Portal cholangiocarcinoma postoperative risk prediction model based on GGT dynamic change trajectory and construction method thereof

Abstract

The invention belongs to the technical field of biological medicine, and discloses a portal cholangiocarcinoma postoperative risk prediction model based on a gamma-glutamyltransferase (GGT) dynamic change track and a construction method thereof. The method comprises the following steps: (1) collecting clinical characteristic data and survival characteristic data (including survival time and survival state) of a patient suffering from porta hepatis cholangiocarcinoma subjected to radical resection; 2) on the basis of GGT detection results at different time points in a perioperative period, dividing patients into different GGT dynamic change track types by using a latent category hybrid model; and 3) taking the clinical characteristic data and the GGT dynamic change track type as independent variables, taking survival data as response variables, screening prediction factors through LASSO regression, and constructing a multi-factor Cox regression model, namely a postoperative risk prediction model. According to the method, the dynamic evolution rule of the GGT is introduced, so that the limitation of single-time static detection is made up, the quantitative evaluation of the survival risk is realized, and a scientific basis is provided for individualized treatment and follow-up visit management.

Description

Technical Field

[0001] This invention belongs to the field of biotechnology and medicine, specifically designing a postoperative risk prediction model for hilar cholangiocarcinoma based on the dynamic change trajectory of γ-glutamyl transferase (GGT) and its construction method. Background Technology

[0002] Perihilar cholangiocarcinoma (pCCA) is a malignant tumor originating at the confluence of the left and right hepatic ducts, accounting for 50%–60% of all cholangiocarcinomas. Its complex anatomical structure and local invasion characteristics make radical surgery extremely challenging. Although surgical resection remains the only potentially curative treatment for pCCA, even with negative resection margins, the long-term prognosis after surgery is still unsatisfactory, with a 5-year overall survival rate generally below 50%. Therefore, accurately assessing the long-term survival risk of pCCA patients after surgery is crucial for optimizing follow-up plans and developing adjuvant therapy regimens.

[0003] Current prognostic assessment methods mainly rely on static clinical indicators such as TNM staging. However, these indicators cannot comprehensively reflect the patient's physiological response and disease progression characteristics during the perioperative period, and their predictive ability is limited. In recent years, gamma-glutamyl transferase (GGT), as a sensitive biomarker of liver function and oxidative stress, has been shown to be associated with poor prognosis in various malignant tumors. However, most current studies only focus on GGT levels at a single time point (such as preoperatively), neglecting the dynamic changes in GGT during the critical perioperative window, and failing to reveal its potential trajectory patterns and clinical value. Therefore, there is an urgent need to develop a method that can predict the prognostic risk of hilar cholangiocarcinoma based on the dynamic trajectory of GGT data at multiple time points. Summary of the Invention

[0004] In view of the problems and shortcomings of the existing technology, the purpose of this invention is to provide a postoperative risk prediction model for hilar cholangiocarcinoma based on the dynamic change trajectory of GGT and its construction method, so as to realize the quantitative assessment of postoperative risk of patients and assist clinical decision-making.

[0005] To achieve the objectives of this invention, the technical solution adopted is as follows:

[0006] The first aspect of this invention provides a method for constructing a trajectory classification prediction model for predicting perioperative GGT level changes in patients with hilar cholangiocarcinoma, comprising the following steps:

[0007] S1: Construct a GGT longitudinal database, which includes GGT longitudinal data of multiple patients with hilar cholangiocarcinoma who underwent radical resection. The GGT longitudinal data are the serum GGT level test results of patients with hilar cholangiocarcinoma who underwent radical resection at different detection time points during the perioperative period.

[0008] S2: A latent class mixed model (LCMM) is used to model the GGT longitudinal data in the GGT longitudinal database to obtain multiple candidate classification prediction models with different model structures and numbers of latent classes. According to the selection criteria for the classification prediction model, the optimal classification prediction model is selected from the multiple candidate trajectory classification prediction models, which is the trajectory classification prediction model. The selection criteria are: model convergence, minimum Bayesian information criterion, entropy value greater than 60%, and the sample proportion of each latent class is not less than 10%.

[0009] According to the above construction method, preferably, in step S2, when using a latent class mixture model to model the GGT longitudinal data in the GGT longitudinal database, the model structure is set to random intercept or random intercept combined with random slope, and the initial estimate of the single-class model is used as the initial estimate of the multi-class model.

