Acute gastrointestinal hemorrhage secondary sepsis risk prediction system and method
By constructing a risk prediction system for sepsis secondary to acute gastrointestinal bleeding based on multi-source data integration, and using machine learning algorithms to screen key features and update them in real time, the system solves the problems of accuracy and timeliness in predicting the risk of sepsis in patients with acute gastrointestinal bleeding in existing technologies, and realizes individualized risk management and early identification.
Patent Information
- Application Number
- CN202511428286.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-30
- Publication Date
- 2026-01-02
AI Technical Summary
Existing technologies lack accurate predictive models for the risk of secondary sepsis in patients with acute gastrointestinal bleeding, and existing methods fail to fully integrate multi-source clinical data, resulting in low prediction accuracy and difficulty in supporting individualized real-time interventions.
A risk prediction system for acute gastrointestinal bleeding followed by sepsis based on multi-source clinical data was constructed. Demographic, clinical symptom, laboratory test results and vital signs data were integrated through machine learning or deep learning algorithms to screen key features, establish a risk prediction model, and adopt a dynamic update mechanism to achieve real-time early warning.
It enables accurate and individualized prediction of sepsis risk in patients with acute gastrointestinal bleeding, supports timely intervention by clinicians, reduces mortality, and improves the timeliness and accuracy of prediction results.
Smart Images

Figure CN121260466A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of medical information processing, and particularly relates to a risk prediction system and method for secondary sepsis of acute gastrointestinal bleeding. BACKGROUND
[0002] Acute gastrointestinal bleeding (AGIB) as a life-threatening emergency brings significant morbidity and mortality burden, especially in patients with comorbidities who often require hospitalization. Patients with acute gastrointestinal bleeding often suffer from intestinal epithelial cell dysfunction, intestinal barrier disruption and bacterial translocation due to stress response, anemia, intestinal nutritional deficiency and immune disorder, and then develop severe infection and sepsis and other complications, which brings major challenges to treatment and increases medical costs. At the same time, sepsis is still a global health burden, and emerging evidence shows that there is a bidirectional relationship between AGIB and sepsis. AGIB patients are highly susceptible to severe infection, and AGIB itself is considered an independent risk factor for bloodstream infection. In clinical practice, the mortality of patients with acute gastrointestinal bleeding secondary to sepsis is significantly increased, the hospitalization time is prolonged, and the prognosis is poor. The current clinical management lacks effective tools for predicting the risk of sepsis in AGIB patients, which often leads to the use of empirical antibiotics without evidence-based guidance, which may either cause over-treatment and increase the risk of multiple drug-resistant infection, or cause under-treatment and lead to severe sepsis.
[0003] The current clinical management lacks effective tools for predicting the risk of sepsis in AGIB patients, which often leads to the use of empirical antibiotics without evidence-based guidance. Although clinical scoring tools such as Glasgow-Blatchford Score (GBS) and Acute Physiology and Chronic Health Evaluation II (APACHE II) are specifically used to assess bleeding risk and severity of systemic disease, respectively, they show limited utility in predicting the development of sepsis. Similarly, although Sequential Organ Failure Assessment (SOFA) score is still the gold standard for stratifying the severity of sepsis in diagnostic cases, it may lack sufficient sensitivity in predicting the risk of sepsis in AGIB population.
[0004] In recent years, with the rapid development of artificial intelligence, big data analysis and machine learning technology in the medical field, some studies have begun to try to use clinical multi-source data to establish a risk prediction model to improve the early identification ability of complications in critically ill patients. For example, the existing patent CN202311523722.0 discloses a sepsis prediction method based on deep learning. The method classifies the sepsis patient sample set according to the patient age threshold, and trains two sepsis prediction models for different age groups. However, the existing technical prediction method mainly faces the sepsis risk assessment of the general population, and lacks a prediction model specially for the secondary sepsis risk of acute gastrointestinal bleeding in this special clinical scenario. Secondly, the existing technology mainly relies on a single dimension such as age for patient grouping, and the prediction dimension is relatively limited, and the multi-source clinical data characteristics specific to acute gastrointestinal bleeding patients cannot be fully integrated. Thirdly, there is currently a lack of an automatic and real-time prediction system for clinical front-line doctors to obtain early warning information of high-risk patients in a timely manner. SUMMARY
[0005] Therefore, the present application provides an acute gastrointestinal bleeding secondary sepsis risk prediction system and method to solve the problem that the existing sepsis risk prediction method stays at the group level, lacks accurate prediction for individual patients, and the prediction dimension is single, which makes it difficult to fully integrate multi-source clinical data and thus affects the prediction accuracy.
[0006] The technical scheme of the present application is as follows: an acute gastrointestinal bleeding secondary sepsis risk prediction method, comprising the following steps: S1, retrospectively collecting multi-source clinical data and sepsis occurrence outcomes of acute gastrointestinal bleeding patients, setting inclusion and exclusion criteria, and constructing an acute gastrointestinal bleeding secondary sepsis data set, including a training set and a validation set; S2, in the training set, after pre-processing the multi-source clinical data, extracting feature vectors, and using feature selection to screen key features related to sepsis risk; S3, in the training set, based on the screened key features, using machine learning algorithms or deep learning algorithms to construct an acute gastrointestinal bleeding secondary sepsis risk prediction model; and evaluating the performance of the prediction model through the validation set; S4, inputting the multi-source clinical data of the acute gastrointestinal bleeding patient to be predicted into the optimized prediction model, and outputting the risk probability value of the acute gastrointestinal bleeding patient to be predicted secondary sepsis.
