Liver cirrhosis ascites sbp double path assisted identification system

Abstract

The application relates to the field of medical data analysis, and particularly discloses a cirrhosis ascites SBP double-path auxiliary identification system, which comprises a data acquisition module, a data standardization module and a path selection module. The data acquisition module is used for acquiring clinical medical history data, hospitalization complication data and ascites test data of a target patient, and acquiring time stamps, units and data sources corresponding to the data to obtain a to-be-processed original data set. The data standardization module is used for performing time window verification and screening on the to-be-processed original data set according to preset time anchor points, and performing unit conversion, repeated record disambiguation and value range verification on the screened data to obtain a standardized data set. The path selection module. The technical scheme of the application can unify time windows, reflect individualized comprehensive risks, has processing capacity for core variable missing, and can output an interpretable result, thereby providing closed-loop decision support for doctors.

Description

Technical Field

[0001] This invention relates to the field of medical data analysis, and in particular to a dual-path auxiliary identification system for SBP in patients with cirrhosis and ascites. Background Technology

[0002] Liver cirrhosis is a common chronic progressive liver disease, and ascites is one of the most prominent clinical manifestations of decompensated cirrhosis. Patients with cirrhotic ascites are highly susceptible to spontaneous bacterial peritonitis (SBP) due to weakened immune function, portosystemic shunting, and intestinal bacterial translocation. SBP is a serious and dangerous clinical complication; if not diagnosed and treated promptly, it can rapidly progress to hepatorenal syndrome, hepatic encephalopathy, septic shock, and even acute decompensation, with an extremely high in-hospital mortality rate. Therefore, early and accurate identification of SBP risk is of crucial clinical significance for improving the prognosis of patients with cirrhotic ascites.

[0003] Currently, the gold standard for clinical diagnosis of severe abdominal bleeding (SBP) primarily relies on invasive diagnostic paracentesis. Specifically, when the polymorphonuclear neutrophil (PMN) count in the ascites reaches or exceeds a certain threshold... When the ascites fluid volume is / L and secondary peritonitis is ruled out, SBP can be diagnosed; further confirmation of the etiology can be achieved by combining the ascites fluid culture results. However, this diagnostic model has many limitations in actual clinical practice. First, some SBP patients have extremely subtle clinical symptoms in the early stages, without typical abdominal pain or fever, which are easily overlooked by clinicians. Second, the positive rate of ascites fluid culture is generally low (usually only 10%~30%), and it is easily affected by factors such as pre-puncture antibiotic exposure, improper sampling timing, non-standard inoculation methods, and low bacterial load. In addition, the patient's clinical history, complication information, and routine laboratory test results are usually scattered in the hospital's electronic medical record (EMR), laboratory information system (LIS), and medical record front page system, making it difficult for clinicians to efficiently and comprehensively integrate the patient's past SBP history, complication status, and the latest test data for rapid and quantitative risk assessment within the limited consultation time.

[0004] In recent years, with the development of medical informatics and artificial intelligence technologies, some SBP prediction models based on machine learning or traditional statistical methods have emerged. These models attempt to predict the occurrence of SBP using patients' clinical characteristics.

[0005] However, existing assisted identification schemes suffer from problems in practical clinical applications, including scattered and untimely data sources, and a lack of standardized preprocessing and time window control. Specifically, existing prediction methods directly read historical data without establishing strict time anchors. This can lead to the introduction of complications or changes in laboratory indicators that occur after a patient is diagnosed with SBP, resulting in a serious "data leakage" problem. Furthermore, data collected from different departments and at different time periods often have inconsistent units, numerous duplicate records, and a lack of standardized disambiguation and conversion mechanisms.

[0006] In addition, a single threshold judgment cannot reflect individualized comprehensive risk. Although the PMN count in ascites is a core indicator, existing auxiliary diagnostic tools often only perform binary classification, ignoring its continuous variation characteristics. They also fail to organically combine clinical background factors with significant prognostic impacts, such as a history of severe bleeding (SBP), upper gastrointestinal bleeding, and liver failure, and thus cannot output continuous individualized risk probabilities and the contribution of each factor.

[0007] Furthermore, the current model suffers from a severe deficiency in handling missing core variables and lacks a flexible path switching mechanism. In real-world clinical scenarios, some patients, due to objective limitations, have not yet undergone paracentesis during their initial clinical assessment, resulting in the absence of the core variable, PMN count, in ascites fluid. If existing predictive models force the input of PMN, they become completely unusable before the puncture, thus losing the value of early triage; if PMN is not introduced at all, this core indicator cannot be used after the puncture, leading to a significant decrease in the model's overall discriminative ability and diagnostic efficacy. Currently, there is a lack of a dual-path auxiliary identification mechanism in clinical practice that can automatically switch calculation paths based on the availability of PMN.

[0008] Most existing scoring systems only output an abstract score or binary result, constituting a "black box" model. They fail to demonstrate to clinicians the specific contribution and magnitude of each variable to the current patient's risk prediction, making it difficult to gain physicians' clinical trust. Furthermore, these systems cannot record the data versions, calculation paths, and times used in the model calculations, failing to meet the auditing requirements of medical quality control. Existing auxiliary tools typically only focus on risk probability calculations, unable to dynamically link high-risk warnings with the hospital's local pathogen and drug resistance knowledge bases, as well as standardized specimen submission guidelines (such as specimen collection before antibiotic use), thus failing to provide physicians with closed-loop decision support.

[0009] Therefore, there is an urgent need for a dual-path auxiliary identification system for SBP in patients with cirrhotic ascites that can unify the time window, reflect individualized comprehensive risk, handle missing core variables, and provide interpretable output to offer closed-loop decision support for physicians. Summary of the Invention

[0010] This invention provides a dual-path auxiliary identification system for SBP in patients with cirrhotic ascites. It can unify the time window, reflect individualized comprehensive risk, handle missing core variables, and provide interpretable output, thus providing physicians with closed-loop decision support.

[0011] To solve the above-mentioned technical problems, this application provides the following technical solution: A dual-pathway assisted identification system for SBP in patients with cirrhotic ascites includes: The data acquisition module is used to acquire the target patient's clinical history data, hospitalization complication data, and ascites test data, and to obtain the timestamp, unit, and data source of each data to obtain the raw dataset to be processed. The data standardization module is used to perform time window verification and filtering on the original dataset to be processed according to a preset time anchor point, and to perform unit conversion, duplicate record disambiguation and value range verification on the filtered data to obtain a standardized dataset. The path selection module is used to identify whether polymorphonuclear neutrophil counts exist in the standardized dataset. If they exist, the polymorphonuclear neutrophil counts are determined to be available, and a first identification path is selected. If they do not exist, the polymorphonuclear neutrophil counts are determined to be missing, and a second identification path is selected. The probability calculation module is used to calculate the first auxiliary identification probability by calling a preset first regression model to perform weighted summation and nonlinear mapping on the history of spontaneous bacterial peritonitis, concurrent upper gastrointestinal bleeding, liver failure status, and polymorphonuclear neutrophil count in the standardized dataset when selecting the first identification path; and to calculate the second auxiliary identification and triage probability by calling a preset second regression model to perform weighted summation and nonlinear mapping on the history of spontaneous bacterial peritonitis, concurrent upper gastrointestinal bleeding, and liver failure status in the standardized dataset when selecting the second identification path. The threshold determination module is used to compare the first auxiliary identification probability with a preset first risk threshold to generate a corresponding auxiliary identification result; or, to compare the second auxiliary identification triage probability with a preset second risk threshold to generate a corresponding pre-puncture triage result. The output and audit module is used to output the selected recognition path, the values ​​of each variable in the standardized dataset, the calculated probability value, the corresponding threshold judgment result, and the clinical review prompt. It also calculates the contribution of each variable to the probability value and displays it in an interpretable manner. At the same time, it records audit information including model version, input data, and operation log.