[0010] According to the above construction method, preferably, when using the LCMM function to model the GGT longitudinal data in the GGT longitudinal database, the fixed effects form of the model is set to a linear function, a quadratic function or a natural spline function, and the latent classes of the model are set to 2 to 4.

[0011] According to the above construction method, preferably, the different detection time points in the perioperative period include once within two weeks before surgery and at least twice within two months after surgery.

[0012] The second aspect of the present invention provides a trajectory classification prediction model for predicting the dynamic changes in GGT levels in patients with hilar cholangiocarcinoma during the perioperative period. The trajectory classification prediction model is constructed using the construction method described in the first aspect above. The model structure of the trajectory classification prediction model is a combination of random intercept and random slope. The fixed effects of the model are natural spline functions, and the latent class of the model is 2.

[0013] According to the above trajectory classification prediction model, preferably, the trajectory classification prediction model divides the dynamic change trajectory of GGT level in patients with hilar cholangiocarcinoma in the perioperative period into two categories, one is a V-shaped trajectory and the other is a stationary trajectory; the GGT level in the perioperative period of the V-shaped trajectory shows a trend of first decreasing and then increasing; the GGT level in the perioperative period of the stationary trajectory fluctuates less overall and does not show a significant increasing or decreasing trend.

[0014] A third aspect of this invention provides a method for constructing a postoperative risk prediction model for hilar cholangiocarcinoma, comprising the following steps:

[0015] (1) Obtain a clinical database, which includes clinical data of multiple patients with hilar cholangiocarcinoma who underwent radical resection. The clinical data includes clinical characteristic data and survival characteristic data. The clinical characteristics include age, sex, perioperative GGT level change trajectory classification, tumor histological differentiation degree, Bismuth classification, surgical margin status, vascular invasion, nerve invasion, and TNM stage. The survival characteristics include survival time and survival status. The perioperative GGT level change trajectory classification is a trajectory classification predicted by the trajectory classification prediction model described in the second aspect above.

[0016] (2) Preprocess the clinical data in the clinical database, and then use clinical characteristic data as independent variables and survival characteristic data as response variables to screen variables through LASSO regression to obtain clinical characteristic variables;

[0017] (3) The clinical characteristic variables obtained in step (2) are incorporated into the multivariate Cox proportional hazards regression model to construct a postoperative risk prediction model.

[0018] According to the above construction method, preferably, the clinical characteristic variables in step (2) are: age, perioperative GGT level change trajectory classification, tumor histological differentiation degree and TNM stage.

[0019] The fourth aspect of this invention provides a postoperative risk prediction model for hilar cholangiocarcinoma, which is constructed using the method described in the third aspect above. The postoperative risk prediction model is used to predict the instantaneous risk function of a patient with hilar cholangiocarcinoma at time t postoperatively, and the function is as follows:

[0020]

[0021] in, (t) represents the baseline risk function; class represents the perioperative GGT level change trajectory classification, which is the trajectory classification predicted using the trajectory classification prediction model described in the second aspect above; age represents the patient's age (years); TNM.L, TNM.Q, and TNM.C represent the linear trend term, quadratic trend term, and cubic trend term obtained after orthogonal polynomial comparison encoding of the patient's TNM stage; Differentiation.L and Differentiation.Q represent the linear trend term and quadratic trend term obtained after orthogonal polynomial comparison encoding of the patient's tumor histological differentiation degree, respectively; wherein, each of the trend terms is a numerical covariate that is determined by orthogonal polynomial comparison encoding of the corresponding ordered categorical variable during the model training phase and remains fixed during the prediction phase.

[0022]

[0023] in, For the first The time of this event; This represents the number of events at that moment; For risk set; For individuals The covariate vector; These are the estimated values ​​of the regression coefficients; The relative risk for an individual.

[0024] As can be seen from the postoperative risk prediction model of this invention, variables such as age, TNM stage, and GGT trajectory category are significantly correlated with patient survival risk. A positive regression coefficient indicates that the factor is a risk factor, while a negative regression coefficient indicates that the factor is a protective factor. By combining the above postoperative risk prediction model with the actual clinical information of individual patients and the perioperative GGT level change trajectory classification input, individualized survival risk prediction values ​​can be calculated, thereby achieving quantitative assessment of patients after surgery and assisting doctors in formulating precise treatment and follow-up strategies, which has important clinical practical value.

