A postoperative biliary fistula risk prediction system for hepatobiliary surgery patients

Abstract

This invention discloses a postoperative bile fistula risk prediction system for hepatobiliary surgery patients, belonging to the field of bile fistula risk data processing technology. The system includes an evidence-driven variable acquisition module, a standardization judgment module, a time-series feature generation module, a feature screening module, and a dynamic prediction module. It performs time-series reconstruction and evolutionary feature extraction for dynamic monitoring indicators of postoperative bile extravasation. Through a structural-biochemical consistency constraint screening method, it fuses and filters features of structural risk channels, biochemical risk channels, and time-series evolution channels. Based on core predictive feature data, it constructs a bile fistula risk prediction model and dynamically predicts the probability of bile fistula occurrence at different postoperative time points through a time-series rolling update mechanism, outputting risk levels and early warning results. This invention enables early identification, dynamic assessment, and individualized intervention support for bile fistula risk, improving the accuracy of prediction results and its clinical application value.

Description

Technical Field

[0001] This invention belongs to the field of bile fistula risk data processing technology, specifically a system for predicting the risk of postoperative bile fistula in hepatobiliary surgery patients. Background Technology

[0002] A postoperative bile fistula risk prediction system for hepatobiliary surgery patients is an intelligent analysis system that uses multi-source perioperative data as input. Through standardized label construction, temporal feature extraction, cross-channel consistency feature screening, and dynamic prediction model calculation, it continuously assesses the probability of bile fistula occurrence at different postoperative time points. The core function of this system is to transform the traditional static judgment method based on a single time point into a dynamic evolution modeling process based on time series. By continuously accessing postoperative monitoring data and updating the risk status, it achieves early identification, trend tracking, and graded early warning of bile fistula risk, providing quantitative data reference for postoperative risk review, trend observation, and graded management.

[0003] However, existing methods for assessing the risk of bile leakage mostly rely on single-time-point indicators, such as the ratio of bilirubin in drainage fluid to serum bilirubin on postoperative day 3, and make judgments by setting fixed thresholds. While these methods have some clinical guiding significance, they still have the following shortcomings: First, they do not uniformly model multi-source data in the perioperative period, and the information before, during, and after surgery is fragmented, making it difficult to form a complete risk chain; second, they lack continuous dynamic modeling of the bile leakage process, and rely solely on static threshold judgments, which cannot reflect the trend of risk evolution; third, feature selection is mostly based on statistical correlation or model importance, lacking constraints on the medical logical consistency between structural damage factors and biochemical responses, resulting in insufficient model stability and interpretability; fourth, existing systems mostly output single assessment results, lacking a dynamic update mechanism based on new data, making it difficult to meet the needs of continuous clinical monitoring and dynamic intervention.

[0004] Therefore, there is an urgent need for a method and system for predicting bile leakage risk that can integrate multi-source perioperative data, characterize the dynamic evolution of bile leakage, and have the ability to constrain feature consistency and dynamically update risk, so as to improve prediction accuracy and enhance clinical application value. Summary of the Invention

[0005] In view of the above situation and to overcome the shortcomings of the prior art, the present invention provides a postoperative bile fistula risk prediction system for hepatobiliary surgery patients, including an evidence-driven variable acquisition module, a standardized judgment module, a time-series feature generation module, a feature screening module, and a dynamic prediction module.

[0006] The evidence-driven variable acquisition module is used to acquire evidence-driven variables of perioperative bile fistula risk factors. By acquiring evidence-driven variables of perioperative bile fistula risk factors, a basic dataset of bile fistula risk is obtained, and the basic dataset of bile fistula risk is sent to the standardization judgment module and the time-series feature generation module respectively.

[0007] The standardized determination module is used for standardized determination of bile fistula outcome. Through standardized determination of bile fistula outcome, bile fistula outcome label data is obtained and the bile fistula outcome label data is sent to the dynamic prediction module.

[0008] The time-series feature generation module is used to generate bile leakage time-series features. Through the generation of bile leakage time-series features, bile leakage evolution feature data is obtained, and the bile leakage evolution feature data is sent to the feature filtering module.

[0009] The feature filtering module is used for multi-channel feature filtering. Through multi-channel feature filtering, the core predictive feature data of bile fistula is obtained, and the core predictive feature data of bile fistula is sent to the dynamic prediction module.

[0010] The dynamic prediction module is used for dynamic prediction of bile fistula risk, and obtains bile fistula risk assessment data through dynamic prediction of bile fistula risk.

[0011] Furthermore, the perioperative bile fistula risk factor evidence-driven variable collection is used to collect structured perioperative candidate risk factors that are directly or indirectly related to the occurrence, severity, and intervention needs of bile fistula. Specifically, it involves the unified collection and standardization of multi-source perioperative data of hepatobiliary surgery patients. Candidate variables from electronic medical record systems, surgical record systems, and laboratory testing systems are screened based on previous clinical studies related to bile fistula, expert consensus, and the mechanism of bile fistula occurrence. Field mapping, data cleaning, and time alignment are then performed to obtain the basic dataset of bile fistula risk.

[0012] The basic dataset for bile leakage risk includes basic demographic variables, preoperative liver function and biliary status variables, surgical structural damage variables, intraoperative trauma variables, and postoperative bile extravasation dynamic variables.

[0013] The basic demographic variables include age, sex, BMI, ASA classification, and whether or not one has diabetes.

[0014] The preoperative liver function and biliary status variables include albumin, total bilirubin, ALP, AST, WBC, ALBI score, Child-Pugh classification, cirrhosis or liver fibrosis, preoperative biliary stent, cholangitis, bile duct dilatation and history of previous biliary intervention.

[0015] The surgical structural damage variables include surgical approach, extent of liver resection, central liver resection, left liver resection, caudate lobe / segment I resection, number of bile duct ends, whether combined bile duct resection was performed, whether choledochoenterostomy was performed, and the method of bile duct reconstruction.

[0016] The intraoperative trauma variables include operation time, intraoperative blood loss, blood transfusion volume, portosystemic occlusion time, intraoperative bile leak test results, and whether bile leakage was detected.

[0017] The dynamic variables of postoperative bile extravasation include the ratio of bilirubin in the drainage fluid to serum bilirubin on the 3rd day after surgery and thereafter, the trend of drainage volume, the duration of continuous drainage, ascites, bile tumor, whether ERCP is needed, whether ENBD is needed, whether ERBD is needed, whether PTBD is needed, and whether reoperation is needed.

[0018] The basic dataset for bile fistula risk is constructed using patient ID as the primary index, and a structured data table is generated according to patient ID, surgery ID, variable name, variable value, variable unit, collection time, and data source. This enables the subsequent standardization judgment module and time-series feature generation module to access the risk variable data of the same patient at different time points.

[0019] Furthermore, the standardized determination of bile fistula outcome is used to standardize the bile fistula occurrence status based on unified bile fistula determination rules. Specifically, it determines the bile fistula occurrence status based on preset bile fistula occurrence determination conditions, and marks whether puncture drainage, endoscopic intervention, or reoperation is performed by introducing clinical intervention behavior as an auxiliary determination basis. At the same time, it locates the time node of bile fistula occurrence by combining the time dimension. Through the fusion of multi-dimensional determination rules, a unified bile fistula outcome label system is constructed to obtain bile fistula outcome label data.

[0020] The preset criteria for determining the occurrence of bile fistula specifically refer to the occurrence of bile fistula when the bilirubin concentration in the drainage fluid reaches or exceeds three times the serum bilirubin concentration on or after the third day after surgery, or when further interventional treatment or reoperation is required due to bile accumulation or bile peritonitis.

[0021] The bile fistula outcome label data specifically includes bile fistula occurrence status labels, bile fistula severity grading labels, intervention need level labels, and bile fistula occurrence time node marker data.