[0007] On the basis of the above technical scheme, preferably, the multi-source clinical data comprises: Demographic information: age, gender, body mass index, history of previous diseases, smoking history and drinking history; Clinical symptom data: hematemesis, melena, hypotension, syncope and bleeding volume assessment; Laboratory test results: blood routine data, liver and kidney function indicators, electrolytes, lactic acid and inflammatory factors; Imaging data: CT, MRI, ultrasound and gastroscopy results; Dynamic physiological monitoring parameters: heart rate, blood pressure, respiratory rate, mean arterial pressure, body temperature and blood oxygen saturation.
[0008] On the basis of the above technical scheme, preferably, step S2 comprises the following steps: S21, the acute gastrointestinal bleeding patients in the training set are marked according to the binary coding of sepsis outcome, and are marked as sepsis group and non-sepsis group respectively through digital grouping algorithm; S22, single variable difference detection is performed on the multi-source clinical data of the two groups of patients, and candidate variables significantly related to sepsis risk are screened out through statistical significance; S23, multivariate correlation analysis is performed, and independent risk factors related to secondary sepsis of acute gastrointestinal bleeding are screened out through independence test as key features.
[0009] On the basis of the above technical scheme, preferably, step S22 specifically comprises: Automatic identification of variable types of multi-source clinical data and classification processing; When it is identified as a discrete type classification variable, a corresponding probability distribution test is automatically selected according to a preset expected frequency threshold parameter for difference evaluation, wherein when the expected frequency is greater than or equal to 5, chi-square distribution fitting test is adopted, and when the expected frequency is less than 5, Fisher hypergeometric distribution test is adopted; When it is identified as a continuous numerical variable, the data distribution characteristics are evaluated through normality test and probability density distribution fitting analysis, if the test statistic is greater than 0.05 and the distribution fitting graph shows the symmetric characteristics of the variable, it is determined that the variable conforms to the normal distribution, parameterized two-sample mean difference test is adopted for inter-group comparison, otherwise non-parametric rank sum distribution test is adopted for analysis; The variables with test statistics less than 0.05 in the two-tailed probability test are marked as candidate variables significantly related to sepsis risk through significance marking algorithm.
[0010] On the basis of the above technical scheme, preferably, step S23 specifically comprises: Through calculating the variance inflation factor, the collinearity between variables is detected, and the high collinearity variables with VIF value greater than or equal to 10 are automatically removed to eliminate the interference of multiple collinearity; The non-linear correlation mode between the variable and sepsis is evaluated by using a restricted cubic spline function fitting, spline node parameters are set at the 5th, 35th, 65th and 95th percentiles of data distribution, and the linear correlation and non-linear correlation variable mode are automatically identified through curve fitting; Performing reverse stepwise iterative deletion to check the statistical quantity less than 0.05 as the significance threshold parameter of variable retention, performing recursive variable screening optimization, and finally outputting the independent risk factor related to acute gastrointestinal hemorrhage secondary sepsis as the key feature.
[0011] Based on the above technical scheme, preferably, in step S23, the key features include a history of chronic kidney disease, respiratory rate, neutrophil-to-lymphocyte ratio, C-reactive protein, creatinine, activated partial thromboplastin time, and fibrinogen level.
[0012] Based on the above technical scheme, preferably, step S3 specifically includes: The machine learning algorithm or the deep learning algorithm performs parameter learning and model training through the training set data, adopts a cross-validation method to perform model selection and parameter optimization, wherein the machine learning algorithm includes a random forest algorithm, a support vector machine algorithm, and a gradient boosting decision tree algorithm, and the deep learning algorithm includes a neural network algorithm. A risk probability calculation formula is established based on multivariate logistic regression, and model parameters are obtained through maximum likelihood estimation; The prediction performance of different algorithms is evaluated through a verification set, and an optimal algorithm is selected as a final prediction model.
[0013] Based on the above technical scheme, preferably, establishing the risk probability calculation formula based on multivariate logistic regression specifically includes: A logistic regression model formula is established: , wherein each regression coefficient βi is obtained through a maximum likelihood estimation method; The following training parameters are used for model training: the convergence criterion is that the likelihood ratio change is less than 1×10 -6 , and the maximum number of iterations is not more than 50 times; A total score-risk probability mapping relationship is constructed: , wherein P is the risk probability of sepsis, S is the total score, and k is a scaling factor, which is obtained through a training set calibration.
[0014] Based on the above technical scheme, preferably, the prediction method further includes a risk grading early warning and a dynamic updating mechanism, specifically including: A three-level risk stratification algorithm is performed on the prediction result, and graphical output is performed through a visual rendering engine, wherein when the risk probability value is greater than 90%, a high-priority alarm signal output is triggered, including a pop-up dialog box display and an audio alarm signal generation; when the risk probability value is in the interval of 70-90%, a medium-priority visual indication signal is activated, and a state bar icon flashing mode is used for prompting; when the risk probability value is less than 70%, a low-priority background identification processing is performed, and a silent risk marking is performed. The patient clinical data is continuously collected and dynamically updated, the multi-source clinical data is updated in real time, the risk prediction model automatically triggers an update mechanism according to the real-time updated multi-source clinical data, the model input features are recalculated, the model output result is dynamically adjusted with time through the fusion of historical data and new data, and real-time early warning is realized.
