Prediction method, model and system for cognitive impairment after liver transplantation

Abstract

The invention aims to develop and verify a risk prediction model for perioperative neurocognitive impairment of liver transplantation recipients after LT operation. According to the research, six machine learning models are developed to predict the risk of perioperative neurocognitive impairment after the LT operation. Research results show that patients suffering from liver transplantation have latent hepatic encephalopathy, platelet count, prothrombin time, glomerular filtration rate, blood calcium concentration and MELD score before liver transplantation; the blood loss amount in the operation; the SOFA score, the hypersensitive C-reactive protein and the glutamic oxalacetic transaminase of a liver transplantation patient after liver transplantation are ten important indexes for influencing the postoperative perioperative neurocognitive impairment. In the six ML models developed in the research, the LR algorithm model shows that the overall performance of predicting the perioperative neurocognitive impairment after the LT operation is optimal.

Description

Technical Field

[0001] The present invention belongs to the field of medical data mining, and specifically relates to a prediction system for cognitive dysfunction after liver transplantation, a prediction model construction method and a kit. Background Art

[0002] In recent years, machine learning technology has been widely used in the field of intelligent medicine, and has important practical significance in clinical decision-making, clinical diagnosis and precision medicine. Machine learning-based models have high accuracy in predicting medical outcomes and identifying high-risk patients by utilizing the massive parameters already present in electronic medical records. However, there is currently no early risk prediction model for neurocognitive dysfunction after liver transplantation based on machine learning.

[0003] Although liver transplantation has become one of the clinical gold standard options for treating end-stage liver disease, it is still a difficult type of surgery to completely overcome due to the complexity of the operation itself and the uncertainty of the patients, especially liver transplant patients will still experience various types of complications after surgery. Complications related to changes in the nervous system and cognitive function after liver transplantation have always been a hot topic and clinical pain point in the field.

[0004] In a retrospective analysis of liver transplant complications based on a perioperative disease database, we found that among all types of complications after liver transplantation, neurocognitive dysfunction had the highest incidence.

[0005] Therefore, based on the current research status of neurocognitive dysfunction after liver transplantation and its adverse effects on patient prognosis, there is an urgent need to explore targeted independent risk factors from perioperative data, and further infer related new potential pathogenesis based on potential risk factors, so as to provide reference strategies for clinical physicians' perioperative decision-making. This is also an important entry point for truly achieving the prevention of such complications and improving the survival rate and quality of life of liver transplant patients.

[0006] Machine learning (ML) is one of the important subfields of artificial intelligence (AI). It is essentially a process of allowing machine learning algorithms to learn from a large amount of data and automatically make corresponding decisions and predictions. By analyzing a large amount of clinical perioperative data, including patient medical records, diagnostic results, medical images, laboratory tests, etc., machine learning algorithms can learn and discover patterns and risk factors associated with specific complications. At present, more and more domestic and foreign studies have confirmed that the use of machine learning algorithms can effectively predict postoperative complications clinically.

[0007] The prognosis of liver transplant patients is affected by complex multi-dimensional confounding factors during the perioperative period. Powerful machine learning algorithm models have been widely recognized in the field of liver transplant research and have good application prospects in clinical scenarios before or after transplantation. The complication of postoperative cognitive dysfunction in liver transplant patients is a clinical pain point that cannot be ignored and directly affects the overall prognosis of patients. Some previous studies have reported some potential risk factors related to such complications, such as excessive alcohol, high Child-Pugh score, high APACHE II score, high MELD score, and high preoperative INR level. Unfortunately, there is currently no suitable machine learning prediction model that can directly and reasonably predict postoperative cognitive dysfunction in liver transplant patients. Therefore, there is an urgent need to establish a suitable prediction model that can be directly used in clinical work and assist clinicians' treatment strategies. Therefore, there is an urgent need for a reliable prediction model for postoperative neurocognitive dysfunction after liver transplantation to guide preventive interventions and treatments. Summary of the Invention

[0008] To overcome the deficiencies in the prior art, the present invention provides a scientific, reliable, highly specific, good model performance, and high-sensitivity model, method, and system for predicting the early stage of postoperative neurocognitive dysfunction after liver transplantation. This study established a logistic regression prediction model based on machine learning. This model incorporates preoperative, intraoperative, and postoperative variables to predict postoperative neurocognitive dysfunction after liver transplantation and is expected to predict postoperative neurocognitive dysfunction in future clinical applications, which helps in making early decisions regarding perioperative neurocognitive dysfunction after liver transplantation in clinical work.

[0009] To achieve the above invention objective, the present invention adopts the following technical solutions:

[0010] Another aspect of the present invention is to provide a method for constructing a prediction model for postoperative neurocognitive dysfunction after liver transplantation, including the following steps:

[0011] S1. Sample extraction for model establishment: Collect clinical sample data of patients undergoing allogeneic liver transplantation surgery. The collected samples are used as the training set and the validation set respectively. Among them, the training set data is mainly used for the development of the machine learning model, and the validation set data is used for the validation of the machine learning model and the comparison of the efficacy between different models. The allocation ratio range of the training set and the validation set is 6:4 to 7:3;

[0012] S2. Model Feature Variable Screening: The features of a prediction model refer to all variable categories that are ultimately used in a machine learning model for outcome prediction. The features of the model can be continuous variables or binary or multi-class variables. Reasonable feature screening is the basis for the effectiveness of the prediction model. In this part of the study, the feature variables included in the machine learning model building adopted the following scheme: collect variables that are screened by univariate analysis to be statistically significant and have no multicollinearity, and finally select the independent risk factors closely related to postoperative neurocognitive dysfunction after liver transplantation based on stepwise regression method in multivariate Logistics regression as the feature variables for final modeling;

[0013] S3. Development of Machine Learning Model: Using the model feature variables obtained in step S2, train the screened model feature variables. Among them, the training model is selected from one or more of 6 classic machine learning model algorithms, such as Logistic Regression (LR), multi-layer perceptron (MLP), support vector machine (SVM), random forest (RF), light gradient boosting (LGB), and extreme gradient boosting (XGB). The training model simultaneously uses the grid search method combined with the K-fold cross-validation method to select the optimal parameter combination. First, set the parameters and parameter value lists that need to be grid searched for each model, perform cross-combinations, and select the best parameters for each model on the internal validation dataset.

[0014] There have been many application examples of machine learning algorithms in the medical field. Therefore, this study selected 6 widely used machine learning algorithms for the subsequent model establishment and screening process. The selected model algorithms and basic principles are as follows:

[0015] 1) Logistic Regression

[0016] Logistic regression (LR) is one of the most widely used algorithms in the field of medical clinical research. Generally speaking, the Logistics regression algorithm also belongs to the category of machine learning. The core of this algorithm is to solve binary or multi-class problems, that is, by constructing a loss function and using optimization techniques to iterate cyclically, so as to determine the optimal values of the model parameters. Although the Logistics regression algorithm is relatively simple and fast, and has good interpretability, its effect in dealing with complex or high-dimensional data and non-linear datasets may be poor.

