Risk assessment models for venous thromboembolism in surgical inpatients: a multicenter retrospective cohort study Authors: Mengying Xie, Zhiliang Zhou, Fan Fei, Yingbin Deng, Chen Liu, Wenyi Jin, Fengyu Chen, Liang Wang, Chan Chen, Zhiyi Wang, Jie Weng, Zhe Xu Published: 2025 ABSTRACT AbstractObjectiveThe objective of this study was to evaluate risk assessment models (RAMs) for venous thromboembolism (VTE) in surgical inpatients.BackgroundVTE significantly contributes to morbidity and mortality among surgical inpatients, with the postoperative period being particularly vulnerable. Accurate risk assessment is essential for guiding thromboprophylaxis. Although various RAMs have been developed, their comparative effectiveness in surgical populations remains unclear.MethodsThis retrospective cohort study, SURG-VTE, was conducted across four centers from 2020 to 2022. This study evaluated the predictive performance of six RAMs: Caprini, Padua, Wells, Geneva, Autar, and IMPROVE scores. The primary outcome was objectively confirmed symptomatic VTE. The association between RAMs and VTE was examined using logistic regression analysis. The area under the receiver operating characteristic curve (AUC) was used to assess each model’s discrimination. Calibration was assessed using the GiViTI calibration belt, while overall performance was quantified by the Brier score. Decision Curve Analysis (DCA) was performed to evaluate the clinical utility of the RAMs. Sensitivity and subgroup analyses were conducted to further assess the models’ performance.ResultsOf the 4851 surgical inpatients, 826 (17.0%) experienced VTE, with significant differences in VTE rates between high- and low-risk groups as classified by most RAMs. However, the overall discriminative performance of the RAMs was poor, with AUCs ranging from 0.594 for the Geneva score to 0.713 for the Padua score. Sensitivity was highest for the Autar score (70.5%) and specificity for the Wells score (99.2%). The clinical utility of RAMs was further questioned as the positive net benefit was suboptimal for all models. Similar results were observed in sensitivity and subgroup analyses.ConclusionsWhile RAMs can distinguish between high- and low-risk groups for VTE, but show poor discrimination, limited precision, and suboptimal calibration (except IMPROVE), suggesting the need for more reliable and clinically relevant VTE risk prediction strategies in surgical care.Trial registration ClinicalTrials.govNCT06502600Supplementary InformationThe online version contains supplementary material available at 10.1186/s40001-025-03575-3.Keywords: Venous thromboembolism, Risk assessment models, Surgical inpatients, Thromboprophylaxis, Predictive performance, Clinical utility Introduction Venous thromboembolism (VTE) is a major and preventable cause of morbidity and mortality in surgical inpatients [1]. The incidence of VTE in the postoperative period is particularly high [2], underscoring the necessity for accurate risk stratification to guide thromboprophylaxis. Although various risk assessment models (RAMs) have been formulated to identify patients at increased risk of VTE, their predictive performance varies across different patient populations and settings [3]. The Caprini [4], Padua [5], Wells [6], Geneva [7], Autar [8], and IMPROVE scores [9] are among the most widely recognized RAMs for VTE. These models assess an individual’s VTE risk by integrating clinical factors such as patient demographics, comorbidities, and surgical characteristics. In a prospective study, Häfliger et al.