Development and external validation of prognostic scoring models for portal vein thrombosis: a multicenter retrospective study

Background Portal vein thrombosis is a common complication of liver cirrhosis and hepatocellular carcinoma; however, few studies have reported its long-term clinical prognosis. This study aimed to establish and validate easy-to-use nomograms for predicting gastrointestinal bleeding, portal vein thrombosis resolution, and mortality of patients with portal vein thrombosis. Methods This multicenter retrospective cohort study included 425 patients with portal vein thrombosis who were divided into training (n = 334) and validation (n = 91) sets. Prediction models were developed using multivariate Cox regression analysis and evaluated using the consistency index and calibration plots. Results Predictors of gastrointestinal bleeding included a history of gastrointestinal bleeding, superior mesenteric vein thrombosis, red color sign observed during endoscopy, and hepatic encephalopathy. Meanwhile, predictors of resolution of portal vein thrombosis included a history of abdominal infection, C-reactive protein and hemoglobin levels, and intake of thrombolytics. Predictors of death included abdominal infection, abdominal surgery, aspartate aminotransferase level, hepatic encephalopathy, and ascites. All models had good discriminatory power and consistency. Anticoagulation therapy significantly increased the probability of thrombotic resolution without increasing the risk of bleeding or death. Conclusions We successfully developed and validated three prediction models that can aid in the early evaluation and treatment of portal vein thrombosis. Supplementary Information The online version contains supplementary material available at 10.1186/s12959-023-00455-w.

important risk factors for PVT in patients with cirrhotic livers [4]. Patients with non-cirrhotic livers also have risk factors for PVT, particularly those associated with a hypercoagulable state such as extrahepatic malignancies, myeloproliferative neoplasms, systemic lupus erythematosus, intra-abdominal infections, abdominal trauma, and intra-abdominal surgery (e.g., splenectomy) [6]. At present, the occurrence of PVT is unpredictable, and several aspects regarding its pathophysiology, prognosis, and treatment remain unknown.
Most patients with PVT do not show specific clinical manifestations and sometimes present with only mild abdominal pain, which is often ignored. In severe and persistent cases, PVT can cause liver injury, mesenteric vein embolism, intestinal perforation, gastrointestinal bleeding (GIB), recurrent thrombosis, and even death [8]. Therefore, early initiation of anticoagulation therapy is recommended for the treatment of PVT [1,2]. However, recent studies have shown that spontaneous thrombotic recanalization can be achieved in 30-50% of PVT cases associated with cirrhosis and other etiologies [9]. Whether anticoagulation and other clinical treatments may affect the prognoses of patients with PVT and how to balance the risks and benefits of treatment remain unresolved [10].
Previous studies have shown that the prognosis of PVT in patients with cirrhosis is associated with the severity of thromboembolism and liver disease [2,11]. Furthermore, the risk factors for GIB in patients with PVT include esophageal varices, superior mesenteric vein (SMV) thrombosis, and ascites [1]. However, these studies only focused on the risk factors for PVT. Additionally, laboratory analysis results or clinical symptoms alone were insufficient when assessing the prognosis. Because an accurate prognostic assessment is critical for physicians and patients, it is necessary to quantify clinical variables to achieve individualized prognosis predictions for patients with PVT. In these cases, a nomogram is an intuitive presentation form for prediction models [12]. To the best of our knowledge, no study has reported the use of prediction models or nomograms for predicting the prognosis of patients with PVT. Therefore, this study aimed to develop and validate nomogram predictive models for assessing the long-term prognoses of patients with PVT to facilitate the accurate evaluation of the clinical statuses of patients and timely initiation of appropriate treatments and interventions.