[0025] The fifth aspect of the present invention provides a method for predicting postoperative risk in patients with hilar cholangiocarcinoma, wherein the clinical characteristic variables of the patients with hilar cholangiocarcinoma are input into the postoperative risk prediction model described in the fourth aspect above, and the instantaneous risk probability of the patients with hilar cholangiocarcinoma at time t is calculated.

[0026] A sixth aspect of the present invention provides an electronic device including a memory and a processor, the memory being configured to store executable instructions capable of running on the processor, the processor being configured to, when running the executable instructions, execute the construction method as described in the first aspect above, or execute the construction method as described in the third aspect above.

[0027] A seventh aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a trajectory classification prediction model described in the second aspect above, which is capable of running on the processor, or the memory stores a postoperative risk prediction model described in the fourth aspect above, which is capable of running on the processor.

[0028] Compared with the prior art, the technical effects achieved by the present invention are as follows:

[0029] (1) This invention provides a method for constructing a trajectory classification prediction model for predicting the perioperative GGT level change trajectory in patients with hilar cholangiocarcinoma. When using the LCMM function to model the longitudinal data of patients' GGT, the model structure is set to random intercept or random intercept plus random slope. The fixed effects form of the model is set to linear function, quadratic function or natural spline function. The initial estimate of the single-class model (ng=1) is used as the initial estimate of the multi-class model (ng=2~4). The latent classes of the model are set to 2~4. After screening, a trajectory classification prediction model is finally obtained. The trajectory classification prediction model is a natural spline function model, which can fully adapt to the nonlinear trend of dynamic changes in GGT. Compared with subjective judgment, it is more objective and reasonable, ensuring the biological rationality, objectivity and statistical stability of trajectory classification, and providing a reliable basis for subsequent individual stratification and prognostic modeling.

[0030] (2) The postoperative risk prediction model for hilar cholangiocarcinoma of the present invention proposes for the first time to combine the classification of perioperative GGT dynamic change trajectory and clinical information to model the postoperative survival outcome of patients with hilar cholangiocarcinoma. This breaks through the limitations of previous prognostic assessments that relied only on single-point GGT values ​​or static indicators, and realizes in-depth mining and effective utilization of GGT time-series information, thereby improving the accuracy and clinical guidance value of the postoperative risk prediction model for postoperative risk prediction.

[0031] (3) In the process of constructing the postoperative risk prediction model of the present invention, LASSO regression is introduced in the variable screening for dimensionality reduction and feature selection, taking into account the simplicity and explanatory power of the model, effectively avoiding the overfitting problem, and combining the Cox proportional hazards model to assess individual survival risk, thereby improving the stability and generalization ability of survival prediction.

[0032] (4) The postoperative risk prediction model modeling method provided by the present invention is applicable to postoperative modeling of other solid tumors with similar perioperative tumor markers. It has good scalability and clinical application prospects, and provides a generalizable paradigm for the modeling and utilization of dynamic tumor biomarkers in the context of precision medicine.

[0033] (5) The postoperative risk prediction model constructed in this invention can calculate individualized survival risk prediction values ​​by inputting the patient's clinical information (age, TNM stage, GGT trajectory category, tumor histological differentiation degree and TNM stage, etc.) into the risk function, thereby realizing the quantitative assessment of the patient after surgery, assisting doctors in formulating precise treatment and follow-up strategies, and has important clinical practical value. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the construction method of the trajectory classification prediction model for predicting the perioperative GGT level changes in patients with hilar cholangiocarcinoma according to the present invention.

[0035] Figure 2 The results show the trajectory classification of the dynamic changes in GGT levels during the perioperative period in 358 patients with hilar cholangiocarcinoma. To facilitate observation of the characteristics of different trajectories, the figure is divided into two comparison charts: the left chart shows the scatter distribution corresponding to the two trajectories and the V-shaped trajectory, and the right chart shows the scatter distribution corresponding to the two trajectories and the stationary trajectory. Class 1 represents the V-shaped trajectory, and Class 2 represents the stationary trajectory.

[0036] Figure 3 The Kaplan-Meier survival curve analysis results of the classification results of the dynamic changes in GGT levels during the perioperative period in patients with hilar cholangiocarcinoma and the corresponding survival outcomes are shown in the figure; where Class 1 represents a V-shaped trajectory and Class 2 represents a stationary trajectory.