[0022] Furthermore, the bile leakage time-series feature generation is used to construct bile leakage evolution features for postoperative dynamic monitoring indicators. Specifically, based on the postoperative bile leakage dynamic variables in the bile fistula risk dataset, the time series reconstruction processing is performed on the drainage fluid volume, drainage fluid bilirubin concentration, and serum bilirubin index. The reconstructed time series data is then subjected to trend fitting, difference calculation, and window statistical analysis to extract feature information reflecting the dynamic evolution process of bile leakage and construct bile leakage evolution feature data.

[0023] The process of generating bile extravasation time-series features uses a bile extravasation phase transition feature modeling method for feature enhancement. By jointly modeling the time-series change gradient and stage fluctuation features, the precursor change features of bile fistula are extracted, including the following steps: postoperative dynamic sequence construction, bilirubin abnormal gradient joint feature construction, bile concentration drainage volume feature construction, inflammatory marker time-delay response feature construction, and risk evolution transition feature construction.

[0024] The postoperative dynamic sequence construction is based on the postoperative bile extravasation dynamic variable in the bile fistula risk dataset. Using the patient number as the index and the postoperative time node as the sequence dimension, the sequence of drainage fluid bilirubin concentration, the same-day serum bilirubin concentration, the drainage fluid volume, and the inflammatory marker sequence are constructed.

[0025] The inflammatory indicators include WBC or other postoperative inflammation-related indicators; furthermore, missing time point data are interpolated or marked for missing information to obtain a continuous postoperative dynamic monitoring sequence.

[0026] The bilirubin abnormal gradient joint feature is constructed by constructing a bilirubin abnormal ratio sequence based on the bilirubin concentration sequence of the drainage fluid and the serum bilirubin concentration sequence at the same time. The first difference of the bilirubin abnormal ratio sequence is calculated to obtain the bilirubin change gradient feature. The bilirubin abnormal ratio is used to characterize the degree of abnormal bile enrichment, and the bilirubin change gradient feature is used to characterize the rate of abnormal bile change.

[0027] The bile concentration drainage volume feature is constructed by performing differential calculation based on the drainage volume sequence to obtain the drainage volume change trend feature. The abnormal bilirubin ratio is coupled with the drainage volume to obtain the bile concentration drainage volume feature. The drainage volume change trend feature is used to characterize the drainage release change trend. The bile concentration drainage volume feature is used to simultaneously characterize the degree of abnormal enrichment of bile components and the intensity of drainage release.

[0028] The construction of the inflammatory marker time-delay response feature is based on the bile concentration drainage volume feature and the inflammatory marker sequence. Specifically, the response intensity at different lag times is calculated within a preset lag time window, and the maximum response intensity is selected as the inflammatory marker time-delay response feature. The inflammatory marker time-delay response feature is used to characterize the delayed correlation between the bile leakage process and the inflammatory response.

[0029] The risk evolution transition feature is constructed based on the abnormal bilirubin ratio, bilirubin change gradient features, drainage volume change trend features, bile concentration drainage volume features, and inflammatory marker time-lag response features. A postoperative risk evolution score is constructed, and the risk change amplitude is calculated based on the difference in postoperative risk evolution scores between adjacent time nodes. The risk change is judged according to a preset transition threshold. When the risk change amplitude is greater than the transition threshold, the corresponding time node is marked as a risk transition node, thus obtaining the postoperative risk evolution transition feature.

[0030] The bile extravasation evolution characteristics data specifically include bilirubin change gradient characteristics, drainage volume change trend characteristics, bile concentration drainage volume characteristics, inflammatory marker time-lag response characteristics, and postoperative risk evolution transition characteristics.

[0031] Furthermore, the multi-channel feature screening is used to screen core predictive features of bile fistula. Specifically, the input features are divided into structural risk channel, biochemical risk channel and temporal evolution channel according to their source. The structural risk channel is used to characterize the risk of surgical structural damage, and the biochemical risk channel is used to characterize the risk of postoperative physiological response. Statistical correlation analysis and predictive contribution evaluation are performed on the features in the three channels respectively. Redundant features and low-contribution features are eliminated. Consistency constraints are applied to the features in the three channels for screening. By analyzing the matching relationship between structural risk features and biochemical risk features in terms of predictive contribution direction and predictive contribution degree, features with significant inconsistencies or weak correlations are screened out or downweighted to obtain the core predictive feature data of bile fistula.

[0032] The multi-channel feature screening process adopts the structural-biochemical consistency constraint screening method for screening and optimization, including the following steps: multi-channel feature attribution division, intra-channel prediction contribution evaluation, structural-biochemical mutual verification relationship construction, consistency constraint score calculation, feature screening and core prediction feature fusion output;

[0033] The multi-channel feature attribution classification, based on the bile extravasation evolution feature data and the bile fistula risk basic dataset, divides the input features into structural risk channels, biochemical risk channels, and temporal evolution channels; wherein, the structural risk channel is used to characterize the risk of surgical structural damage, the biochemical risk channel is used to characterize the risk of postoperative physiological response, and the temporal evolution channel is used to characterize the dynamic evolution risk of bile extravasation.

[0034] The intra-channel prediction contribution assessment evaluates the predictive contribution of features in the structural risk channel, biochemical risk channel, and temporal evolution channel, respectively, to obtain the intra-channel contribution value of each feature. The intra-channel contribution value is composed of the correlation between the feature and the bile fistula occurrence status label, the importance score in the initial prediction model, and the stability score in the cross-validation process.

[0035] The construction of the structural-biochemical mutual verification relationship involves constructing a structural-biochemical mutual verification relationship for structural risk features in the structural risk channel and biochemical risk features in the biochemical risk channel. This relationship is used to determine whether the structural damage risk is reflected in the postoperative biochemical response. For any structural risk feature and biochemical risk feature, the mutual verification strength is determined by the consistency of prediction direction, feature correlation, and contribution value within the channel.

[0036] The consistency constraint score calculation is based on the structural-biochemical mutual verification relationship, and the consistency constraint score is calculated for the structural risk feature, biochemical risk feature and time series evolution feature respectively.

[0037] The consistency constraint score of structural risk features and biochemical risk features is obtained by aggregating their positive mutual evidence strength with all features in another channel;

[0038] The consistency constraint score of the temporal evolution characteristics is calculated based on its correlation with the comprehensive response of structural risk channels and biochemical risk channels;

[0039] The feature filtering is based on the contribution value within the channel and the consistency constraint score to filter or reduce the weight of the input features, thereby obtaining the filtered feature data.

[0040] When the overall screening score is lower than the preset screening threshold, the feature is removed; when the contribution value within the channel is high but the consistency constraint score is lower than the preset consistency threshold, the feature is downweighted to reduce its impact on subsequent prediction models.

[0041] The core predictive feature fusion output, based on the filtered feature data, generates structural risk feature subsets, biochemical risk feature subsets, and temporal evolution feature subsets respectively. The structural risk feature subsets, biochemical risk feature subsets, and temporal evolution feature subsets are then fused according to patient number and postoperative time node to obtain the fused feature subset after channel consistency screening.

[0042] The core predictive features of bile fistula specifically include a subset of structural risk features, a subset of biochemical risk features, a subset of temporal evolution features, and a subset of fused features after channel consistency screening.

[0043] Furthermore, the dynamic prediction of bile fistula risk constructs a bile fistula risk prediction model based on the core prediction feature data of bile fistula, and dynamically calculates the probability of bile fistula occurrence at different time points after surgery. By introducing a time series update mechanism, the newly added monitoring data is rolled and updated to generate risk change trends. The prediction results are graded based on a preset risk threshold, and the risk level and warning status are output to obtain bile fistula risk assessment result data with time dynamics.

[0044] The data on the risk assessment of bile fistula specifically includes the probability of bile fistula occurrence, risk level classification results, risk change trends, and early warning status indicators.