[0015] The application also provides a risk prediction system for secondary sepsis of acute gastrointestinal hemorrhage, which is used for realizing the risk prediction method for secondary sepsis of acute gastrointestinal hemorrhage. The data acquisition module is used for collecting multi-source clinical data of patients with acute gastrointestinal hemorrhage. The data processing module is used for pre-processing the multi-source clinical data. The feature screening module is used for performing single-variable difference detection and multi-variable correlation analysis, and screening key features related to the risk of sepsis. The risk prediction model module is used for constructing a risk prediction model for secondary sepsis of acute gastrointestinal hemorrhage based on the key features and using a machine learning algorithm or a deep learning algorithm. The result output module is used for outputting a risk probability value and performing a three-level risk stratification algorithm, and performing result display and alarm signal processing through a visual rendering engine.
[0016] The risk prediction system and method for secondary sepsis of acute gastrointestinal hemorrhage have the following beneficial effects relative to the prior art: (1) By constructing a risk prediction model for secondary sepsis of acute gastrointestinal hemorrhage based on multi-source clinical data fusion, the precise evaluation of the risk of sepsis for a specific patient group is realized. Different from the prior art which mainly aims at the prediction of sepsis for a general population, the application integrates the demographic information, clinical symptoms, laboratory examination results, imaging features and vital sign monitoring data of patients, combines artificial intelligence and machine learning algorithms, establishes an individualized and intelligent risk dynamic prediction model, realizes early and accurate identification of the risk of secondary sepsis of acute gastrointestinal hemorrhage, and thus assists clinicians to take timely intervention measures, reduces the mortality rate and improves the prognosis of patients.
[0017] (2) The adaptive feature screening adopted by the present application can automatically identify the key features most relevant to the risk of sepsis, effectively solving the problem of strong subjectivity of feature selection in the prior art; the multivariate logistic regression risk probability calculation formula established obtains the regression coefficient through the maximum likelihood estimation method, adopts strict training parameter control, constructs the total score-risk probability mapping relationship, and realizes the conversion from discrete score to continuous probability value. The quantitative prediction method is more accurate than the traditional qualitative evaluation, which can provide specific numerical risk evaluation results for clinicians and support more accurate clinical decision making.
[0018] (3) The dynamic updating mechanism constructed supports high-frequency data acquisition by accessing the data stream of the electronic medical record system, the laboratory information system, the picture archiving and communication system and the vital sign monitoring equipment in real time, and the risk prediction model can automatically trigger updating according to the newly collected data, and realizes the dynamic adjustment of the model output result through the fusion of historical data and new data. This mechanism overcomes the limitation that the existing static scoring system cannot reflect the dynamic changes of the patient's condition, realizes the continuous monitoring and real-time early warning of the risk of sepsis in patients with acute gastrointestinal bleeding, and significantly improves the timeliness and clinical applicability of the prediction result. BRIEF DESCRIPTION OF DRAWINGS
[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.
[0020] Figure 1 Flowchart of the present application for predicting the risk of sepsis secondary to acute gastrointestinal bleeding; Figure 2 Nomogram for predicting the risk of sepsis in AGIB patients according to the embodiment of the present application; Figure 3 ROC curve, calibration curve and DCA curve diagram of the training set and the validation set according to the embodiment of the present application. DETAILED DESCRIPTION
[0021] The technical solutions in the embodiments of the present application will be described in detail below. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.
[0022] As Figure 1As shown, the embodiment of the present application provides a risk prediction method for acute gastrointestinal bleeding secondary sepsis, comprising the following steps: S1, retrospectively collect multi-source clinical data of patients with acute gastrointestinal bleeding and sepsis outcome, set inclusion and exclusion criteria, construct acute gastrointestinal bleeding secondary sepsis data set, including training set and validation set; S2, in the training set, after pre-processing the multi-source clinical data, the feature vector is extracted, and the key features related to the risk of sepsis are screened by feature selection; S3, in the training set, based on the key features screened, a machine learning algorithm or a deep learning algorithm is used to construct an acute gastrointestinal bleeding secondary sepsis risk prediction model; the performance of the prediction model is evaluated by the validation set; S4, input the multi-source clinical data of the patient with acute gastrointestinal bleeding to be predicted into the optimized prediction model, and output the risk probability value of the patient with acute gastrointestinal bleeding secondary sepsis.
[0023] In step S1, the acute gastrointestinal bleeding secondary sepsis prediction data set is constructed by the retrospective data collection method, which lays the data foundation for subsequent model development. The data collection process strictly follows the clinical research standards and ethical requirements to ensure the data quality and integrity.
