A linear clinical prediction model and nomogram evaluating risk of portal vein thrombosis in liver cirrhosis

Abstract

Introduction

As one of the complications of liver cirrhosis, portal vein thrombosis (PVT) brings a significant clinical challenge. To find the predictive clinical variables evaluating the probability of PVT, a prediction model was established and a nomogram was built for visualization.

Materials and methods

Our study selected 245 cirrhosis patients admitted in First Hospital of Jilin University during August 2020 to June 2024, the majority of whom were admitted during July 2022 to June 2023. All patients were randomly separated into a training set and a validation set. Binary logistic and Lasso regression were used to establish predictive models with the occurrence of PVT events within 24 months after discharge as the clinical outcome.

Results

In discovery set, 64 (26.1%) of them had events of PVT within 24 months after their discharge. A linear model “ASP” was constructed with logarithm of alkaline phosphatase (Log₁₀ALP), splenectomy and portal vein width selected for predicting PVT risk, preformed with its AUC reached 0.793, 0.750 and 0.760 in training, internal validation and external validation set, respectively.

Conclusion

We finally screened out three clinical variables identifying high-risk PVT patients in liver cirrhosis beforehand. ALP level might be a potential clinical parameter connected to PVT, but need to be further verified.

Graphical Abstract

Supplementary Information

The online version contains supplementary material available at 10.1186/s12876-026-04939-7.

Introduction

Portal vein thrombosis (PVT) is a complication of liver cirrhosis, with the morbidity of 2 to 4 cases in 100 thousand residents [1]. The main predisposing factors include chronic liver diseases, primary or secondary hepatobiliary malignant tumor, sepsis or abdominal infection, or myeloproliferative disorders [1, 2]. PVT may present either chronically or acutely. In chronic PVT, the thrombus forms slowly it tends to located at left or right branch of portal vein, with no specific symptoms and usually detected by imaging examinations. Conversely, in acute PVT—characterized by thrombus obstructing portal venous inflow, could manifest as an acute abdominal pain, meantime other complications such as severe esophageal gastric variceal bleeding (EGVB), refractory ascites, gut necrosis and hepatic encephalopathy may take place [2].

Liver dysfunction, decreased portal vein velocity (PVV) and portal vein hypertension are independent risk factors of PVT [3, 4], however, the progression of chronic liver disease itself is not the only pathogenesis [4]. Just like pulmonary thromboembolism (PTE) and deep vein thrombosis (DVT) [5], there are shared risk factors including venous blood stasis, endothelium injury and hyper-coagulable status, collectively known as Virchow’s triad [6]. Recent studies have found some representative inflammatory factors such as interleukin-6 (IL-6) elevate in PVT, which can activate leukocytes and trigger inflammatory responses [7].

PVT can be easily diagnosed and graded by imaging examinations such as ultrasound, computerized tomography (CT) and magnetic resonance imaging (MRI) [1], meantime a series of diagnostic models driven by binary logistic [8] and machine learning [9] have been constructed to seeking out clinical parameters for further excavate mechanism of PVT formation. Nevertheless, there is still a challenge for identifying high-risk PVT patients for PVT prevention. Clinical prediction model, which has been utilized for evaluating occurrence and prognosis in diversity types of diseases, has a great potential assisting clinical physicians assess probability of disease occurrence or progression. Therefore, this study aims to develop a model for predicting PVT in patients with liver cirrhosis based on conventional clinical variables, therefore formulate a risk prediction model with a visualized format of model for initial clinical use.

Materials and methods

Patient selection

This retrospective observational study selected liver cirrhosis patients who admitted to the First Hospital of Jilin University between August 2020 and June 2024, and the majority (over 90% in the discovery cohort) of whom were admitted between July 2022 and June 2023. Meantime an external validation set using non‑duplicated samples from same center with partially overlapping time windows was collected for evaluating model efficiency, containing liver cirrhosis patients admitted during Oct, 1 st, 2023 to Jul, 31 st, 2024 in Department of Hepatology of First Hospital of Jilin University, with no patient duplication to discovery set. Inclusion criteria: (1) Diagnosed with cirrhosis according to Guidelines for the Diagnosis and Treatment of Liver Cirrhosis [10]; (2) no PVT was found in abdominal imaging examination (ultrasound, CT or MRI) at first time of admission, including portal vein trunk, left or right branches of portal vein. Exclusion criteria: (1) Patients diagnosed without liver cirrhosis (e.g., viral hepatitis, drug-induced liver injury); (2) diagnosed with PVT at first admission; (3) complicated with malignant tumor, including liver cancer; (4) complicated with chronic respiratory, circulation, kidney or hematological diseases (including chronic obstructive pulmonary disease (COPD), coronary heart disease, hypertension, chronic heart failure, chronic renal failure, leukemia, lymphoma) and current infection (e.g. pneumonia, peritonitis, urinary tract infection, etc.); (5) liver transplantation; (6) lack of necessary information including PVT outcome data. This study was conducted following the ethical principles of the Declaration of Helsinki and its subsequent amendment versions, and has been approved by Ethical Committee of First Hospital of Jilin University before initiation. Except for collecting necessary clinical variables for building prediction model, private and identity information of all patients in either discovery set or external validation set were not disclosed.