[0045] The beneficial effects achieved by adopting the above solution are as follows:

[0046] (1) In view of the technical problems in the existing methods for predicting the risk of postoperative bile leakage in hepatobiliary surgery patients, such as the scattered sources of perioperative variables, the inconsistent criteria for determining the bile leakage outcome label, and the difficulty in jointly modeling preoperative and intraoperative risk factors and postoperative dynamic monitoring indicators under the same time reference, this solution creatively adopts a systematic processing flow of evidence-driven variable collection, standardized determination of bile leakage outcome, generation of bile leakage time sequence characteristics, multi-channel feature screening, and dynamic prediction of bile leakage risk. By taking the end time of surgery as the zero point, the preoperative examination data, intraoperative record data, and postoperative test data are time-aligned, and the bile leakage occurrence status, severity classification, intervention requirement level, and occurrence time node are uniformly coded, so that the structural damage information, biochemical test information, and intervention outcome information of the same patient can form a standardized data chain that is trainable and verifiable, thereby improving the consistency, completeness, and reproducibility of the bile leakage risk prediction data basis.

[0047] (2) In view of the technical problem that existing methods for constructing dynamic characteristics of bile leakage only count the bilirubin ratio or drainage volume of drainage fluid at a single time point, which is difficult to reflect the stage-by-stage changes of postoperative bile leakage from slow accumulation to rapid deterioration, this scheme creatively adopts the bile leakage phase transition characteristic modeling method; by constructing the drainage fluid bilirubin concentration sequence, the concurrent serum bilirubin concentration sequence, the drainage fluid volume sequence and the inflammatory index sequence, and further calculating the abnormal bilirubin ratio, bilirubin change gradient, drainage fluid volume change trend, bile concentration drainage volume characteristics and inflammatory index time delay response characteristics, the model can simultaneously characterize the process of abnormal enrichment of bile components, changes in drainage release intensity and delayed appearance of inflammatory response;

[0048] (3) In view of the technical problem that existing methods for screening predictive features of bile fistula only rely on correlation or model importance for single-dimensional screening, and fail to distinguish the medical logical consistency between surgical structural damage features, biochemical response features and temporal evolution features, resulting in the model retaining accidental related variables or weakening key mutual verification variables, this solution creatively adopts the structural-biochemical consistency constraint screening method. By calculating the intra-channel contribution value of each feature, the mutual verification strength between structural risk features and biochemical risk features, and the consistency constraint scores of structural risk channels, biochemical risk channels and temporal evolution channels, the feature screening process no longer relies solely on statistical correlation, but simultaneously considers prediction direction, cross-channel correlation and contribution stability. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the structure of a postoperative bile fistula risk prediction system for hepatobiliary surgery patients provided by the present invention;

[0050] Figure 2 A flowchart illustrating the steps performed by the system provided for this invention;

[0051] Figure 3 A flowchart illustrating the steps performed by the standardized judgment module;

[0052] Figure 4 A flowchart illustrating the steps performed by the time-series feature generation module;

[0053] Figure 5 This is a flowchart illustrating the steps performed by the feature filtering module.

[0054] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0055] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0056] Example 1, see Figure 1 The technical solution adopted by the present invention is as follows: The present invention provides a postoperative bile fistula risk prediction system for hepatobiliary surgery patients, including an evidence-driven variable acquisition module, a standardized judgment module, a time-series feature generation module, a feature screening module and a dynamic prediction module;

[0057] The evidence-driven variable acquisition module is used to acquire evidence-driven variables of perioperative bile fistula risk factors. By acquiring evidence-driven variables of perioperative bile fistula risk factors, a basic dataset of bile fistula risk is obtained, and the basic dataset of bile fistula risk is sent to the standardization judgment module and the time-series feature generation module respectively.

[0058] The standardized determination module is used for standardized determination of bile fistula outcome. Through standardized determination of bile fistula outcome, bile fistula outcome label data is obtained and the bile fistula outcome label data is sent to the dynamic prediction module.

[0059] The time-series feature generation module is used to generate bile leakage time-series features. Through the generation of bile leakage time-series features, bile leakage evolution feature data is obtained, and the bile leakage evolution feature data is sent to the feature filtering module.

[0060] The feature filtering module is used for multi-channel feature filtering. Through multi-channel feature filtering, the core predictive feature data of bile fistula is obtained, and the core predictive feature data of bile fistula is sent to the dynamic prediction module.

[0061] The dynamic prediction module is used for dynamic prediction of bile fistula risk, and obtains bile fistula risk assessment data through dynamic prediction of bile fistula risk.

[0062] By performing the above operations, this solution addresses the technical problems in existing methods for predicting postoperative bile leakage risk in hepatobiliary surgery patients, such as scattered perioperative variable sources, inconsistent criteria for determining bile leakage outcomes, and the difficulty in jointly modeling preoperative and intraoperative risk factors and postoperative dynamic monitoring indicators on the same time frame. This solution creatively employs a systematic processing flow for evidence-driven variable collection, standardized determination of bile leakage outcomes, generation of bile extravasation time-series characteristics, multi-channel feature screening, and dynamic prediction of bile leakage risk. By using the end of surgery as the zero point, preoperative examination data, intraoperative records, and postoperative test data are time-aligned, and the occurrence status, severity classification, intervention requirement level, and occurrence time of bile leakage are uniformly coded. This allows the structural damage information, biochemical test information, and intervention outcome information of the same patient to form a trainable and verifiable standardized data chain, thereby improving the consistency, completeness, and reproducibility of the bile leakage risk prediction data foundation.

[0063] Example 2, this example is based on the above example, see reference. Figure 1 , Figure 2The perioperative bile fistula risk factor evidence-driven variable collection is used to collect structured perioperative candidate risk factors that are directly or indirectly related to the occurrence, severity and intervention needs of bile fistula. Specifically, it involves the unified collection and standardization of multi-source perioperative data of hepatobiliary surgery patients. Candidate variables from electronic medical record system, surgical record system and laboratory testing system are screened based on previous clinical research on bile fistula, expert consensus and bile fistula occurrence mechanism. Field mapping, data cleaning and time alignment are performed to obtain the basic dataset of bile fistula risk.

[0064] The basic dataset for bile leakage risk includes basic demographic variables, preoperative liver function and biliary status variables, surgical structural damage variables, intraoperative trauma variables, and postoperative bile extravasation dynamic variables.

[0065] The basic demographic variables include age, sex, BMI, ASA classification, and whether or not one has diabetes.

[0066] The preoperative liver function and biliary status variables include albumin, total bilirubin, ALP, AST, WBC, ALBI score, Child-Pugh classification, cirrhosis or liver fibrosis, preoperative biliary stent, cholangitis, bile duct dilatation and history of previous biliary intervention.

[0067] The surgical structural damage variables include surgical approach, extent of liver resection, central liver resection, left liver resection, caudate lobe / segment I resection, number of bile duct ends, whether combined bile duct resection was performed, whether choledochoenterostomy was performed, and the method of bile duct reconstruction.

[0068] The intraoperative trauma variables include operation time, intraoperative blood loss, blood transfusion volume, portosystemic occlusion time, intraoperative bile leak test results, and whether bile leakage was detected.

[0069] The dynamic variables of postoperative bile extravasation include the ratio of bilirubin in the drainage fluid to serum bilirubin on the 3rd day after surgery and thereafter, the trend of drainage volume, the duration of continuous drainage, ascites, bile tumor, whether ERCP is needed, whether ENBD is needed, whether ERBD is needed, whether PTBD is needed, and whether reoperation is needed.