[0024] Firstly, the inclusion and exclusion criteria are established to ensure the homogeneity of the research objects and the data quality. The inclusion criteria include: the patient's age is ≥18 years old; the signs of hemodynamic instability (syncope / shock) appear, such as hematemesis, melena, coffee vomit or blood; the diagnosis of acute gastrointestinal bleeding is confirmed by clinical evaluation or endoscopy. The exclusion criteria include: the hospitalization time is less than 24 hours; the medical record is incomplete; there is infection or sepsis at the time of admission. Through the inclusion and exclusion criteria, it can be ensured that the collected patient samples are all acute gastrointestinal bleeding patients, and the confounding factors that may interfere with the prediction of sepsis risk are excluded.
[0025] The multi-source clinical data collection covers five main dimensions: demographic information includes age, gender, body mass index, and basic information such as medical history, smoking history, and drinking history, which provides basic parameters for individual risk assessment. Clinical symptom data focuses on the direct manifestations of gastrointestinal bleeding, including hematemesis, melena, blood pressure drop, syncope, and bleeding volume assessment, which can reflect the severity of bleeding and hemodynamic status. Laboratory test results cover blood routine data, liver and kidney function indicators, electrolyte levels, lactate concentration, and inflammatory factors, providing objective biomarker information for sepsis risk assessment. Imaging data provides visual information on anatomical structure and pathological changes through CT, MRI, ultrasound imaging, and gastroscopy results. Dynamic physiological monitoring parameters include continuous monitoring data of heart rate, blood pressure, respiratory rate, mean arterial pressure, body temperature, and blood oxygen saturation, which can reflect the real-time physiological state changes of patients.
[0026] The data set construction adopts a hierarchical design to meet the needs of model development and validation. The training set is used for model parameter learning and feature selection, usually accounting for 60-70% of the total sample size, ensuring that there is enough data to support the training of complex models. The validation set is further divided into internal validation set and external validation set. The internal validation set is used for preliminary evaluation of model performance and hyperparameter tuning, and the external validation set comes from different medical institutions, which is used to evaluate the generalization ability and clinical applicability of the model. The determination of sepsis outcome strictly follows the Sepsis-3 standard, ensuring the accuracy and consistency of the definition of outcome variables.
[0027] In this embodiment, the research team recruited 1449 patients with acute gastrointestinal bleeding in four medical centers from January 2020 to July 2024. Among the initial 1750 patients who met the inclusion criteria, 301 were excluded after strict screening, including 182 cases of data missing, 65 cases of short hospital stay, 29 cases of combined sepsis, and 25 cases of withdrawal from the study. The final data set includes 878 patients in the training set, 187 patients in the internal validation set, and 384 patients in the external validation set. Among the 1449 patients, the median age was 65 years old, males accounted for 68.7%, and the incidence of sepsis was 15.4%. Data collection is achieved through interface connection with hospital electronic medical record system, laboratory information system, picture archiving and communication system, and vital sign monitoring equipment, using international standard interface protocols such as HL7, FHIR to ensure the standardization and integrity of data transmission. All clinical data are standardized and stored according to the unified data dictionary, providing a high-quality data foundation for subsequent data preprocessing and model development.
[0028] In step S2, based on the data set constructed in step S1, the multi-source clinical data in the training set is systematically pre-processed and feature screened to identify key features closely related to the risk of acute gastrointestinal hemorrhage secondary to sepsis. This step extracts feature variables with predictive value from the original multi-dimensional clinical data through multi-level data analysis and statistical methods.
[0029] Data preprocessing first standardizes the collected multi-source clinical data to ensure data quality and reliability for subsequent analysis. For missing value processing, the training set adopts a strict exclusion strategy, completely excluding samples with missing key variables such as lactate, blood ammonia, NT-ProBNP, troponin I, or C-reactive protein, ensuring that the training set achieves zero missing rate. For continuous numerical variables, Z-score standardization method is used for normalization, making the mean of all variables 0 and the standard deviation 1, ensuring that features of different orders of magnitude have the same weight contribution in subsequent analysis.
[0030] Step S2 specifically includes: S21, the acute gastrointestinal hemorrhage patients in the training set are marked according to the binary coding of sepsis outcome, and are marked as sepsis group and non-sepsis group respectively through digital grouping algorithm. According to the clearly defined sepsis diagnostic criteria of Sepsis-3 standard, the patients in the training set are digitally grouped according to the sepsis outcome during hospitalization, and the patients in the sepsis group are marked as 1, and the patients in the non-sepsis group are marked as 0.
[0031] S22, single variable difference detection is performed on the multi-source clinical data of the two groups of patients, and candidate variables significantly related to sepsis risk are screened out through statistical significance.
[0032] Specifically, first, the variable types of multi-source clinical data are automatically identified, and the data is classified according to the variable types. When it is identified as a discrete classification variable, the system automatically selects the corresponding probability distribution test according to the preset expected frequency threshold parameter. When the expected frequency is greater than or equal to 5, the chi-square distribution fitting test is used to evaluate the distribution difference between the two groups, and when the expected frequency is less than 5, the Fisher hypergeometric distribution test is used to ensure the accuracy of statistical test. When it is identified as a continuous numerical variable, the system first evaluates the distribution characteristics of the data through normality test and probability density distribution fitting analysis. Specifically, Shapiro-Wilk test is used in combination with Q-Q plot visualization inspection. If the test statistic is greater than 0.05 and the distribution fitting graph shows symmetric characteristics, it is determined that the variable conforms to the normal distribution, and a parametric two-sample mean difference test is used for group comparison; otherwise, a non-parametric rank sum distribution test is used for analysis. Through significance marking, variables with p-value less than 0.05 in the two-tailed probability test are marked as candidate variables significantly related to sepsis risk.