[10]. Verified that simplified Geneva, original Geneva, Padua, and IMPROVE scores had limited accuracy and prognostic performance for VTE in medical inpatients. However, a direct comparative analysis of these models in a surgical inpatient population has been lacking, leaving clinicians without clear guidance on the most effective tool for risk assessment. The Caprini score [9], known for its detailed assessment of surgical type and duration, has been widely adopted in orthopedic and general surgery. The Padua score [5], with its emphasis on perioperative factors, has shown promise in medical and surgical patients. The simplicity of the Wells score makes it an attractive option for rapid clinical decision-making [6]. The Geneva score [6], initially developed for medical inpatients, has been adapted for surgical populations. The Autar score [8], incorporating both clinical and laboratory parameters, offers a comprehensive approach to risk assessment. The IMPROVE score [9], an international registry-based tool, has been validated in diverse patient groups but has seen variable adoption in surgical settings. This study aims to compare the predictive performance of the Caprini, Padua, Wells, Geneva, Autar, and IMPROVE scores in surgical inpatients. We aim to evaluate the prognostic accuracy, sensitivity, specificity, and clinical utility of each model in identifying patients who develop VTE following surgery. This comparative effectiveness analysis will offer insights into the strengths and limitations of current RAMs, guiding evidence-based recommendations for VTE risk assessment in surgical patients. Materials and methods Study design and setting The present retrospective cohort study, termed SURG-VTE (Surgical Venous Thromboembolism Evaluation), was conducted across four tertiary care centers, including the Second Affiliated Hospital and Yuying Children’s Hospital of Wenzhou Medical University, the First Affiliated Hospital of Wenzhou Medical University (both in Wenzhou, Zhejiang Province), Lu’an Hospital of Anhui Medical University, and The Central Health Center of Sanshipu Town (both in Lu’an, Anhui Province). This study reviewed surgical inpatient medical records from January 1, 2020, to December 31, 2022. This study was designed to assess the predictive performance of six established RAMs: Caprini, Padua, Wells, Geneva, Autar, and IMPROVE scores, for the identification of patients at risk for VTE following surgery. Participants Eligible patients were identified retrospectively by querying a hospital system database that logs all surgical procedures from 2020 to 2022, with subsequent matching of demographic and clinical data from electronic medical records across the four centers. The study population comprised adult patients who underwent various elective or emergency surgical procedures. Eligible participants were aged 18 or older, had been hospitalized for over three days, and had no prior anticoagulant medication. Patients are excluded if they have therapeutic anticoagulation, a primary diagnosis of DVT or pulmonary embolism (PE), a life expectancy under 30 days, or are transferred from the intensive care unit or other wards. Complex predictors (dehydration, serious infection, acute infection, and several degrees of immobilization and helplessness, visible varicose veins, etc.) were defined retrospectively through a rigorous process involving over one hour of standardized training for data collectors (clinicians with > 5 years’ experience, led by senior staff), dual review by two independent assessors, and adjudication by a third reviewer when needed. Data sources included electronic medical records, imaging (e.g., ultrasound, CT), laboratory results, physician orders, nursing notes, and consultation records. The institutional review boards of all participating centers approved the study protocol, and the need for written informed consent was waived due to its retrospective nature. Risk assessment models For each patient, the six RAMs scores were documented directly from the medical records when available, or recalculated based on the medical data if necessary. Patients were stratified into low- and high-risk categories for VTE according to the criteria specific to each RAM (e-Table 1). A standardized data collection form was used to extract demographic details, comorbidities, surgery type, and VTE risk factors from medical records. Outcome measures [11]The primary outcome was fatal and non-fatal VTE, including PE as well as deep vein thrombosis (DVT), objectively