Data collection & clinical outcome

The following clinical variables were collected for statistical analysis and model construction: departure time, gender, height, weight; history of splenectomy; blood routine test, including white blood cell count (WBC), hemoglobin (HGB), hematocrit (HCT) and platelet (PLT); liver function test, including aspirate aminotransferase (AST), alanine aminotransferase (ALT), alkaline phosphatase (ALP), gamma-glutamyl transpeptidase (GGT), cholinesterase (CHE) and albumin (ALB); coagulation test, including prothrombin time (PT), activation of partial thromboplastin time (APTT), international normalized rate (INR), prothrombin activity (PTA) and fibrinogen (FBG); blood lipids, including high-density lipoprotein cholesterol (HDL), low-density lipoprotein cholesterol (LDL), triglyceride (TG), cholesterol (CE), fasting glucose and portal vein trunk width. Data of portal vein width were initially collected from CT, ultrasound or MRI were selected if the patient did not go CT examination during inpatient time. All imaging measurements along with examination reports were accomplished by physicians in radiology department or abdominal ultrasonography department in First Hospital of Jilin University and have been uploaded to medical record system. In addition, portal vein width values were not specifically reported in a few patients (approximately 10% of the discovery cohort) who obtained normal portal vein trunk width but were described qualitatively, their width were measured through original imaging archives by experienced investigators who did not know about patients’ medical history.

For all patients matched inclusion criteria, result of first collection upon admission was recorded. All matched patients in discovery cohort were randomly separated into a training set and a validation set in 4:1 rate to construct prediction model and internal validation. The outcomes of each patient were recorded through passive follow-up, that is, patients were assigned into PVT group if thrombosis was found in portal vein trunk or branches in 24 months since their departure, and the rest were assigned into non-PVT group.

Variables selection & model construction

The available variables for prediction model were screened by two main methods, binary logistic regression and Lasso regression, in which the former was conducted by a primary logistic regression with stepwise backward approach method, with a 5-fold cross validation examining correlation strength of each clinical parameter, the latter selected potential variables by addressing penalty factors to each variable and construct the most fitted Lasso model with the minimum lambda (λ) value with a 100 times Bootstrap validation. Least absolute shrinkage and selection operator (Lasso) regression was applied for variable selection and coefficient shrinkage to prevent overfitting. Lasso imposes an “L1 penalty” on the regression coefficients, forcing some coefficients becoming exactly zero and thereby performing continuous variable selection. The penalty parameter λ represents degree of shrinkage, a larger of which leads to more coefficients being shrunk to zero. Variables with non-zero coefficients at λ.min was considered selected by Lasso. The analysis was conducted by “glmnet” package with family = “binomial” and alpha = 1.

Selection criteria for variables are as follows: the intersection of variables combination with AUC decreased no more than 0.1 after 5-fold cross-validation in binary Logistic regression, and the top 10 variables after 100 times of Bootstrap verification in Lasso regression. A prediction model contained shared clinical variables between binary logistic and Lasso was subsequently established and optimized, and being prepared for following validation and visualization.

Evaluation of model accuracy and decision threshold

By using “rms” package in R software, calibration curve was utilized to evaluate the accuracy and robustness of model. Calibration curve is an intuitive chart plotted by Bootstrap method with multiple replications, which contains the deviation between predicted probability calculated by prediction model and the observed probability, therefore reveal the bias between predicted result and actual clinical outcome through main absolute error (MAE). A nomogram was subsequently accomplished for model visualization, in which every variable was scored independently for calculating total points and correspond PVT risk. The decision curve analysis (DCA) built by “rmda” package was involved for net benefit accessing, by comparing with “All-line” (providing intervention for all patients) and “None-line” (providing no intervention at all), the most favorable predicting threshold with correspond net benefit can be properly visualized.

Data management and statistical methods

All patient data in this study were searched in the Real-World Data Application Platform of the First Hospital of Jilin University. For missing data in our primary dataset, the multiple imputation conducted by “mice” package in R was addressed, in which predictive mean matching method was applied and generate 10 data sets (m = 10) and set the maximum number of iterations as 10. The Shapiro–Wilk normality test was applied for examining normal distribution. Categorical variable data were presented as percentages or ratios and were compared across groups using Chi-square test or Fisher’s exact probability method, whereas continuous variable data were presented as mean ± standard error and were compared using t-test if according normal distribution, otherwise were presented as median (interquartile range) and being compared by Wilcox’s test. The statistical analysis and graphics were processed by R software 4.3.3.