[0033] S23, performing multivariate correlation analysis, screening out independent risk factors related to acute gastrointestinal bleeding secondary to sepsis as key features by independence test.
[0034] Specifically, first, the collinearity between variables is detected by calculating the variance inflation factor, and the highly collinear variables with VIF value greater than or equal to 10 are automatically removed to eliminate the multiple collinearity interference between variables and ensure the statistical stability of subsequent modeling. Subsequently, the restricted cubic spline function is used to fit the evaluation variable and sepsis to identify the linear and nonlinear correlation patterns of variables, and the spline node parameters are set at the 5th, 35th, 65th and 95th percentiles of the data distribution. The linear and nonlinear correlation patterns of variables are automatically identified by curve fitting to ensure that the model can capture complex nonlinear relationships. Finally, backward stepwise iteration is performed, and the p value less than 0.05 is used as the significance threshold parameter for variable retention, and through the recursive variable screening optimization process, the variables with no statistical significance are gradually removed, and finally the independent risk factors related to acute gastrointestinal bleeding secondary to sepsis are output as key features.
[0035] In this embodiment, in the 878 training set patients, first, 27 candidate variables significantly related to sepsis risk were screened out by univariate analysis, including anemia history, hypertension, chronic kidney disease, heart failure, heart rate, respiratory rate, esophageal variceal bleeding, neutrophil percentage, neutrophil to lymphocyte ratio, platelet to lymphocyte ratio, C-reactive protein, alanine aminotransferase, aspartate aminotransferase, blood urea nitrogen, creatinine, lactic acid, brain natriuretic peptide precursor, prothrombin time, activated partial thromboplastin time, international normalized ratio, fibrinogen, D-dimer and albumin, etc. as shown in Table 1.
[0036] Table 1 Clinical characteristics of 878 AGIB patients
[0037]
[0038]
[0039] Through multivariate logistic regression analysis and restricted cubic spline function evaluation, 7 independent risk predictors were finally determined as key features, as shown in Table 2 and Figure 2The key features were chronic kidney disease history (OR value 1.797, 95% confidence interval 1.068-3.025), respiratory rate (OR value 1.078, 95% confidence interval 1.020-1.139), neutrophil-to-lymphocyte ratio (OR value 1.042, 95% confidence interval 1.014-1.071), C-reactive protein (OR value 1.010, 95% confidence interval 1.005-1.014), creatinine (OR value 1.002, 95% confidence interval 1.000-1.004), activated partial thromboplastin time (OR value 1.051, 95% confidence interval 1.019-1.084), and fibrinogen level (OR value 1.513, 95% confidence interval 1.238-1.850), as shown. These key features cover the patient's underlying disease state, vital signs, inflammatory response, renal function, coagulation function, and other physiological systems, laying a solid feature foundation for subsequent risk prediction model construction.
[0040] Table 2 Univariate and multivariate logistic regression analysis
[0041]
[0042] In step S3, based on the key features screened out in step S2, a machine learning algorithm or a deep learning algorithm is used to construct a risk prediction model for acute gastrointestinal bleeding secondary to sepsis. This step ensures that the prediction model has the best generalization ability and clinical application value through performance comparison of various algorithms and model optimization.
[0043] Step S3 specifically includes: The machine learning algorithm or the deep learning algorithm learns parameters and trains the model through the training set data, and uses cross-validation method to select the model and optimize the parameters.
[0044] Specifically, the machine learning algorithm construction process first learns parameters and trains the model through the training set data, and uses cross-validation method to select the model and optimize the parameters. The specific machine learning algorithms include random forest algorithm, support vector machine algorithm, gradient boosting decision tree algorithm, and neural network algorithm of deep learning. The random forest algorithm can effectively process high-dimensional feature space and provide feature importance evaluation by integrating the prediction results of multiple decision trees; the support vector machine algorithm can solve the binary classification problem by finding the optimal hyperplane, and has good generalization ability on small sample data; the gradient boosting decision tree algorithm can handle complex nonlinear relationships by iteratively training weak learners and optimizing loss functions; the neural network algorithm can learn deep feature representations in data through nonlinear transformations of multiple perceptrons.
[0045] The risk probability calculation formula is established based on multivariate logistic regression, and the model parameters are obtained by maximum likelihood estimation. The prediction performance of different algorithms is evaluated by the validation set, and the optimal algorithm is selected as the final prediction model.
[0046] The core of establishing a risk probability calculation formula based on multivariate logistic regression is to build a mathematical model to map the relationship between input features and the probability of sepsis. The established logistic regression model formula is: , Where each regression coefficient βi is obtained by maximum likelihood estimation method; In this embodiment, the established logistic regression model formula is: , Where CKD is a history of chronic kidney disease, RR is respiratory rate, NLR is neutrophil-to-lymphocyte ratio, CRP is C-reactive protein, Creatinine, APTT is activated partial thromboplastin time, and FIB is fibrinogen; The following training parameters are used for model training: the convergence criterion is that the likelihood ratio change is less than 1×10 -6 , and the maximum number of iterations is not more than 50; The total score-risk probability mapping relationship is constructed: , Where P is the risk probability of sepsis, S is the total score, and k is the scaling factor, which is obtained by training set calibration.