Study setting and participants
This retrospective clinical study was performed at Shanghai East Hospital, Shanghai, China and Shanghai Tongji Hospital, Shanghai, China from January 2012 to May 2021. Overall, 471 patients with PVT were enrolled; 46 patients with insufficient clinical information were excluded. The internal cohort included 334 patients from Shanghai East Hospital, and the external validation cohort consisted of 91 patients from Shanghai Tongji Hospital. Patients with insufficient clinical information were excluded from this study.
The inclusion criteria were as follows: (1) sufficient clinical information, including medical history, clinical manifestations, physical signs, and laboratory examination results; and (2) diagnosis of PVT using at least two diagnostic methods, including portal vein Doppler ultrasound, contrast-enhanced computed tomography (CT) of the abdomen, contrast-enhanced magnetic resonance imaging (MRI) of the abdomen, and portal vein angiography. The following imaging information were utilized to diagnose PVT (all with sensitivities and specificities ranging from 90 to 100%) [13]: (1) portal vein Doppler ultrasound scan showing solid isoechoic or hypoechoic material within the portal vein that either partially or completely fills the lumen and with absent or reduced portal venous flow; (2) contrast-enhanced CT or MRI scan of the abdomen showing a either a benign thrombus that is typically seen as a low-density non-enhancing defect within the portal vein or a tumor thrombus that is enhanced following administration of a contrast agent with distension of the vessel wall or intra-thrombus contrast enhancement; and (3) portal vein angiography showing either a filling defect or complete absence of contrast agent within the portal vein, and distal obstruction and dilatation, tortuosity, lengthening, and displacement of the splenic vein.
A flow chart of the screening process is shown in Fig. 1. All procedures were performed following the ethical standards of the responsible committee on human experimentation (institutional and national) and in accordance with the Helsinki Declaration of 1975 as revised in 2008. The study protocol was reviewed and approved by the Human Participants Ethics Committees of the two hospitals ([2022] Research Approval No. 012). Informed consent was submitted by all subjects when they were enrolled.

Endpoints
The primary endpoints were GIB, PVT resolution, and death. GIB includes upper and lower GIB. Resolution or progression of PVT was determined only in patients who underwent imaging assessments more than 3 months after the date of diagnosis. PVT resolution was defined as partial thrombus resorption or complete recanalization of the PV lumen, whereas PVT progression was defined as an extension of the area of portal vein embolism or extension of the thrombus to other lumens (e.g., splenic and superior mesenteric veins). Death events were defined as all-cause mortality events.

Data collection
Clinical data, which were recorded from the date of diagnosis to either the date of occurrence of the different clinical endpoints, the end of the study (July 16, 2021), or the date of the last follow-up visit, were collected. The baseline characteristics and pre-treatment endpoint data of the patients, including demographic information, body mass index, comorbidities, history of smoking, alcohol consumption status, GIB, concomitant use of medications, history of abdominal surgery and infection, laboratory and endoscopic characteristics, imaging features, treatment protocol with or without anticoagulation therapy, endoscope ligation, splenectomy, or abdominal surgery, were obtained. The occurrence of any of the three clinical endpoints was retrospectively recorded before the last visit.

Statistical analysis Univariate analysis
Continuous variables were expressed as means ± standard deviation or as median (interquartile range) based on the data distribution, and categorical data were presented as frequencies (percentages). The Cox regression model was used to conduct univariate analysis to identify potential predictors of the outcomes of PVT.

Development and validation of the prognostic model
Variables with P-values < 0.1 were included in the multivariate analysis. A multivariate Cox regression analysis was performed to develop the final prediction models. Calibration plots were generated to assess the similarity between the predicted and actual outcomes. Nomograms were used to visualize the prediction models. The concordance index (C-index) was calculated to evaluate the predictive ability of each model; a higher C-index indicated better discrimination. All statistical analyses were performed using SAS version 9.4 (SAS Institute Inc, Cary, NC, USA). The nomogram and calibration plots were constructed using R software version 3.6.0 (http:// cran.r-proje ct. org).

