodiments of the present invention, Navicat Premium v15.0.12 and SQL were used to organize and analyze patient data in detail. Specific ICD-9 codes, including 431, I610-I616, and I618-I619, were used to highlight patients with ICH as the primary diagnosis. In embodiments of the present invention, individuals under the age of 18 or those with less than 24 hours of first ICU stay were excluded. For patients with repeated ICU admissions, only the data from the first admission was used. The MIMIC data was split in an 8:2 ratio to form the training and internal validation datasets, with the eICU-CRD data serving as the external validation dataset. Embodiments of the present invention primarily focused on the mortality rate within 30 days for ICH patients admitted to the ICU.

[0038] After strict filtering based on criteria, 2478 individuals from MIMIC and 1033 individuals from eICU-CRD were included in the study for further analysis. In ICH patients treated in the ICU, the 30-day mortality rate was 29.14% in MIMIC and 21.88% in eICU-CRD. The average time to death was 6.68 days (standard deviation: 6.77) and 6.53 days (standard deviation: 5.44), respectively. In MIMIC, the average age of patients who died within 30 days after ICU was 76.00 years (interquartile range: 64.00-83.00), of which 49.17% were female. The average age of survivors was 67.00 years (interquartile range: 55.75-78.00), of which 44.59% were female. In eICU-CRD, the average age of non-survivors was 71.00 years (interquartile range: 59.00-79.75), of which 48.23% were female. In contrast, the average age of survivors was 68.00 years (interquartile range: 57.00-80.00), of which 46.22% were female.

[0039] After determining the study subjects, the relevant data were collected in various tables using their hadm_id, icustay_id and patientunitstay_id in the embodiments of the present application, including demographic details, medical conditions, clinical indicators, ventilation details, assessment indicators, survival outcomes and hospital transfer records. Their earliest vital signs and laboratory results during the first ICU stay were analyzed. The systemic immune inflammation (SII) index was determined using the formula: platelets multiplied by neutrophils divided by lymphocytes. The severity indicators in the embodiments of the present application include Acute Physiology Score III (APS III) and Glasgow Coma Scale (GCS).

[0040] In view of the difference in the magnitude of the features, a standard scaler was used to normalize the data in the embodiments of the present application, which was carefully executed after data segmentation to reduce the risk of data overflow and enhance the universal applicability of the model.

[0041] In the embodiments of the present application, three feature selection methods were used for the sample data in the training set: Lasso regularization, mutual information and recursive feature elimination (RFE) to determine the relevant features related to the 30-day mortality rate after ICU admission. Lasso regularization is a regression method that penalizes complex models, thereby reducing the coefficients of non-essential features to zero. Mutual information quantifies the interdependence between two variables, which helps to identify the importance of features to the expected outcome. RFE is an iterative technique that ranks features based on their relevance and systematically removes the least important features. The meaningful indicators selected by the three methods are cross-applied to the machine learning models subsequently established.

[0042] Specifically, among the 45 primary potential indicators of 30-day mortality in ICU patients, embodiments of the present application ensured data integrity by imputing missing values. After imputation, methods including Lasso regularization, mutual information, and RFE reduced the potential indicators to 14 key factors. Including age, systolic blood pressure (SBP), acute physiology score III (APSI I I), mean blood pressure (MBP), Glasgow coma scale (GCS), temperature, oxygen saturation (Sp02), blood glucose level (BGL), lymphocyte count, neutrophil count, albumin, chloride, platelet count, and white blood cell count (WBC), these factors represent a combination of demographic data, clinical statistics, and laboratory results.

[0043] In embodiments of the present application, five machine learning architectures were employed to predict 30-day mortality, including: logistic regression, multilayer perceptron (MLP), random forest, KNN, and XGBoost. Gaussian process-based Bayesian optimization was used to fine-tune the hyperparameters of each model, with the goal of maximizing the AUC score through cross-validation on the training data. Once the optimal hyperparameters were determined, the models were trained on the selected features and evaluated on the validation set.