Results

Characteristics of patient’s data

A total of 336 cirrhosis patients whose diagnosis accord with criteria were successfully searched from the platform, in which 32 were already diagnosed with PVT by abdominal CT, 40 were lack of outcome data and 19 were duplicated (shown in Fig. 1). After removing excluded patients, the Shapiro–Wilk normality test showed that age (p = 0.0960), HCT (p = 0.0802) and PTA (p = 0.1221) presented normal distribution in discovery set. Univariate analysis indicated that splenectomy, blood routine test (WBC, HGB and HCT), liver enzymes (AST, ALT, ALP, GGT), blood lipids (LDL, TG) and portal vein width presented statistical difference in discovery set (p < 0.05, Table 1. Characteristics of patients in Discovery Set (Overall)). Blood routine test, liver enzymes, LDL, TG, CE and portal vein width also reached statistical difference in training set (p < 0.05, Supplementary Table 1), meanwhile ALT and FBG reached statistical difference in validation set (p < 0.05, Supplementary Table 1). The proportion and pattern of missing data is displayed in Fig. 2a, meantime a correlation analysis visualizing relationship among various clinical variables is shown in Fig. 2b.

Fig. 1.

Flowchart of matched patients in discovery cohort. Totally 1229 cirrhosis patients with complete diagnosis and laboratory test information were searched from the Real-World Data Application Platform of the First Hospital of Jilin University, and 893 patients were excluded when exclusion criteria addressed, 91 were then excluded for being diagnosed with PVT at admission, ending data missing or duplication

Table 1.

Characteristics of patients in discovery set (overall)

Varibles	Overall (n = 245)	Total		p
		PVT group (n = 64)	non-PVT group (n = 181)
Male (n, %)	169 (69.0%)	46 (71.9%)	123 (68.0%)	0.671
Age (years)	53.1 ± 10.9	53.8 ± 10.4	52.8 ± 11.0	0.414
Height (cm)	170.0 (162.0, 173.0)	168.5 (163.8, 172.3)	170.0 (162.0, 173.0)	0.981
Weight (kg)	66.5 (59.0, 76.0)	65.0 (57.8, 75.0)	67.5 (60.0, 76.0)	0.221
Splenectomy (n, %)	12 (4.9%)	7 (10.9%)	5 (2.8%)	0.023*
WBC (× 109/L)	3.62 (2.56, 5.39)	3.16 (1.84, 4.58)	4.08 (2.72, 5.70)	0.004*
HGB (× 109/L)	113.0 (80.0, 131.0)	97.0 (75.5, 123.3)	116.0 (88.0, 136.0)	0.003*
HCT (L/L)	0.324 ± 0.093	0.303 ± 0.087	0.332 ± 0.094	0.028*
PLT (× 109/L)	72.0 (47.0, 112.0)	61.5 (44.5, 94.8)	76.0 (50.0, 114.0)	0.072
AST (U/L)	40.0 (27.4, 69.9)	31.6 (21.6, 43.6)	44.9 (30.5, 79.7)	< 0.001*
ALT (U/L)	27.2 (19.2, 43.5)	20.7 (15.2, 30.3)	31.2 (21.5, 54.4)	< 0.001*
ALP (U/L)	102.8 (74.4, 144.5)	82.3 (62.2, 112.3)	110.6 (81.9, 149.6)	< 0.001*
GGT (U/L)	51.4 (25.9, 109.6)	34.4 (19.1, 65.2)	58.9 (27.4, 118.8)	< 0.001*
CHE (U/L)	3566 (2384, 4841)	3622 (2880, 4712)	3502 (2342, 4947)	0.546
ALB (g/L)	31.9 (27.7, 37.0)	34.1 (28.5, 37.0)	31.4 (27.6, 36.9)	0.145
PT (s)	14.4 (13.2, 16.4)	14.4 (13.2, 15.7)	14.4 (13.3, 16.7)	0.442
APTT (s)	28.7 (26.7, 31.9)	27.8 (25.9, 31.0)	28.9 (26.8, 32.2)	0.075
INR	1.24 (1.12, 1.42)	1.23 (1.12, 1.36)	1.25 (1.12, 1.44)	0.596
PTA (%)	68.2 ± 18.5	68.3 ± 15.4	68.1 ± 19.5	0.713
FBG (g/L)	1.75 (1.29, 2.16)	1.83 (1.40, 2.15)	1.73 (1.25, 2.16)	0.652
HDL	0.89 (0.57, 1.20)	0.96 (0.70, 1.16)	0.86 (0.53, 1.21)	0.292
LDL	1.93 (1.53, 2.52)	1.80 (1.36, 2.23)	2.07 (1.60, 2.57)	0.017*
TG	0.87 (0.66, 1.25)	0.79 (0.62, 1.03)	0.91 (0.68, 1.28)	0.023*
CE	3.26 (2.48, 4.04)	3.13 (2.41, 3.55)	3.31 (2.60, 4.10)	0.102
Glucose (mmol/L)	5.35 (4.61, 6.84)	5.44 (4.94, 7.04)	5.32 (4.54, 6.74)	0.108
Portal vein width (cm)	1.50 (1.20, 1.70)	1.60 (1.40, 1.90)	1.40 (1.20, 1.60)	< 0.001*

Significant node: p < 0.05: ‘*’

Fig. 2.