[0047] The prediction performance of different algorithms is evaluated by the validation set, and the optimal algorithm is selected as the final prediction model. Performance evaluation is comprehensively evaluated by multiple indicators, including the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, accuracy, positive predictive value and negative predictive value, etc. The model also evaluates the consistency of the predicted probability and the actual incidence rate through the calibration curve, and uses the mean square error (MSE) to quantify the calibration performance. Decision curve analysis (DCA) is used to evaluate the clinical net benefit of the model at different probability thresholds, to ensure the clinical value of the model in practical application.
[0048] Step S4 realizes risk prediction and stratified management of patients with acute gastrointestinal bleeding. The multi-source clinical data of the patient to be predicted are input into the optimized prediction model, and the system automatically extracts the corresponding feature values, including a history of chronic kidney disease, respiratory rate, neutrophil-to-lymphocyte ratio, C-reactive protein, creatinine, activated partial thromboplastin time and fibrinogen level. The prediction model calculates the risk probability value based on these input features, and classifies and manages through a three-level risk stratification algorithm.
[0049] The risk stratification system employs a probability threshold-based grading strategy, combining clinical experience to determine the probability cutoff points for different risk levels. When the risk probability value is greater than 90%, the system classifies the patient as high-risk, triggering a high-priority alarm signal output mechanism, including pop-up dialog box display and audio alarm signal generation, ensuring that clinical medical staff can promptly identify and take emergency intervention measures. When the risk probability value is in the 70-90% range, the system activates a medium-priority visual indicator signal, prompting medical staff to strengthen monitoring and consider preventive measures through a flashing status bar icon. When the risk probability value is less than 70%, the system performs low-priority background identification processing, silently marking the risk and recording the patient's status without actively alerting, avoiding excessive alarms that could disrupt normal work processes.
[0050] The prediction model constructed in this embodiment exhibits excellent prediction performance on the training set, internal validation set, and external validation set. For example... Figure 3 As shown, where Figure 3 Figure A shows the ROC curve in the training set; Figure B shows the ROC curve in the internal validation set; Figure C shows the ROC curve in the external validation set; Figure D shows the calibration curve in the training set; Figure E shows the calibration curve in the internal validation set; Figure F shows the calibration curve in the external validation set; Figure G shows the DCA curve in the training set; Figure H shows the DCA curve in the internal validation set; and Figure I shows the DCA curve in the external validation set. The AUC value of the training set is 0.827 (95% confidence interval 0.759–0.888), the AUC value of the internal validation set is 0.836 (95% confidence interval 0.776–0.896), and the AUC value of the external validation set is 0.884 (95% confidence interval 0.816–0.952), all of which are significantly better than traditional scoring systems such as GBS, APACHE II, and SOFA scoring. The calibration curves show good consistency between the predicted and actual probabilities. The mean squared errors for the training, internal, and external validation sets are 0.00094, 0.00791, and 0.00045, respectively, indicating good calibration capability. Decision curve analysis confirms that the model provides significant net clinical benefits across a wide range of probability thresholds, offering valuable support for clinical decision-making. The final predictive model was integrated into an online early warning system, achieving a fully automated process from data input and risk calculation to early warning output, providing an intelligent clinical decision support tool for sepsis risk management in patients with acute gastrointestinal bleeding.
[0051] The prediction method of this invention also includes a dynamic update mechanism. The system supports continuous collection and updating of patient clinical data, and the risk prediction model can dynamically adjust its output based on the latest data to achieve real-time early warning. This dynamic update mechanism ensures the timeliness and accuracy of the prediction results through two key steps: continuous data collection and model updates.
[0052] The continuous data collection mechanism realizes continuous monitoring of patient data throughout the entire hospitalization period. For vital sign type data, the system supports access at a sampling frequency of seconds or minutes, realizing high timeliness monitoring. The dynamic update algorithm mechanism realizes real-time response and adjustment of the prediction model. The risk prediction model can automatically trigger the update mechanism according to the newly collected data, and recalculate the model input features. When the system detects changes in key feature data, it automatically starts the feature re-extraction and risk re-evaluation process, and recalculates the risk probability based on the latest chronic kidney disease history, respiratory rate, neutrophil-to-lymphocyte ratio, C-reactive protein, creatinine, activated partial thromboplastin time, and fibrinogen levels. The system integrates the patient's time series data features using sliding window technology, so that the model output results can be dynamically adjusted over time, avoiding false positives caused by single point data. The update frequency can be differentiated according to the data type, with vital sign data supporting real-time updates, laboratory test data triggering updates based on test completion time, and imaging data updating based on report release time.
[0053] The real-time warning mechanism is based on the dynamic update of the risk assessment results. When the patient's latest risk assessment result exceeds the preset threshold, the system automatically generates a warning signal and notifies the clinical medical staff through a visual interface, mobile terminal reminder, or hospital information system message push. The warning level is divided into three levels: high, medium, and low, to facilitate medical staff to take different intervention measures according to the risk level. The high-risk warning trigger condition is that the risk probability is greater than 90% for two consecutive calculations, and the system immediately sends an emergency warning notification; the medium-risk warning trigger condition is that the risk probability is in the interval of 70-90% and shows an upward trend, and the system sends an attention reminder; the low-risk marker is recorded silently when the risk probability is less than 70%, without active push but preserving the historical trajectory for clinical review. Through this dynamic update and real-time warning mechanism, the present application can provide timely and accurate sepsis risk assessment information for clinical medical staff, supporting dynamic optimization and precise intervention of clinical decision-making.
