Development of A Risk Assessment Model for Venous Thromboembolism among Patients with Cholangiocarcinoma

Abstract

Venous thromboembolism (VTE) has a high incidence among patients with cholangiocarcinoma (CCA). However, appropriate risk assessment models (RAMs) for CCA are limited. This study aimed to develop a RAM for predicting VTE in patients with CCA. We conducted a retrospective study at a single university-based hospital in Thailand. We included consecutive patients who were newly diagnosed with CCA between January 2018 and December 2022. A total of 694 CCA patients were included in the study. A 12-month follow-up period, 91 patients (13.11%) developed VTE. The logistic regression analysis initially considered 18 potential clinical predictors, and after using backward elimination, five key predictors were identified: ECOG score ≥ 2, intrahepatic CCA, stage IV CCA, total bilirubin ≤ 13 mg/dL, and CA19-9 > 1600 U/mL. These predictors formed the basis of the scoring system. The resulting RAM achieved an area under the receiver operating characteristic curve (AuROC) of 0.70 (95% confidence interval [CI] 0.64-0.75), indicating acceptable discrimination. The scoring system, ranging from 0 to 6, was categorized into three groups. For the validation group, the AuROC curve was 0.68 (95% CI: 0.63-0.73). In addition, the model demonstrated consistently high NPV across all risk categories, indicating strong performance in ruling out VTE. Therefore, this score is most useful for identifying non-high-risk patients. In conclusion, a RAM for VTE in CCA was developed by incorporating five critical risk factors. This model may assist clinicians in identifying individuals for risk of VTE. External validation is warranted to confirm its generalizability.

Introduction

Venous thromboembolism (VTE) is a significant and potentially life-threatening complication frequently observed in cancer patients.^1,2 VTE in cancer patients is caused by a various mechanism, including direct coagulation pathway activation, induction of inflammatory responses, inhibition of fibrinolytic activity,^3,4 surgery and cancer treatment.^5,6 The hypercoagulable state associated with malignancy, particularly in hepatobiliary tumors, such as cholangiocarcinoma (CCA), further increases the risk of VTE.^7,8

CCA is a highly aggressive cancer originating from bile duct epithelial cells, characterized by a poor prognosis and high mortality. ⁹ The study, which analyzed age-standardized mortality rates extracted from databases for CCA across 32 countries in Europe, the Americas, and Australia, reported that intrahepatic CCA mortality rates reached 1.5–2.5/100 000 in men and 1.2–1.7/100 000 in women, while extrahepatic CCA rates were generally below 1/100 000, with Japan being the notable exception, registering 2.8/100 000 in men and 1.4/100 000 in women. ¹⁰ In patients with CCA, the incidence of VTE is increasing, with reported rates ranging from 14.7% to 24.8%,^11–13 rising to 35.5% in advanced-stage disease ⁷ and 19.4 to 36.2% in surgically treated patients.^14–16

The occurrence of VTE further exacerbates adverse clinical outcomes and complicates management strategies.^11,12,17 Identification of patients with CCA who are at heightened risk of VTE is crucial for implementing timely preventative measures.^12,13 Current clinical guidelines recommend prophylactic anticoagulation for selected patients, including those with active cancer,^2,18 but its application remains limited due to the associated risk of bleeding and uncertainty about which patients derive the most benefit.^19,20 Therefore, a reliable and specific Risk Assessment Model (RAM) tailored to CCA is essential to better stratify patients based on their VTE risk and to optimize personalized care. ³

Several general RAMs have been developed to estimate the risk of VTE in cancer patients and are recommended in clinical guidelines. ²¹ However, these models are not always suited to capture the unique characteristics and tumor biology of specific cancers, such as CCA. ¹² In addition, the specific RAM developed for patients with CCA was not originally designed to predict risk of VTE in this patient population. ²² Due to CCA being a relatively understudied malignancy, it is not consistently included in risk scores, and its relative VTE risk is therefore not clearly established. ²³ Chen et al developed a nomogram to predict VTE in patients undergoing surgery for distal CCA; however, its applicability to other CCA subtypes remains uncertain, and the timing of predictor assessment was not clearly defined. Therefore, identifying an accurate RAM for VTE in patients with CCA is crucial to determine individuals who at high risk of developing VTE.