Demographic and clinical characteristics of patients in the internal cohort
A total of 334 patients with PVT (n = 220; men, 65.9%), with an average age of 60.8 ± 11.6 years, were enrolled in the internal cohort. We retrieved all patients' medical records during hospitalization, which ranged from 3 days to 110. 3      A total of 187 patients underwent assessments to determine progression or resolution of PVT. PVT progressed, resolved, and was unchanged in 39 (20.9%), 85 (45.5%), and 63 (33.7%) patients, respectively; the mortality rate of PVT progression was 2.6% (1/39). One patient who did not receive anticoagulation therapy had a fatal intestinal infarction.

Development of the prognostic nomogram
The univariate predictors of PVT resolution events are shown in Table 2, whereas the univariate predictors of other clinical endpoints are outlined in Supplementary Tables 1 and 2. After purposefully removing nonsignificant variables from the multivariate analysis, we identified predictors for the three clinical endpoints (Table 3) 19-3.27), and ascites (HR = 1.95, 95% CI = 1.09-3.51). The final Cox regression models were presented as nomograms incorporating the aforementioned independent predictors (Fig. 2). The model was applied as follows: the total scores were obtained by calculating the sum of the scores of the predictors, and the scores were used to predict the incidence of the three clinical endpoints at 1, 3, and 5 years, or at 1, 2, and 3 years.

Validation and clinical use of the nomogram
During internal validation, the C-indices of GIB, PVT resolution, and mortality were 0.69, 0.73, and 0.75, respectively (Table 4). During external validation, the C-indices of the three clinical endpoints were 0.82, 0.68, and 0.78, respectively, which was mostly better than the C-indices of the internal cohort and indicated the satisfactory diagnostic power of the nomogram models.
The results of the calibration plots for the nomograms showed consistency between the predicted and observed values (Fig. 3). The X-axis represents the occurrence of different clinical outcomes as predicted by the nomogram, whereas the Y-axis represents the actual outcomes. The red line represents the perfect prediction of an ideal model. The black line represents the performance of our nomogram model. The closer the black line is to the gray line, the better the predicted value.
The predictive power of the nomograms can be determined through practical applications. For example, regarding the calibration curve of the nomogram prediction model for GIB events, the model will underestimate the risk of GIB when the predicted probability of GIB ranges from 60 to 75%. Therefore, appropriate treatment for patients with PVT may be deferred in clinical settings. However, the prediction is remarkably close to the actual outcome during this period, which makes no difference in clinical settings. When the prediction of GIB is between 75 and 80%, the nomogram model will overestimate the risk of GIB, which indicates the need for the timely initiation of treatment for patients at high risk for bleeding. Therefore, the implementation of this nomogram model has strong clinical value.