[0017] 2) Multi-layer perceptron

[0018] The multi-layer perceptron (MLP) algorithm can capture and model complex non-linear relationships in data by introducing one or more hidden layers. It has powerful non-linear data processing capabilities. At the same time, the structure of the MLP can be adjusted according to the complexity of the problem, adding more hidden layers or neurons to improve the performance of the model. Therefore, it is applicable to various machine learning tasks such as classification, regression, and feature learning, and has applications in many fields such as image processing, speech recognition, and natural language processing. The main disadvantage of the MLP algorithm is that the model training has a large computational amount and is relatively difficult to train. In addition, the MLP model contains a large number of parameters (weights and biases), and multiple hyperparameters (such as learning rate, number of layers, number of neurons in each layer) need to be adjusted, which makes the fine-tuning of the model relatively complex and time-consuming.

[0019] 3) Support vector machine

[0020] The core principle of the support vector machine (SVM) is to determine an optimal separation plane (hyperplane) in the data space to distinguish different training sample data. In this process, the SVM algorithm uses the function margin as a constraint condition and maximizes the geometric margin as the objective function. By adjusting the function margin to obtain the maximum geometric margin, the purpose of optimizing the objective function is achieved. The main advantage of the SVM is that it still has stable working efficiency for high-dimensional data sets, and the incidence of generalization errors of the model is relatively low. Its main disadvantage is that the interpretability of the model variables is relatively poor, and it is mainly effective for binary classification problems and not suitable for multi-classification problems.

[0021] 4) Random forest

[0022] The random forest (RF) belongs to a type of ensemble learning method. It constructs multiple decision trees by randomly selecting samples and features from the data set multiple times. After all these decision trees are trained, the final prediction model is formed by integrating the prediction results of the "branches" of these decision trees. This algorithm can effectively reduce the overfitting problem of the model. The main disadvantage of the RF algorithm is that it also does not have good model interpretability.

[0023] 5) Light Gradient Boosting Machine

[0024] Light Gradient Boosting (LGB) is an algorithm belonging to the boosting category. Its core principle lies in gradually iterating and constructing multiple relatively weak learners, which focus on fitting the negative gradients of the residuals of the previous model. With the addition of these weaker learners, the LGB algorithm can reduce the loss of the overall model in the direction of the negative gradient. The algorithm linearly combines these base learners by assigning different weights, so that the learners with better performance obtain more weights in the model. The main disadvantage of the LGB algorithm is that, like other gradient boosting methods, it may suffer from overfitting, especially when the dataset is small or the model is too complex. In addition, such algorithms are sensitive to outliers and require more thorough data cleaning and preprocessing before model training.

[0025] 6) Extreme Gradient Boosting

[0026] Extreme Gradient Boosting (XGB) is an advanced machine learning algorithm derived from the basic boosting algorithm. Its core principle is to construct a relatively optimal model by calculating the minimization of the loss function. Compared with traditional boosting algorithms, the XGB algorithm has higher data processing efficiency. The main disadvantages of the XGB algorithm are that the interpretability of the model is not as good as that of linear models, especially when the model is constructed very complexly. In addition, as a tree-based ensemble learning method, the XGB algorithm is less sensitive to small datasets and may even perform worse than simple models, and is prone to overfitting.

[0027] Preferably, the training model in step S3 is selected from the Logistic Regression algorithm LR, which has certain advantages.

[0028] S4. Evaluation of the effect of the prediction model: The model verifies its effectiveness through one or more of the following 3 methods: ① Model evaluation indicators based on the confusion matrix: All feature variables are respectively imported into the candidate machine learning prediction model algorithms to construct the prediction model, the model is trained with the data of the training set, and the internal validation and effect evaluation of the model are carried out with the data of the internal validation set. Finally, the AUC value, accuracy, sensitivity, specificity, and F1 score of each model are calculated according to the principle of the confusion matrix for a comprehensive effect evaluation; ② Model evaluation based on the calibration curve: Based on the R software, the calibration curve of the constructed machine learning prediction model is generated and drawn. The closer it is to the optimal schematic calibration curve, the better the calibration degree of the model algorithm. ③ Model evaluation based on the decision curve: The DCA decision curve of postoperative neurocognitive dysfunction after liver transplantation is drawn through the DCA decision analysis algorithm for evaluation.

[0029] Another aspect of the present invention is to provide a system for constructing a prediction model for postoperative neurocognitive disorder after liver transplantation. The system includes one or more computer processors and a computer-readable medium. The computer-readable medium stores a plurality of instructions, and the plurality of instructions direct the one or more computer processors to execute the above-mentioned method for constructing a prediction model for postoperative neurocognitive disorder after liver transplantation.

[0030] Another aspect of the present invention is to provide a kit for predicting postoperative neurocognitive disorder after liver transplantation. The kit contains detection reagents and detection instruments for detecting the following 10 indicators; these 10 indicators are respectively: occult hepatic encephalopathy before liver transplantation in liver transplant patients, platelet count, prothrombin time, glomerular filtration rate, blood calcium concentration, MELD score; intraoperative blood loss; and SOFA score, high-sensitivity C-reactive protein, and aspartate aminotransferase in liver transplant patients after liver transplantation.

[0031] Another aspect of the present invention is to provide the application of detection reagents and detection instruments for detecting 10 indicators in the preparation of products or kits for establishing an early prediction model for postoperative neurocognitive disorder after liver transplantation or for predicting postoperative neurocognitive disorder after liver transplantation; these 10 indicators are respectively: occult hepatic encephalopathy before liver transplantation in liver transplant patients, platelet count, prothrombin time, glomerular filtration rate, blood calcium concentration, MELD score; intraoperative blood loss; and SOFA score, high-sensitivity C-reactive protein, and aspartate aminotransferase in liver transplant patients after liver transplantation. The detection reagents and instruments are conventional detection reagents and instruments clinically used.

[0032] Another aspect of the present invention provides a method for predicting postoperative neurocognitive disorder after liver transplantation; the method includes the following steps: a) obtaining the measurement results of the following 10 indicators of liver transplant patients: occult hepatic encephalopathy before liver transplantation in liver transplant patients, platelet count, prothrombin time, glomerular filtration rate, blood calcium concentration, MELD score, intraoperative blood loss, and SOFA score, high-sensitivity C-reactive protein, and aspartate aminotransferase in liver transplant patients after liver transplantation; b) inputting the measurement result parameters of the 10 indicators in step a) into the total risk score calculation formula to obtain the total risk score of postoperative neurocognitive disorder after liver transplantation, and then calculating the risk prediction value according to the total risk score. The calculation formulas for the total risk score and the risk prediction value are as follows:

[0033] Total risk score = (-0.578) + (1.721) * X a + (-0.719) * Xb + (-0.315) * X c + (0.393) * X d + (0.257) * X e + (0.17) * X f+(0.364)*X g +(0.382)*X h +(0.34)*X i +(0.335)*X j ;

[0034]

[0035] wherein, X a represents whether there is CHE (cryptogenic hepatic encephalopathy) before surgery, 0 means no, 1 means yes; X b represents the last PLT (platelet) before surgery, and the unit is 10^9 / L; X c represents eGFR (glomerular filtration rate), and the unit is ml / (min*1.73m^2); X d represents PT (prothrombin time) within 1 day after surgery, and the unit is s; X e represents the last Ca 2+ (serum calcium ion) before surgery, and the unit is mmol / L; X f represents the MELD score; X g represents EBL (estimated blood loss during surgery), and the unit is mL; X h represents the SOFA score; X i represents AST (aspartate aminotransferase) within 1 day after surgery, and the unit is U / L; X j represents the first hsCRP (high-sensitivity C-reactive protein) after surgery, and the unit is mg / L.