[0044] In embodiments of the present application, each prediction model was rigorously validated using both the internal validation dataset of MIMIC and the external validation dataset of eICU-CRD. In embodiments of the present application, metrics such as accuracy, sensitivity, specificity, F1 score, and AUC were used to comprehensively evaluate the predictive ability of the models. The discriminative ability of the models was visually represented through ROC curves.

[0045] In embodiments of the present application, SHapley Additive exPlanation (SHAP) values were used to systematically determine the feature importance of each model, which helped to understand the importance and direction of the influence of each feature.

[0046] In embodiments of the present application, based on the above-mentioned 14 key factors, ML models including random forest, KNN, MLP neural network, XGBoost, and logistic regression were trained. Bayesian optimization was used to fine-tune the parameters of these models, with the goal of improving the average AUC through 5-fold cross-validation. The performance of these optimized ML models was rigorously evaluated through internal and external validation queues. The random forest model showed significant performance, with an AUC of 0.877 and an accuracy of 82.7%. In addition, its sensitivity, specificity, and F1 score were 53.1%, 94.6%, and 63.9%, respectively. This superior performance was further confirmed through external validation, where the AUC of the random forest was 0.838, the accuracy was 82.4%, the sensitivity was 49.6%, the specificity was 91.6%, and the F1 score was 55.2%. For detailed comparisons, please refer to Table 1 andFigure 1 In Figure 1 the left graph represents the ROC curve of the machine learning algorithm based on the internal validation set, and the right graph represents the ROC curve of the machine learning algorithm based on the external validation set. The blue line represents the ROC curve of the random forest model, the yellow line represents the ROC curve of KNN, the green line represents the ROC curve of the MLP neural network, the red line represents the ROC curve of XGBoost, and the purple line represents the ROC curve of the logistic regression model.

[0047] Table 1. Performance of the prediction model

[0048]

[0049] To illustrate the main determinants within the random forest model, in the embodiments of the present application, a SHAP summary plot is used, as shown in Figure 2 , wherein the color gradient from blue to red represents the feature value from low to high, and the abscissa represents the SHAP value. According to Figure 2 , it can be seen that the features in the order of contribution of each key factor to the model are: APS III, age, WBC, albumin, GCS, lymphocyte count, BGL, platelet count, SpO2, chlorine, body temperature, SBP, neutrophil count and MBP. This representation clearly outlines the association between individual feature values and their associated SHAP values. It is worth noting that features with higher SHAP values indicate an increased risk of death within 30 days in ICH patients after ICU admission. The SHAP dependence plot examines the influence of individual features on the model prediction result. SHAP values greater than zero indicate an increased risk of death. Figure 3 The SHAP dependence plots of the four most influential predictors are shown, wherein part A represents the SHAP dependence plot of the APS III score factor; part B represents the SHAP dependence plot of the age factor; part C represents the SHAP dependence plot of the white blood cell count factor; and part D represents the SHAP dependence plot of the albumin factor. In this case, APS III and age are positively correlated with SHAP values, meaning that an increase in these factors can indicate a higher probability of the event occurring. On the contrary, albumin is negatively correlated with SHAP values, indicating that a lower albumin level can result in a lower probability of the event. The WBC level shows a V-shaped relationship with the SHAP value, meaning that both too high and too low WBC levels can increase the risk of death within 30 days in ICU ICH patients.

[0050] To verify the robustness of the prediction model trained by the embodiments of the present application, comparative analysis with benchmark models recorded by related technologies is also carried out. The Glasgow Coma Scale (GCS) scoring system is particularly important due to its simplicity and reliability, and has been widely used to predict the mortality risk of ICU patients. Similarly, Acute Physiology Score III (APS III) containing 17 physiological variables is also often used to predict the prognosis of ICU patients. In order to compare the model of the embodiments of the present application with the above benchmark models, the excellent prediction effect of the prediction model trained by the embodiments of the present application is illustrated.