Discussion
We comprehensively analyzed the long-term clinical outcomes of patients with PVT, identified predictors of the three clinical outcomes, and established three simple, intuitive, and rapid prognostic prediction models. Additionally, we evaluated the effects of anticoagulation therapy on PVT. The results indicated that anticoagulation therapy may increase the rate of recanalization without increasing the risk of GIB and death.
Recently, the incidence of PVT, which is estimated to range from 1.6 to 15.8% in patients with cirrhosis or portal hypertension, has gradually increased due to improvements in diagnostics brought about by advancements in imaging modalities [4]. PVT increases the risk of portal vein hypertension and related complications, such as bleeding, thrombus progression, and death [8,14,15]. Therefore, monitoring the prognosis of patients with PVT is crucial for clinical decision-making. According to recent studies, the incidence rates of bleeding, PVT  [14][15][16][17]. In the present study, the overall incidence rates of GIB and death were 9.9, and 15.6%, respectively, while 45.5 and 20.9% of patients who have been evaluated for morphological change of PVT were observed to have PVT resolution and progression, respectively, which are similar to the previously reported rates.
Previous studies have confirmed that anticoagulation therapy, an interval of less than 6 months between a diagnosis of thrombosis and initiation of anticoagulation therapy, and splenic thickness may be positively associated with portal vein recanalization [7,18,19]. Anticoagulation therapy is a crucial treatment option for patients with PVT; however, it was rarely implemented previously as clinicians and patients were concerned about the risk of complications such as GIB [16]. Moreover, among the patients who were evaluated for morphological PVT changes in the present study, the PVT recanalization rate (45.5%) was higher than the proportion of patients who received thrombolytic therapy (38.0%), suggesting that spontaneous recanalization occurred in a small number of patients with PVT, similar to the findings of previous studies [6,17]. Notably, portal vein hypertension is a predictor of nonresponse to anticoagulation therapy [9,20]. In our study, 63.3% of patients treated with anticoagulation therapy achieved thrombus resolution; this result is consistent with those of previous studies (30-80%) [8,21] and also confirmed that anticoagulation therapy was a significant predictor of PVT resolution. Therefore, we suggest that most patients with PVT should receive anticoagulation therapy unless there is a high risk of bleeding.
Hemoglobin is essential for maintaining cellular bioenergetic homeostasis and modulating cell functions (inflammation and redox status of cells) through its ability to bind and transport oxygen to tissues, which may  decrease the incidence of thrombosis [22]. However, excessively high levels of hemoglobin can lead to local inflammation and even tissue damage among patients with hemoglobinemia, which further accelerates the formation of thromboses [23]. The C-reactive protein level is a common indicator of inflammation and infection and is often used to assess their severity [24]. Darzi et al. demonstrated that a C-reactive protein level > 10 mg/L was positively associated with venous thromboembolism and that it could lead to a transient hypercoagulable state [25]. Additionally, the 2020 Chinese consensus regarding PVT indicated that inflammation or infection of the abdominal cavity may be an important risk factor for PVT in patients with cirrhosis [3]. PVT may be a potential consequence of any inflammatory intra-abdominal process (including cholecystitis, pancreatitis, or inflammatory bowel disease), and its risk will increase in a setting of acute infection and recurrent infections [26]. Therefore, patients with a higher C-reactive protein level and a history of abdominal infection may have higher levels of inflammation, which lowers the probability of PVT resolution. Portal vein hypertension is the main determinant of esophagogastric variceal bleeding [11]. Previous studies have demonstrated that esophageal varices, the red color sign observed during endoscopy, advanced stage of liver disease (Child-Turcotte-Pugh class C patients), and ascites were possible predictors of bleeding events [16,27]. Although the same conclusions were reached in the present study, we also demonstrated that a history of GIB and thrombus extension into the mesenteric veins were significant predictors of GIB. Portal vein hypertension results in redistribution and increased blood flow through the short gastric and coronary veins, causing esophagogastric varices. Esophagogastric varices begin to form at a pressure gradient of 8-10 mmHg, and bleeding risk increases at a gradient of at least 12 mmHg [27]. Certain endoscopic variceal stigma, collectively referred to as "red color sign" (red-whale markings, nipple symptoms, cherry-red spots), correlated with a significantly  higher risk of acute variceal bleeding and re-bleeding [28]; hence, early preventive endoscopic treatment and shortening of the prothrombin time may decrease the occurrence of GIB [3]. Additionally, when hepatic encephalopathy occurs in patients with advanced liver disease due to liver failure, and imaging