[0036] Another aspect of the present invention provides a prediction system for postoperative neurocognitive dysfunction after liver transplantation. The system includes an input device, a processor, and a computer-readable medium; the input device is used to obtain the measured values of the relevant detection indexes of the measured liver transplantation patient; the processor is connected to the input device, and the processor is used to process the data obtained by the input device and output the predicted value of the risk of neurocognitive dysfunction; the computer-readable medium stores a plurality of instructions, and the plurality of instructions instruct the input device and the processor to execute the prediction method for postoperative neurocognitive dysfunction after liver transplantation; the method includes the following steps: a) obtaining the measurement results of the following 10 indexes of the liver transplantation patient: cryptogenic hepatic encephalopathy, platelet count, prothrombin time, glomerular filtration rate, blood calcium concentration, MELD score, blood loss during surgery, and SOFA score, high-sensitivity C-reactive protein, aspartate aminotransferase of the liver transplantation patient after liver transplantation; b) inputting the measurement result parameters of the 10 indexes in step a) into the total risk score calculation formula to obtain the total risk score of postoperative neurocognitive dysfunction after liver transplantation, and then calculating the risk prediction value according to the total risk score. The calculation formulas of the total risk score and the risk prediction value are as follows:

[0037] Total risk score = (-0.578) + (1.721) * X a + (-0.719) * X b + (-0.315) * X c + (0.393) * X d + (0.257) * X e + (0.17) * X f + (0.364) * X g + (0.382) * X h + (0.34) * X i + (0.335) * X j ;

[0038] where, X a represents whether there is CHE (cryptogenic hepatic encephalopathy) before surgery, 0 for no, 1 for yes; X b represents the last PLT (platelet) before surgery, with the unit of 10^9 / L; X c represents eGFR (glomerular filtration rate), with the unit of ml / (min*1.73m^2); X d represents PT (prothrombin time) within 1 day after surgery, with the unit of s; X e represents the last Ca 2+ (serum calcium ion) before surgery, with the unit of mmol / L; X f represents the MELD score; X g represents EBL (estimated blood loss during surgery), with the unit of mL; X h represents the SOFA score; X i represents AST (aspartate aminotransferase) within 1 day after surgery, with the unit of U / L; X j represents the first hsCRP (high-sensitivity C-reactive protein) after surgery, with the unit of mg / L.

[0039]

[0040] The present invention finally obtains 10 important indicators for risk assessment of predicting neurocognitive dysfunction after liver transplantation in the early stage, namely preoperative cryptogenic hepatic encephalopathy, platelet count, prothrombin time, glomerular filtration rate, blood calcium concentration, MELD score, intraoperative blood loss, and SOFA score, high-sensitivity C-reactive protein, aspartate aminotransferase of liver transplant patients after liver transplantation. Among the 6 machine learning models developed in the study, the LR algorithm model has the best overall performance in predicting neurocognitive dysfunction after liver transplantation.

[0041] The present invention fills the gap in the prior art of lacking effective prediction methods and prediction models for postoperative neurocognitive disorders after liver transplantation. In addition, the present invention uses a variety of model evaluation methods to evaluate the developed models. Finally, the selected LR algorithm model shows more stable performance, more accurate prediction results, and can reliably predict the postoperative neurocognitive disorder prediction model, method and system for liver transplant recipients. It has better performance than the existing prediction neurocognitive disorder scoring system. This prediction model, method and system incorporate preoperative, intraoperative and postoperative variables to predict postoperative neurocognitive disorders after liver transplantation, and are expected to predict postoperative neurocognitive disorders in future clinical applications, which helps early decision-making on postoperative neurocognitive disorders in clinical work. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 Flow chart for screening the study population;

[0043] Figure 2 Bar chart of feature contributions of the LR model for machine learning prediction of PND after liver transplantation;

[0044] Figure 3 ROC curve of the machine learning prediction model for PND after liver transplantation;

[0045] Figure 4 Calibration curve of the LR model for machine learning prediction of PND after liver transplantation;

[0046] Figure 5 Decision analysis curve of the LR model for machine learning prediction of PND after liver transplantation;

[0047] Figure 6 ROC curve for horizontal comparison of the LR machine learning prediction model;

[0048] Figure 7 ROC curve of the time-series external validation set of the machine learning prediction model;

[0049] Figure 8 Web calculator for the PND prediction model after liver transplantation (Example 1);

[0050] Figure 9 Web calculator for the PND prediction model after liver transplantation (Example 2). DETAILED DESCRIPTION OF THE EMBODIMENTS

[0051] To more clearly illustrate the embodiments of the present invention, the specific embodiments of the present invention will be described below with reference to the accompanying drawings.

[0052] Terms and definitions:

[0053] The preoperative, intraoperative, and postoperative times of the present invention are based on the time of the first liver transplantation surgery, and are respectively the last time before surgery, during surgery, and within 1 - 3 days after surgery.

[0054] The perioperative neurocognitive disorder involved in the present invention is determined according to the following facts. In 2018, the Perioperative Cognitive Function Assessment Expert Panel recommended that any change in the cognitive function status occurring after all surgeries and anesthesia be uniformly described as perioperative neurocognitive disorders (PND). Briefly summarize this important research content, that is, this comprehensive term includes various types of symptoms related to the change or decline in the cognitive ability of patients occurring after surgery, specifically including postoperative delirium (POD) or emergence delirium (emergency delirium, emergence delirium), delayed neurocognitive recovery (DNR), neurocognitive disorder (NCD), postoperative cognitive dysfunction (POCD), etc. (see Table 1 for details).

[0055] Table 1. Main Diagnostic Categories and Classifications of PND

[0056]

[0057]

[0058] The secondary indicators refer to the relevant indicators formed after secondary calculation through a generally recognized calculation formula or table in the medical field based on the original primary indicators. The main types of relevant secondary indicators included in this study are as follows:

[0059] 1) Liver function Child Pugh classification system

[0060] The assessment of liver function before and after surgery is crucial for ensuring patient safety. The classification of liver function, which is a grading system formed by selecting a combination of key indicators, is extremely important for evaluating the degree of liver damage and prognosis of patients with liver malignancies, and is also an important basis for formulating treatment plans. Among them, the Child-Pugh liver function classification system (abbreviated as Child-Pugh classification) is the most widely used and classic method currently. This classification comprehensively considers indicators such as the patient's hepatic encephalopathy status, total bilirubin, albumin level, and prothrombin time (as shown in Table 2), and divides the patient's liver function reserve capacity into three grades: A (5-6 points), B (7-9 points), and C (≥10 points). The higher the grade, the worse the prognosis. Studies have shown that the Child-Pugh classification can effectively predict the postoperative outcomes of patients with liver malignancies and end-stage liver diseases.