[0051] In the embodiments of the present application, whether the variable is continuous or categorical, the median and quartile [M(Q1, Q3)] or count and proportion [N(%) ] are used. Any missing values in the dataset are comprehensively processed using multiple imputation techniques. All analysis procedures are carefully performed using Python 3.11.

[0052] The embodiments of the present application compare the performance of the GCS and APS III scoring systems based on logistic regression with the random forest prediction model of the embodiments of the present application. The random forest prediction model performs better in F1 score, with internal and external validation scores of 0.639 and 0.552, respectively, exceeding the GCS and APS III systems based on logistic regression. It is worth noting that the lower sensitivity of GCS (internal validation 0.133 and external validation 0.084) affects its overall F1 score. At the same time, the random forest exhibits more consistent AUC values on the dataset (internal 0.877, external 0.838), compared with the significant variation in AUC values of GCS (internal 0.564, external 0.882). Although APS III is slightly better than GCS in some indicators, it does not reach the level of the random forest model overall. Overall, the random forest shows stronger performance indicators in prediction accuracy and stability than GCS and APS III.

[0053] Through the exploratory research of the embodiments of the present application, the embodiments of the present application explore the incidence and related risk factors of death within 30 days after ICU admission in ICH patients. The embodiments of the present application establish a prediction model to identify the key determinants leading to these deaths. The main findings include: the embodiments of the present application introduce five ML models to predict the 30-day mortality rate, among which the random forest model shows higher universality. In the random forest model, APS III, age, WBC and albumin are identified as the most influential features, and further analysis is carried out on their influence on the model. The prediction model of the embodiments of the present application is superior to several previously published models.

[0054] Previous studies have mainly focused on mortality in ICU patients with ICH based on the MIMIC-III database. The present embodiment of the invention, based on the MIMIC database, combined with eICU-CRD data, achieved a credible prediction performance (AUC = 0.838) in external validation. In addition, the direct comparison of the random forest model of the present embodiment of the invention with two published models highlights the superiority of the method of the present embodiment of the invention.

[0055] In the field of prognosis prediction for ICU patients with cerebral hemorrhage and stroke, the random forest algorithm shows enhanced prediction ability, and the present embodiment of the invention not only demonstrates the effectiveness of random forest in clinical scenarios, but also demonstrates its substantial impact on clinical decision-making, which can improve the prognosis accuracy and treatment methods in cerebrovascular care.

[0056] To understand the impact of the predictors in more detail, the present embodiment of the invention assessed the importance of features through SHAP values. The present embodiment of the invention highlights the importance of the APS III score in predicting the 30-day mortality risk of ICU patients with ICH, and the possible prognosis threshold is about 60 points. The APS III score integrates 17 physiological variables, such as vital signs, inflammatory markers, and internal environment indicators, providing a comprehensive and detailed quantification of the patient's multi-system injury risk, severity, and overall condition. Higher APS III scores are associated with increased short-term mortality risk in patients with gastrointestinal bleeding, sepsis-related encephalopathy, and atrial fibrillation in the ICU. WBC count is a common indicator of systemic inflammation and immune response, showing a V-shaped correlation with 30-day mortality in ICU patients with ICH, indicating that both abnormally low and high counts are associated with increased mortality risk. Leukopenia may be associated with immune deficiency, making patients susceptible to complications such as infection, thereby increasing the risk of death. In contrast, leukocytosis indicates systemic stress and excessive inflammation, often exacerbating brain injury and possibly indicating secondary infection, a major cause of death in ICU patients with ICH. It was found that a decrease in serum albumin levels was significantly associated with potential indicators of malnutrition or increased inflammation and increased mortality risk.

[0057] The present embodiment of the invention provides a method for taking a comprehensive risk assessment of patients with ICH in the ICU. However, it must be emphasized that the results of the prediction model should be used as a supplementary reference, not a substitute for clinical decision-making. In the case of a high mortality risk prediction, medical personnel can be reminded to conduct a more thorough assessment and potential treatment adjustment, rather than prematurely leading to end-of-life care decisions.