Given the limited data available on predicting VTE risk specifically for CCA, it is important to develop a specific RAM to address this gap in clinical care. ¹² This study aimed to develop a novel RAM tailored to assess the risk of VTE in patients with CCA. This model is intended to assist physicians in making informed decisions regarding VTE risk stratification and to identify patients at high risk who may benefit from VTE prophylaxis.

Materials and Methods

Study Design and Setting, Participants, and Data Collection

This was a single-center, retrospective study conducted at a university-based hospital in Thailand. The study included consecutive patients newly diagnosed with CCA between January 2018 and December 2022. Patients were identified using the International Statistical Classification of Diseases and Related Health Problems, 10^th Revision (ICD-10). This study was approved by the Institutional Review Board (Approval No. 107/2567), with a waiver of informed consent.

Eligible participants were adults aged ≥18 years who received treatment and follow-up at the Surgery or Oncology Clinics. Patients were excluded if they had a history of VTE, received anticoagulant therapy prior to CCA diagnosis for any indication, were not newly diagnosed with CCA, or had incomplete medical records, defined as missing all clinical data or having only a single hospital visit without subsequent follow-up (Figure 1).

Figure 1.

Participant flow diagram.

Clinical data were reviewed from electronic medical records, which included information on patient demographics and clinical characteristics at the time of diagnosis, including age, sex at birth, body mass index (BMI), and Eastern Cooperative Oncology Group (ECOG) performance status. Laboratory parameters collected at diagnosis included hemoglobin levels, white blood cell (WBC) count, platelet count, renal function tests, serum albumin, total bilirubin, direct bilirubin, D-dimer, CA19-9, alpha-fetoprotein (AFP), and carcinoembryonic antigen (CEA) levels. In addition, data on cancer characteristics (tumor site and stage) were recorded at baseline, along with details of all treatments received by the patients.

The diagnosis of CCA was confirmed based on pathological findings, imaging studies, and tumor markers, including CA19-9, AFP, and CEA levels. CCA was classified into three subtypes according to tumor location: intrahepatic, hilar, and extrahepatic.^9,24 Patients were considered to have VTE if the diagnosis was objectively confirmed within 12 months after initiation of CCA management.

VTE was characterized by the development of deep vein thrombosis (DVT) in the extremities, pulmonary embolism (PE), portal vein thrombosis, hepatic vein thrombosis, inferior vena cava (IVC) thrombosis, or intra-abdominal vein thrombosis following the initiation of CCA management. These events were diagnosed based on clinical presentation and confirmed through imaging studies, including Doppler ultrasound for detecting DVT, computed tomography (CT) scans, and pulmonary angiography for PE.^25,26

Study Size Estimation

Sample size estimation for prediction model development was conducted using the pmsampsize package in Stata version 16.1 (StataCorp, TX, USA), following the guidelines from the clinical prediction model framework by Richard D. Riley et al ²⁷ The minimum sample size required to develop a multivariable prediction model for a binary outcome was estimated based on an anticipated outcome prevalence of 0.362 (36.2%), 16 candidate predictor parameters, an expected receiver operating characteristic (ROC) of 0.80 and assuming an anticipated Cox-Snell R² of 0.246. ¹⁴ A Minimum sample size required for new model development and small degree of overfitting was defined as an expected shrinkage of predictor effects by 10% or less (Criteria 1) was 501 participants. This corresponds to approximately 182 expected events, resulting in an events per parameter (EPP) of 11.31.

Predictors

A total of 18 candidate predictors were considered for the model, including age, sex, ECOG score, history of hypertension, diabetes, chronic kidney disease, cirrhosis, type of CCA, CCA stage, hemoglobin level, WBC count, platelet count, serum creatinine level, total bilirubin level, CA19-9 level, as well as cancer treatments, including chemotherapy, surgery, and radiotherapy. Continuous variables were dichotomized to improve clinical interpretability and facilitate construction of a practical risk assessment model. Optimal cut-off values were determined using ROC curve analysis, with selection based on the Youden index, which maximizes the sum of sensitivity and specificity. ²⁸ Predictor selection was guided by logistic regression analysis, with results presented as beta-coefficient. Predictive significance of each predictor was justified by p-value. Predictive variables with a p-value of less than .05 in univariable analysis were carried forward for inclusion in multivariable logistic regression analysis. All statistical analyses and model derivation procedures were carried out using Stata version 16.1 (StataCorp, TX, USA).