Pattern of missing data and correlation analysis. A Rectangle heatmap displays the ratio of each missing data in discovery cohort. Numbers on left side shows the number of individuals with different missing patterns. Numbers on right side shows the number of missing variables in specific missing pattern. Numbers on lower side showed the proportion of missing data in each variable. B Bubble plot displays correlations among clinical variables in this study. Purple circles indicate two variables are positively related while green represents negative. Depth of color reflects the strength of correlation (calculated by spearman correlation analysis)

Binary logistic regression with 5-folds cross validation

In order to adjust potential confounding factors and ensuring fair selection of variables, all clinical related variables were entered into our initial variable pool, except for height and weight due to their high missing proportion (9.39%) and involving subjective collection in medical records during actual clinical practice. The binary logistic regression with cross validation altogether selected splenectomy (β = 0.370, p = 0.005), HGB (β = −0.002, p = 0.054), ALB (β = 0.009, p = 0.063), ALP (β = −0.001, p = 0.023) and portal vein width (β = 0.433, p < 0.001) predicting the probability of post-discharge PVT occurrence (Supplementary Table 2), and obtained an acceptable predicting effect (AUC = 0.806, training set; AUC = 0.797, validation set). However, results of 5-folds cross validation showed an unstable effect (Mean AUC = 0.779 ± 0.108), as shown in Supplementary Table 3.

Lasso regression & bootstrap cross validation

A Lasso regression model (minimum λ = 0.1045, Fig. 3a, b) was constructed. Lasso regression screened out HGB, ALP, GGT and LDL negatively correlated to PVT risk, meanwhile ALB, splenectomy and portal vein width were positively correlated (Supplementary Table 4). A cross-validation result showed portal vein width, GGT, ALP, splenectomy, HGB, LDL and ALB all in top ten variables in Bootstrap (Supplementary Table S5).

Fig. 3.

Parameters and predicting effects of Lasso regression models. A Mean squared error of Lasso-1 under different lambda values. B Coefficients of variables in different lambda values. C C-indexes of 63 different Lasso models, in which red break line represents training set and blue line for validation set, with five rectangle color blocks added to distinguish different counts of variables. D Bar plot shows ALP levels in different pathogenesis (PBC, viral or alcohol). E Bar plot shows the correlation between ALP and PVT outcome in three pathogenesis sets

Variable necessity analysis

The five shared clinical variables (portal vein width, ALP, splenectomy, HGB and ALB) were selected as potential variables for following analysis. Next, a variable necessity analysis was addressed by removing five potential variables successively by setting the penalty factor of each variable as infinite in Lasso, in order to evaluate the proportion of each variable and compare the predicting effect among different combination of variables. The AUC displayed a decreasing trend as count of variables reduced in training set, but in validation set the AUC fluctuated visibly. Furthermore, the broken line chart also showed predicting efficiency obviously weakened when splenectomy or portal vein width were deleted, indicating creditable correlation between these two variables and PVT events (Fig. 3c, Supplementary Table 6), as result portal vein width and splenectomy were defined as major variables, while HGB, ALP and ALB were regarded as minor variables. Among 31 independent variable combinations, the C-index ranged from 0.532 to 0.806 in training set and 0.502 to 0.799 in validation set.

Perfection of the final prediction model

Our study subsequently modified the primary model (Supplementary Table 7) by adjusting variables according to statistical and clinical value of each variable. After further adjusting parameters of variables, a linear prediction model containing logarithm of ALP (A), history of splenectomy (S) and portal vein width (P) was defined as our final model, named “ASP” by combining the initials of three independent variables. Binary logistic regression showed statistical significance of log₁₀ALP [OR: 0.075 (0.012, 0.486), p = 0.007], splenectomy [OR: 6.299 (1.394, 28.460), p = 0.017] and portal vein width [OR: 14.983 (4.330, 51.849), p < 0.001]. The coefficients of each clinical variables in “ASP” model are shown in Table 2. Parameters of “ASP” model and Fig. 4e.

Table 2.

Parameters of “ASP” model

Varibles	β	Standard error	Wald z	p	OR (95% CI)
(Intercept)	−0.027	2.238	−0.012	0.990	0.973 (0.012, 78.216)
Log10ALP	−2.592	0.954	−2.718	0.007**	0.075 (0.012, 0.486)
Splenoctomy	1.840	0.769	2.392	0.017*	6.299 (1.394, 28.460)
Portal vein width	2.707	0.633	4.274	< 0.001***	14.983 (4.330, 51.849)

Significant nodes:

p < 0.001: ‘***’

0.001 ≤ p < 0.01: ‘**’

0.01 ≤ p 0.05: ‘*’

Fig. 4.