confirmed by imaging. DVT was systematically screened via ultrasound at admission and discharge, while PE was investigated only upon clinical suspicion (e.g., dyspnea, chest pain) using multidetector computed tomographic pulmonary angiography (CTPA). Statistical analysis Data distribution and preliminary analysis In the SURG-VTE study, the distribution of the primary outcome, in-hospital VTE, was characterized to understand the baseline risk within the study population. Descriptive statistics are computed as mean (standard deviation) and median (interquartile range) for continuous variables, and frequency (percentages) for categorical variables. The distribution of RAM scores and baseline patient characteristics were examined between VTE and non-VTE groups. Regression analysis We employed logistic regression analysis to examine the RAMs’ performance and their correlation with VTE. Odds ratios (OR) and their 95% confidence intervals (CIs) were calculated, both unadjusted and adjusted for pharmacologic thromboprophylaxis use and study site in a multivariate analysis. RAMs discrimination The area under the receiver operating characteristic curve (AUC) was calculated to assess each model’s ability to distinguish between patients who developed VTE and those who did not. The AUC values were interpreted following conventional benchmarks, with values ranging from 0.5 (no discrimination) to 1.0 (perfect discrimination). Sensitivity and specificity were determined for each RAM to assess their ability to correctly identify patients at risk of VTE. The positive and negative predictive values were also calculated to understand the models’ predictive accuracy in the given population. Likelihood ratios were computed to provide an estimate of the increase or decrease in the odds of an event per unit increase in the predictor variable, and to understand the strength of the association between the RAM scores and the outcome. RAMs calibration Although the H–L fit test and calibration plots are standard tools for evaluating model calibration, the H–L fit test’s sensitivity is constrained by sample size, as it categorizes patients into deciles rather than assessing them individually. The GiViTI calibration belt was used to visualize the relationship between predicted and observed VTE probabilities, offering a nuanced calibration assessment with a confidence band reflecting the curve’s positional uncertainty [12, 13]. The calibration plot displayed 80% CI (light gray) and 95% CI (dark gray). A statistically significant deviation is noted if the 95% CI does not intersect the bisector. A P-value < 0.05 indicates poor model calibration, whereas a large P-value signifies a well-fitting model. Brier score The prediction model’s precision was evaluated using the Brier score, which measures the average squared difference between actual and predicted VTE, incorporating both discrimination and calibration. Brier scores range from 0 to 1, with lower scores indicating better model performance. Decision curve analysis To further assess the clinical utility of each of the RAMs, we have incorporated Decision Curve Analysis (DCA). DCA quantifies the impact of model predictions on patient management by assessing the net benefit of treatment decisions across different probability thresholds. This approach allows for a direct comparison of the six RAMs—Caprini, Padua, Wells, Geneva, Autar, and IMPROVE scores—regarding their clinical efficacy in VTE prediction. DCA provides a visual representation of the net benefit, offering clinicians an intuitive measure of each model’s practical value in decision-making processes. Subgroup analyses Patients were analyzed in subgroups according to the type of surgery (orthopedic, abdominal and neurosurgery) and those who did not receive pharmacologic thromboprophylaxis during hospitalization. Sensitivity analysis A sensitivity analysis was conducted to assess the impact of different RAMs on the predictive performance of PE. All statistical tests were two tailed with a significance threshold set