Multiple Imputation

Predictor parameters which were identified having missing values could potentially introduce bias in the estimates of the prognostic model. To address this, multiple imputation using chained equations was carried out via the mi impute chained command to handle the missing data before deriving the model. All potential predictor variables were included in the multivariable logistic regression analysis using the mi estimate commands.

Derivation of the Model

The full multivariable model initially included nine predictors, which were reduced using a backward stepwise elimination method. The selection of potential predictors was based on statistical significance and prior literature.^12,13 The final reduced model comprised five predictors.

The simplified risk score was developed by assigning specific points to each predictor, based on the beta coefficients from the multivariable logistic model. Each coefficient was divided by the smallest coefficient and then rounded to the nearest whole number. The total score was subsequently grouped into three risk categories including low, moderate, and high risk for practical use in clinical settings.

Discrimination and Calibration

The reduced multivariable logistic model was evaluated for its calibration and discrimination. Discrimination was analyzed by assessing the model's predictive performance, which was visualized using a distributional plot and quantified by the area under the receiver operating characteristic curve (AuROC). Calibration was assessed through the calibration-in-the-large (CITL), Expected/observed ratio (O:E), and calibration slope.

Internal Validation

Internal validation was conducted using bootstrapping with 500 replications to evaluate model stability and performance, implemented with the bsvalidation command.

Results

Baseline and Clinical Characteristics of Patients

Of the 694 consecutive patients with newly diagnosed CCA included in the model development cohort, 443 (63.83%) were male, age of which 555 (79.79%) were >55 years, and 78 (11.24%) had an ECOG performance status score ≥2. During the 12-month follow-up period, 91 patients (13.11%) developed VTE. In addition, 515 patients (74.21%) were diagnosed with intrahepatic CCA. Baseline demographic and clinical characteristics are summarized in Table 1.

Table 1.

Baseline Characteristics and Univariable Logistic Regression Analysis in Venous Thromboembolism with Cholangiocarcinoma Patients at 12 Months.

Characteristics	Total, n (%)	Missing, n (%)	Coefficients	95% CI	p-value	AuROC
Age, n (%)
>55	555 (79.97)	0 (0)	Reference	Reference
≤55	139 (20.03)		0.42	−0.09–0.93	.107	0.54 (0.49-0.58)
Gender, n (%)
Male	443 (63.83)	0 (0)	Reference	Reference
Female	251 (36.17)		0.32	−0.12–0.77	.155	0.54 (0.48-0.59)
ECOG score, n (%)
< 2	616 (88.76)	0 (0)	Reference	Reference
≥ 2	78 (11.24)		0.62	0.02–1.22	.042	0.54 (0.50-0.58)
Co-morbidities, n (%)
Hypertension	277 (39.91)	0 (0)	-0.28	−0.75–0.18	.223	0.53 (0.48-0.59)
Diabetes	138 (19.88)	0 (0)	0.22	−0.31–0.75	.414	0.52 (0.47-0.56)
Chronic kidney disease	27 (3.89)	0 (0)	1.90	−0.88–4.68	.180	0.50 (0.49-0.52)
Cirrhosis	45 (6.48)	0 (0)	0.21	−0.62–1.05	.616	0.51 (0.48-0.54)
Site of tumor, n (%)
Non-intrahepatic	179 (25.79)	0 (0)	Reference	Reference
Intrahepatic	515 (74.21)		1.41	0.66–2.16	<.001	0.60 (0.56-0.63)
Stage CCA (TNM system), n (%)
Stage I-III	346 (49.86)	0 (0)	Reference	Reference
Stage IV	348 (50.14)		0.86	0.39–1.33	<.001	0.60 (0.55-0.66)
Hemoglobin<10 g/dL, n (%)	159 (22.98)	2 (0.29)	−0.07	−0.60–0.47	.808	0.51 (0.46-0.55)
WBC count>10 × 109/L, n (%)	275 (39.74	2 (0.29)	0.46	0.01–0.90	.044	0.56 (0.50-0.61)
Platelet count > 350 × 109/L, n (%)	229 (33.09)	2 (0.29)	0.05	−0.41–0.52	.832	0.51 (0.45-0.59)
Creatinine > 0.6 (mg/dL), n (%)	589 (85.12)	2 (0.29)	0.82	0.02–1.62	.044	0.54 (0.51-0.57)
Serum albumin (g/dL), n (%)
> 3.5	396 (58.49)	17 (2.45)	Reference	Reference
≤ 3.5	281 (41.51)		−0.49	−0.96–(−0.01)	.042	0.56 (0.51-0.61)
Total bilirubin, (mg/dL), n (%)
> 13	183 (26.83)	12 (1.73)	Reference	Reference
≤ 13	499 (73.17)		1.19	0.51–1.87	.001	0.59 (0.56-0.63)
CA19-9, (U/mL), n (%)
≤ 1600	183 (26.83)	101 (14.55)	Reference	Reference
> 1600	499 (73.17)		0.50	0.04–0.96	.032	0.56 (0.50-0.62)
Chemotherapy, n (%)	371 (53.46)	0 (0)	0.33	−0.12–0.77	.153	0.54 (0.49-0.59)
Surgery, n (%)	281 (40.49)	0 (0)	−0.67	−1.15–(−0.18)	.007	0.57 (0.52-0.62)
Radiotherapy, n (%)	27 (3.89)	0 (0)	−0.65	−2.11–0.80	.379	0.51 (0.49-0.53)