[0061] Table 2. Child-Pugh Classification System

[0062]

[0063] 2) Sequential Organ Failure Assessment Score Related to Sepsis

[0064] As shown in Table 3, the establishment of the SOFA score is mainly to meet the needs of the critical care medicine field for an objective and simple assessment tool that can continuously track the degree of dysfunction or failure of a single organ and can conveniently and intuitively describe the whole process from mild organ dysfunction to severe organ failure. The purpose of the SOFA score is to measure and evaluate the progression characteristics of single or multiple organ failures in clinical scenarios and accurately describe the characteristics of organ dysfunction or failure.

[0065] Table 3 Sequential Organ Failure Assessment Scores Tab.3 Sequential Organ Failure Assessment Scores

[0066]

[0067] 3) Glasgow Coma Scale

[0068] The Glasgow Coma Scale (GCS), that is, the Glasgow Coma Scoring Method (see Table 4 for details), sets corresponding scoring criteria for eye opening, language, and movement respectively, and represents the degree of the patient's consciousness disorder by the sum of the three. The highest score of the GCS is 15 points, indicating clear consciousness. A score below 8 is coma, and the lowest score is 3 points.

[0069] Table 4 Glasgow Coma Scale

[0070]

[0071] 4) Model for End - Stage Liver Disease Score

[0072] The Model for End - Stage Liver Disease Score (MELD score) is a classic system for evaluating the urgency of liver transplantation in patients with chronic liver disease. The MELD score has been extended to evaluate the severity of all patients with end - stage liver disease. Currently, the calculation of the MELD score follows the latest Kamath version formula, which is as follows:

[0073] MELD score = 3.8 * In[bilirubin (mg / dl)] + 11.2 * In(INR) + 9.6 * In[creatinine (mg / dl)] + 6.4 * (etiology: 0 for biliary or alcoholic, 1 for others).

[0074] 5) Triglyceride - Glucose Index

[0075] An increasing number of clinical studies have shown that the TyG index is an indicator that can reliably evaluate the insulin status of individuals at high risk, especially in predicting high - risk groups of diabetes, and it performs better than traditional indicators. Subsequently, other related studies have also successively revealed the potential relationship between the TyG index and various cardiovascular diseases. The calculation formula of the TyG index is:

[0076] TyG index = In[fasting triglyceride (mg / dl) * blood glucose (mg / dl)] / 2.

[0077] 6) Body Mass Index

[0078] The body mass index (BMI) is a common indicator for measuring the obesity degree of patients, which is commonly seen in routine physical examinations and the collection of basic information of in - patients. However, in recent years, an increasing number of studies have also begun to focus on the relationship between BMI and peri - operative complications; in addition, a retrospective study of 18,000 surgical patients conducted by Liu et al. showed that preoperative underweight increases the risk of in - hospital death after craniotomy for brain tumors, and it is recommended that appropriate nutritional management before craniotomy may reduce the risk of all - cause death within 30 days after surgery in patients with abnormal BMI. The calculation formula of the body mass index is:

[0079] Body mass index = weight (kg) / height (m) 2 .

[0080] Example 1: The flow chart for screening the study population of the present invention is as Figure 1As shown, 39 patients with overt hepatic encephalopathy (OHE) after surgery, 28 patients who could not complete cognitive function assessment after surgery, 42 patients with postoperative stroke, and 15 patients with re - liver transplantation / multiple liver transplantations were excluded. Finally, 957 patients were included in the study cohort. Among them, 751 patients who underwent surgery during the period from January 2015 to January 2021 were used as the training cohort for the prediction model, and 206 patients from February 2021 to March 2023 were used as the temporal external validation cohort of the prediction model.

[0081] The method for constructing an early prediction model for neurocognitive dysfunction after liver transplantation in this application includes the following steps:

[0082] S1. Sample extraction for model establishment: The dataset samples are from the liver transplantation records on the big data platform of the Third Affiliated Hospital of Sun Yat - sen University (Guangzhou, China), including patients who received allogeneic liver transplantation from January 2015 to March 2023. All liver transplant recipients are registered in the Chinese Organ Transplant Response System (www.cot.org.cn). The inclusion criteria adopted are: (1) age ≥ 18 years old; (2) allogeneic liver transplantation. Patients with the following conditions are excluded from this study: (1) patients with pre - existing overt hepatic encephalopathy (OHE); (2) patients undergoing re - liver transplantation in this operation; (3) patients who cannot complete cognitive function assessment after surgery; (4) patients with postoperative stroke. Clinical sample data of patients who received allogeneic liver transplantation surgery are collected and used as the training set and the validation set respectively. Among them, the training set data is mainly used for the development of machine learning models, and the validation set data is used for the validation of machine learning models and the comparison of efficacy between different models. The allocation ratio range of the training set and the validation set is from 6:4 to 7:3.

[0083] S2. Screening of model feature variables

[0084] The features of the prediction model refer to all variable categories that are ultimately used in the machine learning model for outcome prediction. The features of the model can be continuous variables or binary or multi-class variables. Reasonable feature screening is the basis for the effectiveness of the prediction model. In this part of the study, the feature variables included in the machine learning model construction were selected by collecting univariate variables that were statistically significant and had no multicollinearity through univariate analysis. Multivariate analysis was performed on the perioperative variables with statistical significance (P<0.05) obtained from all univariate analyses of postoperative neurocognitive dysfunction after liver transplantation using binary Logistic stepwise regression. The independent risk factors with significant statistical differences (P<0.05) in the final multivariate analysis results were the modeling feature variables of the machine learning model, and then they were all included in the construction of their respective model algorithms.

[0085] A total of 162 perioperative-related indicators were included in the univariate analysis of the present invention, which can be divided into preoperative indicators, intraoperative indicators and postoperative indicators according to the perioperative period. Among them, 75 preoperative indicators and 40 intraoperative and postoperative indicators were included. According to the results of univariate analysis, among the variables related to the perioperative period (preoperative, intraoperative, postoperative) of the patients, with P<0.05 as the determination threshold for statistical significance, a total of 66 indicators were correlated with the occurrence of postoperative neurocognitive dysfunction in liver transplant patients.