[0037] Figure 4 Internal validation ROC curves for the postoperative risk prediction model constructed in this invention to predict the survival outcomes of patients with hilar cholangiocarcinoma;

[0038] Figure 5 The external validation ROC curves for the postoperative risk prediction model constructed in this invention to predict the survival outcomes of patients with hilar cholangiocarcinoma are shown. Detailed Implementation

[0039] To enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention will be described in detail below with reference to specific embodiments.

[0040] The following detailed description is exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0041] It should be noted that the terminology used herein is for the purpose of describing specific embodiments only and is not intended to limit the exemplary embodiments of the present invention. Furthermore, experimental methods in the following embodiments that do not specify specific conditions employ conventional techniques in this art or follow the conditions recommended by the manufacturer; reagents or instruments whose manufacturers are not specified are all commercially available conventional products.

[0042] Example 1:

[0043] A method for constructing a trajectory classification prediction model for predicting perioperative GGT level changes in patients with hilar cholangiocarcinoma, such as... Figure 1 As shown, the specific steps are as follows:

[0044] S1: Construct a GGT longitudinal database, which includes GGT longitudinal data of multiple patients with hilar cholangiocarcinoma who underwent radical resection. The GGT longitudinal data are the serum GGT level test results of patients with hilar cholangiocarcinoma who underwent radical resection at different detection time points during the perioperative period. The different detection time points during the perioperative period include at least one test within two weeks before surgery and at least two tests within two months after surgery.

[0045] The sample source for constructing the GGT longitudinal database was: clinical data from 765 patients with hilar cholangiocarcinoma (pCCA) who underwent surgical treatment at Henan Provincial People's Hospital and Southwest Hospital of Army Medical University. The inclusion criteria for the sample were: pathologically confirmed pCCA patients, with clinical data including perioperative GGT values ​​and testing time, demographic information, tumor histological differentiation, Bismuth classification, resection margin status, vascular and nerve invasion, TNM stage, and follow-up survival data.

[0046] Patient samples were progressively screened from 765 samples according to the following criteria: (1) 15 patients with distant metastases were excluded; (2) 5 patients who underwent related surgeries again after surgery were excluded; (3) 11 patients with a history of malignant tumors were excluded; (4) 66 patients who did not meet the criteria for radical resection were excluded. After screening, 566 patient samples with GGT test results were obtained. The interquartile range (IQR) method was used to remove 12 extreme GGT outliers, i.e., observations exceeding the first or third quartile ± 2 times the IQR. In addition, 159 patients without GGT test data within two weeks before surgery and 37 patients with fewer than three GGT tests within two weeks before surgery to two months after surgery were excluded. Finally, 358 patient samples were included, and the longitudinal GGT data of the 358 selected samples were used for trajectory classification prediction model modeling and analysis.

[0047] S2: Trajectory modeling is performed based on the distribution selection function of the GGT longitudinal data in the GGT longitudinal database. If the original GGT data distribution is approximately normal, the hlme function in the LCMM package is directly used for modeling. If the data does not conform to normality but can be normalized through conventional transformations (such as logarithmic transformations), the hlme function is still used after the data transformation. If normalization cannot be achieved, the lcmm function is used for modeling. Analysis shows that the distribution of the GGT longitudinal data in the GGT longitudinal database constructed in step S1 of this invention cannot be normalized. Therefore, the LCMM function in the LCMM (latent class mixed model) package of R language (version 4.5.1) is used to model the GGT longitudinal data in the GGT longitudinal database constructed in step S1. When using the LCMM function to model the GGT longitudinal data in the GGT longitudinal database, the model structure is set to random intercept or random intercept plus random slope to fully consider individual-level variability. Furthermore, to improve the convergence stability of the multi-class model fitting, it is preferable to use the initial estimate of the single-class model (ng=1) as the initial estimate of the multi-class model (ng=2~4), specified through the B parameter. Additionally, to evaluate the fitting effect of different model structures, the fixed effects form of the model is set to a linear function, a quadratic function, or a natural spline function, and the latent classes of the model are set to 2~4. Based on the above model settings, a total of 18 classification prediction models were fitted, named as follows:

[0048] Linear fixed effects + random intercepts: m2a, m3a, m4a;

[0049] Linear fixed effects + random intercept + random slope: m2b, m3b, m4b;

[0050] Quadratic fixed effects + random intercepts: m2c, m3c, m4c;

[0051] Quadratic fixed effects + random intercept + random slope: m2d, m3d, m4d;

[0052] Natural spline fixed effects + random intercepts: m2e, m3e, m4e;

[0053] Natural spline fixed effect + random intercept + random slope: m2f, m3f, m4f.