Abbreviations: ECOG, Eastern Cooperative Oncology Group; CCA, cholangiocarcinoma; WBC, white blood cell; CA19-9, carbohydrate antigen 19-9; AuROC, area under the receiver operating characteristic curve; CI, confidence interval.

Beta Coefficients of Candidate Predictors

The univariable analysis identified nine clinical characteristics as potential predictors for patients with VTE in CCA. These include an ECOG score of ≥2 (Coefficients 0.54, 95% confidence interval [CI] −0.09-1.18, p = .095), and intrahepatic CCA (coefficients 0.96, 95% CI 0.17-1.75, p = .017). Additionally, stage IV CCA (coefficients 0.52, 95% CI 0.01-1.03, p = .047) and WBC count exceeding 10 × 10⁹/L (coefficients 0.40, 95% CI −0.07-0.88, p = .094) were predictive. Other identified factors include creatinine levels >0.6 mg/dL (coefficients 0.76, 95% CI −0.07-1.58, p = .072), Serum albumin ≤3.5 (coefficients −0.30, 95% CI −0.82-0.22, p = .252) and total bilirubin levels ≤13.0 mg/dL (coefficients 0.90, 95% CI 0.16-1.64, p = .018). Moreover, a CA19-9 level above 1600 U/mL (coefficients 0.30, 95% CI −0.19-0.79, p = .232) and a history of surgery (coefficients 0.07, 95% CI −0.48-0.63, p = .795) were identified as significant predictors (Table 2).

Table 2.

Logistic Regression Coefficients in the Full and Reduced Multivariable.

Predictors	Full Model			Reduced Model
	Coefficients	95% CI	p-value	Coefficients	95% CI	p-value
ECOG score
< 2	1.00	Reference		1.00	Reference
≥ 2	0.54	−0.09–1.18	.095	0.50	−0.13–1.12	.118
Intrahepatic
No	1.00	Reference		1.00	Reference
Yes	0.96	0.17–1.75	.017	1.01	0.24–1.79	.010
CCA stage
Stage I-III	1.00	Reference		1.00	Reference
Stage IV	0.52	0.01–1.03	.047	0.53	0.05–1.03	.032
WBC count
≤10 × 109/L	1.00	Reference			Not included
>10 × 109/L	0.40	−0.07–0.88	.094
Creatinine (mg/dL)
≤ 0.6	1.00	Reference			Not included
> 0.6	0.76	−0.07–1.58	.072
Serum albumin (g/dL), n (%)
> 3.5	1.00	Reference			Not included
≤ 3.5	−0.30	−0.82–0.22	.252
Total bilirubin (mg/dL)
> 13	1.00	Reference		1.00	Reference	.007
≤ 13	0.90	0.16–1.64	.018	0.96	0.26–1.66
CA19-9 (U/mL)
≤ 1600	1.00	Reference	.232	1.00	Reference
> 1600	0.30	−0.19–0.79		0.32	−0.15–0.81	.182
Surgery
Yes	1.00	Reference			Not included
No	0.07	−0.48–0.63	.795

Abbreviations: ECOG, Eastern Cooperative Oncology Group; CCA, cholangiocarcinoma; WBC, white blood cell; CA19-9, carbohydrate antigen 19-9.