[0012] The basic principles and beneficial effects of this invention are as follows: This invention acquires clinical history data (past SBP history), inpatient complication data (upper gastrointestinal bleeding, liver failure status), and ascites test data (PMN count) of target patients from different data sources such as electronic medical records, laboratory information systems, and medical record systems through a data acquisition module. Simultaneously, it acquires the timestamp, unit, and data source of each data point, forming a raw dataset to be processed. A unified access and structured mapping mechanism for multi-source heterogeneous clinical data is established. By centrally associating fragmented data scattered across different business systems and fully preserving its time dimension and source attributes, a standardized data foundation is provided for subsequent time-series alignment and quality control. This effectively solves the problems of scattered clinical information, low efficiency of manual extraction, and easy omissions in existing technologies, achieving seamless integration of clinical history, complications, and laboratory test data, laying the foundation for timely risk assessment.

[0013] The data standardization module performs rigorous time window verification and filtering on the original dataset based on preset time anchors (time of first diagnostic paracentesis or time of first determination of SBP), and performs unit conversion (e.g., converting cells / mm³ to...). The module performs data cleaning and quality control, including duplicate record disambiguation and value range verification, to output a standardized dataset. The core mechanism of this module lies in constructing data cleaning and quality control rules based on "causal time series." By establishing a reference time point as a time anchor, it restricts medical history data to be earlier than this anchor point, complication data to be earlier than or on the same day and earlier than antibiotic adjustments, and routine test indicators to prioritize the most recent results within 24 hours before the anchor point. This completely eliminates data leakage, a technical defect that seriously affects the reliability of clinical prediction models. By strictly filtering clinical feature changes that occur after diagnosis, the causal time series logic between input variables and output risks is ensured. Simultaneously, through unit conversion and disambiguation verification, computational biases caused by inconsistent formats of multi-source data and duplicate testing are eliminated, significantly improving the quality and consistency of input data.

[0014] The path selection module dynamically identifies the presence of PMN counts in the standardized dataset for ascites. If present, it determines the status as available and triggers the first identification path (with-PMN path); if absent, it determines the status as missing and automatically switches to the second identification path (non-PMN path). This module implements a conditional routing mechanism based on the availability of core features. Instead of the traditional "one-size-fits-all" deletion of missing values ​​or brute-force mean filling, the system dynamically distributes computation tasks based on the real-time status of the core diagnostic indicator, PMN, directing the computation to either a four-variable model supporting post-puncture diagnosis or a three-variable model for rapid pre-puncture triage. This effectively addresses the pain point of existing models that either fail to run or experience significant performance degradation when core variables are missing. Through the dual-path automatic switching mechanism, the system can provide highly available triage prompts via the second identification path during the "pre-puncture" stage (before the patient undergoes paracentesis); and immediately switch to the first identification path during the "post-puncture" stage (after puncture and PMN data feedback), fully leveraging the core diagnostic value of PMN. This ensures high system availability while maximizing the accuracy of diagnostic support.

[0015] The probability calculation module, under the first identification path, calls the first regression model to perform a weighted summation and logistic nonlinear mapping on the patient's history of severe bloating (SBP), concurrent upper gastrointestinal bleeding, liver failure status, and PMN count in ascites, calculating the first auxiliary identification probability. Under the second identification path, it calls the second regression model to perform a weighted summation and logistic nonlinear mapping on the first three clinical variables, calculating the second auxiliary identification and triage probability. This module utilizes a multivariate logistic regression algorithm optimized through fitting large-sample clinical data to perform multi-dimensional weighted fusion of discrete binary clinical features and continuous laboratory indicators. Through a nonlinear mapping function, the linear prediction value of the synergistic effect of multiple factors is transformed into a continuous risk probability between 0 and 1. This overcomes the limitation of traditional clinical practice relying solely on a single PMN threshold for binary classification. The probability calculation module can accurately quantify the cumulative effect of significant clinical background factors such as past medical history, upper gastrointestinal bleeding, and liver failure on the current SBP risk. The output continuous probability value is more informative than the traditional "yes / no" dichotomy, and can more delicately reflect the true risk level of an individual patient.

[0016] The threshold determination module compares the first auxiliary identification probability with a preset first risk threshold to generate an auxiliary identification result; or, it compares the second auxiliary identification triage probability with a preset second risk threshold to generate a pre-puncture triage result. This module transforms complex mathematical probabilities into intuitive classification decision signals with clear clinical guidance by comparing a continuous probability distribution with an optimal cutoff point determined based on a clinical validation set (e.g., data from 488 patients). This provides clinicians with clear and standardized decision-making support. In the first path, using an optimized threshold (e.g., 0.391) can achieve a sensitivity of up to 91.91%, significantly reducing the missed diagnosis rate of SBP, making it highly suitable as an auxiliary identification tool. In the second path, using an optimized threshold (e.g., 0.599) for pre-puncture triage can accurately identify high-risk individuals requiring priority for paracentesis and infection assessment, optimizing the treatment process in emergency and general wards.

[0017] The output and audit module outputs the selected path, variable values, probability values, threshold judgment results, and clinical review prompts. It also calculates the contribution of each variable to the probability value and provides an interpretable display, while recording audit information including model version, input data, and operation logs. This module introduces interpretable computation (such as additive interpretive models) and a structured audit trail mechanism. By calculating the contribution direction and weight of each input variable at the current individual patient level, the complex model calculation logic is made transparent. Simultaneously, the input, output, model version, and operation context of each calculation are logged immutably. This eliminates "black box" concerns, enhances clinical trust, and allows physicians to not only see the risk probability but also clearly understand which clinical features (such as liver failure or past history) dominated the high-risk judgment, facilitating physician review based on professional experience.

[0018] Meeting medical quality control and audit compliance requirements, the complete audit logs ensure that the auxiliary identification results are traceable, auditable, and reproducible, providing solid technical support for medical dispute prevention, model version iteration, and multi-center quality control.

[0019] This invention does not rely on expensive genomics or special equipment for testing. It only requires four readily available routine clinical variables, has minimal computational overhead, and can be easily promoted in primary hospitals or resource-limited areas, significantly reducing the economic and time costs of early screening.

[0020] The combination of the dual-path architecture and the audit module enables the present invention to perform localized parameter calibration in different medical centers very conveniently. Without changing the core software architecture, it can quickly adapt to the differences in patient characteristics in different regions, demonstrating excellent scalability.

[0021] Furthermore, the data standardization module, when performing time window verification and filtering, specifically uses the following methods: establishing a fixed time anchor within the system, using the time of the first diagnostic paracentesis or the first diagnosis of spontaneous bacterial peritonitis as a reference time point; limiting the clinical history data to information known at admission and recorded earlier than the reference time point; limiting the inpatient complication data to complications occurring before the reference time point, or on the same day as the first diagnostic paracentesis and earlier than the antibiotic adjustment period; limiting the ascites test data to the test results of the first diagnostic paracentesis specimen, and recording its sampling time and testing unit; for other routine laboratory indicators, prioritizing the most recent test results within 24 hours before the first diagnostic paracentesis and no later than the reference time point; and automatically eliminating complications and laboratory indicator changes occurring after the reference time point through the above time window verification, preventing data leakage and ensuring the causal time sequence logic of the calculation.