[0070] Preferably, the field mapping process specifically involves mapping variables from different data sources that have the same clinical meaning but different field names to a unified field name. For example, "intraoperative blood loss" and "bleeding" are uniformly mapped to intraoperative bleeding, and "TBIL" and "total bilirubin" are uniformly mapped to total bilirubin. The field type, unit of measurement, and allowed value range are configured for each unified field.

[0071] The data cleaning process specifically involves unifying units, identifying outliers, and marking missing values ​​for continuous variables; and encoding and converting categorical variables and removing invalid values. Among these, albumin, total bilirubin, ALP, AST, WBC, bilirubin in drainage fluid, and drainage volume are treated as continuous variables, while surgical approach, extent of liver resection, method of bile duct reconstruction, and intervention method are treated as categorical variables.

[0072] For missing values, continuous variables with a missing proportion below a preset proportion threshold can be filled by interpolation using the median of patients in the same group or adjacent time points. For variables with a missing proportion above a preset proportion threshold, a missing value field can be generated. Unidentified categorical variables can be set to unknown category to avoid sample removal due to missing data.

[0073] The time alignment process specifically involves using the end time of the surgery as the zero point, mapping preoperative examination data to the preoperative time window, mapping intraoperative recording data to the intraoperative time window, and aggregating postoperative test data according to postoperative day 1, day 3, day 5, and day 7 to obtain the postoperative time series; when multiple test results exist at the same time point, the test result closest to that time point is selected first.

[0074] More preferably, the dynamic variables of postoperative bile extravasation can be aggregated, specifically by associating and storing the bilirubin concentration of the drainage fluid, the serum bilirubin concentration at the same time point, and the drainage fluid volume according to the same postoperative time point, and calculating the ratio of bilirubin in the drainage fluid to serum bilirubin at the same time point as one of the indicators for dynamic monitoring of bile extravasation.

[0075] The basic dataset for bile fistula risk is constructed using patient ID as the primary index, and a structured data table is generated according to patient ID, surgery ID, variable name, variable value, variable unit, collection time, and data source. This enables the subsequent standardization judgment module and time-series feature generation module to access the risk variable data of the same patient at different time points.

[0076] Example 3, this example is based on the above examples, see reference. Figure 1 , Figure 2 and Figure 3 The standardized determination of bile fistula outcome is used to standardize the bile fistula occurrence status according to unified bile fistula determination rules. Specifically, it determines the bile fistula occurrence status based on preset bile fistula occurrence determination conditions, and marks whether puncture drainage, endoscopic intervention or reoperation is performed by introducing clinical intervention behavior as an auxiliary determination basis. At the same time, it locates the time node of bile fistula occurrence by combining the time dimension. Through the fusion of multi-dimensional determination rules, a unified bile fistula outcome label system is constructed to obtain bile fistula outcome label data.

[0077] The preset criteria for determining the occurrence of bile fistula specifically refer to the occurrence of bile fistula when the bilirubin concentration in the drainage fluid reaches or exceeds three times the serum bilirubin concentration on or after the third day after surgery, or when further interventional treatment or reoperation is required due to bile accumulation or bile peritonitis.

[0078] Preferably, the standardized determination of bile fistula outcome includes the following steps: candidate determination data extraction, construction of basic determination indicators, preliminary determination of bile fistula occurrence, intervention behavior auxiliary determination, time node positioning, severity grading determination, intervention need level marking, and multi-dimensional label fusion construction.

[0079] The candidate decision data extraction is based on the bile fistula risk dataset, extracting key variable data for determining the outcome of bile fistula. The key variable data includes the bilirubin concentration of drainage fluid at each postoperative time point, the serum bilirubin concentration at the same time, the drainage fluid volume, the results of abdominal imaging examinations, records of biliary peritonitis, and clinical intervention record data. Among them, the clinical intervention record data includes whether puncture drainage was performed, whether ERCP was performed, whether ENBD was performed, whether ERBD was performed, whether PTBD was performed, and whether reoperation was performed.

[0080] The basic judgment index is constructed based on postoperative time series data. The ratio of bilirubin concentration in drainage fluid to serum bilirubin concentration at each time point from the 3rd day after surgery is calculated to obtain a bilirubin ratio sequence. When multiple test values ​​exist at the same time point, the test result closest to that time point is selected for calculation. When data is missing at a certain time point, the time point is marked as missing or data from adjacent time points is interpolated to fill in the missing data.

[0081] In one implementation, the specific formula for calculating the bilirubin ratio at the t-th time point after surgery is as follows:

[0082] ;

[0083] In the formula, It is the bilirubin ratio at the t-th time point after surgery. It represents the bilirubin concentration in the drainage fluid at time point t. This refers to the serum bilirubin concentration during the same period. It is a minimal smoothing constant used to avoid abnormal amplification of the ratio when the serum bilirubin concentration is too low during the same period. Preferably, Desirable ;

[0084] The initial determination of bile fistula occurrence is carried out for each patient on the 3rd day after surgery and at each time point. When the bilirubin ratio corresponding to any time point is greater than or equal to 3, the time point is marked as a candidate time point for bile fistula determination and the patient is marked as a candidate state for bile fistula occurrence. If the above conditions are not met at all time points, the patient is marked as a non-candidate state for bile fistula and enters the intervention behavior auxiliary determination.

[0085] The intervention behavior-assisted determination is performed in the non-cholesterol leakage candidate state. When the patient has a record of biliary peritonitis or abdominal imaging examinations indicate bile accumulation, the clinical intervention record data is retrieved. If there is a clinical intervention behavior for bile leakage, the time point of the first implementation of the clinical intervention behavior is marked as the time point of bile leakage occurrence, and the patient's bile leakage occurrence status is marked as "occurred". If there is no clinical intervention behavior, the patient's bile leakage occurrence status is marked as "not occurred".

[0086] In the specific implementation process, when there are records of puncture drainage, ERCP, ENBD, ERBD, PTBD, or reoperation, the indications, disease course descriptions, or operation record summaries of the corresponding intervention records are further read; if the intervention record contains descriptions related to bile accumulation, bile leakage, bile peritonitis, bile tumor, or biliary decompression, then the intervention is marked as a bile leakage-related intervention; if several pre-records are only related to infection control, general ascites management, or other non-bile leakage causes, they are not used as the basis for confirming the occurrence of bile fistula.

[0087] If there are interventions related to bile leakage, the time point when clear evidence of bile leakage first appears will be marked as the time point of bile fistula occurrence; when the time when clear evidence of bile leakage cannot be determined, the time point when the first intervention related to bile leakage is implemented will be taken as the time point of bile fistula occurrence; if there are no interventions related to bile leakage and the bilirubin ratio evidence item does not meet the preset judgment criteria, the patient's bile fistula occurrence status will be marked as not occurring.

[0088] The time node positioning involves determining the time node of bile fistula occurrence for patients whose bile fistula occurrence status label is "occurred". Specifically, if there are candidate time nodes for bile fistula determination, the earliest appearing candidate time node is selected as the bile fistula occurrence time node, resulting in bile fistula occurrence time node marker data. If the occurrence is determined through intervention behavior, the time node of the first implementation of the clinical intervention behavior is selected as the bile fistula occurrence time node. For patients whose bile fistula occurrence status label is "not occurred", their bile fistula occurrence time node marker data is marked as empty.

[0089] In the specific implementation process, if there are candidate time points for determining bile fistula, the earliest candidate time point that appears and passes data quality control is selected as the time point for bile fistula occurrence; if there is no effective bilirubin ratio confirmation time, but there are time points where imaging examinations suggest bile accumulation, bile tumors, or medical records clearly record bile peritonitis, the earliest time point in the corresponding record is selected as the time point for bile fistula occurrence; if neither of the above two types of time points can be determined, but the bile fistula occurrence status is determined to be "occurring" through intervention behavior, the time point of the first implementation of bile leakage-related intervention behavior is selected as the time point for bile fistula occurrence.