[0086] (2) Multicollinearity diagnosis

[0087] The 60 perioperative-related indicators with statistical significance screened by univariate analysis were further included in the subsequent multicollinearity diagnosis, including:

[0088] 1) Preoperative indicators: ASA classification, history of alcohol consumption, preoperative fever, occult hepatic encephalopathy, coronary heart disease, acute respiratory distress syndrome, pulmonary hypertension, pulmonary nodules, acute renal insufficiency, chronic renal failure, Child-Pugh score, MELD score, liver malignancy, hepatorenal syndrome, ascites, esophageal and gastric varices, hypersplenism, hemodialysis, plasma exchange, mechanical ventilation, tracheal intubation, hemoglobin, hematocrit, white blood cell count, platelet count, alanine aminotransferase, aspartate aminotransferase, gamma-glutamyl transferase, albumin, total bilirubin, free bilirubin, glomerular filtration rate, activated partial thromboplastin time, prothrombin time, fibrinogen, international normalized ratio, triglyceride, total cholesterol, high-density lipoprotein, low-density lipoprotein, blood amylase;

[0089] 2) Intraoperative indicators: anesthesia time, donor type;

[0090] 3) Postoperative indicators: MELD score, SOFA score, hematocrit, hemoglobin, white blood cell count, platelet count, aspartate aminotransferase, glutamyl transferase, albumin, total bilirubin, blood urea nitrogen, prothrombin time, international normalized ratio, high-sensitivity C-reactive protein, procalcitonin, blood ammonia, serum osmotic pressure;

[0091] Taking VIF > 10 as the threshold for judging the possibility of multicollinearity, the main results are shown in Table 5. According to the actual statistical results and the experience of previous literature, after excluding two variables, preoperative hematocrit and postoperative hematocrit, there is no multicollinearity among the remaining 58 variables, and all can be included in the subsequent multivariate Logistics regression analysis.

[0092] Table 5 Results of multicollinearity diagnosis

[0093]

[0094]

[0095] (3) Multivariate analysis

[0096] The above 58 perioperative variables that were screened by univariate analysis and had statistical significance (P < 0.05) and no multicollinearity were included in the multivariate Logistics regression analysis. The multivariate Logistics regression analysis using the stepwise forward regression method was used to screen variables and correct relevant confounding factors. Finally, the influencing factors related to postoperative neurocognitive dysfunction after liver transplantation were obtained through screening. After screening by multivariate Logistics regression, a total of 10 perioperative indicators (variables) were finally obtained as independent risk factors for postoperative neurocognitive dysfunction complications after liver transplantation (see Table 6 for details):

[0097] Preoperative occult hepatic encephalopathy (OR = 8.755, 95% CI: 4.997 - 12.118, P < 0.001), platelet count (OR = 0.826, 95% CI: 0.715 - 0.949, P < 0.001), prothrombin time (OR = 1.522, 95% CI: 1.241 - 1.869, P = 0.001), glomerular filtration rate (OR = 0.796, 95% CI: 0.667 - 0.926, P = 0.001), serum calcium concentration (OR = 1.941, 95% CI: 1.366 - 2.770, P < 0.001), MELD score (OR = 2.109, 95% CI: 1.421 - 3.185, P = 0.001); intraoperative blood loss (OR = 1.226, 95% CI: 1.075 - 1.469, P < 0.001); postoperative SOFA score (OR = 1.799, 95% CI: 1.389 - 2.166, P = 0.001), high-sensitivity C-reactive protein (OR = 1.336, 95% CI: 1.125 - 1.578, P = 0.001), aspartate aminotransferase (OR = 1.255, 95% CI: 1.079 - 1.496, P = 0.008).

[0098] Table 6 Multivariate analysis of the PND cohort after liver transplantation

[0099]

[0100]

[0101] As one of the schemes for the global interpretation of the SHAP model, the bar chart quantifies the contribution degree (positive or negative) of each feature to the predicted outcome in the simplest and most direct way. Figure 2 These are the ten feature variables and their respective contribution degrees in a partial LR prediction model of this study. Based on the lengths of their respective bar strips, it can be intuitively reflected that the variable with the greatest contribution to the outcome prediction in this model is preoperative occult hepatic encephalopathy, followed by preoperative platelet count and postoperative SOFA score, which are the three most important prediction variables in the model; secondly, the remaining feature variables are ranked by importance as postoperative high-sensitivity C-reactive protein level, preoperative prothrombin time, preoperative glomerular filtration rate, postoperative aspartate aminotransferase, intraoperative blood loss, preoperative serum calcium concentration, and preoperative MELD score.

[0102] S3. Development of machine learning models: Use the model feature variables screened in step S2 to train the model with the 10 important variables obtained from the screening. Here, important variables refer to the variables ranked at the top. The training models are selected from one or more of the following 6 classic machine learning model algorithms: Logistic Regression (LR), multi-layer perceptron (MLP), support vector machine (SVM), random forest (RF), light gradient boosting (LGB), and extreme gradient boosting (XGB). The model training simultaneously uses the grid search method combined with the K-fold cross-validation method to select the optimal parameter combination. First, set the parameters and parameter value lists that need to be grid-searched for each model, perform cross-combinations, and select the best parameters for each model on the internal validation dataset. Calculate the average value of the evaluation metrics of the corresponding model as the model score of this parameter combination. By comparing the model scores of each parameter combination, the best parameters of each model can be finally obtained. The final hyperparameter situations of the 6 different machine learning algorithms constructed in this study are shown in Table 7.

[0103] Table 7 Details of Hyperparameters of Machine Learning Prediction Models

[0104]

[0105]

[0106]

[0107] Note: LR, logistics regression; MLP, multi-layer perceptron; SVM, support vector machine; RF, random forest; LGB, light gradient boosting; XGB, extreme gradient boosting.

[0108] The present invention compares six modeling prediction methods: Logistic Regression (LR), multi-layer perceptron (MLP), support vector machine (SVM), random forest (RF), light gradient boosting (LGB), and extreme gradient boosting (XGB).

[0109] S4. Evaluation of the effect of the prediction model: The model verifies its effectiveness through one or more of the following three methods: ① Model evaluation indicators based on the confusion matrix: All feature variables are respectively imported into the candidate machine learning prediction model algorithms to construct the prediction model. The model is trained with the data of the training set, and the internal validation and effect evaluation of the model are carried out with the data of the internal validation set. Finally, the AUC value, accuracy, sensitivity, specificity, and F1 score of each model are calculated according to the principle of the confusion matrix for a comprehensive effect evaluation; ② Model evaluation based on the calibration curve: The calibration curve of the constructed machine learning prediction model is generated and drawn based on R software. The closer it is to the optimal schematic calibration curve, the better the calibration degree of the model algorithm. ③ Model evaluation based on the decision curve: The DCA decision curve graph of the neurocognitive dysfunction after liver transplantation is drawn through the DCA decision analysis algorithm for evaluation.

[0110] 1) The effect of the six prediction models of the present invention is evaluated based on the model evaluation indicators of the confusion matrix, and the results are shown in Table 8:

[0111] Table 8 Evaluation of the effect of the PND cohort prediction model after liver transplantation

[0112]

[0113]

[0114] Note: LR, logistics regression; MLP, multi-layer perceptron; SVM, support vector machine; RF, random forest; LGB, light gradient boosting; XGB, extreme gradient boosting.