[0054] The following is a sample code for model fitting (taking linear fixed effects + random intercept as an example):

[0055] # Load lcmm package

[0056] >library(lcmm)

[0057] # Construct a uniclass model as an initial estimate for the multiclass model.

[0058] >m1a <- lcmm(ggt ~ time,

[0059] random = ~1,

[0060] subject = "ID",

[0061] ng = 1,

[0062] data = data_ggt)

[0063] # Building a binary class model

[0064] >m2a <- lcmm(ggt ~ time,

[0065] random = ~1,

[0066] mixture = ~time,

[0067] subject = "ID",

[0068] ng = 2,

[0069] data = data_ggt,

[0070] B = m1a)

[0071] illustrate:

[0072] ggt ~ time specifies that the fixed-effects form of the GGT value is a linear function;

[0073] random = ~1 indicates that the model structure is a random intercept structure;

[0074] mixture = ~time indicates that different categories can have different time-fixed effects;

[0075] ng = 2 indicates that two potential trajectory categories are set;

[0076] B = m1a indicates that the estimation results of the single-class model are used as the initial values ​​of the multi-class model to improve the stability of the model.

[0077] In the natural spline function fitting model, the fixed effects function takes the form:

[0078] >ggt ~ ns(time, df = 2)

[0079] The ns() function represents a natural spline, and df = 2 specifies the degrees of freedom.

[0080] The fitting statistics of the 18 classification prediction models are shown in Table 1.

[0081]

[0082] Based on the selection criteria for classification prediction models, the optimal classification model was selected from 18 models. These criteria included: good model convergence (conv=1), minimum Bayesian Information Criterion (BIC), entropy greater than 60% to ensure classification accuracy, and a sample proportion of each latent class not less than 10%. Ultimately, model m2f was selected as the optimal classification model (i.e., the trajectory classification prediction model). The model structure of m2f (i.e., the trajectory classification prediction model) is a random intercept plus a random slope, with a fixed effects form of a natural spline function, and two latent classes. Model m2f (i.e., the trajectory classification prediction model) converged on the 26th iteration and could distinguish two statistically significant GGT trajectory patterns: the first type is a "V-shaped trajectory," characterized by a high preoperative GGT level, a decrease during surgery, and a re-increase postoperatively; the second type is a "stationary trajectory," with relatively stable overall GGT levels and small fluctuations.

[0083] After selecting the trajectory classification prediction model—m2f—from 18 classification prediction models, the dynamic trajectory of perioperative GGT level changes in patients with hilar cholangiocarcinoma can be predicted and classified. The trajectory prediction and classification method specifically involves inputting the longitudinal data of perioperative GGT from patients with hilar cholangiocarcinoma into model m2f to obtain the posterior probability of each potential trajectory category. Using the maximum a posteriori rule, the patient is assigned to the potential trajectory category with the highest probability, thus obtaining the trajectory classification prediction result of the perioperative GGT level change trajectory for patients with hilar cholangiocarcinoma.

[0084] The m2f model was used to classify the dynamic changes in perioperative GGT levels in 358 patients with hilar cholangiocarcinoma in the GGT longitudinal database constructed in this invention. The classification results are as follows: Figure 2 As shown, the potential trajectories are divided into two categories: V-shaped trajectories (48.32% of the total sample) and stationary trajectories (51.68% of the total sample). Figure 2As shown in the left-middle figure, the V-shaped trajectory exhibits a trend of high preoperative GGT levels, a significant short-term decrease in GGT levels postoperatively (β = -14.56, where β is the parameter of the natural spline function; SE = 0.93, p < 0.001), and a significant increase in GGT levels three weeks postoperatively (β = 8.17, where β is the parameter of the natural spline function; SE = 1.07, p < 0.001). The estimated values ​​of natural spline function 1 and natural spline function 2 for the V-shaped trajectory are -14.56 and 8.17, respectively. Figure 2 As shown in the right-middle figure, the overall GGT level of the stationary trajectory fluctuates less and shows a relatively stable trend. This is characterized by a higher preoperative GGT level, a slight intraoperative decrease (β = -1.53, where β is the parameter of the natural spline function; SE = 0.68, p = 0.025), and a slight postoperative increase (β = 1.39, where β is the parameter of the natural spline function; SE = 0.66, p = 0.034). Specifically, the estimated values ​​of natural spline basis 1 and natural spline basis 2 for the stationary trajectory are -1.53 ​​and 1.39, respectively. The variance estimates of the random intercept and random slope for the V-shaped trajectory and the stationary trajectory are 1.99 and 0.01, respectively, and the covariance between the random intercept and random slope is -0.07.