[0090] For patients whose bile fistula occurrence status is marked as "not occurred", their bile fistula occurrence time marker data is set to null; for patients whose bile fistula occurrence time marker data is marked as "to be reviewed" due to conflicts between different judgment criteria, their bile fistula occurrence time marker data is set to "to be reviewed" and the type of evidence that triggered the conflict is recorded for subsequent manual review or training sample screening.

[0091] The severity grading is performed under the premise that the bile fistula occurrence status label is "occurring". The severity of the bile fistula is graded based on the intervention type and intensity in the clinical intervention record data to obtain a bile fistula severity grading label. Specifically, if only drainage is extended or no invasive intervention is performed, it is marked as low grade; if endoscopic or percutaneous drainage intervention is performed, it is marked as medium grade; if reoperation is required or serious complications occur, it is marked as high grade.

[0092] The intervention need level label is based on the number of interventions and intervention methods in the clinical intervention record data to classify the patient's intervention needs into levels, resulting in intervention need level labels. Specifically, patients who have not undergone clinical intervention or have only received routine treatment are labeled as having a low need level; patients who have undergone a single clinical intervention are labeled as having a medium need level; and patients who have undergone multiple clinical interventions or have undergone reoperation are labeled as having a high need level.

[0093] The multidimensional label fusion construction involves uniformly encoding the bile fistula occurrence status label, severity grading label, intervention requirement level label, and bile fistula occurrence time node marker data to generate bile fistula outcome label data. The bile fistula occurrence status label is represented in binary form, the severity grading label and intervention requirement level label are represented in multi-category form, and the bile fistula occurrence time node marker data is represented in time index or timestamp form. This forms a standardized label system that corresponds one-to-one with the bile fistula risk baseline dataset, used for the subsequent training and evaluation of the dynamic prediction model.

[0094] The bile fistula outcome label data specifically includes bile fistula occurrence status labels, bile fistula severity grading labels, intervention need level labels, and bile fistula occurrence time node marker data.

[0095] Example 4, this example is based on the above examples, see below. Figure 1 , Figure 2 and Figure 4 The bile leakage time-series feature generation is used to construct bile leakage evolution features for postoperative dynamic monitoring indicators. Specifically, based on the postoperative bile leakage dynamic variables in the bile fistula risk dataset, the time series reconstruction processing of drainage fluid volume, drainage fluid bilirubin concentration and serum bilirubin index is performed, and the reconstructed time series data is subjected to trend fitting, difference calculation and window statistical analysis to extract feature information reflecting the dynamic evolution process of bile leakage and construct bile leakage evolution feature data.

[0096] Preferably, the bile leakage time-series feature generation process employs a bile leakage phase transition feature modeling method for feature enhancement processing. By jointly modeling the time-series change gradient and stage fluctuation features, the precursor change features of bile fistula are extracted, including the following steps: postoperative dynamic sequence construction, bilirubin abnormal gradient joint feature construction, bile concentration drainage volume feature construction, inflammatory marker time-delay response feature construction, and risk evolution transition feature construction.

[0097] The postoperative dynamic sequence construction is based on the postoperative bile extravasation dynamic variable in the bile fistula risk dataset. Using the patient number as the index and the postoperative time node as the sequence dimension, the sequence of drainage fluid bilirubin concentration, the same-day serum bilirubin concentration, the drainage fluid volume, and the inflammatory marker sequence are constructed.

[0098] The inflammatory indicators include WBC or other postoperative inflammation-related indicators; furthermore, missing time point data are interpolated or marked for missing information to obtain a continuous postoperative dynamic monitoring sequence.

[0099] The bilirubin abnormal gradient joint feature is constructed by constructing a bilirubin abnormal ratio sequence based on the bilirubin concentration sequence of the drainage fluid and the serum bilirubin concentration sequence at the same time. The first difference of the bilirubin abnormal ratio sequence is calculated to obtain the bilirubin change gradient feature. The bilirubin abnormal ratio is used to characterize the degree of abnormal bile enrichment, and the bilirubin change gradient feature is used to characterize the rate of abnormal bile change.

[0100] In this embodiment, specifically, the set of effective postoperative time points can be set as follows: ,in It is the first A valid postoperative time point is set. Indicates time node The bilirubin concentration in the drainage fluid, Indicates time node The serum bilirubin concentration during the same period, The minimum smoothing constant is preferably 0.1 to avoid the denominator being zero or close to zero; then the time nodes... The corresponding abnormal bilirubin ratio The calculation formula is:

[0101] ;

[0102] Furthermore, the bilirubin change gradient characteristics It is calculated by the difference in the abnormal bilirubin ratio between adjacent valid time points, and the calculation formula is as follows:

[0103] ;

[0104] In the formula, It is the abnormal bilirubin ratio at the current effective time point. It is the abnormal bilirubin ratio at the previous effective time point. It is the time interval between adjacent effective time nodes, and the abnormal bilirubin ratio is used to characterize the degree of abnormal bile enrichment.

[0105] The bile concentration drainage volume feature is constructed by performing differential calculation based on the drainage volume sequence to obtain the drainage volume change trend feature. The abnormal bilirubin ratio is coupled with the drainage volume to obtain the bile concentration drainage volume feature. The drainage volume change trend feature is used to characterize the drainage release change trend. The bile concentration drainage volume feature is used to simultaneously characterize the degree of abnormal enrichment of bile components and the intensity of drainage release.

[0106] Specifically, set Indicates time node The characteristics of traffic generation trends The calculation formula is:

[0107] ;

[0108] In the formula, It is a time node Traffic generation It is the drainage volume at the previous effective time point; the drainage volume change trend feature is used to characterize the trend of postoperative drainage volume release intensity changing over time;

[0109] Furthermore, the abnormal bilirubin ratio was coupled with the drainage volume to calculate the bile concentration drainage volume characteristics. The calculation formula is:

[0110] ;

[0111] In the formula, It is the abnormal bilirubin ratio at the current effective time point. It is a time node Traffic generation Overall, it is used to reduce the excessive amplification effect of extreme drainage volume on coupling characteristics;

[0112] The construction of the inflammatory marker time-delay response feature is based on the bile concentration drainage volume feature and the inflammatory marker sequence. Specifically, the response intensity at different lag times is calculated within a preset lag time window, and the maximum response intensity is selected as the inflammatory marker time-delay response feature. The inflammatory marker time-delay response feature is used to characterize the delayed correlation between the bile leakage process and the inflammatory response.

[0113] Specifically, set Indicates time node The inflammatory marker values, wherein the inflammatory markers include at least one of WBC, CRP, or PCT; Let... For the current time node The previously preset lag time window, preferably selecting 1 to 2 effective time points before the current time point, will reveal the time-lag response characteristics of inflammatory indicators. The calculation formula is:

[0114] ;

[0115] In the formula, It is a lagging time node Corresponding bile concentration drainage characteristics. This is the inflammation marker value at the current time point. These are inflammatory marker values ​​at a lagging time point. The overall value is used to extract the positive increase of inflammatory markers; when the inflammatory markers decrease or do not increase, this value is set to 0; by calculating the response intensity at different lag times within a preset lag time window and selecting the maximum response intensity as the time-lag response feature of the inflammatory markers, the delayed correlation between bile extravasation and inflammatory response can be characterized.

[0116] The risk evolution transition feature is constructed based on the abnormal bilirubin ratio, bilirubin change gradient features, drainage volume change trend features, bile concentration drainage volume features, and inflammatory marker time-lag response features. A postoperative risk evolution score is constructed, and the risk change amplitude is calculated based on the difference in postoperative risk evolution scores between adjacent time nodes. The risk change is judged according to a preset transition threshold. When the risk change amplitude is greater than the transition threshold, the corresponding time node is marked as a risk transition node, thus obtaining the postoperative risk evolution transition feature.