[0058] The embodiments of the present application successfully develop and verify a machine learning model for predicting 30-day mortality of ICH patients admitted to ICU using MIMIC and eICU-CRD databases. The random forest model, as a reliable prediction model, shows solid performance in both datasets. The embodiments of the present application emphasize the importance of integrating advanced analysis into medical prognosis, identifying APS III, age, WBC and albumin level as key factors as significant predictors. By incorporating these models into clinical practice, the embodiments of the present application aim to promote early intervention, possibly improve patient outcomes and improve the overall quality of ICU care for ICH patients.

[0059] Based on the above research, the embodiments of the present application further provide a 30-day mortality prediction device for ICH patients, as shown in Figure 4 The 30-day mortality prediction device for ICH patients 400 provided by the embodiments of the present application includes:

[0060] The acquisition module 401 is configured to acquire key mortality prediction indicators of ICH patients admitted to ICU, the key mortality prediction indicators at least including: acute physiology score III, age, white blood cell level and albumin level;

[0061] The prediction module 402 is configured to predict 30-day mortality of the patient based on the key mortality prediction indicators, based on a 30-day mortality prediction model, the 30-day mortality prediction model being trained based on sample key mortality prediction indicator data of ICH patients who died within 30 days after being admitted to ICU;

[0062] Among them, APS III and age are positively correlated with 30-day mortality, albumin level is negatively correlated with 30-day mortality, and white blood cell level shows a V-shaped relationship with 30-day mortality, and excessively high and low white blood cell level both increase the risk of 30-day mortality.

[0063] Optionally, the device further includes a training module, which includes:

[0064] The sample acquisition submodule is configured to acquire a plurality of medical indicator data of ICH patients who died within 30 days after being admitted to ICU from a plurality of public databases;

[0065] The feature determination submodule is configured to determine key features related to 30-day mortality after being admitted to ICU from the plurality of medical indicators respectively based on a feature selection method; the key features including: age, systolic blood pressure SBP, APS III, mean blood pressure MBP, Glasgow coma scale GCS, body temperature, oxygen saturation SpO2, blood glucose level BGL, lymphocyte count, neutrophil count, albumin, chlorine, platelet count and white blood cell count WBC.

[0066] a training submodule configured to train a preset model based on medical index data corresponding to the key features, and to fine-tune hyperparameters of the preset model by combining Bayesian optimization of a Gaussian process in a training process to obtain a prediction model;

[0067] an analysis submodule configured to analyze SHAP values of the prediction model, to analyze influences and correlations of each key feature in the preset model based on the SHAP values, and to determine key prediction indexes related to 30-day mortality.

[0068] Optionally, the preset model is multiple, and the training submodule is specifically configured to train multiple preset models based on medical index data corresponding to the key features, to fine-tune hyperparameters of each preset model by combining Bayesian optimization of a Gaussian process in a training process to obtain multiple prediction models, and the device further comprises:

[0069] a first evaluation module configured to evaluate the multiple prediction models based on accuracy, sensitivity, specificity, F1 score and AUC respectively, to determine a target prediction model based on an evaluation result, and to take a preset model corresponding to the target prediction model as a target model;

[0070] a second training module configured to train the target model based on key prediction index data of ICH patients who died within 30 days after entering an ICU to obtain a 30-day mortality prediction model.

[0071] Optionally, the preset model is multiple, and the training submodule is specifically configured to train multiple preset models based on medical index data corresponding to the key features, to fine-tune hyperparameters of each preset model by combining Bayesian optimization of a Gaussian process in a training process to obtain multiple prediction models, and the device further comprises:

[0072] a second evaluation module configured to evaluate the multiple prediction models based on accuracy, sensitivity, specificity, F1 score and AUC respectively, and to determine a target prediction model based on an evaluation result;

[0073] a third training module configured to perform parameter updating on the target prediction model based on key prediction index data of ICH patients who died within 30 days after entering an ICU to obtain a 30-day mortality prediction model.