[0022] Furthermore, the data standardization module, when performing unit conversion, duplicate record disambiguation, and value range verification, specifically performs the following: It reads the original detection unit of the ascites polymorphonuclear neutrophil count from the original dataset to be processed; if the original detection unit is cells per cubic millimeter, it converts it to a continuous value in units of 10 to the power of 9 per liter according to a preset conversion rule; it performs non-negativity verification on all variables in the original dataset to be processed, determining whether the ascites polymorphonuclear neutrophil count is non-negative; if it is negative, it triggers an outlier warning; when multiple duplicate routine laboratory test records are detected within the same time window, it selects the most recent record closest to the reference time point as valid data according to a preset disambiguation rule, and automatically discards the remaining duplicate records; when an outlier exceeding the preset normal range is identified, it triggers a system review prompt, and retains the original value, revised value, revision time, and the identification information of the revision operator in the system interface, maintaining the original traceability and integrity of the data.

[0023] Furthermore, when identifying whether polymorphonuclear neutrophil counts exist in the standardized dataset, the path selection module is specifically used to: retrieve a specific data field in the standardized dataset corresponding to the polymorphonuclear neutrophil count in the ascites; if the specific data field contains a valid value and is not marked as missing, then the polymorphonuclear neutrophil count in the ascites is determined to be available, the system controls the system to enter the first identification path, and marks the path as a clinical decision support path containing ascites test indicators in the output interface; if the specific data field is empty, contains a preset placeholder indicating a missing value, or fails to retrieve the corresponding test report within a preset time window, then the polymorphonuclear neutrophil count in the ascites is determined to be missing, the system controls the system to automatically switch to the second identification path, and prominently displays a prompt for the purpose of pre-puncture triage in the output interface, clearly indicating that this path does not replace ascites examination, so as to prevent clinicians from mistakenly taking the triage result as a basis for diagnosis.

[0024] Furthermore, when selecting the first identification path, the probability calculation module is specifically used to: convert the history of spontaneous bacterial peritonitis, concurrent upper gastrointestinal bleeding, and liver failure status in the standardized dataset into corresponding binary coded values, where zero is the code for no occurrence and one is the code for occurrence; multiply the converted binary coded values ​​and the ascites polymorphonuclear neutrophil count (as a continuous value) with the first regression coefficients corresponding to each variable in the first regression model to obtain the weighted score of each variable; accumulate the weighted scores of all variables with the first intercept term in the first regression model to obtain the first linear prediction value; and use the logistic function to perform a nonlinear mapping on the first linear prediction value to calculate a continuous probability value between zero and one, which serves as the first auxiliary identification probability, wherein the logistic function uses the base of the natural logarithm as the basis for operation to map the linear prediction value into a probability value.

[0025] Furthermore, when selecting the second identification path, the probability calculation module is specifically used to: convert the three variables in the standardized dataset—history of spontaneous bacterial peritonitis, concurrent upper gastrointestinal bleeding, and liver failure status—into corresponding binary coded values, where zero is the code for no occurrence and one is the code for occurrence; multiply the binary coded values ​​of these three variables by the second regression coefficients corresponding to each variable preset in the second regression model to obtain corresponding weighted scores; accumulate these three weighted scores with the second intercept term preset in the second regression model to obtain a second linear prediction value; and use the logistic function to perform a nonlinear mapping on the second linear prediction value to calculate another continuous probability value between zero and one, which serves as the second auxiliary identification and triage probability to provide a risk triage reference before puncture when polymorphonuclear neutrophil counts in ascites are lacking, wherein the logistic function uses the base of the natural logarithm as the basis for calculation to map the linear prediction value to a probability value.

[0026] Furthermore, the threshold determination module also includes a parallel comparison submodule, which is used to: display in parallel on the system interface the probability determination results calculated based on the first regression model or the second regression model, and the single-indicator comparison results based solely on the ascites polymorphonuclear neutrophil count; the single-indicator comparison results include: a first comparison result comparing the ascites polymorphonuclear neutrophil count with a data-driven screening cutoff point determined based on historical research data, and a second comparison result comparing the ascites polymorphonuclear neutrophil count with a diagnostic threshold recommended by clinical guidelines; the parallel comparison submodule displays the data-driven screening cutoff point and the diagnostic threshold recommended by clinical guidelines independently in the system interface by region or color, and outputs text prompts to clearly indicate that the two have different clinical uses, so as to avoid clinicians confusing the screening cutoff point with the guideline diagnostic criteria, and to ensure the accuracy and safety of the auxiliary decision-making.

[0027] Furthermore, when calculating the interpretability of the contribution of each variable to the probability value, the output and audit module specifically performs the following: It calls the additive explanation model to calculate the Shapley additive explanation value of each input variable in the standardized dataset relative to the currently calculated probability value; based on the sign of the calculated Shapley additive explanation value, it determines the direction of contribution of each input variable to the current patient's risk of spontaneous bacterial peritonitis, where a positive value indicates that the variable increases the risk of the disease, and a negative value indicates that the variable reduces the risk of the disease; based on the absolute value of the Shapley additive explanation value, it determines the contribution weight of each input variable to the risk impact, and dynamically displays the contribution distribution of each variable at the individual level in the system interface in the form of a graphical beehive diagram, waterfall diagram, or bar chart, intuitively presenting to clinicians the specific combination of clinical characteristics leading to the current high or low risk and its explanatory mechanism, thereby assisting physicians in reviewing and auditing clinical decisions.

[0028] Furthermore, the system also includes a local pathogen linkage module, which is used to: maintain and regularly update a local pathogen spectrum and drug sensitivity knowledge base. The knowledge base records common bacterial species isolated from spontaneous bacterial peritonitis in patients with cirrhosis and ascites in the local region or hospital in recent years, the detection rate of each bacterial species, and the drug resistance rate data against commonly used antimicrobial drugs. When the auxiliary identification result output by the threshold determination module or the pre-puncture triage result reaches the corresponding risk threshold, the local pathogen linkage module is triggered. It displays the common bacterial species and drug resistance rate information in the knowledge base in the system interface in a descriptive, non-mandatory medication instruction manner, and generates clinical specimen submission prompt information. The clinical specimen submission prompt information includes a suggestion to prioritize the collection of specimens for diagnostic peritoneal puncture, ascites culture, and blood culture before adjusting or using antimicrobial drugs, so as to realize the closed-loop linkage between risk identification and local pathogen diagnosis and treatment decision-making, and assist physicians in formulating reasonable empirical anti-infection treatment plans.

[0029] Furthermore, the output and audit module also includes a multi-center dynamic calibration and security audit sub-module. This sub-module is used to: when the system receives validation datasets from different medical institutions, call the logistic calibration algorithm to recalibrate the intercept and slope of the first regression model and the second regression model, generate localized model parameters adapted to specific medical institutions, and assign a unique model version number to each localized model parameter; the multi-center dynamic calibration and security audit sub-module is also used to: each time a probability calculation is performed, structurally encapsulate the currently used model version number, the input standardized dataset, the calculation path, the calculated probability value, the output judgment result, and the identity and timestamp of the currently operating physician, and write them into an immutable system audit log database for subsequent medical quality control, model performance auditing, and traceability query of historical diagnostic auxiliary records, ensuring the security and auditability of medical data. Attached Figure Description

[0030] Figure 1 This is a general architecture diagram of an embodiment of the dual-path auxiliary identification system for SBP in cirrhotic ascites. Figure 2 A flowchart illustrating the data standardization and time window verification process for an embodiment of the dual-path assisted identification system for SBP in cirrhotic ascites. Figure 3 The dynamic routing logic diagram of the path selection module in the embodiment of the dual-path assisted identification system for SBP in cirrhosis ascites. Figure 4 A comparison chart of ROC curves between the dual-path model and the single-index model in an embodiment of the dual-path auxiliary identification system for SBP in cirrhosis ascites. Figure 5 This is a schematic diagram of the overall contribution of model features (SHAP beehive graph) in an embodiment of the SBP dual-path assisted identification system for ascites in liver cirrhosis. Figure 6 This is a schematic diagram of the individual feature contribution (SHAP waterfall plot) in an embodiment of the SBP dual-path assisted identification system for ascites caused by liver cirrhosis. Detailed Implementation