Calibration test and nomogram of “ASP” model. A Calibration curve of Lasso-1. B Calibration curve of “ASP”. C A nomogram visualizing “ASP” with its linear predict value, correlated to its total point and predicted probability. D Decision curve for “ASP” and Lasso-1, in which blue and orange bold lines represent two models respectively, while green and purple lines are for conditions of all-treat and none-treat. E Forest plot showed coefficients of each clinical parameter

Sensitivity analysis of etiology subgroup

To explore whether the inverse association between ALP and PVT was confounded by cholestatic liver disease, we performed a sensitivity analysis stratified by cirrhosis etiology (viral, alcohol-related, and cholestatic) in the training set. In all three etiological subgroups, ALP showed all negative correlations with PVT occurrence: viral hepatitis (rho = −0.320, p = 0.002), alcohol-related cirrhosis (rho = −0.287, p = 0.213), and PBC (primary biliary cholangitis) (rho = −0.458, p = 0.016), with univariate analysis in different subgroups displayed similar tendency (Fig. 3e). Notably, ALP levels differed significantly across subgroup, among which the highest levels were observed in PBC subgroup (median [IQR]: 153.1 [101.5–227.2] U/L) compared to viral (90.8 [69.9–139.6] U/L) and alcohol (114.8 [74.9–152.2] U/L) groups (p < 0.001 and p = 0.030, Wilcox’s test, respectively), with no statistic difference between viral hepatitis and alcohol-related cirrhosis group (p = 0.116, Wilcox’s test) (Fig. 3d). Despite the higher ALP levels in PBC patients, the inverse correlation remained consistent across all etiologies, although it appeared strongest in the cholestatic subgroup.

Calibration curve

Two calibration curves of Lasso-1 (MAE = 0.048) and “ASP” (MAE = 0.033) were then depicted by setting 1000 repetitions through Bootstrap method (Fig. 4a, b). Two calibration curves both displayed the tendency that apparent and bias-corrected lines are close to ideal at the beginning, then diverted towards upper left area when predicted probability is between 0.2 and 0.4, and diverted to lower right when predicted probability was over 0.4, indicating an over-estimating risk exists in high-risk area, for instance if one patient is evaluated obtaining PVT risk exceed 0.4 calculated by current model, the result should be considered as going through precise assessment but not directly define the patient as high PVT risk.

DCA curve & nomogram

The DCA curves of Lasso-1 and “ASP” depicted same tendency of decreasing, in which reached best net benefit with risk threshold around 0.3 (Fig. 4d). DCA analysis quantifies the net benefit of adopting a risk-based “intervention” strategy across a range of risk thresholds. In this study, the “intervention” was defined as going through hypercoagulability test and intensified regular imaging surveillance (e.g., abdominal ultrasound every 6–12 months) for early precaution for PVT rather than directly initiate anticoagulation therapy, given the latter’s bleeding risk in cirrhosis. The net benefit was compared against two extremes: “all intervention” (all patients receive intensified surveillance) and “none intervention” (no patient receives surveillance). The threshold probability at which the model yields the highest net benefit was identified as optimal cutoff for clinical decision-making.

A nomogram was then built for visualizing the “ASP” model. Three clinical parameters in nomogram are scored independently to predict the probability of PVT (Fig. 4c), to use the nomogram, firstly locate the value of each variable on correspond axis, then draw a vertical line upward intersect to “Points” axis to obtain three parts of point, then sum up the points on “Total Points” axis and draw a vertical line downward intersect to “Risk of PVT” axis to obtain the predicted probability. The predicted probability will reach over 30% when the total score exceeds 65, indicating liver cirrhosis patients scored over 65 points may have potential PVT risk in 24 months.

Model efficiency & external validation

We finally examined the predicting effect of “ASP” model (Table 3. Predictive performance of Lasso regression and final model; Fig. 5a-c). In training set “ASP” showed a comparative effect with Lasso-1 (AUC = 0.793 vs 0.806), with higher specificity (0.752 vs 0.593) and lower sensitivity (0.706 vs 0.961), and performed less accurate than Lasso-1 in internal validation (AUC = 0.750 vs 0.797). The external validation set contains 82 liver cirrhosis patients, in which 9 were diagnosed with PVT and 73 were without PVT after discharge (Supplementary Table 8). ROC curve analysis showed “ASP” model achieved a favorable predicting effect (AUC = 0.760) competing with Lasso-1 (AUC = 0.760), indicating a satisfied predicting effect in analyzing PVT risk.

Table 3.

Predictive performance of Lasso regression and final model

Datasets	Models	AUROC	Threshold	Sensitivity	Specificity	Positive predictive value (PPV)	Negative predictive value (NPV)
Training set (n = 196)	Lasso-1	0.806	−0.071	0.961	0.593	0.454	0.977
	“ASP” Model	0.793	−1.029	0.706	0.752	0.500	0.879
Validation set (n = 49)	Lasso-1	0.797	0.286	0.769	0.750	0.526	0.900
	“ASP” Model	0.750	−0.384	0.769	0.806	0.588	0.906
External validation set (n = 82)	Lasso-1	0.760	0.316	0.778	0.712	0.250	0.963
	“ASP” Model	0.760	−1.209	1.000	0.507	0.200	1.000

Fig. 5.