[0117] In one specific implementation, it is possible to , , , and Perform max-min normalization to obtain the following results: , , , and And calculate the postoperative risk evolution score. The calculation formula is:

[0118] ;

[0119] Among them, the abnormal bilirubin ratio and bile concentration drainage volume characteristics are used to characterize the degree of abnormal bile enrichment and its coupling state with the intensity of drainage release, so the weights are set relatively high; the bilirubin change gradient characteristics, drainage volume change trend characteristics and inflammatory marker time-lag response characteristics are used to reflect the speed of risk change, drainage release trend and delayed inflammatory response, and to correct the risk evolution score.

[0120] Furthermore, the magnitude of risk change can be calculated based on the difference in postoperative risk evolution scores at adjacent time points. The calculation formula is:

[0121] ;

[0122] when At that time, the corresponding time node is marked as a risk transition node, and the postoperative risk evolution transition characteristics are obtained; among them, The preset transition threshold is preferably set to 0.20; the postoperative risk evolution transition characteristics include the risk change magnitude, risk transition node markers, and corresponding postoperative time nodes;

[0123] The bile extravasation evolution characteristics data specifically include bilirubin change gradient characteristics, drainage volume change trend characteristics, bile concentration drainage volume characteristics, inflammatory marker time-lag response characteristics, and postoperative risk evolution transition characteristics.

[0124] By performing the above operations, this solution addresses the technical problem in existing methods for constructing dynamic characteristics of bile extravasation that only statistically analyze the bilirubin ratio or drainage volume of drainage fluid at a single time point, making it difficult to reflect the phased changes in postoperative bile extravasation from slow accumulation to rapid deterioration. This solution creatively employs a bile extravasation phase transition characteristic modeling method. By constructing drainage fluid bilirubin concentration sequences, concurrent serum bilirubin concentration sequences, drainage fluid volume sequences, and inflammatory marker sequences, and further calculating abnormal bilirubin ratios, bilirubin change gradients, drainage fluid volume trends, bile concentration and drainage volume characteristics, and inflammatory marker time-delay response characteristics, the model can simultaneously characterize the abnormal enrichment of bile components, changes in drainage release intensity, and the delayed appearance of inflammatory responses.

[0125] For example, when the patient's bilirubin ratio has not reached the three-fold threshold on the third day after surgery, but the bilirubin ratio gradient increases significantly from the third to the fifth day, the drainage volume decreases insufficiently, and the WBC increases within the lag window, the method can generate a higher risk evolution score and mark risk transition nodes, thereby improving the ability to identify the precursor stage of bile fistula.

[0126] Example 5, this example is based on the above examples, see below. Figure 1 , Figure 2 and Figure 5 The multi-channel feature screening is used to screen the core predictive features of bile fistula. Specifically, the input features are divided into structural risk channel, biochemical risk channel and temporal evolution channel according to their source. The structural risk channel is used to characterize the risk of surgical structural damage, and the biochemical risk channel is used to characterize the risk of postoperative physiological response. Statistical correlation analysis and predictive contribution evaluation are performed on the features in the three channels respectively. Redundant features and low-contribution features are eliminated. Consistency constraints are applied to the features in the three channels. By analyzing the matching relationship between structural risk features and biochemical risk features in terms of predictive contribution direction and predictive contribution degree, features with significant inconsistencies or weak correlations are screened out or downweighted to obtain the core predictive feature data of bile fistula.

[0127] The multi-channel feature screening process adopts the structural-biochemical consistency constraint screening method for screening and optimization, including the following steps: multi-channel feature attribution division, intra-channel prediction contribution evaluation, structural-biochemical mutual verification relationship construction, consistency constraint score calculation, feature screening and core prediction feature fusion output;

[0128] The multi-channel feature attribution classification, based on the bile extravasation evolution feature data and the bile fistula risk basic dataset, divides the input features into structural risk channels, biochemical risk channels, and temporal evolution channels; wherein, the structural risk channel is used to characterize the risk of surgical structural damage, the biochemical risk channel is used to characterize the risk of postoperative physiological response, and the temporal evolution channel is used to characterize the dynamic evolution risk of bile extravasation.

[0129] The intra-channel prediction contribution assessment evaluates the predictive contribution of features in the structural risk channel, biochemical risk channel, and temporal evolution channel, respectively, to obtain the intra-channel contribution value of each feature. The intra-channel contribution value is composed of the correlation between the feature and the bile fistula occurrence status label, the importance score in the initial prediction model, and the stability score in the cross-validation process.

[0130] For any feature, the formula for calculating the corresponding intra-channel contribution value is:

[0131] ;

[0132] In the formula, It is a feature The contribution value within the channel, It is the label relevance weight. It is a feature The correlation between the status label y and the occurrence of bile fistula is preferably calculated using Pearson correlation coefficient or point-bivariate correlation coefficient. It is the importance weight. It is a feature In the initial prediction model, the importance score can be based on feature importance based on information gain or number of splits when using a random forest or XGBoost model; when using a logistic regression model, the importance score can be based on the absolute value of the standardized regression coefficients. It is a contribution to stability weight. It is a feature The contribution stability during cross-validation is specifically the percentage of times the feature enters the set of the top N important features in K-fold cross-validation.

[0133] Preferably, , , satisfy In one embodiment, it is preferable to , , ;

[0134] Preferably, the initial prediction model is at least one of a random forest model, an XGBoost model, or a logistic regression model;

[0135] The construction of the structural-biochemical mutual verification relationship involves establishing a mutual verification relationship between structural risk features in the structural risk channel and biochemical risk features in the biochemical risk channel. This relationship is used to determine whether the structural damage risk is reflected in the postoperative biochemical response. For any structural risk feature and biochemical risk feature, the mutual verification strength is determined by the consistency of prediction direction, feature correlation, and contribution value within the channel. The calculation formula is as follows:

[0136] ;

[0137] In the formula, It is a structural risk characteristic Biochemical risk characteristics The strength of mutual verification between them and These are the signs of the regression coefficients for the corresponding features in the initial prediction model, used to characterize the prediction direction. To represent the correlation between structural risk characteristics and biochemical risk characteristics, the Pearson correlation coefficient is preferred. This is used to enhance the weighting of high-contribution feature pairs, where... It is the channel contribution value of structural risk characteristics. It is the in-channel contribution value of biochemical risk characteristics, which strengthens the mutual verification relationship between high-importance characteristics;

[0138] In one embodiment, it is possible to Set a minimum effective threshold, such as 0.1. When the value is below this threshold, it is considered a weak correlation and its mutual evidence strength is recorded as 0, in order to reduce the impact of noise correlation.

[0139] The consistency constraint score calculation is based on the structural-biochemical mutual verification relationship, and the consistency constraint score is calculated for the structural risk feature, biochemical risk feature and time series evolution feature respectively.

[0140] Preferably, the consistency constraint score of structural risk features and biochemical risk features is obtained by aggregating their positive cross-validation strength with all features in another channel. The formula for calculating the consistency constraint score of the structural risk features is as follows:

[0141] ;

[0142] In the formula, It is a consistency constraint score of structural risk characteristics. 'b' represents the number of features in the biohazard risk channel, and 'b' represents the feature index in the biohazard risk channel. Used to retain only the mutual verification relationship with consistent prediction direction, and to eliminate the negative impact of features with conflicting directions;

[0143] The formula for calculating the consistency constraint score of the biochemical risk characteristics is as follows:

[0144] ;

[0145] In the formula, It is a consistency constraint score for biochemical risk characteristics. 'a' represents the number of features in the structural risk channel, and 'a' represents the feature index in the structural risk channel.