[0031] The following detailed description illustrates the specific implementation method: A dual-pathway assisted identification system for SBP in patients with cirrhotic ascites includes: The data acquisition module is used to acquire the target patient's clinical history data, hospitalization complication data, and ascites test data, and to obtain the timestamp, unit, and data source of each data to obtain the raw dataset to be processed. The data standardization module is used to perform time window verification and filtering on the original dataset to be processed according to a preset time anchor point, and to perform unit conversion, duplicate record disambiguation and value range verification on the filtered data to obtain a standardized dataset. The path selection module is used to identify whether polymorphonuclear neutrophil counts exist in the standardized dataset. If they exist, the polymorphonuclear neutrophil counts are determined to be available, and a first identification path is selected. If they do not exist, the polymorphonuclear neutrophil counts are determined to be missing, and a second identification path is selected. The probability calculation module is used to calculate the first auxiliary identification probability by calling a preset first regression model to perform weighted summation and nonlinear mapping on the history of spontaneous bacterial peritonitis, concurrent upper gastrointestinal bleeding, liver failure status, and polymorphonuclear neutrophil count in the standardized dataset when selecting the first identification path; and to calculate the second auxiliary identification and triage probability by calling a preset second regression model to perform weighted summation and nonlinear mapping on the history of spontaneous bacterial peritonitis, concurrent upper gastrointestinal bleeding, and liver failure status in the standardized dataset when selecting the second identification path. The threshold determination module is used to compare the first auxiliary identification probability with a preset first risk threshold to generate a corresponding auxiliary identification result; or, to compare the second auxiliary identification triage probability with a preset second risk threshold to generate a corresponding pre-puncture triage result. The output and audit module is used to output the selected recognition path, the values ​​of each variable in the standardized dataset, the calculated probability value, the corresponding threshold judgment result, and the clinical review prompt. It also calculates the contribution of each variable to the probability value and displays it in an interpretable manner. At the same time, it records audit information including model version, input data, and operation log.

[0032] In practical use, it can be deployed on in-hospital servers, cloud servers, personal computers, mobile terminals, or embedded medical information terminals. It can standardize, select paths, calculate probabilities, and output interpretable data by reading ex vivo data generated from clinical diagnosis and treatment processes.

[0033] The data acquisition module is used to acquire the target patient's clinical history data, inpatient complication data, and ascites test data, and to obtain the timestamp, unit, and data source of each data to obtain the raw dataset to be processed.

[0034] The clinical history data refers to the patient's confirmed past medical history upon admission; the inpatient complication data refers to the concurrent disease status that occurred during the patient's hospitalization and is recorded in the medical record; the ascites test data refers to the quantitative indicators obtained from laboratory testing of ascites samples extracted by paracentesis. Optional implementation types of the data acquisition module include: an automated data capture tool based on the HL7 (HealthLevelSeven) or FHIR (FastHealthcareInteroperabilityResources) standard medical interface; a structured query module based on SQL database connections; or a manual input interactive interface based on a graphical user interface (GUI).

[0035] At the data architecture level, the data acquisition module maps multi-source heterogeneous data into standard relational data tables by configuring a standard data dictionary. Input variables are defined. This indicates a history of spontaneous bacterial peritonitis (SBP). This indicates concurrent upper gastrointestinal bleeding. This indicates a state of liver failure. This indicates the polymorphonuclear neutrophil (PMN) count in the ascites fluid.

[0036] In this embodiment, the data acquisition module is deployed on the hospital information system (HIS) server. Through a standard SQL interface, when a patient triggers an assisted identification request, it automatically reads historical diagnostic fields from the electronic medical record (EMR), clinical progress records, and routine ascites reports from the laboratory information system (LIS). The specific variables and encoding rules acquired are as follows: (Past SBP history): None = 0, Yes = 1; Data is derived from admission medical history or previous hospitalization records.

[0037] (Upper gastrointestinal bleeding): None = 0, Yes = 1; Data sourced from medical records or emergency room records.

[0038] (Liver failure status): None = 0, Yes = 1; Confirmed based on inpatient medical records and pre-set liver failure diagnostic criteria (such as ACLF diagnostic criteria).

[0039] (PMN count in ascites): Continuous values, unit is... / L; Data is derived from the laboratory report of the first diagnostic paracentesis specimen.

[0040] In addition, the system synchronously obtains the generation timestamp, detection unit, and data source system identifier for each variable.

[0041] The data standardization module is used to perform time window verification and filtering on the original dataset to be processed according to a preset time anchor point, and to perform unit conversion, duplicate record disambiguation, and value range verification on the filtered data to obtain a standardized dataset.

[0042] The time anchor point refers to the reference benchmark time used to align all clinical features in time sequence; the time window verification refers to limiting the maximum allowable time difference between the collection time of each variable and the time anchor point; the duplicate record disambiguation refers to the process of retaining a unique valid value according to preset rules when the same variable has multiple test results within the same time window.

[0043] The data standardization module, when performing time window verification and filtering, specifically: establishes a fixed time anchor point within the system, using the time of the first diagnostic paracentesis or the first determination of spontaneous bacterial peritonitis as a reference time point; limits the clinical history data to medical history information known at admission and recorded earlier than the reference time point; limits the inpatient complication data to complication information occurring before the reference time point, or on the same day as the first diagnostic paracentesis and earlier than the antibiotic adjustment period; limits the ascites test data to the test results of the first diagnostic paracentesis specimen, and records its sampling time and testing unit; for other routine laboratory indicators, prioritizes the most recent test results within 24 hours before the first diagnostic paracentesis and no later than the reference time point; through the above time window verification, automatically eliminates complications and changes in laboratory indicators occurring after the reference time point, preventing data leakage and ensuring the causal time sequence logic of the calculation.

[0044] The data standardization module, when performing unit conversion, duplicate record disambiguation, and value range verification, specifically performs the following: It reads the original detection unit of the ascites polymorphonuclear neutrophil count from the original dataset to be processed; if the original detection unit is cells per cubic millimeter, it converts it to a continuous value in units of 10 to the power of 9 per liter according to a preset conversion rule; it performs non-negativity verification on all variables in the original dataset to be processed, determining whether the ascites polymorphonuclear neutrophil count is non-negative; if negative, it triggers an outlier warning; when multiple duplicate routine laboratory test records are detected within the same time window, it selects the most recent record closest to the reference time point as valid data according to a preset disambiguation rule, and automatically discards the remaining duplicate records; when an outlier exceeding the preset normal range is identified, it triggers a system review prompt and retains the original value, revised value, revision time, and the identification information of the revision operator in the system interface, maintaining the original traceability and integrity of the data.

[0045] In this embodiment, as shown in the appendix Figure 2 As shown, the specific execution steps of the data standardization module are as follows: Time anchor points are established, and the system reads the time of the first diagnostic abdominal paracentesis. .like If it exists, then it is... As a time anchor If not, the time of the first suspected SBP diagnosis in clinical practice shall be used. As .

[0046] Time window filtering, clinical history variables Recording time Must meet .

[0047] Hospitalization complication variables and Recording time Must meet If with On the same day, it must be earlier than the first adjustment time of the antibacterial drug.

[0048] Ascites test data The test values ​​must be from the first diagnostic paracentesis specimen, and the sampling time must be specified. satisfy Hour.

[0049] The time window for other routine laboratory indicators (such as serum creatinine, total bilirubin, etc.) is limited to... .

[0050] Unit conversion and disambiguation, reading The unit field. If the unit is cells / mm³ or cells / μL, the system will automatically apply the conversion formula: .