Final Model

The final selection predictors were ECOG score ≥2 (coefficients 0.50, 95% CI −0.13-1.12, p = .118), intrahepatic CCA (Coefficients 1.01, 95% CI 0.24-1.79, p = 0.010), stage IV CCA (coefficients 0.53, 95% CI 0.05-1.03, p = .032), total bilirubin ≤13 mg/dL (coefficients 0.96, 95% CI 0.26-1.66, p = .007), and CA19-9 > 1600 U/mL (coefficients 0.32, 95% CI −0.15-0.81, p = .182). Among all clinical predictors, intrahepatic CCA demonstrated the highest predictive performance (Table 2).

Score Transformation

After selecting predictors by using multivariable model was assigned with specific score derived from logistic regression coefficient (Table 3). The scoring scheme with a total score ranging from 0 to 6 was then further categorized into 3 risk subcategories. This categorization was developed based on the calibration plot, low risk group with score ranging from 0 to 1, moderate risk group with score ranging from 2 to 3 and high-risk group with score ranging from 4 to 6 (Table 3).

Table 3.

Simplified Risk Score of Risk Assessment Model.

Variables	Coefficients	Point
ECOG score ≥ 2	0.50	1
Intrahepatic CCA	1.02	2
CCA stage IV	0.54	1
Total bilirubin ≤13 mg/dL	0.96	1
CA19-9 > 1600 U/mL	0.34	1
Interpretation		Total score
Low risk	Reference	0–1
Moderate risk	1.61	2–3
High risk	2.55	4–6

Abbreviations: ECOG, Eastern Cooperative Oncology Group; CCA, cholangiocarcinoma; WBC, white blood cell; CA19-9, carbohydrate antigen 19-9.

Model Discrimination and Calibration

For apparent performance in discrimination, the AuROC for the score-based logistic regression model was 0.70 (95% CI 0.64-0.75), indicating acceptable performance (Figure 2). From the plot, the predicted probability of VTE increased as the score increased, demonstrating a high level of agreement between the actual and predicted VTE risks. The CITL was 0.00, indicating that the predicted probabilities matched the observed number of VTE events (O:E = 1.00). The calibration slope was 1.01, showing that the model fit the data well (Figures 3 and 4).

Figure 2.

Performance of the risk assessment model, area under the receiver operating characteristics curve (AuROC).

Figure 3.

Calibration plot for predicted probability of VTE in CCA.

Figure 4.

Observed risk (circle) versus score predicted risk (solid line) of VTE.

The mean total score was significantly different between VTE group and None-VTE group (4.04 ± 1.02 vs 3.13 ± 1.46, p < .001) (Table 4). The positive predictive values (PPV) were 1.9 (95% CI 0.2-6.5) for low risk, 8.7 (95% CI 5.4-13.1) for moderate risk, and 19.4 (95% CI 15.4-23.9) for high risk (Table 4). The negative predictive values (NPV) were 84.8 (95% CI 81.6-87.6) for low risk, 84.7 (95% CI 81.1-87.8) for moderate risk, and 93.5 (95%CI 90.3-95.9) for high risk (Table 4).

Table 4.

Comparing Different of VTE and non-VTE with Level of Risk at 12 Months (Low Risk, Moderate Risk, and High Risk), Positive Predictive Value (PPV) with 95% Confidence Interval, Negative Predictive Value (NPV) with 95% Confidence Interval, and p-Value.

Risk Categories	Score	VTE (n = 91)	none-VTE (n = 603)	PPV (95% CI)	NPV (95% CI)	p-value
		n (%)	n (%)
Low	0–1	2 (2.20)	106 (17.58)	1.9 (0.2-6.5)	84.8 (81.6-87.6)	<.001
Moderate	2–3	20 (21.98)	211 (34.99)	8.7 (5.4-13.1)	84.7 (81.1-87.8)	.014
High	4–6	69 (75.82)	286 (47.43)	19.4 (15.4-23.9)	93.5 (90.3-95.9)	<.001
Mean ± SD		4.04 ± 1.02	3.13 ± 1.46			<.001

Abbreviations: VTE, venous thromboembolism; PPV, positive predictive value; NPV, negative predictive value; CI, confidence interval; SD, standard deviation.