[0054] The present application also provides an acute gastrointestinal bleeding secondary sepsis risk prediction system, which is used to realize an acute gastrointestinal bleeding secondary sepsis risk prediction method as described above, comprising: The data acquisition module is used for acquiring multi-source clinical data of patients with acute gastrointestinal hemorrhage; the module realizes unified input of multi-source data by data interface docking with an electronic medical record system (EMR), a laboratory information system (LIS), a picture archiving and communication system (PACS) and vital sign monitoring equipment. The interface docking is realized by using international standard interface protocols such as HL7 and FHIR to dock with the EMR, LIS and PACS, image data is acquired through the DICOM protocol, and vital sign monitoring equipment is docked through serial ports, Bluetooth, Wi-Fi or Internet of Things interfaces, so as to realize automatic and real-time data transmission.
[0055] The data processing module is used for pre-processing the multi-source clinical data. The feature screening module is used for performing single variable difference detection and multivariate correlation analysis, and screening key features related to the risk of sepsis. The risk prediction model module is used for constructing a risk prediction model of secondary sepsis of acute gastrointestinal hemorrhage based on key features by using a machine learning algorithm or a deep learning algorithm. The result output module is used for outputting a risk probability value and performing a three-level risk stratification algorithm, and performing result display and alarm signal processing through a visual rendering engine.
[0056] In addition, the system also supports a dynamic update mechanism, which can continuously collect and update patient clinical data. The system accesses data streams from various information systems and monitoring equipment in real time, ensuring that key parameters during the entire hospitalization process of the patient can be continuously collected. The risk prediction model can automatically trigger the update mechanism according to the newly collected data, recalculate the model input features, and through the fusion of historical data and new data, the model output result can be dynamically adjusted over time. When the latest risk assessment result of the patient exceeds the preset threshold, the system automatically generates an early warning signal, and notifies the clinical medical staff through a visual interface, a mobile terminal or an HIS message push, and the warning level is divided into three levels of high, medium and low, so that the medical staff can take different intervention measures according to the risk level.
[0057] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.
Claims
1. A method for predicting the risk of sepsis secondary to acute gastrointestinal bleeding, characterized in that, Includes the following steps: S1. Retrospectively collect multi-source clinical data and sepsis outcomes of patients with acute gastrointestinal bleeding. After setting inclusion and exclusion criteria, construct a dataset of acute gastrointestinal bleeding followed by sepsis, including a training set and a validation set. S2. In the training set, feature vectors are extracted after preprocessing multi-source clinical data, and feature selection is used to screen key features related to sepsis risk. S3. Within the training set, based on the selected key features, a machine learning algorithm or deep learning algorithm is used to construct a risk prediction model for acute gastrointestinal bleeding followed by sepsis; the performance of the prediction model is evaluated using a validation set. S4. Input the multi-source clinical data of patients with acute gastrointestinal bleeding to be predicted into the optimized prediction model, and output the risk probability value of secondary sepsis in patients with acute gastrointestinal bleeding to be predicted.
2. The method for predicting the risk of sepsis secondary to acute gastrointestinal bleeding as described in claim 1, characterized in that, The multi-source clinical data includes: Demographic information: age, sex, body mass index, medical history, smoking history, and alcohol consumption history; Clinical symptom data: hematemesis, melena, hypotension, syncope, and assessment of bleeding volume; Laboratory test results: complete blood count, liver and kidney function indicators, electrolytes, lactate, and inflammatory factors; Imaging data: CT, MRI, ultrasound images and gastroscopy results; Dynamic physiological monitoring parameters: heart rate, blood pressure, respiratory rate, mean arterial pressure, body temperature, and blood oxygen saturation.
3. The method for predicting the risk of sepsis secondary to acute gastrointestinal bleeding as described in claim 2, characterized in that, Step S2 includes the following steps: S21. Patients with acute gastrointestinal bleeding in the training set are labeled with binary codes according to the sepsis outcome, and then labeled as sepsis group and non-sepsis group respectively using a digital grouping algorithm; S22. Perform univariate difference detection on the multi-source clinical data of the two groups of patients, and screen out candidate variables that are significantly associated with the risk of sepsis through statistical significance screening; S23. Perform multivariate association analysis and use independence tests to screen out independent risk factors associated with sepsis secondary to acute gastrointestinal bleeding as key features.