[0051] If in If multiple blood creatinine tests are performed, the system automatically runs a disambiguation algorithm to retain the closest result. The detected value is discarded, and the rest of the records are discarded.

[0052] Value range validation, validation .like The system detects an anomaly and does not automatically delete the data. Instead, it displays a red warning on the interface: "Abdominal fluid PMN count is abnormal, please check." After an authorized physician manually reviews the data and enters the revised value, the system writes the original value, the revised value, the operator ID "User_083", and the timestamp to the backend database.

[0053] The path selection module is used to identify whether polymorphonuclear neutrophil counts exist in the standardized dataset. If they exist, the polymorphonuclear neutrophil counts are determined to be available, and a first identification path is selected. If they do not exist, the polymorphonuclear neutrophil counts are determined to be missing, and a second identification path is selected.

[0054] The first identification path refers to the complete variable calculation path that includes the core diagnostic indicator of ascites PMN count (i.e., the with-PMN path); the second identification path refers to the calculation path that relies only on clinical background variables before ascites puncture or when the puncture results are not available (i.e., the non-PMN path). Optional path routing implementation methods include: a logical splitter based on conditional branching statements (If-Else), a workflow engine based on directed acyclic graphs (DAGs), or a microservice gateway based on a dynamic loading mechanism.

[0055] When identifying whether polymorphonuclear neutrophil (Pneumocystis) count exists in the standardized dataset, the path selection module is specifically used to: retrieve a specific data field in the standardized dataset corresponding to the Pneumocystis count in the ascites; if the specific data field contains a valid value and is not marked as missing, the Pneumocystis count in the ascites is determined to be available, the system enters the first identification path, and marks the path as a clinical decision support path containing ascites test indicators in the output interface; if the specific data field is empty, contains a preset placeholder indicating a missing value, or fails to retrieve the corresponding test report within a preset time window, the Pneumocystis count in the ascites is determined to be missing, the system automatically switches to the second identification path, and prominently displays a pre-puncture triage indication in the output interface, clearly stating that this path does not replace ascites examination, to prevent clinicians from mistaking triage results for diagnostic evidence.

[0056] In this embodiment, the path selection module is implemented using a Python logical splitter. The system retrieves the ascites_pmn field from the standardized dataset.

[0057] If the ascites_pmn field value is a valid real number (e.g.) The system determines that the path is available, executes the code route_path = "with_PMN", enters the first identification path, and renders a green label in a prominent position on the front-end interface: "Currently running: with-PMN clinical decision support path".

[0058] If the `ascites_pmn` field is NaN, None, or an empty string, the system determines it to be missing, executes the code `route_path = "non_PMN"`, automatically switches to the second identification path, and displays a yellow warning label on the front-end interface: "Currently running: non-PMN pre-puncture triage path (Note: This result is only for pre-puncture triage and does not replace ascites examination!)". See attached... Figure 3 As shown.

[0059] The probability calculation module is used to calculate the first auxiliary identification probability by calling a preset first regression model to perform weighted summation and nonlinear mapping on the history of spontaneous bacterial meningitis, concurrent upper gastrointestinal bleeding, liver failure status, and polymorphonuclear neutrophil count in the standardized dataset when selecting the first identification path; and to calculate the second auxiliary identification and triage probability by calling a preset second regression model to perform weighted summation and nonlinear mapping on the history of spontaneous bacterial meningitis, concurrent upper gastrointestinal bleeding, and liver failure status in the standardized dataset when selecting the second identification path.

[0060] The first and second regression models refer to mathematical prediction models based on multivariate logistic regression; the nonlinear mapping refers to using the logistic cumulative distribution function (i.e., the sigmoid function) to map the linear prediction values ​​in the real domain to... The interval is used to output continuous risk probability values ​​that conform to statistical significance.

[0061] This embodiment constructs a first regression model (with-PMN model) and a second regression model (non-PMN model) based on a multivariate logistic regression model. Specifically, we configured a generalized linear model (GLM) with a binomial distribution for its random error term and a Logit link function. To eliminate the effects of multicollinearity, we performed variance inflation factor (VIF) testing on candidate variables during model construction, ensuring that the VIF of all included variables was less than 5 and that no coagulation function indicators with strong collinearity were introduced, thereby guaranteeing the stability and biological rationality of the regression coefficients.

[0062] The first and second regression models were trained using a clinical training sample of 488 patients with cirrhosis and ascites. The spontaneous bacterial peritonitis (SBP) group consisted of 272 patients, and the non-SBP group consisted of 216 patients. Maximum likelihood estimation was used to fit the sample data, and the regression coefficients and intercepts of each input variable were obtained. To evaluate the model's generalization performance and avoid overfitting, the system underwent 500 bootstrap resampling iterations for internal validation. The receiver operating characteristic (AUC), Brier score, calibration intercept, and calibration slope were calculated, and the model parameters were initialized and the configuration file was saved.

[0063] When selecting the first identification path, the probability calculation module is specifically used to: convert the history of spontaneous bacterial peritonitis, concurrent upper gastrointestinal bleeding, and liver failure status in the standardized dataset into corresponding binary coded values, where zero is the code for no occurrence and one is the code for occurrence; multiply the converted binary coded values ​​and the ascites polymorphonuclear neutrophil count (a continuous value) with the first regression coefficients corresponding to each variable in the first regression model to obtain the weighted score of each variable; sum the weighted scores of all variables with the first intercept term in the first regression model to obtain the first linear prediction value; and use the logistic function to perform a nonlinear mapping on the first linear prediction value to calculate a continuous probability value between zero and one, which serves as the first auxiliary identification probability, wherein the logistic function uses the base of the natural logarithm as the basis for operation to map the linear prediction value to the probability value.

[0064] When selecting the second identification path, the probability calculation module specifically performs the following steps: It converts the three variables in the standardized dataset—history of spontaneous bacterial peritonitis, concurrent upper gastrointestinal bleeding, and liver failure status—into corresponding binary coded values, where zero is the code for no occurrence and one is the code for occurrence; it multiplies the binary coded values ​​of these three variables by the second regression coefficients corresponding to each variable in the second regression model to obtain corresponding weighted scores; it accumulates these three weighted scores with the second intercept term in the second regression model to obtain a second linear prediction value; and it uses a logistic function to perform a non-linear mapping on the second linear prediction value to calculate another continuous probability value between zero and one, which serves as the second auxiliary identification and triage probability. This provides a risk triage reference before puncture in the absence of polymorphonuclear neutrophil counts in ascites, wherein the logistic function uses the base of the natural logarithm as the basis for calculation, mapping the linear prediction value to a probability value.

[0065] In this embodiment, after training with the aforementioned 488 samples, the specific regression parameters stored in the system configuration file are as follows: First regression model (with-PMN model): First intercept term

[0066] Previous SBP history First regression coefficient

[0067] Upper gastrointestinal bleeding First regression coefficient

[0068] liver failure First regression coefficient

[0069] PMN count in ascites First regression coefficient

[0070] First linear prediction value Calculation formula:

[0071] First auxiliary recognition probability Calculation formula (Logistic nonlinear mapping):

[0072] Second regression model (non-PMN model): Second intercept term

[0073] Previous SBP history Second regression coefficient

[0074] Upper gastrointestinal bleeding Second regression coefficient

[0075] liver failure Second regression coefficient

[0076] Second linear prediction value Calculation formula:

[0077] Second auxiliary identification triage probability Calculation formula:

[0078] The threshold determination module is used to compare the first auxiliary identification probability with a preset first risk threshold to generate a corresponding auxiliary identification result; or, to compare the second auxiliary identification triage probability with a preset second risk threshold to generate a corresponding pre-puncture triage result.