[0085] It should be noted that each potential trajectory does not correspond to a fixed analytical function formula. The specific trajectory shape is automatically determined during the model training phase by the latent class mixture model in conjunction with natural spline functions. During the model application phase, the perioperative GGT measured time-point data of the patient to be predicted are input into the trained latent class mixture model. Based on the fixed effects parameters, random effects structure, and natural spline basis function forms determined during the training phase, the likelihood function value of the patient under each potential trajectory category is calculated, and the posterior probability of belonging to each potential trajectory category is further obtained, thus completing the trajectory classification prediction.

[0086] The above prediction process can be implemented through statistical modeling software or a programming environment, which at least supports latent class mixture models, natural spline function construction, and posterior probability calculation.

[0087] Furthermore, based on the classification results of the perioperative GGT level dynamic change trajectory of 358 patients with hilar cholangiocarcinoma in the GGT longitudinal database constructed in this invention using the m2f model, and the corresponding survival outcomes of the patients, Kaplan-Meier survival curve analysis was performed. The results are as follows: Figure 3 As shown.

[0088] Depend on Figure 3It can be seen that the survival rate of patients with a stable trajectory in the perioperative GGT horizontal trajectory classification is significantly higher than that of patients with a V-shaped trajectory classification, suggesting that there is a significant correlation between GGT trajectory classification and postoperative survival, and verifying the effectiveness and applicability of the model constructed in this invention in postoperative risk assessment.

[0089] Example 2:

[0090] A method for constructing a postoperative risk prediction model for hilar cholangiocarcinoma, the specific steps of which are as follows:

[0091] (1) Construction of a clinical database: The clinical database includes clinical data of multiple patients with hilar cholangiocarcinoma who underwent radical resection. The clinical data includes clinical characteristic data and survival characteristic data. The clinical characteristics include age, sex, perioperative GGT level change trajectory classification, tumor histological differentiation degree, Bismuth classification, surgical margin status, vascular invasion, nerve invasion, and TNM stage. The survival characteristics include survival time and survival status. Among them, the perioperative GGT level change trajectory classification is obtained by predicting the trajectory classification prediction model (model m2f) constructed in Example 1.

[0092] The sample source for constructing the clinical database was as follows: based on the 358 samples included in Example 1, patients lacking follow-up information were further excluded, resulting in a final sample of 286. The 286 included samples showed a balanced distribution in terms of age, sex, perioperative GGT level trajectory classification, tumor histological differentiation, Bismuth classification, surgical margin status, vascular invasion, neurological invasion, and TNM stage. However, postoperative survival time and outcomes differed among the different trajectory groups, suggesting that the differences in postoperative survival outcomes among different trajectory groups are related to the GGT trajectory, rather than the effect of other confounding factors.

[0093] (2) Preprocessing of clinical data: The clinical data in the clinical database were preprocessed, including outlier identification and removal, and missing value imputation. Outlier identification and removal used the IQR method, specifically, outliers were defined as values ​​more than twice the IQR value from the first or third quartile. Missing value imputation used the chain equation multiple interpolation (MICE) method, where categorical variables used polyreg regression, ordered variables used proportional dominance logistic regression (polr), and continuous variables used predictive mean matching (pmm). The MICE process was completed using the mice package in R, with 5 imputation iterations and a random seed of seed = 123 to ensure repeatability. After preprocessing, the dataset was randomly divided into a training set and a validation set in a 7:3 ratio, where the training set was used for model building and the validation set was used for model performance verification.