[0115] As can be seen from the results in Table 8, among the 6 types of machine learning prediction models constructed in this part of the study: the AUC values of all machine learning model algorithms are greater than 0.75, with the LR algorithm being the highest (0.799, 95% CI: 0.709 - 0.877), followed by the MLP algorithm (0.797, 95% CI: 0.708 - 0.879); the accuracy of the MLP algorithm is the highest, and the LR algorithm is slightly lower than the MLP algorithm (0.788, 95% CI: 0.722 - 0.854; 0.748, 95% CI: 0.675 - 0.821); in terms of model sensitivity, the SVM algorithm is the highest (0.725, 95% CI: 0.593 - 0.837), followed by the LR algorithm (0.715, 95% CI: 0.575 - 0.835); in terms of model specificity, the MLP algorithm is the best (0.882, 95% CI: 0.812 - 0.942), followed by the LR algorithm (0.772, 95% CI: 0.686 - 0.851); the F1 score is the harmonic mean composed of sensitivity and accuracy. Among the constructed model algorithms, the F1 score of the MLP algorithm is the highest (0.663, 95% CI: 0.537 - 0.760), and the LR algorithm is slightly lower than the MLP algorithm (0.655, 95% CI: 0.547 - 0.755). Based on this, it can be concluded that the LR algorithm and the MLP algorithm both perform relatively well in the overall model evaluation. And because the LR algorithm has higher sensitivity, LR is a relatively optimal choice for the effective recognition rate of patients with positive outcomes.

[0116] In addition, Figure 3 visually and in detail shows the ROC curves of the models constructed by 6 machine learning algorithms. The size of the area corresponding to each curve is the AUC value corresponding to the model. Among them, the area under the ROC curve of the LR algorithm is the largest, that is, the model with the highest AUC value.

[0117] 2) Based on the model evaluation based on the calibration curve, evaluate the effects of the six prediction models of the present invention. Essentially, the calibration curve is a scatter plot composed of the probability of actual events occurring and the probability of predicted events occurring. Generate and plot the calibration curves of the 6 machine learning prediction models constructed in this part of the study based on R software (see details in Figure 4 ), and the dotted line that grows evenly at 45 degrees in the figure is the optimal schematic calibration curve. The closer it is to this dotted line, the better the calibration degree of the model algorithm. It can be seen from Figure 4 that the LR algorithm is an algorithm with relatively good calibration degree among the 6 machine learning algorithms constructed in this part of the study.

[0118] 3) Model evaluation based on the decision curve

[0119] As described above, the LR algorithm demonstrated relatively the most stable predictive efficacy in terms of the model evaluation metrics (AUC value, accuracy, sensitivity, specificity, F1 score) based on the confusion matrix and the comparison results of the model calibration curve. Therefore, the LR algorithm was finally selected as the prediction model construction scheme in this study. To verify the true benefit effect of the LR algorithm in clinical prediction, this study further plotted the DCA decision curve of stroke after liver transplantation through the DCA decision analysis algorithm (see Figure 5 ), in order to further explore the applicable range of the screened LR algorithm prediction model in clinical practice. In the DCA decision curve, the black horizontal line indicates that none of the liver transplant patients predicted by the prediction model had the postoperative PND outcome, and the net benefit of the model was "0" at this time; while the gray dotted line in the curve indicates that all of the liver transplant patients predicted by the prediction model had the postoperative PND outcome. As Figure 5 can be seen, the PND prediction model for liver transplantation constructed based on the LR algorithm had good clinical net benefit in the range of 18% - 77%.

[0120] Example 2: Horizontal Comparison of the Optimal Model LR

[0121] Both the SOFA score and the MELD score are widely used scoring systems in the medical field, often used to evaluate the disease severity and prognosis of patients, especially in critically ill patients, and there have been many relevant research literature reports on their applications.

[0122] This invention analyzed the predictive effects of the two independent risk factors, the preoperative MELD score and the postoperative SOFA score, screened by the multi-factor Logistics regression algorithm, on the postoperative neurocognitive dysfunction after liver transplantation, and made a horizontal comparison with the finally constructed LR algorithm prediction model. The results of the horizontal comparison of the models are shown in Table 9, where the AUC value, accuracy, sensitivity, specificity, and F1 score of the LR model were significantly higher than those of using the SOFA score and the MELD score alone; Figure 6 Figure

[0123] is the ROC curve in the horizontal comparison model. Therefore, for the complication of postoperative neurocognitive dysfunction after liver transplantation, compared with using the SOFA score and the MELD score alone, the LR machine learning algorithm constructed in this study had relatively better predictive efficacy compared with the traditional clinical scoring system, and could more effectively assist clinicians in identifying high-risk patients and the decision-making system.

[0124]

[0125] Note: LR, logistics regression; SOFA, sequential organ failure assessment; MELD, model for end-stage liver disease.

[0126] Example 3: External Validation Based on a Temporal Cohort

[0127] In this part of the study, a dataset consisting of 206 patients who underwent liver transplantation under general anesthesia at the Third Affiliated Hospital of Sun Yat-sen University from February 2021 to March 2023 was selected as the temporal external validation cohort for the machine learning prediction model of postoperative neurocognitive disorder after liver transplantation. The results of the comparison of the model prediction efficacy between the temporal external validation set and the internal validation set for model building are shown in Table 10. The prediction model of postoperative neurocognitive disorder after liver transplantation based on the LR algorithm constructed in this study also has relatively stable model prediction performance in the temporal validation set. Among them, the AUC value of the model (0.806, 95% CI: 0.755 - 0.871), accuracy (0.806, 95% CI: 0.690 - 0.812), sensitivity (0.726, 95% CI: 0.621 - 0.801), specificity (0.743, 95% CI: 0.667 - 0.811), and F1 score (0.671, 95% CI: 0.599 - 0.756) all show stable performance consistent with the modeling results of the internal validation set. Further, the ROC curve of the temporal external validation cohort was plotted using R software (version 4.3.2) ( Figure 7 ), and the ROC curve shows that the area under the curve AUC value of this prediction model in the external validation set is 0.806, proving that the constructed prediction model of postoperative neurocognitive disorder after liver transplantation has relatively stable prediction efficacy.