also shows PVT extension into the superior mesenteric vein, further reductions in the flow velocity and increases in the portal vein pressure, even GIB, may occur. In the long-term, approximately 70% of patients with GIB may experience further variceal bleeding because of superficial varices and a thinner vessel wall [29]. Therefore, a history of GIB may increase the risk of re-bleeding. These new findings may help clinicians identify patients at high risk for GIB, and they may also facilitate the timely initiation of anticoagulation therapy. Patients with cirrhosis or hepatocellular carcinoma are in a state of imbalanced coagulation function that can promote the propensity for bleeding or thrombosis [4,30], thus making it challenging for clinicians to initiate anticoagulation therapy for PVT. Mohan et al. reported that the incidence of GIB for patients receiving anticoagulation therapy was 7.8% [31], which is very close to the rate (8.8%) observed in the present study. Furthermore, the univariate analysis performed in the present study showed that anticoagulation therapy did not promote GIB events, thus demonstrating the safety and efficacy of anticoagulation therapy for PVT. Therefore, our conclusion is consistent with those of existing studies. Qi et al. confirmed that prophylactic anticoagulation therapy for deep venous thrombosis in hospitalized patients with cirrhosis and without active bleeding was safe and did not increase the incidence of GIB or death [18]. Furthermore, Ageno et al. found that the duration of anticoagulation therapy was associated with a reduced risk of bleeding [21]. However, this variable was not statistically significant in the univariate analysis for GIB events in the present study, which may be due to its dual effect on GIB. Prolonging anticoagulation therapy will increase the risk of bleeding by preventing coagulation. Simultaneously, anticoagulation therapy may reduce the severity of esophagogastric varices due to PVT resolution, which decreases the incidence of GIB.
Regarding the analysis of the overall death of patients in our study, most of the fatal events were related to the underlying disease. In one case, the progression of PVT into the superior mesenteric vein caused intestinal obstruction, resulting in death. In a large prospective study of 178 patients with PVT, few deaths occurred during follow-up, and the 5-year survival rate was 96% [11]. Moreover, overall 5-year survival rates ranging from 70 to 78% have been reported by a previous study [10]. In the present study, 23.1% of all deaths were caused by fatal GIB events; however, GIB caused only 3% of the deaths in a study of 120 non-cirrhotic patients with PVT [32]. Most patients in our study had cirrhosis, hepatocellular carcinomas, and coagulation disorders, which led to a higher mortality rate. Significantly relevant factors for PVT-related mortality events include the Child-Turcotte-Pugh score, age, and ascites [32]. In partial agreement with the results of previous studies, the univariate analysis performed during the present study confirmed that the Child-Turcotte-Pugh score was related to survival (P = 0.0002). Because multiple variables of the Child-Turcotte-Pugh score may interact with each other during statistical analyses, these variables were analyzed separately in our study. We found that hepatic encephalopathy and ascites were independent predictors of PVT progression in the Cox regression analysis of death. Both hepatic encephalopathy and ascites are important signs of liver failure, which has a high risk for mortality and allows for a more accurate assessment of a patient's prognosis. In addition, abdominal infections, history of abdominal surgery, and aspartate aminotransferase level > 35 U/L were positively associated with death in the present study. Intra-abdominal infections and surgery pose serious clinical challenges, and may result in wide variety of conditions ranging from uncomplicated cases to fulminant septic shock and multi-organ dysfunction, further increasing the risk of death [26].
This study had some potential limitations. First, the effectiveness of the data analysis was limited because of the retrospective nature of the study. Additionally, the etiology ratio of PVT was not compared with those reported by other studies to confirm the external implementation performance of the nomograms. Second, some data related to PVT (e.g., portal vein velocity and morphological changes in the thrombus) were not included in this study. Moreover, many patients underwent only one blood or endoscopic examination. Therefore, a dynamic follow-up could not be performed. Future prospective, multicenter, randomized clinical trials with larger sample sizes are needed to corroborate the findings of this study.
In conclusion, we developed three easy-to-use nomogram prediction models to evaluate the prognosis and assist with the initiation of early intervention for patients with PVT. Additionally, the results of this study suggested that most patients with PVT should undergo anticoagulation therapy. Furthermore, these findings provided evidence of the benefits and risks of anticoagulation therapy for patients with PVT, which will help clinicians balance the benefit-to-risk ratio of anticoagulation therapy.

PVT
Portal vein thrombosis GIB Gastrointestinal bleeding