Internal Validation

Internal validation of the developed prognostic model was conducted using a bootstrap resampling method with 500 replicates. Validation of the full model across validation cohorts demonstrated a AuROC of 0.68 (95% CI 0.63-0.73), indicating fair effective at distinguishing between patients at high risk and low risk for VTE. ²⁹ The calibration slope was 0.87, suggesting an overestimation of VTE risk in high-risk patients, while the CITL was 0.00, reflecting that the predicted probabilities closely matched the observed number of VTE events (O:E ratio = 1.00) (Figure 5).

Figure 5.

Internal validation.

Discussion

The occurrence of VTE in patients with CCA is substantially high, with VTE complications being linked to an increased risk of mortality in this population. ¹³ In our study, we observed that 13.11% of CCA patients experienced VTE events at 12-month follow-up period. Similarly, the study from Korea reported a 14.7% incidence of VTE in CCA patients over a 48-month follow-up period. ¹¹ Several other reports have demonstrated comparably high rates. A study conducted in Germany reported a markedly elevated incidence of thromboembolic events at 29.3%. ¹² Moreover, the incidence of VTE exceeded 35.5% among patients with advanced-stage CCA, ⁷ and among patients undergoing surgery for distal CCA in China, the reported incidence was 36.2%. ¹⁴ These findings emphasize the necessity of implementing VTE risk assessments to facilitate early detection and prophylactic intervention.

While RAMs have been developed for VTE risk stratification, they were standard tools in clinical practice for many institutions, although most are unable to adequately identify VTE risk in CCA patients. ¹² Pfrepper et al evaluated the risk of thromboembolism in CCA patients using RAMs such as the Khorana, ONKOTEV, and Protecht scores. Their study found that only the ONKOTEV score independently predicted thromboembolism in these patients. ¹² In a previous study, Chen et al created a nomogram using independent risk factors identified by multivariate logistic regression to predict the likelihood of VTE in patients after distal CCA surgery. They assessed the nomogram's performance with ROC and calibration curves, finding AUC values of 0.80 (95% CI: 0.72-0.88) in the training group and 0.79 (95% CI: 0.73-0.89) in the validation group. However, their research focused only on patients who had distal CCA surgery and involved a small sample size, raising questions about its applicability to other CCA patient groups. ¹⁴

In this study, we designed a specific RAM aimed at early prediction of VTE in patients with CCA. Multivariable analysis identified three independent predictors significantly associated with VTE risk: intrahepatic CCA, stage IV disease, and total bilirubin levels ≤13 mg/dL. Additionally, two clinically relevant predictors (ECOG performance status ≥2 and CA19-9 levels >1600 U/mL) were included in the model despite demonstrating borderline statistical significance, based on their potential clinical impact and evidence from previous studies.^12,13 These variables were carefully selected and synthesized from the existing literature.

In the study by Blasi et al, a perihilar tumor location was associated with a higher incidence of VTE compared to non-perihilar locations (p = .005). This contrasts with our findings, which showed that intrahepatic CCA was associated with VTE. The difference may be due to the fact that their analysis focused on CCA patients who had undergone liver resection. ³⁰ The study by Jeon et al demonstrated an association between tumor progression and VTE occurrence (p = .022), indicating that locally advanced tumors may have metastatic potential. ¹¹ Additionally, patients with advanced CCA are more likely to exhibit poorer functional performance and experience more complications, such as cholangitis and cholecystitis. ¹¹ Other factors associated with VTE included elevated C-reactive protein (CRP) levels (p = .006) and chemotherapy treatment (p = .014). ¹¹ Furthermore, a study by Chen et al revealed that an elevated TNM stage was predictive of a higher risk of VTE in patients following distal CCA surgery (OR 2.57, 95% CI 1.24-5.36, p = .01). ¹⁴ Moreover, a study by Duman et al demonstrated that bilirubin levels are independently correlated with the occurrence of VTE ³¹ Niprapan et al also found that lower total bilirubin levels were associated with VTE. ¹³ Additionally, Pfrepper et al observed that CA19-9 levels were significantly elevated in patients who experienced thromboembolism within the first year compared to those without thromboembolic events (p = .040). ¹² This is associated with tumor burden and more aggressive disease, highlighting a potential relationship between hypercoagulability and tumor biology. ³² In addition, a study by Kim et al reported several factors associated with the development of VTE in advanced CCA, including intrahepatic tumor location (p = .008), receipt of anticancer treatment (p = .005), postoperative recurrence (p < .001), major vessel invasion (p < .001), and elevated baseline CA19-9 levels (p = .022). ⁷ It is important to note that, there are no international guideline-recommended cut-off values for total bilirubin or CA19-9 specifically for predicting VTE risk in patients with CCA. We selected cut-offs using the Youden index to derive thresholds that best discriminate VTE risk within our cohort. We believe this approach is appropriate when guideline-directed values are unavailable for the clinical endpoint of interest.