[0128] Table 10 Temporal External Validation Results of the Prediction Model

[0129]

[0130] Example 4: A method for early prediction of postoperative neurocognitive disorder after liver transplantation. This method predicts the risk of postoperative neurocognitive disorder after liver transplantation through an LR model. The method includes the following steps: a) Obtain the measurement results of the following 10 indicators of liver transplant patients: occult hepatic encephalopathy before liver transplantation, platelet count, prothrombin time, glomerular filtration rate, blood calcium concentration, MELD score, intraoperative blood loss, and SOFA score, high-sensitivity C-reactive protein, and aspartate aminotransferase after liver transplantation in liver transplant patients; b) Input the measurement result parameters of the 10 indicators in step a) into the total risk score calculation formula of the LR model to obtain the total risk score of postoperative neurocognitive disorder after liver transplantation, and then calculate the risk prediction value of the LR model according to the total risk score. The calculation formulas for the total risk score and the risk prediction value are as follows:

[0131] Total risk score = (-0.578) + (1.721) * Xa + (-0.719) * Xb + (-0.315) * X c + (0.393) * X d + (0.257) * X e + (0.17) * X f + (0.364) * X g + (0.382) * X h + (0.34) * X i + (0.335) * X j ;

[0132]

[0133] where, X a represents whether there is CHE (occult hepatic encephalopathy) before surgery, 0 for no, 1 for yes; X b represents the last PLT (platelet) before surgery, and the unit is 10 9 / L; X c represents eGFR (glomerular filtration rate), and the unit is ml / (min * 1.73m 2 ); X d represents PT (prothrombin time) within 1 day after surgery, and the unit is s; X e represents the last Ca 2+ (serum calcium ion) before surgery, and the unit is mmol / L; X f represents the MELD score; X g represents EBL (intraoperative estimated blood loss), and the unit is mL; X h represents the SOFA score; X i represents AST (aspartate aminotransferase) within 1 day after surgery, and the unit is U / L; X jIndicates the first postoperative hsCRP (high-sensitivity C-reactive protein), with the unit of mg / L.

[0134] Example 5: Online Web Calculator for the Risk Prediction Model of Neurocognitive Dysfunction after Liver Transplantation

[0135] The prediction model of neurocognitive dysfunction after liver transplantation based on the LR algorithm constructed in this research part showed relatively good prediction efficacy after model evaluation, internal validation, external validation with multiple datasets, and horizontal comparison with traditional scores. Since the 10 important indicators included in the LR model of the present invention can be obtained in clinical practice, facilitating the calculation of the risk of neurocognitive dysfunction after LT, the present invention also developed an online risk calculator according to the method of predicting the risk of neurocognitive dysfunction after liver transplantation by the LR model. The algorithm of the online risk calculator is fixed. This online risk calculator allows anesthesiologists and peers around the world to use this model. Obtain the measurement results of the following 10 indicators of liver transplant patients: occult hepatic encephalopathy before liver transplantation, platelet count, prothrombin time, glomerular filtration rate, blood calcium concentration, MELD score, intraoperative blood loss, and SOFA score, high-sensitivity C-reactive protein, and aspartate aminotransferase after liver transplantation of liver transplant patients; input the result parameters into the online risk calculator to obtain the predicted value of the risk of neurocognitive dysfunction after LT of liver transplant patients.

[0136] The page situation of the web calculator is as Figure 8 、 Figure 9 shown. When using it, input the 10 perioperative data values of the patient to be predicted into the corresponding dialog box, and click the "Calculate" button to generate the prediction result. The result output format is "Positive" (positive), "Negative" (negative), and the corresponding predicted occurrence probability. Figure 8 、 Figure 9 respectively show the prediction process and results of two example patients:

[0137] As Figure 8 shown, this patient (Example 1) had preoperative occult hepatic encephalopathy (fill in Yes if yes, No if no), preoperative MELD score was 34 points, preoperative platelet count was 13*10 9 / L, preoperative prothrombin time was 23.5 s, preoperative blood calcium concentration was 2.76 mmol / L, preoperative glomerular filtration rate was 54.2 mL / min, intraoperative blood loss was 3000 mL, postoperative SOFA score was 10.9 points, postoperative aspartate aminotransferase was 1530 U / L, and postoperative high-sensitivity C-reactive protein was 33.4 mg / L. The calculation result of the prediction model was 0.94 (94%), and the prediction result was "positive", indicating that the patient had a relatively high risk of postoperative neurocognitive dysfunction;

[0138] As Figure 9 shown, the patient (Example 2) did not have occult hepatic encephalopathy before surgery, the preoperative MELD score was 11, the preoperative platelet count was 81×10 9 / L, the preoperative prothrombin time was 16.6 s, the preoperative serum calcium concentration was 2.23 mmol / L, the preoperative glomerular filtration rate was 105.6 mL / min, the intraoperative blood loss was 1845.9 mL, the postoperative SOFA score was 11, the postoperative aspartate aminotransferase was 1031.3 U / L, the postoperative high-sensitivity C-reactive protein was 2 mg / L, the calculation result of the prediction model was 0.17 (17%), and the prediction result was "negative", indicating that the risk of postoperative neurocognitive disorder in this patient was relatively low. The result calculated by the early prediction method for postoperative neurocognitive disorder after liver transplantation in Example 4 was consistent with the result of the online risk calculator.

[0139] Example 6: The present invention provides a system for constructing an early prediction model for postoperative neurocognitive disorder after liver transplantation. The system includes one or more computer processors and a computer-readable medium. The computer-readable medium stores a plurality of instructions, and the plurality of instructions instruct the one or more computer processors to execute the above-mentioned method for constructing a prediction model for neurocognitive disorder.

[0140] Example 7: The present invention provides an early prediction system for postoperative neurocognitive disorder after liver transplantation.

[0141] The system includes an input device, a processor, and a computer-readable medium; the input device is used to obtain the measurement results of relevant detection indexes of the measured liver transplantation patient; the processor is connected to the input device, and the processor is used to process the data obtained by the input device and output a predicted value of the risk of neurocognitive disorder; the computer-readable medium stores a plurality of instructions, and the plurality of instructions instruct the input device and the processor to execute the prediction method for postoperative neurocognitive disorder after liver transplantation; the method includes the following steps: a) obtaining the measurement results of the following 10 indexes of the liver transplantation patient: occult hepatic encephalopathy before liver transplantation surgery, platelet count, prothrombin time, glomerular filtration rate, serum calcium concentration, MELD score, intraoperative blood loss, and SOFA score, high-sensitivity C-reactive protein, and aspartate aminotransferase after liver transplantation of the liver transplantation patient; b) inputting the measurement result parameters of the 10 indexes in step a) into the total risk score calculation formula to obtain the total risk score of postoperative neurocognitive disorder after liver transplantation, and then calculating a risk prediction value according to the total risk score. The calculation formulas for the total risk score and the risk prediction value are as follows:

[0142] Total risk score = (-0.578) + (1.721) * X a + (-0.719) * Xb + (-0.315) * X c + (0.393) * X d + (0.257) * X e + (0.17) * X f + (0.364) * X g + (0.382) * X h + (0.34) * X i + (0.335) * X j ;

[0143]

[0144] wherein, X a represents whether there is CHE (cryptogenic hepatic encephalopathy) before surgery, 0 means no, 1 means yes; X b represents the last PLT (platelet) before surgery, and the unit is 10 9 / L; X c represents eGFR (glomerular filtration rate), and the unit is ml / (min * 1.73m 2 ); X d represents PT (prothrombin time) within 1 day after surgery, and the unit is s; X e represents the last Ca 2+ (serum calcium ion) before surgery, and the unit is mmol / L; X f represents MELD score; X g represents EBL (estimated blood loss during operation), and the unit is mL; X h represents SOFA score; X i represents AST (aspartate aminotransferase) within 1 day after surgery, and the unit is U / L; X j represents the first postoperative hsCRP (high-sensitivity C-reactive protein), and the unit is mg / L.