[0146] The consistency constraint score of the temporal evolution characteristics is calculated based on its correlation with the combined response of structural risk channels and biochemical risk channels, and the calculation formula is as follows:

[0147] ;

[0148] In the formula, It is a consistency constraint score for temporal evolution characteristics. It is the structural risk channel constraint weight. The correlation between temporal evolution characteristics and the comprehensive response of structural risk channels can be obtained by weighted summation of structural risk channel characteristics. It is the biochemical risk channel constraint weight. It is a comprehensive response to biochemical risk pathways;

[0149] In one embodiment, it is preferable , To appropriately increase the weight of biochemical responses in temporal consistency;

[0150] The feature filtering is based on the contribution value within the channel and the consistency constraint score to filter or reduce the weight of the input features, thereby obtaining the filtered feature data.

[0151] When the overall screening score is lower than the preset screening threshold, the feature is removed; when the contribution value within the channel is high but the consistency constraint score is lower than the preset consistency threshold, the feature is downweighted to reduce its impact on subsequent prediction models.

[0152] The formula for calculating the comprehensive screening score is as follows:

[0153] ;

[0154] In the formula, It is a comprehensive screening score based on features. It is the contribution value weight. It is the contribution value within the channel. It is a constraint on the scoring weight. It is a consistency constraint score;

[0155] Preferably, and satisfy In one embodiment, it is preferable to , To ensure that basic forecasting capabilities are given priority, while consistency constraints are introduced for adjustment;

[0156] when If the feature is below a preset screening threshold, it is removed; preferably, the screening threshold can be 0.2 to 0.3.

[0157] When the contribution value within the channel The contribution score is higher than the preset contribution threshold (preferably 0.5), but the consistency constraint score is higher. If the value is below a preset consistency threshold (preferably 0.2), the feature is downweighted.

[0158] The formula for calculating the weighting coefficient in the weight reduction process is:

[0159] ;

[0160] In the formula, It is the weighting factor. This is a very small constant, preferably 0.01, to avoid the denominator being zero; the weighting process involves multiplying the eigenvalue by... accomplish;

[0161] The core predictive feature fusion output, based on the filtered feature data, generates structural risk feature subsets, biochemical risk feature subsets, and temporal evolution feature subsets respectively. The structural risk feature subsets, biochemical risk feature subsets, and temporal evolution feature subsets are then fused according to patient number and postoperative time node to obtain the fused feature subset after channel consistency screening.

[0162] The core predictive features of bile fistula specifically include a subset of structural risk features, a subset of biochemical risk features, a subset of temporal evolution features, and a subset of fused features after channel consistency screening.

[0163] By performing the above operations, this solution addresses the technical problem in existing bile fistula prediction feature screening methods that rely solely on correlation or model importance for single-dimensional screening, failing to distinguish the medical logical consistency between surgical structural damage features, biochemical response features, and temporal evolution features. This results in the model retaining accidental correlated variables or weakening key mutually corroborating variables. This solution creatively adopts a structural-biochemical consistency constraint screening method. By calculating the intra-channel contribution value of each feature, the mutual corroboration strength between structural risk features and biochemical risk features, and the consistency constraint scores of structural risk channels, biochemical risk channels, and temporal evolution channels, the feature screening process no longer relies solely on statistical correlation but simultaneously considers prediction direction, cross-channel correlation, and contribution stability.

[0164] For example, for feature combinations where "increased number of bile duct ends" and "postoperative abnormal bilirubin ratio" are in the same predictive direction and have stable correlation, their mutual evidence strength is high and they are more likely to be retained. For isolated variables that are highly important but lack structural or biochemical response support, their weight is reduced by consistency scoring, thereby improving the stability, interpretability and noise resistance of the core predictive feature data of bile fistula.

[0165] Example 6, this example is based on the above examples, see below. Figure 1 The dynamic prediction of bile fistula risk is based on the core prediction feature data of bile fistula to construct a bile fistula risk prediction model, and dynamically calculates the probability of bile fistula occurrence at different time points after surgery. By introducing a time series update mechanism, the newly added monitoring data is updated on a rolling basis to generate risk change trends. The prediction results are graded based on a preset risk threshold, and the risk level and warning status are output to obtain bile fistula risk assessment result data with time dynamics.

[0166] The bile fistula risk prediction model includes at least one of the following: Logistic regression model, random forest model, XGBoost model, or long short-term memory network model.

[0167] The data on the risk assessment of bile fistula specifically includes the probability of bile fistula occurrence, risk level classification results, risk change trends, and early warning status indicators.

[0168] Preferably, the risk assessment results are automatically updated on the 1st, 3rd and 5th postoperative days, and the probability of bile fistula occurrence, low, medium and high risk stratification results, risk change trend and warning status are output.

[0169] More preferably, risk review prompts are generated based on the risk level and risk change trend. These prompts include drainage status review prompts, abdominal imaging data review prompts, biliary decompression related data review prompts, and high-risk status prompts, which provide data references for medical personnel to make subsequent judgments.

[0170] By performing the above operations, this solution addresses the technical problems in existing postoperative bile fistula risk assessment systems, which often output prediction results only once, cannot continuously adjust the risk probability based on newly added postoperative monitoring data, and have insufficient correlation between risk level and clinical intervention prompts. It creatively adopts a dynamic prediction method for bile fistula risk based on rolling time series updates. By continuously inputting newly added core predictive feature data of bile fistula at time points such as postoperative day 1, day 3, and day 5, the system dynamically calculates the probability of bile fistula occurrence using a model, and outputs risk level, risk change trend, and warning status indicator based on risk thresholds, enabling the system to reflect the continuous changes in the patient's postoperative risk status.

[0171] For example, if a patient is classified as medium risk on postoperative day 1, shows an increasing trend in abnormal bilirubin ratio and drainage volume on day 3, and exhibits enhanced time-lag response of inflammatory markers on day 5, the system can adjust the patient's risk level from medium to high risk and generate prompts for image data review, drainage status review, or biliary decompression-related data review, thereby improving the predictive results' support for dynamic clinical management.

[0172] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process or method.

[0173] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0174] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A system for predicting the risk of postoperative bile leakage in hepatobiliary surgery patients, characterized in that: It includes an evidence-driven variable collection module, a standardization judgment module, a time-series feature generation module, a feature selection module, and a dynamic prediction module; The evidence-driven variable acquisition module is used to acquire evidence-driven variables of perioperative bile fistula risk factors. By acquiring evidence-driven variables of perioperative bile fistula risk factors, a basic dataset of bile fistula risk is obtained, and the basic dataset of bile fistula risk is sent to the standardization judgment module and the time-series feature generation module respectively. The standardized determination module is used for standardized determination of bile fistula outcome. Through standardized determination of bile fistula outcome, bile fistula outcome label data is obtained and the bile fistula outcome label data is sent to the dynamic prediction module. The time-series feature generation module is used to generate bile extravasation time-series features. Based on the postoperative bile extravasation dynamic variables in the bile fistula risk baseline dataset, it performs time-series reconstruction processing on drainage volume, drainage bilirubin concentration, and serum bilirubin index. The reconstructed time-series data is then subjected to trend fitting, difference calculation, and window statistical analysis to extract feature information reflecting the dynamic evolution of bile extravasation and construct bile extravasation evolution feature data. This includes the following steps: postoperative dynamic sequence construction, bilirubin abnormal gradient joint feature construction, bile concentration drainage volume feature construction, inflammatory index time-lag response feature construction, and risk evolution transition feature construction. The bile extravasation evolution feature data is then sent to the feature filtering module. The feature filtering module is used for multi-channel feature filtering. It divides the input features into structural risk channels, biochemical risk channels, and temporal evolution channels according to their sources. It performs statistical correlation analysis and predictive contribution evaluation on the features in each of the three channels, eliminates redundant and low-contribution features, and performs consistency constraint filtering on the features in the three channels. By analyzing the matching relationship between structural risk features and biochemical risk features in terms of predictive contribution direction and predictive contribution degree, it obtains the core predictive feature data of bile fistula and sends the core predictive feature data of bile fistula to the dynamic prediction module. The dynamic prediction module is used for dynamic prediction of bile fistula risk, and obtains bile fistula risk assessment data through dynamic prediction of bile fistula risk.