[0074] Optionally, the device further comprises a reminding module configured to output reminding information according to the 30-day mortality.

[0075] In the embodiment of the present application, after determining the key prediction indicators related to the 30-day mortality, the sample mortality key prediction indicator data can be obtained based on the medical indicator data corresponding to the key prediction indicators related to the 30-day mortality, and the target model (random forest model) is trained and verified to obtain the 30-day mortality prediction model.

[0076] In the embodiment of the present application, after the prediction model is trained based on the medical indicator data corresponding to the key features, the prediction model is further fine-tuned based on the medical indicator data corresponding to the key prediction indicators, and the 30-day mortality prediction model is obtained through parameter update training.

[0077] In the fine-tuning process, the data of the features missing in the key prediction indicators of the fine-tuning samples relative to the key features can be filled with zero.

[0078] Based on the same inventive concept, the embodiment of the present application also provides a 30-day mortality prediction method for a cerebral hemorrhage patient, which comprises:

[0079] The key prediction indicators of the mortality of the ICH patients entering the ICU are obtained, and the key prediction indicators of the mortality at least include acute physiology score III, age, white blood cell level and albumin level.

[0080] According to the key prediction indicators of the mortality, the 30-day mortality of the patient is predicted based on the 30-day mortality prediction model, and the 30-day mortality prediction model is obtained based on the sample mortality key prediction indicator data of the ICH patients who died within 30 days after entering the ICU;

[0081] Among them, APS III and age are positively correlated with 30-day mortality, albumin level is negatively correlated with 30-day mortality, and white blood cell level has a V-shaped relationship with 30-day mortality, and too high or too low white blood cell level increases the risk of 30-day mortality.

[0082] Based on the same inventive concept, the embodiment of the present application also provides an electronic device, which comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor executes the computer program to realize the 30-day mortality prediction method for a cerebral hemorrhage patient as described in the above embodiment.

[0083] Based on the same inventive concept, the embodiment of the present application also provides a computer readable storage medium having a computer program stored thereon, and the computer program is executed by a processor to realize the 30-day mortality prediction method for a cerebral hemorrhage patient as described in the above embodiment.

[0084] Based on the same inventive concept, the embodiment of the present application also provides a computer program product comprising computer programs / instructions which, when executed by a processor, implement the steps of the brain hemorrhage patient 30-day mortality rate prediction method described in the above embodiment.

[0085] Although the preferred embodiments of the embodiments of the present application have been described, those skilled in the art can make additional changes and modifications to the embodiments once they know the basic inventive concept. Therefore, the appended claims are intended to be interpreted as including all the changes and modifications falling within the scope of the embodiments of the present application.

[0086] The above describes the brain hemorrhage patient 30-day mortality rate prediction device, method, equipment, medium and program product provided by the present application in detail, the principle and implementation mode of the present application are described in this paper by applying specific examples; the above embodiment is only used to help understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed; in view of the above, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A device for predicting the 30-day mortality rate of patients with cerebral hemorrhage, characterized in that, The device includes: The acquisition module is used to acquire key predictive indicators of mortality for ICH patients admitted to the ICU. The key predictive indicators of mortality include at least: Acute Physiology Score III, age, white blood cell count, and albumin level. The prediction module is used to predict the 30-day mortality rate of patients based on the key mortality prediction indicators and a 30-day mortality prediction model. The 30-day mortality prediction model is trained based on sample mortality key prediction indicator data of ICH patients who died within 30 days after admission to the ICU. Among them, APS III and age were positively correlated with 30-day mortality, albumin level was negatively correlated with 30-day mortality, and white blood cell level showed a V-shaped relationship with 30-day mortality. Both excessively high and excessively low white blood cell levels increased the risk of 30-day mortality.