ROC curves of “ASP” and Lasso-1 model. A ROC in training set. B ROC in validation set. C ROC in external validation set. D ROC comparation between “ASP” model and nested model (“ASP” plus D-dimer)

Efficacy of combined predicting ability with D-dimer and ASP model

In the external validation, we further filtered out 50 patients (43 non-PVT, 7 PVT) with complete D-dimer data to evaluate the incremental value of D-dimer beyond the “ASP” model. The “ASP” model alone achieved an AUC of 0.718, while adding D-dimer to the fixed “ASP” linear predictor increased the AUC to 0.777 (p for difference = 0.275, Delong’s test). The continuous NRI was 0.738 (95% CI: 0.139—1.336, p = 0.016) and IDI was 0.044 (95% CI: 0.000—0.087, p = 0.047), suggesting a modest but not statistically significant improvement (Supplementary Table 9). These results should be interpreted as exploratory due to small number of events.

Discussion

As a complication of liver cirrhosis, PVT could further aggravate portal hypertension and cause cavernous transformation, even could induce other complications like EGVB, hepatic encephalopathy, acute peritonitis etc. [2], increase long-term mortality rate [1, 11], therefore earlier identifying high PVT risk patients is important for preventing liver cirrhosis progressing. Our study initially analyzed clinical characteristics in discovery cohort, then a model conducted by binary Logistic and Lasso regression eventually screened out three variables for PVT risk predicting.

As a critical independent risk factor, portal vein width showed strong positive correlation with PVT, both in binary Logistic and Lasso models, in accord with the PVT diagnostic model built by Li et al. [9]. The changes in structure of portal vein are critical factors influencing occurrence of PVT [3], recent research suggests that widen diameter of portal vein [12] and PVV beneath 15 cm/s are both apparent predicting factors of PVT, whether the patient undergoes splenectomy [9, 13]. Progression of portal hypertension could induce continually expansion of portal vein vessels, which can further injury endothelial cells, slow down the blood flow and cause unstable hemodynamics [14]. Besides, a recent study focused on mechanism of PVT by Anton et.al pointed that eccentric hyperplasia of portal vein endothelium is driven by ECM accumulation and infiltration of inflammatory and myofibroblast-like cells, indicating inflammatory factors out of coagulation system acting on endothelium cells also involve in PVT process [15]. Additionally, an observational study conducted by Ibrahim EH et.al found serum vWF positively correlated with esophageal varices grade and EGVB probability, and pointed out that increased portal vein width could lead to elevated sheer stress on endothelium in portal vein system, which may trigger the release of vWF [16]. Compared to PVV, portal vein width can be measured conveniently on abdominal CT for initially screening high risk patients.

Splenectomy also act as a significant predicting factor for PVT, which aligns with the study analyzing PVT risk factors by Kinjo et al. [17]. Previous studies have showed splenectomy may increase at least 10 times of PVT risk in liver cirrhosis independent of liver function [18], a recent meta-analysis conducted by Lu et al. also suggested preventive anticoagulation therapy for Child–Pugh level A-B cirrhosis patients with portal hypertension after splenectomy [19]. Several reasons could explain the phenomenon of splenectomy elevating PVT risk, firstly, the surgery splenectomy or embolization therapy of spleen artery could injury endothelium cells and expose collagenous fiber then activate coagulation system; secondly, even though hypersplenism can be recovered after spleen excision, this operation simultaneously decrease effect of platelet phagocytosis, which could re-elevate the platelet level and disturb the balance of coagulation and anti-coagulation [20]; last, the remaining spleen vein can form into anatomic blind end without blood flow, increasing blood flow resistance and cause blood statis or turbulence [21].

In addition to splenectomy, we also found HGB negatively correlated with PVT risk both statistically significant in univariate and multivariate analysis, in accord with results from previous studies [22]. It is known that anemia and pancytopenia are common manifestations in decompensated cirrhosis owing to hypersplenism caused by portal hypertension, however, blood cell parameters (HGB, WBC and PLT) were not involved in “ASP” model due to their statistical insignificance in final regression model, simultaneously for avoiding model overfitting.