2. The postoperative bile fistula risk prediction system for hepatobiliary surgery patients according to claim 1, characterized in that: The perioperative bile fistula risk factor evidence-driven variable collection involves the unified collection and standardized processing of multi-source data from hepatobiliary surgery patients. Candidate variables from electronic medical record systems, surgical record systems, and laboratory testing systems are screened based on previous clinical studies related to bile fistula, expert consensus, and the mechanism of bile fistula occurrence. Field mapping, data cleaning, and time alignment are then performed to obtain the basic dataset of bile fistula risk. The basic dataset for bile leakage risk includes basic demographic variables, preoperative liver function and biliary status variables, surgical structural damage variables, intraoperative trauma variables, and postoperative bile extravasation dynamic variables.

3. The postoperative bile fistula risk prediction system for hepatobiliary surgery patients according to claim 2, characterized in that: The standardized determination of bile fistula outcome is based on preset bile fistula occurrence criteria to determine the status of bile fistula occurrence. Clinical intervention behaviors are introduced as auxiliary determination criteria to mark whether puncture drainage, endoscopic intervention or reoperation is performed. At the same time, the time node of bile fistula occurrence is located by combining the time dimension. Through the fusion of multi-dimensional determination rules, a unified bile fistula outcome labeling system is constructed to obtain bile fistula outcome label data.

4. The postoperative bile fistula risk prediction system for hepatobiliary surgery patients according to claim 3, characterized in that: The generation of bile leakage time-series features employs a bile leakage phase transition feature modeling method for feature enhancement. By jointly modeling the time-series change gradient, change acceleration, and stage fluctuation features, the precursor change features of bile fistula are extracted, including the following steps: postoperative dynamic sequence construction, bilirubin abnormal gradient joint feature construction, bile concentration drainage volume feature construction, inflammatory marker time-delay response feature construction, and risk evolution transition feature construction. The postoperative dynamic sequence construction is based on the postoperative bile extravasation dynamic variable in the bile fistula risk dataset. Using the patient number as the index and the postoperative time node as the sequence dimension, the sequence of drainage fluid bilirubin concentration, the same-day serum bilirubin concentration, the drainage fluid volume, and the inflammatory marker sequence are constructed. The construction of the bilirubin abnormal gradient joint feature is based on the bilirubin concentration sequence of the drainage fluid and the serum bilirubin concentration sequence at the same time. A bilirubin abnormal ratio sequence is constructed, and the first difference is calculated on the bilirubin abnormal ratio sequence to obtain the bilirubin change gradient feature. The bilirubin abnormal ratio is used to characterize the degree of abnormal bile enrichment, and the bilirubin change gradient feature is used to characterize the rate of abnormal bile change.

5. The postoperative bile fistula risk prediction system for hepatobiliary surgery patients according to claim 4, characterized in that: The bile concentration drainage volume feature is constructed by performing differential calculation based on the drainage volume sequence to obtain the drainage volume change trend feature. The abnormal bilirubin ratio is coupled with the drainage volume to obtain the bile concentration drainage volume feature. The drainage volume change trend feature is used to characterize the drainage release change trend. The bile concentration drainage volume feature is used to simultaneously characterize the degree of abnormal enrichment of bile components and the intensity of drainage release. The construction of the inflammatory marker time-delay response feature is based on the bile concentration drainage volume feature and the inflammatory marker sequence. Specifically, the response intensity at different lag times is calculated within a preset lag time window, and the maximum response intensity is selected as the inflammatory marker time-delay response feature. The inflammatory marker time-delay response feature is used to characterize the delayed correlation between the bile leakage process and the inflammatory response. The risk evolution transition feature is constructed based on the abnormal bilirubin ratio, bilirubin change gradient features, drainage volume change trend features, bile concentration drainage volume features, and inflammatory marker time-lag response features. A postoperative risk evolution score is constructed, and the risk change amplitude is calculated based on the difference in postoperative risk evolution scores between adjacent time nodes. The risk change is judged according to a preset transition threshold. When the risk change amplitude is greater than the transition threshold, the corresponding time node is marked as a risk transition node, thus obtaining the postoperative risk evolution transition feature.

6. The postoperative bile fistula risk prediction system for hepatobiliary surgery patients according to claim 5, characterized in that: The bile extravasation evolution characteristics data specifically include bilirubin change gradient characteristics, drainage volume change trend characteristics, bile concentration drainage volume characteristics, inflammatory marker time-lag response characteristics, and postoperative risk evolution transition characteristics.

7. The postoperative bile fistula risk prediction system for hepatobiliary surgery patients according to claim 6, characterized in that: The multi-channel feature screening adopts a structural-biochemical consistency constraint screening method for screening and optimization, including the following steps: multi-channel feature attribution division, intra-channel prediction contribution evaluation, structural-biochemical mutual verification relationship construction, consistency constraint score calculation, feature screening and core prediction feature fusion output; The multi-channel feature attribution classification, based on the bile extravasation evolution feature data and the bile fistula risk basic dataset, divides the input features into structural risk channels, biochemical risk channels, and temporal evolution channels; wherein, the structural risk channel is used to characterize the risk of surgical structural damage, the biochemical risk channel is used to characterize the risk of postoperative physiological response, and the temporal evolution channel is used to characterize the dynamic evolution risk of bile extravasation. The intra-channel prediction contribution assessment evaluates the predictive contribution of features in the structural risk channel, biochemical risk channel, and temporal evolution channel, respectively, to obtain the intra-channel contribution value of each feature. The intra-channel contribution value is composed of the correlation between the feature and the bile fistula occurrence status label, the importance score in the initial prediction model, and the stability score in the cross-validation process.

8. The postoperative bile fistula risk prediction system for hepatobiliary surgery patients according to claim 7, characterized in that: The construction of the structural-biochemical mutual verification relationship involves constructing a structural-biochemical mutual verification relationship for structural risk features in the structural risk channel and biochemical risk features in the biochemical risk channel. This relationship is used to determine whether the structural damage risk is reflected in the postoperative biochemical response. For any structural risk feature and biochemical risk feature, the mutual verification strength is determined by the consistency of prediction direction, feature correlation, and contribution value within the channel. The consistency constraint score calculation is based on the structural-biochemical mutual verification relationship, and the consistency constraint score is calculated for the structural risk feature, biochemical risk feature and time series evolution feature respectively. The feature filtering is based on the contribution value within the channel and the consistency constraint score to filter or reduce the weight of the input features, thereby obtaining the filtered feature data. The core predictive feature fusion output, based on the filtered feature data, generates structural risk feature subsets, biochemical risk feature subsets, and temporal evolution feature subsets respectively. The structural risk feature subsets, biochemical risk feature subsets, and temporal evolution feature subsets are then fused according to patient number and postoperative time node to obtain the fused feature subset after channel consistency screening. The core predictive features of bile fistula specifically include a subset of structural risk features, a subset of biochemical risk features, a subset of temporal evolution features, and a subset of fused features after channel consistency screening.

9. A system for predicting the risk of postoperative bile leakage in hepatobiliary surgery patients according to claim 8, characterized in that: The dynamic prediction of bile fistula risk is based on the core predictive feature data of bile fistula to construct a bile fistula risk prediction model, and dynamically calculates the probability of bile fistula occurrence at different time points after surgery. By introducing a time series update mechanism, the newly added monitoring data is updated on a rolling basis to generate risk change trends. The prediction results are graded based on preset risk thresholds, and the risk level and warning status are output to obtain bile fistula risk assessment results data with time dynamics.

Images