[0079] The first risk threshold and the second risk threshold refer to probability thresholds determined based on the principle of maximizing the Youden's Index of the ROC curve or based on screening for high-sensitivity clinical needs.

[0080] The threshold determination module further includes a parallel comparison submodule, which is used to: display in the system interface the probability determination results calculated based on the first regression model or the second regression model, and the single-indicator comparison results based solely on the ascites polymorphonuclear neutrophil count; the single-indicator comparison results include: a first comparison result comparing the ascites polymorphonuclear neutrophil count with a data-driven screening cutoff point determined based on historical research data, and a second comparison result comparing the ascites polymorphonuclear neutrophil count with a diagnostic threshold recommended by clinical guidelines; the parallel comparison submodule displays the data-driven screening cutoff point and the diagnostic threshold recommended by clinical guidelines independently in the system interface by region or color, and outputs text prompts to clearly indicate that their clinical uses are different, so as to avoid clinicians confusing the screening cutoff point with the guideline diagnostic criteria, and to ensure the accuracy and safety of the auxiliary decision-making.

[0081] In this embodiment, the system preset decision parameters are as follows: First risk threshold (with-PMN path): preset to .when When the system determines that the assisted identification threshold has been reached, a high-risk warning is triggered; when When the preset threshold is not reached, the output will be "Not reached".

[0082] Second risk threshold (non-PMN path): preset to .when When the system determines that the "high-risk threshold for triage has been reached", it outputs the message: "It is recommended to prioritize completing diagnostic abdominal paracentesis and infection assessment".

[0083] Parallel comparison configuration: A separate panel on the right side of the system interface displays parallel comparisons. Data-driven screening cutoff points: / L (displayed in blue and labeled "Data-driven screening cutoff point for high-sensitivity screening").

[0084] Clinical guideline diagnostic thresholds: / L (shown in bold red, marked as "International Guideline Diagnostic Threshold, used as a basis for diagnosis").

[0085] As attached Figure 4 As shown, the ROC curve comparison chart intuitively demonstrates the performance differences between the with-PMN model (AUC=0.818), non-PMN model (AUC=0.633), and PMN-only comparison modules.

[0086] The output and audit module is used to output the selected recognition path, the values ​​of each variable in the standardized dataset, the calculated probability value, the corresponding threshold judgment result, and the clinical review prompt. It also calculates the contribution of each variable to the probability value and displays it in an interpretable manner. At the same time, it records audit information including model version, input data, and operation log.

[0087] The interpretability demonstration refers to using an additive feature attribution method to quantify the positive or negative pull of a single clinical feature on the current risk probability prediction value of a specific patient; the audit information refers to the structured record of the complete input, output and operation context of the software during operation in order to meet medical compliance and traceability requirements.

[0088] The output and audit module, when calculating the interpretability of the contribution of each variable to the probability value, specifically involves: calling the additive explanation model to calculate the Shapley additive explanation value of each input variable in the standardized dataset relative to the currently calculated probability value; determining the direction of contribution of each input variable to the current patient's risk of spontaneous bacterial peritonitis based on the sign of the calculated Shapley additive explanation value, where a positive value indicates that the variable increases the risk of the disease, and a negative value indicates that the variable reduces the risk of the disease; determining the contribution weight of each input variable to the risk based on the absolute value of the Shapley additive explanation value, and dynamically displaying the contribution distribution of each variable at the individual level in the system interface in the form of a graphical beehive diagram, waterfall diagram, or bar chart, intuitively presenting to clinicians the specific combination of clinical characteristics leading to the current high or low risk and its explanatory mechanism, thereby assisting physicians in reviewing and auditing clinical decisions.

[0089] The output and audit module also includes a multi-center dynamic calibration and security audit sub-module. This sub-module is used to: when the system receives validation datasets from different medical institutions, call the logistic calibration algorithm to recalibrate the intercept and slope of the first regression model and the second regression model, generate localized model parameters adapted to specific medical institutions, and assign a unique model version number to each localized model parameter; the multi-center dynamic calibration and security audit sub-module is also used to: each time a probability calculation is performed, structurally encapsulate the currently used model version number, the input standardized dataset, the calculation path, the calculated probability value, the output judgment result, and the identity and timestamp of the currently operating physician, and write them into an immutable system audit log database for subsequent medical quality control, model performance auditing, and traceability query of historical diagnostic auxiliary records, ensuring the security and auditability of medical data.

[0090] In this embodiment, the output and auditing module is implemented as follows: SHAP interpretability computation: The system backend integrates the Python shap library. Using 488 training samples as the background dataset, a shap.LinearExplainer interpreter is constructed. For the patient currently being computed, the SHAP values ​​of each variable are calculated. .

[0091] Overall variable contribution (as shown in the appendix) Figure 5 The SHAP colony diagram shown indicates that the PMN count in ascites ( ) and previous SBP history ( () is the core source of the model's discriminative ability.

[0092] Contribution of individual variables (as shown in the appendix) Figure 6 The SHAP waterfall plot shown visually illustrates the linear predictive value of each patient's current variables. The cumulative contribution process: .

[0093] Security Audit and Version Control: The system uses a PostgreSQL database to store audit logs. After each calculation, a structured JSON record is automatically inserted into the `audit_log` table, as shown in the example below: { "log_id":"LOG_20250803_0019", "timestamp":"2025-08-03T10:15:30Z", "operator_id":"MD_Zhang66", "patient_id":"PAT_992817", "model_version":"with-PMN_v1.0.2_siteA", "inputs":{"X1":1,"X2":0,"X3":1,"X4":0.005}, "path_selected":"with-PMN", "linear_predictor": 1.780, "probability": 0.856, "threshold":0.391, "decision":"ABOVE_THRESHOLD" } The system also includes a local pathogen linkage module, which is used to: maintain and regularly update a local pathogen spectrum and drug sensitivity knowledge base. The knowledge base records common bacterial species isolated from spontaneous bacterial peritonitis in patients with cirrhosis and ascites in the local region or hospital in recent years, the detection rate of each species, and the drug resistance rate data against commonly used antimicrobial drugs. When the auxiliary identification result or pre-puncture triage result output by the threshold determination module reaches the corresponding risk threshold, the local pathogen linkage module is triggered. It displays the common bacterial species and drug resistance rate information in the knowledge base in the system interface in a descriptive, non-mandatory medication instruction manner, and generates clinical specimen submission prompts. The clinical specimen submission prompts include suggestions to prioritize the collection of specimens for diagnostic peritoneal puncture, ascites culture, and blood culture before adjusting or using antimicrobial drugs, so as to realize the closed-loop linkage between risk identification and local pathogen diagnosis and treatment decision-making, and assist physicians in formulating reasonable empirical anti-infection treatment plans.

[0094] In this embodiment, the local pathogen linkage module achieves decision linkage by reading the pathogen_stats table in the local SQLite database.

[0095] Knowledge base data configuration: Based on the ascites culture data of historical SBP patients in our hospital, a total of 264 cases were submitted for testing, with 49 positive cases (positive rate of 18.56%); a total of 55 pathogenic bacteria were isolated, including 30 Gram-negative bacteria (accounting for 54.6%, mainly Escherichia coli and Klebsiella pneumoniae), 24 Gram-positive bacteria (accounting for 43.6%, mainly Enterococci and Streptococci), and 1 fungus (accounting for 1.8%).