An unnoticed serological parameter in liver function test—ALP, performed both significant in binary logistic and Lasso was also involved in final model. It is a glycoprotein usually elevates concurrently with total bilirubin when cholestasis occurred, also increases in malignant tumor patients with bone metastasis, however, models constructed by logistic and Lasso regression both showed a negative correlation between ALP and PVT. ALP is frequently elevated in cholestatic liver diseases such as PBC, which could theoretically confound its association with PVT, however, the sensitivity analysis displayed inverse correlation not only in PBC subgroup but in viral and alcohol-related cirrhosis subgroups, suggesting that the association is not solely driven by cholestasis. The stronger correlation in PBC subgroup may be explained by the wider ALP range, which increases statistical power to detect an association, rather than indicating a distinct biological mechanism of ALP impact on thrombosis. Initially, we intended to treat ALP as interfering factor for this negative relation difficult to explain in clinical circumstance, yet a recent retrospective cohort study concentrating on functions of ALP in VTE among pregnant women attracted us [23], in which discovered lower ALP level in late pregnancy period related to higher VTE risk, and explained that ALP usually increases for supporting fatal growth and development in late pregnancy, consequently a lower ALP may reflect potential thromboembolism hazard, which might be mediated by decreased hemoglobin. Another study also depicted the relationship between ALP and VTE, but found contrary results that elevated or abnormal ALP levels contribute to VTE risk [24]. A primary assumption for this negative correlation is that liver enzymes like ALP could elevate mildly in progression of cirrhosis for intrahepatic cholestasis, thus relatively decreased ALP level may reflect underlying thrombosis risk. However, the negative correlation of ALP and PVT could be interfered by various confounding factors except for cholestasis, such as liver failure, skeletal disease, which can all cause an obvious elevation of ALP level. The biological mechanisms underlying this inverse association remain unclear, future prospective studies with serial ALP measurements and adjustment for markers of cholestasis and other confounders are needed to validate this finding. Regardless of the underlying mechanism, ALP may serve as an easily accessible clinical parameter that helps identify higher PVT risk patients along with other specific risk factors.

Notably, coagulation parameters including PT, APTT and INR were all not involved in “ASP” model, with no significant difference observed between PVT and non-PVT, which showed different result comparing with formal diagnostic model study [9]. Patients diagnosed with cirrhosis are tend to be considered with a increasing bleeding risk for clotting factors reduced due to liver dysfunction, and lower platelet level caused by hypersplenism [25], however, a retrospective case–control study based on coagulation function among 113 inpatient cirrhosis cases conducted by Patrick G et.al showed the abnormal coagulation in cirrhosis is actually a rebalanced status with bleeding and thrombogenesis risk increased simultaneously [26], which is accord with today’s insights that, the weakened coagulation system can induce bleeding while reduced anti-coagulation level could cause thrombosis events [27, 28]. Therefore, common coagulation parameters such as PT, INR are eligible for evaluating clotting factors decrease but not reliable for predicting bleeding or thrombosis events directly. For D-dimer, a coagulation marker which has been verified correlated to PVT [29] and other types of venous thrombosis, showed a synergistic effect with “ASP” model, indicating that adding D-dimer as a continuous variable could improve the predictive performance in identifying high-risk patients in cirrhosis. Nonetheless, due to the small sample size of the external validation set, the overall effect of D-dimer combining “ASP” model still needs to be further verified in a larger cohort.

Based on DCA curve, a predicted probability over 0.3 was identified as an optimal risk threshold for further diagnostic evaluation. For cirrhosis patients who are receiving outpatient or being about to discharge from inpatient ward, we propose a two-step workflow: first, a fibrinolytic marker such as D-dimer or fibrinogen degradation products (FDP) can be measured, if the fibrinolytic markers also perform positive (according to local laboratory’s upper limit of normal), the patient should be suggested for regular abdominal imaging surveillance (e.g., ultrasound every 6–12 months) to enable early detection of PVT. Nevertheless, due to the increased bleeding hazard in chronic liver disease [26], especially who appears prolonging PT and APTT, or abdominal coagulation factors activity, it is crucial to emphasize that this model is not intended to directly guide prophylactic anticoagulation, any decision for anticoagulation must be based on patients’ individual bleeding risks (such as inpatient bleeding score for internal medicine patients). The proposed workflow also requires prospective validation before implementation.

Compared with previous models concentrated on PVT, the “ASP” model constructed in this study is more concise and entirely based on common clinical data. It can be applied in hospitals that do not routinely perform additional thrombosis-related tests (such as von Willebrand Factor (vWF), etc.), and can serve as a simplified supplementary tool for the existing PVT prediction models. Meanwhile, the “ASP” model is inferior to the other models on predictive performance (Supplementary Table 10), it can also be observed that portal vein velocity (PVV) was involved in all three published predictive models [8, 9, 29], which accord with the clinical experience that blood stasis plays an important role in venous thromboembolism. In future research, we will further collect data such as PVV and D-dimer of hospitalized patients to construct more robust predictive model.