[0145] The prediction results of the patients obtained by using the system and the online risk calculator developed based on the method for predicting postoperative neurocognitive dysfunction after liver transplantation are consistent.

[0146] Example 7: Another aspect of the present invention is to provide the application of detection reagents and detection instruments for detecting 10 indicators in the preparation of products or kits for establishing an early prediction model for postoperative neurocognitive dysfunction after liver transplantation or for preparing a method for predicting postoperative neurocognitive dysfunction after liver transplantation; these 10 indicators are respectively: cryptogenic hepatic encephalopathy before liver transplantation surgery of liver transplantation patients, platelet count, prothrombin time, glomerular filtration rate, blood calcium concentration, MELD score; blood loss during surgery; and SOFA score, high-sensitivity C-reactive protein, aspartate aminotransferase after liver transplantation of liver transplantation patients. Wherein the detection reagents and instruments are detection reagents and instruments routinely used clinically.

[0147] This application mainly focuses on the statistical operation platform and software

[0148] 1) R software (version 4.3.2);

[0149] 2) Python (version 3.11);

[0150] 3) PostgreSQL (version 15.3);

[0151] 4) Navicate Premium (version 16.0);

[0152] 5) XGBoost Package (version 0.24.1);

[0153] 6) LightGBM Package (version 3.1.3);

[0154] 7) Scikit-learn Package (version 1.3.3);

[0155] 8) SHAP (version 0.40.3);

[0156] The above is only a detailed description of the preferred embodiments and principles of the present invention. For those of ordinary skill in the art, according to the idea provided by the present invention, there will be changes in the specific implementation manners, and these changes should also be regarded as the protection scope of the present invention.

Claims

1. A prediction system for neurocognitive dysfunction after liver transplantation, the system comprising an input device and a processor; and a computer-readable medium, the computer-readable medium storing a plurality of instructions, the input device being used to obtain the measurement results of the relevant detection indicators of the tested liver transplant patient; the processor being connected to the input device, the processor being used to process the data obtained by the input device and output the predicted value of the risk of neurocognitive dysfunction; the instructions instructing the input device and the processor to execute a method for predicting neurocognitive dysfunction after liver transplantation; the method comprising the following steps: a) obtaining the measurement results of the liver transplant patient; b) inputting the measurement result parameters of the 10 indicators in step a) into the total risk score calculation formula to obtain the total risk score of neurocognitive dysfunction after liver transplantation, and then calculating the risk prediction value according to the total risk score. The calculation formulas of the total risk score and the risk prediction value are as follows: Total risk score = (-0.578) + (1.721) * X a +(-0.719)*X b +(-0.315)*X c +(0.393)*X d +(0.257)*X e +(0.17)*X f +(0.364)*X g +(0.382)*X h +(0.34)*X i +(0.335)*X j ; X a Indicates whether there is CHE (latent hepatic encephalopathy) before surgery, 0 for no, 1 for yes; X b Indicates the last PLT (platelet) before surgery, the unit is 10 9 / L;X c Represents eGFR (glomerular filtration rate), the unit is ml / (min*1.73m 2 );X d It represents PT (prothrombin time) within 1 day after surgery, in seconds; X e The last Ca 2+ (Serum calcium ion) concentration, unit is mmol / L; X f represents MELD score; X g Indicates EBL (estimated intraoperative blood loss), unit: mL; X h represents SOFA score; X i Indicates the AST (aspartate aminotransferase) concentration within 1 day after surgery, in U / L; X j It represents the first hsCRP (high-sensitivity C-reactive protein) measurement value after surgery, in mg / L.

2. A method for constructing a prediction model for neurocognitive dysfunction after liver transplantation, characterized in that: The prediction model is an early prediction model, and the method comprises the following steps: S1. Sample extraction for model establishment: Clinical sample data of patients undergoing allogeneic liver transplantation were collected, and the collected samples were used as training sets and validation sets respectively. The training set data was mainly used for the development of machine learning models, and the validation set data was used for the validation of machine learning models and the comparison of the performance between different models. The distribution ratio of the training set and the validation set ranged from 6:4 to 7:3; S2. Model characteristic variable screening: The characteristic variables included in the machine learning model are collected by collecting variables that are statistically significant and free of multicollinearity through univariate analysis. Finally, the independent risk factors closely related to neurocognitive dysfunction after liver transplantation that are included in multivariate logistics regression and screened based on stepwise regression are used as the final characteristic variables for modeling. S3. Development of machine learning model: Using the model feature variables obtained in step S2, the model feature variables screened are subjected to model training, wherein the training model is selected from one or more of six classic machine learning model algorithms, including logistic regression algorithm LR (Logistic Regression), multi-layer perceptron (MLP), support vector machine (SVM), random forest (RF), light gradient boosting (LGB), and extreme gradient boosting (XGB); the training model simultaneously uses a grid search method combined with a K-fold cross-validation method to select the optimal parameter combination, first setting the parameters and parameter value lists that need to be grid searched for each model, performing cross-combinations, and selecting the optimal parameters for each model on an internal validation data set.

3. The method for constructing a neurocognitive dysfunction prediction model according to claim 2, characterized in that: It also includes the effect evaluation step of the prediction model. The model verifies its effectiveness through one or more of the following three methods: ① model evaluation indicators based on confusion matrix; ② model evaluation based on calibration curve; ③ model evaluation based on decision curve.

4. The method for constructing a neurocognitive dysfunction prediction model according to claim 2 or 3, characterized in that: The method also includes a step of evaluating the effect of the prediction model, wherein the effect of the model is evaluated using one or more of AUC, accuracy, sensitivity, specificity or F1 score as evaluation indicators.

5. A system for constructing a neurocognitive dysfunction prediction model, the system comprising one or more computer processors and a computer-readable medium, wherein the computer-readable medium stores a plurality of instructions, wherein the plurality of instructions instruct the one or more computer processors to execute the method for constructing a neurocognitive dysfunction prediction model as described in any one of claims 2-4.

6. Use of a detection reagent and a detection instrument for detecting 10 indicators in the preparation of a product or a kit for establishing an early prediction model for neurocognitive dysfunction after liver transplantation or a method for predicting neurocognitive dysfunction after liver transplantation, characterized in that: The 10 indicators are: latent hepatic encephalopathy, platelet count, prothrombin time, glomerular filtration rate, blood calcium concentration, MELD score of liver transplant patients before liver transplantation; blood loss during surgery; and SOFA score, high-sensitivity C-reactive protein, and aspartate aminotransferase of liver transplant patients after liver transplantation.

7. A kit for predicting the risk of neurocognitive dysfunction after liver transplantation, characterized in that: The kit contains detection reagents and detection instruments for detecting 10 indicators, and the 10 indicators are: Latent hepatic encephalopathy, platelet count, prothrombin time, glomerular filtration rate, blood calcium concentration, MELD score of liver transplant patients before liver transplantation; blood loss during surgery; and SOFA score, high-sensitivity C-reactive protein, and aspartate aminotransferase of liver transplant patients after liver transplantation.

Images