[0096] Triggering Mechanism: When the system determines that the auxiliary identification threshold or the high-risk triage threshold has been reached, a descriptive reference information will automatically pop up in the lower right corner of the front-end interface: "[SBP Etiology Reference in Our Hospital] In recent years, the positive rate of SBP ascites culture in our hospital has been 18.56%, the most common isolated bacteria are Gram-negative bacteria (54.6%), and the resistance rate to third-generation cephalosporins is XX%. [Submission Tips] It is strongly recommended to collect bilateral blood cultures and ascites cultures in a standardized manner before the first use or adjustment of antibiotics (it is recommended to use blood culture bottles to inoculate ascites to improve the positive rate)." The process of auxiliary identification using a patient who has undergone paracentesis and whose PMN data has been returned is as follows: Data Acquisition and Standardization: The data acquisition module reads the data of patient A (ID: PAT_001): There is a history of SBP occurrences, and the system coding... ; No upper gastrointestinal bleeding was reported during this admission. (System code) ; Clinical assessment indicates liver failure; system coding ; The routine ascites test report showed a PMN count of 5 cells / mm³.

[0097] The data standardization module identified The original unit is cells / mm³, and the unit conversion will be performed automatically: ; After time window verification, all the above data were within the time frame of the first diagnostic paracentesis. Data sets taken before or on the same day as the medication are considered valid standardized datasets.

[0098] Path selection: The path selection module retrieved... For a valid value, determine It can be obtained that the control system enters the first identification path (with-PMN path).

[0099] Probability Calculation: The probability calculation module calls the first regression model, substitutes standardized data into the formula for calculation: Assuming linear prediction values calculate: ; Assuming auxiliary recognition probability calculate: ; Threshold determination: The threshold determination module will... Compared with the first risk threshold Comparison. Because The system determined that the threshold for with-PMN assisted identification had been reached, triggering a high-risk warning.

[0100] Output, interpretability, and auditability: Interface display. The system outputs the following on the clinical decision support terminal: Selected path is with-PMN; Current risk probability is 85.6%; Threshold is determined to be high risk.

[0101] Explainability visualization (as attached) Figure 6 As shown in the diagram), the SHAP waterfall plot visually demonstrates that although the patient's ascites PMN count was extremely low ( / L has a downward pulling effect on risk. However, due to the patient's "history of SBP (positive pull)", ) and "liver failure state" (positive pull) The combined effect of the two high-risk clinical backgrounds far exceeds the pulling effect of low PMN, ultimately leading to an extremely high overall risk probability.

[0102] Parallel comparison, displayed in parallel on the right side of the interface: This patient / L, far below the data-driven screening cutoff point / L and guideline diagnostic thresholds The system message reads: "Although this patient's ascites PMN does not meet the guideline diagnostic criteria, due to the high-risk clinical background, the overall risk calculated by the system is extremely high, suggesting that the patient may be in the very early stage of SBP. Physicians should closely consider the clinical manifestations for a comprehensive judgment and should not easily rule out the diagnosis based on a negative PMN." For audit retention, the system automatically writes the inputs, outputs, model version with-PMN_v1.0.2, and operation logs of this calculation to the PostgreSQL audit database.

[0103] The process of rapid triage for an emergency patient who has just been admitted and has not yet undergone paracentesis (no PMN data) is as follows: Data Acquisition and Standardization: The data acquisition module reads the data of patient B (ID: PAT_002): There is a history of SBP occurrences, and the system coding... ; No upper gastrointestinal bleeding was reported during this admission. (System code) ; Clinical assessment indicates liver failure; system coding ; Ascites test data Missing (puncture not yet performed).

[0104] Data standardization module Perform time window alignment and validation, generate a standardized dataset, and then... Marked as missing (NaN).

[0105] Path selection: The path selection module recognizes... If the status is missing, the control system automatically switches to the second identification path (non-PMN path).

[0106] Probability Calculation: The probability calculation module calls the second regression model, assuming that the values ​​are substituted into the formula for calculation: Linear Predicted Values calculate: Assuming we get

[0107] Assuming triage probability calculate: Assuming we get

[0108] Threshold determination: The threshold determination module will With the second risk threshold Comparison. Because The system determined that the "high-risk threshold for triage" had been reached.

[0109] Output and decision-making linkage: Interface display: The system prominently displays a yellow warning on the emergency triage terminal: "High risk of non-PMN pre-puncture triage! Risk probability: 88.8%." It also prominently states: "Note: This path does not include PMN indicators and is only used for pre-puncture triage. The internal validation AUC is 0.633 and it cannot be used as a basis for diagnosis." Decision-making linkage prompt: The system automatically triggers the local etiology linkage module and pops up a prompt: "The risk of triage before SBP puncture for this patient is extremely high. It is recommended to contact the attending physician immediately, prioritize the diagnosis of abdominal paracentesis, and collect ascites and blood culture specimens in accordance with regulations before the first use of antibiotics." Audit Retention: The system will write the complete calculation log of the non-PMN path to the audit database and mark the model version as non-PMN_v1.0.2.

[0110] The above are merely embodiments of the present invention. The invention is not limited to the fields covered by these embodiments. Commonly known structures and characteristics in the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are able to access all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A dual-path auxiliary identification system for SBP in patients with cirrhotic ascites, characterized in that, include: The data acquisition module is used to acquire the target patient's clinical history data, hospitalization complication data, and ascites test data, and to obtain the timestamp, unit, and data source of each data to obtain the raw dataset to be processed. The data standardization module is used to perform time window verification and filtering on the original dataset to be processed according to a preset time anchor point, and to perform unit conversion, duplicate record disambiguation and value range verification on the filtered data to obtain a standardized dataset. The path selection module is used to identify whether polymorphonuclear neutrophil counts exist in the standardized dataset. If they exist, the polymorphonuclear neutrophil counts are determined to be available, and a first identification path is selected. If they do not exist, the polymorphonuclear neutrophil counts are determined to be missing, and a second identification path is selected. The probability calculation module is used to calculate the first auxiliary identification probability by calling a preset first regression model to perform weighted summation and nonlinear mapping on the history of spontaneous bacterial peritonitis, concurrent upper gastrointestinal bleeding, liver failure status, and polymorphonuclear neutrophil count in the standardized dataset when selecting the first identification path; and to calculate the second auxiliary identification and triage probability by calling a preset second regression model to perform weighted summation and nonlinear mapping on the history of spontaneous bacterial peritonitis, concurrent upper gastrointestinal bleeding, and liver failure status in the standardized dataset when selecting the second identification path. The threshold determination module is used to compare the first auxiliary identification probability with a preset first risk threshold to generate a corresponding auxiliary identification result; or, to compare the second auxiliary identification triage probability with a preset second risk threshold to generate a corresponding pre-puncture triage result. The output and audit module is used to output the selected recognition path, the values ​​of each variable in the standardized dataset, the calculated probability value, the corresponding threshold judgment result, and the clinical review prompt. It also calculates the contribution of each variable to the probability value and displays it in an interpretable manner. At the same time, it records audit information including model version, input data, and operation log.

2. The dual-path auxiliary identification system for SBP in cirrhotic ascites according to claim 1, characterized in that, The data standardization module, when performing time window verification and filtering, specifically: establishes a fixed time anchor point within the system, using the time of the first diagnostic paracentesis or the first determination of spontaneous bacterial peritonitis as a reference time point; limits the clinical history data to medical history information known at admission and recorded earlier than the reference time point; limits the inpatient complication data to complication information occurring before the reference time point, or on the same day as the first diagnostic paracentesis and earlier than the antibiotic adjustment period; limits the ascites test data to the test results of the first diagnostic paracentesis specimen, and records its sampling time and testing unit; for other routine laboratory indicators, prioritizes the most recent test results within 24 hours before the first diagnostic paracentesis and no later than the reference time point; through the above time window verification, automatically eliminates complications and changes in laboratory indicators occurring after the reference time point, preventing data leakage and ensuring the causal time sequence logic of the calculation.