There are also several shortages in our study. Firstly, data missing was observed in our primary data, which could cause systematic bias though multiple imputation was involved, meanwhile, the passive follow-up for collecting PVT ending data in 24 months from departure was restricted by actual inpatient information, for example, not all patients who have been discharged will undergo regular check-ups or be hospitalized in the same hospital again in two years, consequently the model may bias towards identifying more severe PVT patients or those who exhibit obvious clinical symptoms, resulting in an overestimation of sensitivity, also, the PVT diagnosis based on different measures (CT, MRI or ultrasound) can also introduce heterogeneity in endpoint ascertainment. These shortages reminding us necessary to establish prospective cohorts for active follow-up and to unify the procedure or examining tool for endpoint identification to strengthen model verification and optimize the final model. Besides, different examination modalities for portal vein width, and retrospective measurement based on original imaging archives could also introduce extra heterogeneities in portal vein width value. Even when the same appliance was used, the reported width values could still vary due to the experience and measurement habits of investigators, therefore this potential bias should be considered during data collection period, and reminds us for measurement standardization especially for collecting this type of imaging data. Then, though all involved patients were diagnosed with cirrhosis, the heterogeneities of pathogenesis, stage of liver decompensation, other cirrhosis complications and history of anticoagulation therapies are also need to be considered during model perfection. Furthermore, for a tiny proportion of discovery cohort (4 patients, 1.63%) obtained a time overlap with external validation set, this study did not strictly accomplish a time-based external validation but simply validated the model efficiency in a relatively small-size independent external cohort, meanwhile the high proportion of excluded patients may cause selecting bias, leading the final model tend to be appropriate for patients who obtained complete inpatient data and without loss to follow-up. The external validation based on 82 hospitalized patients is a preliminary validation result which requires a space external validation cohort to examine the predicting effort and credibility in other cirrhosis patient groups. Eventually, the critical theory of thromboembolism—Virchow’s triad was not adequately embodied in our final model, in which only portal vein width was discussed to reflect abnormal vessel structure, some specific serum biomarkers proved with stronger correlation to thromboembolism such as IL-6 [7], D-dimer and vWF factor [8] have not been involved when constructing our model due to limitation of retrospective medical history data, which were not conventionally collected in previous hospitalized patients, though an supplementary analysis concentrated on D-dimer impact on PVT was accomplished in external validation set. We acknowledge that the selected variables splenectomy and portal vein width could partially reflect the susceptibility of PVT according to published researches, however, an increasing number of literatures have mentioned that serum biomarkers like vWF and imaging parameters such as PVV play a crucial role in PVT risk prediction, the “ASP” model should be initially regarded as a basic model and needed to be further validated and strengthened in a more standardized prospective cohort, approximately other aforementioned factors like PVV, vWF and IL-6 were encouraged to be integrated for enhancing predictive accuracy. To sum up, compared with previous published models on PVT, there is still a room in prediction accuracy for simple “ASP” model composed of current three variables, the current “ASP” model is suitable for the initial screening high-risk PVT patients but not a replacement for diagnosis right now, a more accurate prediction model is expected based on current model.

Conclusion

In conclusion, our study constructed a prediction model and screened out three independent clinical variables for screening out high-risk PVT patients in cirrhosis initially, which can assist clinical physicians conducting more appropriate follow-up measures for cirrhosis patients, in order to prevent progression of PVT or other complications. Meanwhile, ALP might be a potential clinical variable connected to PVT risk in cirrhosis, but need to be further verified.

Abbreviations

ALB

Albumin

ALP

Alkaline phosphatase

ALT

Alanine aminotransferase

APTT

Activation of partial thromboplastin time

AST

Aspirate aminotransferase

AUC

Area under curve

CE

Cholesterol

CHE

Cholinesterase

DCA

Decision curve analysis

DVT

Deep vein thrombosis

EGVB

Esophageal gastric variceal bleeding

FBG

Fibrinogen

FDP

Fibrinogen degradation products

GGT

Gamma-glutamyl transpeptidase

HCT

Hematocrit

HDL

High-density lipoprotein cholesterol

HGB

Hemoglobin

INR

International normalized rate

LDL

Low-density lipoprotein cholesterol

MAE

Main absolute error

PBC

Primary biliary cholangitis

PLT

Platelet

PT

Prothrombin time

PTA

Prothrombin activity

PTE

Pulmonary thromboembolism

PVT

Portal vein thrombosis

PVV

Portal vein velocity

ROC

Receiver operating characteristic

TG

Triglyceride

WBC

White blood cell count

vWF

Von Willebrand Factor

Authors’ contributions

Zian Wang, Feng Jiang and Jinglan Jin wrote the main manuscript text; Yaya Li and Jintao Jiang provided original data for our research; Zian Wang provided research methods and prepared figures; Jinglan Jin supervised and revised the final manuscript; All authors reviewed the manuscript.

Funding

This study was funded by Jilin Province Science and Technology Development Plan Project (Grant ID: YDZJ202201ZYTS122).

Data availability

Data will be available with a request.

Declarations

Ethics approval and consent to participate

This study was conducted following the ethical principles of the Declaration of Helsinki and its subsequent amendment versions, and has been approved by Ethical Committee of First Hospital of Jilin University before initiation (Approval ID: 2024–989). Except for collecting necessary clinical variables for building prediction model, private and identity information of all patients were not disclosed. As all original data were extracted from routinely collected clinical medical records, prior to study initiation, formal exemption from the requirement for informed consent was granted by Ethics Committee of the First Hospital of Jilin University, in accordance with national regulations governing retrospective analyses of anonymized health data.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.