[0094] (3) Screening of clinical characteristic variables: Clinical data (age, sex, perioperative GGT level change trajectory classification, tumor histological differentiation degree, Bismuth classification, surgical margin status, vascular invasion, nerve invasion, and TNM stage) from the training set were used as independent variables. All independent variables were appropriately coded and converted into numerical matrix format to form an independent variable matrix x. At the same time, response variables y were constructed using total survival time (os) and survival status (event), where y represents the survival object Surv(os, event) of the Cox proportional hazards model. The variable screening process adopted the LASSO (Least Absolute Shrinkage and Selection Operator) regression method and modeled using the glmnet R package. The parameter α=1 was set, and regularized regression analysis was performed using the Cox regression framework, which is suitable for high-dimensional feature selection. During the modeling process, 10-fold cross-validation was performed on the training set using the cv.glmnet function to calculate the mean squared error (MSE) under different penalty coefficients λ, and error curves were plotted to assist in selecting the optimal λ value. The λ(lambda.min) corresponding to the minimum MSE obtained from cross-validation is used as the optimal parameter. The corresponding non-zero regression coefficients are extracted, and the model is independently validated in the validation set to evaluate its predictive performance and generalization ability.

[0095] The variables ultimately selected with independent predictive power include: GGT trajectory classification, TNM stage, tumor histological differentiation degree, and age. These four variables were incorporated into a multivariate Cox proportional hazards regression model to construct a postoperative risk prediction model (the model was constructed using the R language's `survival` package, with the function `coxph()`). The postoperative risk prediction model is used to predict the instantaneous risk function of patients with hilar cholangiocarcinoma at time t postoperatively, and the function is:

[0096]

[0097] in, (t) represents the baseline risk function, class represents the trajectory classification of perioperative GGT level changes predicted by the trajectory classification prediction model constructed in Example 1, age represents the patient's age (years); TNM.L, TNM.Q, and TNM.C represent the linear, quadratic, and cubic trend terms of TNM staging after orthogonal polynomial comparison coding, respectively; Differentiation.L and Differentiation.Q represent the linear and quadratic trend terms of tumor histological differentiation after orthogonal polynomial comparison coding, respectively; wherein, each trend term is a numerical covariate obtained by orthogonal polynomial comparison coding of the corresponding ordered categorical variable.

[0098]

[0099] Used as the benchmark risk function; For the first The time of this event; This represents the number of events at that moment; For risk set; For individuals The covariate vector; These are the estimated values ​​of the regression coefficients; The relative risk for an individual.

[0100] By inputting the clinical characteristic data of patients with hilar cholangiocarcinoma into the postoperative risk prediction model, the survival probability of patients with hilar cholangiocarcinoma at time t after surgery can be predicted through calculation.

[0101] To validate the model's predictive performance, time-dependent ROC curves were plotted using the R language's timeROC package on both the training and validation sets, combining the predicted survival probabilities. This evaluated the constructive postoperative risk prediction model's ability to predict the 5-year overall survival of patients with hilar cholangiocarcinoma, and the corresponding AUC (Area Under Curve) values ​​were reported. Internal validation (training set) results are as follows: Figure 4As shown, the external validation (validation set) results are as follows: Figure 5 As shown.

[0102] Depend on Figure 4 and Figure 5 As can be seen, the postoperative risk prediction model constructed in this invention exhibits good discriminative ability across different datasets. Specifically, in internal validation (training set), the model's 5-year overall survival prediction AUC for patients with hilar cholangiocarcinoma was 82.4% (95% confidence interval: 73.0%–90.4%); in external validation (validation set), the corresponding 5-year prediction AUC was 75.8% (95% confidence interval: 6.04%–90.6%). These results demonstrate that the model's predictive performance is relatively stable across different samples, exhibiting good extrapolation ability and clinical application value.

[0103] Finally, it should be noted that the above embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may use the above technical content as inspiration to make changes or modifications. These are equivalent embodiments with variations. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the technical concept of the present invention still fall within the protection scope of the claims of the present invention.

Claims

1. A method for constructing a trajectory classification prediction model for predicting perioperative GGT level changes in patients with hilar cholangiocarcinoma, characterized in that, Includes the following steps: S1: Construct a GGT longitudinal database, which includes GGT longitudinal data of multiple patients with hilar cholangiocarcinoma who underwent radical resection. The GGT longitudinal data are the serum GGT level test results of patients with hilar cholangiocarcinoma who underwent radical resection at different detection time points during the perioperative period. S2: A latent class mixture model is used to model the GGT longitudinal data in the GGT longitudinal database to obtain multiple candidate classification prediction models with different model structures and the number of latent classes; according to the selection criteria for the classification prediction model, the optimal classification prediction model is selected from the multiple candidate trajectory classification prediction models, which is the trajectory classification prediction model; wherein, the selection criteria are: model convergence, minimum Bayesian information criterion, entropy value greater than 60%, and the sample proportion of each latent class is not less than 10%.