3. The dual-path auxiliary identification system for SBP in cirrhotic ascites according to claim 1, characterized in that, The data standardization module, when performing unit conversion, duplicate record disambiguation, and value range verification, is specifically used for: reading the original detection unit of the ascites polymorphonuclear neutrophil count in the original dataset to be processed; if the original detection unit is cells per cubic millimeter, converting it into a continuous value in units of 10 to the power of 9 per liter according to a preset conversion rule; performing non-negativity verification on all variables in the original dataset to be processed, determining whether the ascites polymorphonuclear neutrophil count is non-negative, and triggering an outlier warning if it is negative; When multiple duplicate routine laboratory test records are detected within the same time window, the most recent record closest to the reference time point is selected as valid data according to the preset disambiguation rules, and the remaining duplicate records are automatically discarded. When an abnormal value exceeding the preset normal range is identified, the system is triggered to prompt for review, and the original value, revised value, revision time, and identification information of the revision operator are retained in the system interface to maintain the original traceability and integrity of the data.

4. The dual-path auxiliary identification system for SBP in cirrhotic ascites according to claim 1, characterized in that, When identifying whether polymorphonuclear neutrophil (Pneumocystis) count exists in the standardized dataset, the path selection module is specifically used to: retrieve a specific data field in the standardized dataset corresponding to the Pneumocystis count in the ascites; if the specific data field contains a valid value and is not marked as missing, the Pneumocystis count in the ascites is determined to be available, the system enters the first identification path, and marks the path as a clinical decision support path containing ascites test indicators in the output interface; if the specific data field is empty, contains a preset placeholder indicating a missing value, or fails to retrieve the corresponding test report within a preset time window, the Pneumocystis count in the ascites is determined to be missing, the system automatically switches to the second identification path, and prominently displays a pre-puncture triage indication in the output interface, clearly stating that this path does not replace ascites examination, to prevent clinicians from mistaking triage results for diagnostic evidence.

5. The dual-path auxiliary identification system for SBP in cirrhotic ascites according to claim 1, characterized in that, When selecting the first identification path, the probability calculation module is specifically used to: convert the history of spontaneous bacterial peritonitis, concurrent upper gastrointestinal bleeding, and liver failure status in the standardized dataset into corresponding binary coded values, where zero is the code for no occurrence and one is the code for occurrence; multiply the converted binary coded values ​​and the ascites polymorphonuclear neutrophil count (a continuous value) with the first regression coefficients corresponding to each variable in the first regression model to obtain the weighted score of each variable; sum the weighted scores of all variables with the first intercept term in the first regression model to obtain the first linear prediction value; and use the logistic function to perform a nonlinear mapping on the first linear prediction value to calculate a continuous probability value between zero and one, which serves as the first auxiliary identification probability, wherein the logistic function uses the base of the natural logarithm as the basis for operation to map the linear prediction value to the probability value.

6. The dual-path auxiliary identification system for SBP in cirrhotic ascites according to claim 1, characterized in that, When selecting the second identification path, the probability calculation module specifically performs the following steps: It converts the three variables in the standardized dataset—history of spontaneous bacterial peritonitis, concurrent upper gastrointestinal bleeding, and liver failure status—into corresponding binary coded values, where zero is the code for no occurrence and one is the code for occurrence; it multiplies the binary coded values ​​of these three variables by the second regression coefficients corresponding to each variable in the second regression model to obtain corresponding weighted scores; it accumulates these three weighted scores with the second intercept term in the second regression model to obtain a second linear prediction value; and it uses a logistic function to perform a non-linear mapping on the second linear prediction value to calculate another continuous probability value between zero and one, which serves as the second auxiliary identification and triage probability. This provides a risk triage reference before puncture in the absence of polymorphonuclear neutrophil counts in ascites, wherein the logistic function uses the base of the natural logarithm as the basis for calculation, mapping the linear prediction value to a probability value.

7. The dual-path auxiliary identification system for SBP in cirrhotic ascites according to claim 1, characterized in that, The threshold determination module further includes a parallel comparison submodule, which is used to: display in the system interface the probability determination results calculated based on the first regression model or the second regression model, and the single-index comparison results based only on the ascites polymorphonuclear neutrophil count; The single-indicator comparison results include: a first comparison result comparing the ascites polymorphonuclear neutrophil count with a data-driven screening cutoff point determined based on historical research data, and a second comparison result comparing the ascites polymorphonuclear neutrophil count with a diagnostic threshold recommended by clinical guidelines; the parallel comparison submodule displays the data-driven screening cutoff point and the diagnostic threshold recommended by clinical guidelines independently in the system interface by region or color, and outputs text prompts to clearly indicate that the two have different clinical uses, so as to avoid clinicians confusing the screening cutoff point with the guideline diagnostic criteria, and to ensure the accuracy and safety of the assisted decision-making.

8. The dual-path auxiliary identification system for SBP in cirrhotic ascites according to claim 1, characterized in that, The output and audit module, when calculating the interpretability of the contribution of each variable to the probability value, specifically involves: calling the additive explanation model to calculate the Shapley additive explanation value of each input variable in the standardized dataset relative to the currently calculated probability value; determining the direction of contribution of each input variable to the current patient's risk of spontaneous bacterial peritonitis based on the sign of the calculated Shapley additive explanation value, where a positive value indicates that the variable increases the risk of the disease, and a negative value indicates that the variable reduces the risk of the disease; determining the contribution weight of each input variable to the risk based on the absolute value of the Shapley additive explanation value, and dynamically displaying the contribution distribution of each variable at the individual level in the system interface in the form of a graphical beehive diagram, waterfall diagram, or bar chart, intuitively presenting to clinicians the specific combination of clinical characteristics leading to the current high or low risk and its explanatory mechanism, thereby assisting physicians in reviewing and auditing clinical decisions.

9. The dual-path auxiliary identification system for SBP in cirrhotic ascites according to claim 1, characterized in that, The system also includes a local pathogen linkage module, which is used to: maintain and regularly update a local pathogen spectrum and drug sensitivity knowledge base. The knowledge base records common bacterial species isolated from spontaneous bacterial peritonitis in patients with cirrhosis and ascites in the local region or hospital in recent years, the detection rate of each species, and the drug resistance rate data against commonly used antimicrobial drugs. When the auxiliary identification result or pre-puncture triage result output by the threshold determination module reaches the corresponding risk threshold, the local pathogen linkage module is triggered. It displays the common bacterial species and drug resistance rate information in the knowledge base in the system interface in a descriptive, non-mandatory medication instruction manner, and generates clinical specimen submission prompts. The clinical specimen submission prompts include suggestions to prioritize the collection of specimens for diagnostic peritoneal puncture, ascites culture, and blood culture before adjusting or using antimicrobial drugs, so as to realize the closed-loop linkage between risk identification and local pathogen diagnosis and treatment decision-making, and assist physicians in formulating reasonable empirical anti-infection treatment plans.

10. The dual-path auxiliary identification system for SBP in cirrhotic ascites according to claim 1, characterized in that, The output and audit module also includes a multi-center dynamic calibration and security audit sub-module. This sub-module is used to: when the system receives validation datasets from different medical institutions, call the logistic calibration algorithm to recalibrate the intercept and slope of the first regression model and the second regression model, generate localized model parameters adapted to specific medical institutions, and assign a unique model version number to each localized model parameter; the multi-center dynamic calibration and security audit sub-module is also used to: each time a probability calculation is performed, structurally encapsulate the currently used model version number, the input standardized dataset, the calculation path, the calculated probability value, the output judgment result, and the identity and timestamp of the currently operating physician, and write them into an immutable system audit log database for subsequent medical quality control, model performance auditing, and traceability query of historical diagnostic auxiliary records, ensuring the security and auditability of medical data.

Images