2.1 Patients
A total of 1505 patients diagnosed with cirrhosis were retrospectively enrolled at the Beijing Ditan Hospital of Capital Medical University in Beijing, China, from January 2011 to January 2016. Eligibility criteria included being a first-time diagnosis of cirrhosis, age between 18 and 80 years, and confirmation of cirrhosis through liver biopsy and/or compatible clinical, laboratory, and imaging data. Patients who died within a 3-year or 5-year period or were lost to follow-up were excluded. Exclusion criteria encompassed individuals with known hepatocellular carcinoma (HCC), pregnant women, those who have undergone previous orthotopic liver transplantation (OLT), unwillingness to provide informed consent, and use of anticoagulation therapy or prior surgical or transjugular intrahepatic portosystemic shunt (TIPS) procedures. To ensure a representative sample, 986 patients were randomly assigned, with 70% (n = 685) allocated to the training cohort and the remaining 30% (n = 301) assigned to the validation cohort (Fig. 1). This study received ethical approval from the Ethics Committee of Beijing Ditan Hospita
2.2 Clinical definition and follow-up
Compensatory cirrhosis was determined through the following methods: (1) On biopsy, the presence of pathological findings indicating F4 stage cirrhosis; (2) During endoscopy, the presence of esophageal varices and exclusion of noncirrhotic portal hypertension; (3) In the absence of histological evidence and endoscopic findings, at least two out of three criteria should be met:①Imaging techniques such as ultrasonography, computed tomography, or magnetic resonance imaging show changes in liver morphology, such as nodules in the liver tissue and uneven texture on the liver surface; ②Platelet count lower than 100×109 cells/L, without any other identifiable causes; ③Serum albumin lower than 35.0g/L, international normalized ratio higher than 1.3, or prothrombin time prolonged by more than 3 seconds. The diagnosis of decompensated cirrhosis is based on the presence of cirrhosis along with complications related to portal and venous hypertension and/or liver dysfunction. (1) The diagnosis requires evidence of cirrhosis; (2) Presence of complications associated with portal hypertension such as ascites, bleeding from esophageal and gastric varices, and hepatic encephalopathy (22). PVT (portal vein thrombosis) is diagnosed by the visualization of non-tumoral thrombosis within the portal vein or its branches. Such diagnosis and assessment of its extent are always confirmed through computed tomography or magnetic resonance imaging. PVT is categorized as occlusive if there is a complete absence of blood flow in the vein, and partial if the lumen is only partially occluded while blood flow is still present
The starting point of this research was the initial cirrhosis diagnosis at the hospital. The conclusion of this study focused on the recent identification of PVT within a one-year span and the subsequent follow-up of five years. The collection of clinical data encompassed various categories such as demographics (age, gender), complications (bleeding gastrointestinal varices, ascites, hepatic encephalopathy [HE]), biochemical indicators (alanine aminotransferase [ALT], aspartate aminotransferase [AST], total bilirubin [TBIL], serum albumin [ALB], ɣ-glutamyl transpeptidase [GGT], white blood cell count [WBC], neutrophil count [NC], lymphocyte count [LC], platelet count [PLT], creatinine [CREA], prothrombin time [PT], international normalized ratio [INR], Alpha-fetoprotein [AFP]), and routine laboratory tests which involved computed tomography or magnetic resonance imaging, conducted every 3–6 months
2.3 Construction of ANN
The artificial neural network (ANN) is composed of complex and interconnected processing units called neurons. These neurons are connected through weighted connections and organized into an input layer, output layer, and one or more hidden layers (23, 24). Some advantages of using an ANN include its ability to self-learn, self-adapt, and perform inference processes. The ANN learns by examining examples and adjusting the weights of the connections between neurons to establish a relationship between input and output. When applied to data, the input is passed through the layers of neurons until an output is generated. Following this output, a process of self-adaptation occurs. The produced output is compared to the desired output, and if there is a difference, an error signal is generated. This error signal is then used in a back propagation (BP) method to modify the weights of the connections between neurons. This modification aims to minimize the overall error of the network. Throughout the learning process, the error between the produced outputs and the desired outputs gradually decreases until it reaches a minimum, indicating convergence of the network. Once this convergence is achieved, the ANN can perform an inference process. During this process, new input data can be used to generate outputs or predictions based on the knowledge gained during the training process. This allows the ANN to accurately predict outcomes on different data sets (25, 26, 27)
In this investigation, the 5-year progression of PVT in the 637 layers of cirrhosis input contained neurons that imported the available data, encompassing a variety of clinical, demographic, and laboratory information. The output layers consisted of neurons that exported the corresponding predictive results. The hidden layers were utilized to facilitate intricate interactions between the input and output neurons. Factors significantly linked to PVT in patients with cirrhosis were used to construct artificial neural networks (ANNs) employing Mathematica 11.1.1 for Microsoft Windows (64-bit), a graphical tool for neural network development. A total of 912 patients were assigned to either a training group (n = 637, 70%) or a validation group (n = 275, 30%). The backpropagation (BP) algorithm was employed in the learning process of this ANN, which involved assessing the errors between the generated and desired outputs. The connections between neurons were adjusted by modifying the weights to minimize the overall network errors. The learning (training) process would be terminated once the sum of squared errors reached a minimum compared to the cross-validation dataset. Ultimately, the final model provided the development of PVT risks over a 3/5-year period for each patient.
2.4Statistical analysis
The data were expressed in the form of median (range) or n (%) as applicable. To assess the statistical significance of differences among continuous and categorical variables, we employed Student's t-test (or Mann-Whitney test if appropriate) and chi-squared test (or Fisher's exact test if appropriate). Once the relationship between demographic, biochemical, and clinical variables (inputs) and prognosis (outputs) was determined, we selected the variables with statistically significant differences or important clinical characteristics as the input layers for constructing artificial neural networks (ANNs) to predict the development of portal vein thrombosis (PVT) in patients over a period of 3 years and 5 years. We presented hazard ratios (HR) and their corresponding 95% confidence intervals (CI), along with p values. To evaluate the discriminatory performance, we utilized receiver-operating characteristic (ROC) curves. The area under the ROC curve was computed to generate Harrell's c-index. Furthermore, we compared the performance of the ANN model with that of Model for End-Stage Liver Disease (MELD) in the ROC curves [28,29,30]. The scores for MELD were calculated based on the published scoring formula. To visually assess the agreement between the predicted probability of PVT over 3/5 years by the model and the observed probability, we employed a calibration plot. Additionally, we conducted decision curve analysis (DCA) to compare the clinical net benefits of the new model compared to previous models. For all statistical tests, a p-value of less than 0.05 was considered to indicate a statistically significant difference. We performed the statistical analysis using SPSS 22 (IBM, Armonk, NY, USA) and R version 3.3.2 (R core development team, 2010).