After developing the RAM for evaluating the performance of the prediction model, the results demonstrated discrimination and calibration, supporting its identifying patients with CCA who are at risk of VTE. The AuROC was 0.70 (95% CI 0.64-0.75), indicating acceptable discrimination. ²⁹ This result suggests that the model has a modest ability to differentiate between patients with and without VTE based on the identified predictors. An AUC of 0.69 reflects a clinically meaningful level of predictive accuracy, considering the complex and multifactorial nature of VTE risk in cancer patients.^29,33 The model's calibration was exemplary, with the CITL = 0.00, indicating a perfect match between the observed and predicted probabilities of VTE events (O:E = 1.00). Additionally, the calibration slope of 1.01 further highlights that the predicted risk aligns closely with actual outcomes.^34,35 The overall AuROC of the model was acceptable. In addition, the model demonstrated consistently high NPV across all risk categories, indicating strong performance in ruling out VTE. Therefore, this score is most useful for identifying non-high-risk patients who are potentially unlikely to benefit from thromboprophylaxis.

The internal validation of the RAM using the bootstrap resampling method provided critical insights into its predictive performance and reliability. With 500 replicated samples, the validation demonstrated a AuROC of 0.68 (95% CI 0.63-0.73), indicating the model's fair discriminatory ability to differentiate between patients at high and low risk for VTE.^29,33 The CITL of 0.00 suggests that the overall predicted probabilities closely matched the observed incidence of VTE within the cohort, achieving a perfect O:E ratio of 1.00. However, the calibration slope of 0.87 underscores some degree of overestimation, particularly in high-risk patients.^29,35

In comparison to our model, the model by Chen et al demonstrated higher discriminative ability, as it was derived from patients who underwent distal CCA surgery. However, our model showed good calibration, with match between the observed and predicted probabilities of VTE events, and is applicable to a broader population of general CCA patients.

A strength of our study is that the RAM effectively distinguishes differences in VTE probabilities among various risk groups. This RAM incorporates specific predictors tailored to CCA patients to assess their risk of VTE. By integrating the scoring system into routine clinical workflows, healthcare providers might identify patients who would potentially benefit most from intensive VTE prophylaxis, such as high- or moderate-risk patients, while avoiding unnecessary interventions in low-risk individuals. The ability to tailor preventive and therapeutic interventions based on individual risk stratification is particularly vital in CCA patients, who often face a high burden of both disease and comorbidities.

A key clinical challenge in cholangiocarcinoma is balancing VTE prevention against bleeding risk from anticoagulant prophylaxis. Patients with CCA frequently have liver dysfunction, thrombocytopenia, and undergo invasive procedures, increasing hemorrhagic risk, while evidence for routine prophylaxis in ambulatory biliary tract cancer is limited.^36,37 Given the modest VTE incidence in our cohort, universal prophylaxis may expose many patients to unnecessary harm. Our RAM therefore supports a risk-stratified approach to guide individualized prophylaxis decisions.

Our study has several limitations. First, it was a retrospective study, which may introduce bias and missing data. However, multiple imputation was applied to estimate the missing data and reduce bias. Second, the model showed slight overestimation in the high-risk groups identified, highlighting the need for external validation in more diverse patient populations. Such validation could improve the model's calibration and enhance its performance across various risk categories. This step is crucial to ensure that the model is both generalizable and applicable across different healthcare settings, ultimately contributing to better outcomes and higher quality care for patients with CCA.

Conclusion

In conclusion, a RAM for VTE in CCA was developed by incorporating five critical risk factors. This model may assist clinicians in identifying individuals at high risk for VTE, thereby supporting clinical decision-making. appropriate preventive measures. The prediction model demonstrated acceptable discrimination and good calibration, with clear differentiation between patients of varying VTE risk. External validation is warranted to confirm its generalizability.