2. The device for predicting the 30-day mortality rate of patients with cerebral hemorrhage according to claim 1, characterized in that, The device further includes: a training module, the training module comprising: The sample acquisition submodule is used to obtain multiple medical indicator data of ICH patients who died within 30 days of being admitted to the ICU from multiple public databases. The feature determination submodule is used to determine key features related to mortality within 30 days of admission to the ICU from the multiple medical indicators based on feature selection methods. The key features include: age, systolic blood pressure (SBP), APS-III, mean blood pressure (MBP), Glasgow Coma Scale (GCS), body temperature, oxygen saturation (SpO2), blood glucose level (BGL), lymphocyte count, neutrophil count, albumin, chloride, platelet count, and white blood cell count (WBC). The training submodule is used to train the preset model based on the medical indicator data corresponding to the key features. During the training process, the hyperparameters of the preset model are fine-tuned by combining the Bayesian optimization of the Gaussian process to obtain the prediction model. The analysis submodule is used to analyze the SHAP value of the prediction model, and based on the SHAP value, analyze the impact and correlation of each key feature in the preset model to determine the key predictive indicators related to the mortality rate within 30 days.

3. The device for predicting the 30-day mortality rate of patients with cerebral hemorrhage according to claim 2, characterized in that, The preset models are multiple, and the training submodule is specifically used to train the multiple preset models respectively based on the medical indicator data corresponding to the key features. During the training process, the hyperparameters of each preset model are fine-tuned by combining Gaussian process Bayesian optimization to obtain multiple prediction models. The device further includes: The first evaluation module is used to evaluate multiple prediction models based on accuracy, sensitivity, specificity, F1 score and AUC respectively, determine the target prediction model based on the evaluation results, and use the preset model corresponding to the target prediction model as the target model. The second training module is used to train the target model based on key predictive indicator data of ICH patients who die within 30 days of admission to the ICU to obtain a 30-day mortality prediction model.

4. The 30-day mortality prediction device for patients with cerebral hemorrhage according to claim 2, characterized in that, The preset models are multiple, and the training submodule is specifically used to train the multiple preset models respectively based on the medical indicator data corresponding to the key features. During the training process, the hyperparameters of each preset model are fine-tuned by combining Gaussian process Bayesian optimization to obtain multiple prediction models. The device further includes: The second evaluation module is used to evaluate multiple prediction models based on accuracy, sensitivity, specificity, F1 score and AUC, and to determine the target prediction model based on the evaluation results. The third training module is used to update the parameters of the target prediction model based on key predictive indicator data of ICH patients who die within 30 days of admission to the ICU, so as to obtain a 30-day mortality prediction model.

5. The device for predicting the 30-day mortality rate of patients with cerebral hemorrhage according to any one of claims 1-4, characterized in that, The device further includes a reminder module for outputting reminder information based on the mortality rate within 30 days.

6. A method for predicting the 30-day mortality rate of ICU patients with cerebral hemorrhage, characterized in that, The method includes: Key predictive indicators of mortality for ICH patients admitted to the ICU were obtained, including at least: Acute Physiology Score III, age, white blood cell count, and albumin level. Based on the aforementioned key mortality prediction indicators, and using a 30-day mortality prediction model, the 30-day mortality rate of patients is predicted. The 30-day mortality prediction model is trained based on sample mortality key prediction indicator data of ICH patients who died within 30 days of admission to the ICU. Among them, APS III and age were positively correlated with 30-day mortality, albumin level was negatively correlated with 30-day mortality, and white blood cell level showed a V-shaped relationship with 30-day mortality. Both excessively high and excessively low white blood cell levels increased the risk of 30-day mortality.

7. An electronic device, characterized in that, The invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the 30-day mortality prediction method for patients with cerebral hemorrhage as described in claim 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the 30-day mortality prediction method for patients with cerebral hemorrhage as described in claim 6.

9. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps in the 30-day mortality prediction method for patients with cerebral hemorrhage as described in claim 6.

Images