4. The method for predicting the risk of sepsis secondary to acute gastrointestinal bleeding as described in claim 3, characterized in that, Step S22 specifically includes: Automatically identify and classify variable types in multi-source clinical data; When identified as a discrete categorical variable, the appropriate probability distribution test is automatically selected for difference assessment based on the preset expected frequency threshold parameter. Specifically, the chi-square distribution fit test is used when the expected frequency is greater than or equal to 5, and the Fisher hypergeometric distribution test is used when the expected frequency is less than 5. When identified as a continuous numerical variable, the data distribution characteristics are evaluated through normality test and probability density distribution fitting analysis. If the test statistic is greater than 0.05 and the distribution fitting plot shows symmetrical characteristics, the variable is determined to conform to normal distribution. The parametric two-sample mean difference test is used for inter-group comparison. Otherwise, the nonparametric rank sum distribution test is used for analysis. Variables with a test statistic less than 0.05 in the two-tailed probability test were marked as candidate variables that are significantly associated with the risk of sepsis using a significance labeling algorithm.
5. The method for predicting the risk of sepsis secondary to acute gastrointestinal bleeding as described in claim 3, characterized in that, Step S23 specifically includes: Multicollinearity detection among variables is performed by calculating the variance inflation factor, and highly collinear variables with a VIF value greater than or equal to 10 are automatically removed to eliminate multicollinearity interference. Restricted cubic spline function fitting was used to assess the nonlinear association pattern between variables and sepsis. Spline node parameters were set at the 5th, 35th, 65th and 95th percentiles of the data distribution. Curve fitting was used to automatically identify variable patterns of linear and nonlinear association. A reverse stepwise iterative elimination process was performed, with a test statistic less than 0.05 as the significance threshold parameter for variable retention. Recursive variable screening and optimization were carried out, and finally, the independent risk factors associated with acute gastrointestinal bleeding secondary to sepsis were output as key features.
6. The method for predicting the risk of sepsis secondary to acute gastrointestinal bleeding as described in claim 3, characterized in that, In step S23, the key features include a history of chronic kidney disease, respiratory rate, neutrophil to lymphocyte ratio, C-reactive protein, creatinine, activated partial thromboplastin time, and fibrinogen level.
7. The method for predicting the risk of sepsis secondary to acute gastrointestinal bleeding as described in claim 1, characterized in that, Step S3 specifically includes: Machine learning algorithms or deep learning algorithms learn parameters and train models using training set data, and use cross-validation methods for model selection and parameter optimization. Machine learning algorithms include random forest algorithms, support vector machine algorithms, and gradient boosting decision tree algorithms, while deep learning algorithms include neural network algorithms. A risk probability calculation formula is established based on multivariate logistic regression, and model parameters are obtained through maximum likelihood estimation. The prediction performance of different algorithms is evaluated using a validation set, and the optimal algorithm is selected as the final prediction model.
8. The method for predicting the risk of sepsis secondary to acute gastrointestinal bleeding as described in claim 7, characterized in that, The specific formula for calculating risk probability based on multivariate logistic regression includes: Formula for establishing a logistic regression model: , Among them, each regression coefficient β i (i=0~7) were obtained using the maximum likelihood estimation method. CKD represents a history of chronic kidney disease, RR represents respiratory rate, NLR represents the neutrophil-to-lymphocyte ratio, and CRP represents C-reactive protein. For creatinine, APTT is activated partial thromboplastin time, and FIB is fibrinogen; The following training parameters were used for model training: the convergence criterion was a likelihood ratio change of less than 1 × 10⁻⁶. -6 The maximum number of iterations shall not exceed 50. Construct a mapping relationship between total score and risk probability: , Where P is the probability of developing sepsis, S is the total score, and k is the scaling factor, which is obtained through calibration of the training set.
9. The method for predicting the risk of sepsis secondary to acute gastrointestinal bleeding as described in claim 1, characterized in that, The prediction method also includes a risk grading and early warning mechanism and a dynamic update mechanism, specifically including: A three-level risk stratification algorithm is applied to the prediction results, and the results are graphically output through a visualization rendering engine. When the risk probability value is greater than 90%, a high-priority alarm signal is triggered, including a pop-up dialog box and an audio alarm signal. When the risk probability value is between 70% and 90%, a medium-priority visual indicator signal is activated, which is indicated by a flashing status bar icon. When the risk probability value is less than 70%, a low-priority background identification process is performed, and a silent risk label is applied. The system continuously collects and dynamically updates patient clinical data, providing real-time updates of multi-source clinical data. The risk prediction model automatically triggers an update mechanism based on the real-time updated multi-source clinical data, recalculates the model input features, and dynamically adjusts the model output results over time through the fusion of historical data and new data, thereby achieving real-time early warning.
10. A risk prediction system for sepsis secondary to acute gastrointestinal bleeding, characterized in that, The prediction system is used to implement the method for predicting the risk of sepsis secondary to acute gastrointestinal bleeding as described in any one of claims 1-9, comprising: The data acquisition module is used to collect multi-source clinical data from patients with acute gastrointestinal bleeding. The data processing module is used for preprocessing multi-source clinical data; The feature screening module is used to perform univariate difference detection and multivariate association analysis to screen key features associated with sepsis risk; The risk prediction model module is used to construct a risk prediction model for sepsis secondary to acute gastrointestinal bleeding based on key features using machine learning or deep learning algorithms. The results output module is used to output risk probability values and execute a three-level risk stratification algorithm. The results are displayed and alarm signals are processed through a visualization rendering engine.
Citation Information
Patent Citations
A sepsis prediction method and system based on deep learning
CN117238510B
Cited By
Dispersive intravascular coagulation risk prediction method and device, terminal equipment and medium
CN121687517A