2. The construction method according to claim 1, characterized in that, In step S2, when modeling the GGT longitudinal data in the GGT longitudinal database using a latent class mixture model, the model structure is set to random intercept or random intercept combined with random slope, and the initial estimate of the single-class model is used as the initial estimate of the multi-class model.

3. The construction method according to claim 2, characterized in that, When using the LCMM function to model the GGT longitudinal data in the GGT longitudinal database, the fixed effects form of the model is set to a linear function, a quadratic function, or a natural spline function, and the latent classes of the model are set to 2 to 4.

4. The construction method according to any one of claims 1-3, characterized in that, The perioperative testing time points include one test within two weeks before surgery and at least two tests within two months after surgery.

5. A trajectory classification prediction model for predicting the dynamic changes in GGT levels during the perioperative period in patients with hilar cholangiocarcinoma, characterized in that, The trajectory classification prediction model is constructed using any of the construction methods described in claims 1-4; the model structure of the trajectory classification prediction model is a combination of random intercept and random slope, the fixed effects of the model are natural spline functions, and the latent class of the model is 2; the trajectory classification prediction model divides the dynamic change trajectory of GGT level in patients with hilar cholangiocarcinoma during the perioperative period into two categories, one of which is a V-shaped trajectory, in which the GGT level in the perioperative period shows a trend of first decreasing and then increasing; the other is a stationary trajectory.

6. A method for constructing a postoperative risk prediction model for hilar cholangiocarcinoma, characterized in that, Includes the following steps: (1) Obtain a clinical database, which includes clinical data of multiple patients with hilar cholangiocarcinoma who underwent radical resection. The clinical data includes clinical characteristic data and survival characteristic data. The clinical characteristics include age, sex, perioperative GGT level change trajectory classification, tumor histological differentiation degree, Bismuth classification, surgical margin status, vascular invasion, nerve invasion, and TNM stage. The survival characteristics include survival time and survival status. The perioperative GGT level change trajectory classification is a trajectory classification predicted using the trajectory classification prediction model described in claim 5. (2) Preprocess the clinical data in the clinical database, and then use clinical characteristic data as independent variables and survival characteristic data as response variables to screen variables through LASSO regression to obtain clinical characteristic variables; (3) The clinical characteristic variables obtained in step (2) are used to construct a model using the multivariate Cox regression method to obtain the postoperative risk prediction model.

7. The construction method according to claim 6, characterized in that, The clinical characteristic variables mentioned in step (2) are: age, perioperative GGT level change trajectory classification, tumor histological differentiation degree and TNM stage.

8. A postoperative risk prediction model for hilar cholangiocarcinoma, characterized in that, The postoperative risk prediction model is constructed using the method described in claim 6 or 7; the postoperative risk prediction model is used to predict the instantaneous risk function of patients with hilar cholangiocarcinoma at time t after surgery, and the function is as follows: ,in, (t) represents the baseline risk function; class represents the perioperative GGT level change trajectory classification, which is a trajectory classification predicted using the trajectory classification prediction model described in claim 5; age represents the patient's age; TNM.L, TNM.Q, and TNM.C represent the linear trend term, quadratic trend term, and cubic trend term of TNM staging after orthogonal polynomial comparison coding, respectively; Differentiation.L and Differentiation.Q represent the linear trend term and quadratic trend term of tumor histological differentiation degree after orthogonal polynomial comparison coding, respectively. ,in, For the first The time of this event; This represents the number of events at that moment; For risk set; For individuals The covariate vector; These are the estimated values ​​of the regression coefficients; The relative risk for an individual.

9. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory is used to store executable instructions that can run on the processor, and the processor is used to execute, when running the executable instructions, the construction method as described in any one of claims 1 to 4, or the construction method as described in claim 6 or 7, or the memory stores the trajectory classification prediction model of claim 5 that can run on the processor, or the memory stores the postoperative risk prediction model of claim 8 that can run on the processor.

Images