at P < 0.05. Missing data, occurring in less than 10% of key variables, were addressed using multiple imputation to maintain analytical integrity. Statistical analysis was conducted using R software (version 3.6.1, R Foundation for Statistical Computing, Vienna, Austria) with packages pROC, givitiR, and rmda. Results Cohort In the SURG-VTE cohort, out of 6966 patients screened, 4851 were ultimately included. The median age was 63 years (IQR, 49–71 years). There were 2675 male patients (55.1%) and 2176 female patients (44.9%). The primary surgeries performed were orthopedic (1765, 36.4%), abdominal (1541, 31.8%), neurosurgery (616, 12.7%), and other (929, 19.2%). A total of 698 patients (14.4%) had cancer, and 121 (2.49%) had a history of VTE. Immobilization for three or more days was noted in 1237 patients (25.5%). Pharmacological and mechanical VTE prophylaxis were prescribed for 1906 (39.3%) and 33 (0.68%) patients, respectively. The proportion of patients categorized as high risk for VTE based on the Caprini score was significantly higher at 4062 (83.7%) compared to the Autar score with 2627 (54.2%), Padua score with 1739 (35.8%), IMPROVE score with 1694 (34.9%), Geneva score with 1312 (27.0%), and Wells score with 155 (3.20%). A total of 826 patients (17.0%) experienced VTE, and 44 (0.91%) had PE. Figure 1 and Table 1 present the study flow diagram and cohort characteristics. The study flow diagram Table 1Baseline patient characteristics, stratified by venous thromboembolism eventVariablesNon-VTE (N = 4025)VTE (N = 826)P-ValueDemographics Age (years), Median (IQR)63 (49–70)65 (52–76) < 0.001 Male sex, n (%)2214 (55.0)461 (55.8)0.700 BMI (kg/m2), Median (IQR)23.5 (21.2–26.1)23.7 (21.2–26.2)0.607Chronic/acute comorbidities Diabetes mellitus, n (%)848 (21.1)165 (20.0)0.511 Hypertension, n (%)1361 (33.8)293 (35.5)0.381 Chronic heart disease, n (%)540 (13.4)67 (8.11) < 0.001 Atrial fibrillation, n (%)140 (3.48)31 (3.75)0.775 Cancer, n (%)541 (13.4)157 (19.0) < 0.001 Stroke, n (%)243 (6.04)70 (8.47)0.012 Diabetes insipidus, n (%)31 (0.77)0 (0.00)0.022 Autoimmune disease, n (%)15 (0.37)7 (0.85)0.083 Nephrotic syndrome, n (%)0 (0.00)2 (0.24)0.029 Inflammatory bowel disease, n (%)0 (0.00)3 (0.36)0.005 Sepsis, n (%)81 (2.01)47 (5.69) < 0.001 Thrombophilia, n (%)0 (0.00)48 (5.81) < 0.001Vital signs on admission T (℃), Median (IQR)36.6 (36.5–36.7)36.6 (36.5–36.7)0.042 RR (breaths/min), Median (IQR)17 (16–18)17 (16–18)0.329 SBP (mmHg), Median (IQR)121 (109–133)121 (109–132)0.531 DBP (mmHg), Median (IQR)66 (59–73)67 (59–74)0.214 HR (bpm), Median (IQR)75 (70–80)74 (70–80)0.257Symptoms Paralysis, n (%)0 (0.00)31 (3.75) < 0.001 Lower limb swelling, n (%)148 (3.68)198 (24.0) < 0.001 Varicose veins, n (%)240 (5.96)51 (6.17)0.879Medication history Glucocorticoid, n (%)306 (7.60)68 (8.23)0.585 Contraceptive drug, n (%)27 (0.67)8 (0.97)0.487 Chemotherapy drug, n (%)129 (3.20)33 (4.00)0.296Dehydrant, n (%)353 (8.77)179 (21.7) < 0.001Surgery-related History of surgery (≤ 1, month), n (%)2259 (56.1)317 (38.4) < 0.001 History of trauma (≤ 1, month), n (%)677 (16.8)362 (43.8) < 0.001 Artificial instrument implantation, n (%)366 (9.09)143 (17.3) < 0.001 Cardiac pacemaker, n (%)31 (0.77)13 (1.57)0.044Others Personal or family history of VTE, n (%)33 (0.82)88 (10.7) < 0.001 Unexplained habitual abortion, n (%)16 (0.40)4 (0.48)0.764 Pregnancy or puerperium, n (%)36 (0.89)32 (3.87) < 0.001 Strict bed rest (> 3, day), n (%)804 (20.0)433 (52.4) < 0.001 Central venous catheter, n (%)545 (13.5)142 (17.2)0.007 Menopause1340 (33.3)280 (33.9)0.767 Mechanical ventilation, n (%)116 (2.88)79 (9.56) < 0.001 Hypertonic dehydration, n (%)36 (0.89)16 (1.94)0.014Laboratory findings CRP10.8 (3.13–39.9)12.2 (3.13–44.4)0.134 WBC7.78 (6.10–10.2)8.16 (6.05–10.5)0.094 HB125 (111–138)114 (100–132) < 0.001 PLT219 (175–271)217 (170–265)0.274 Cr61.9 (52.3–75.9)62.0 (51.9–76.4)0.881 D-dimer0.64 (0.28–2.57)0.68 (0.28–2.66)0.972 FIB3.42 (2.87–4.32)3.48 (2.87–4.28)0.868VTE prophylaxis at baseline, (%) Pharmacological prophylaxis1432 (35.6)474 (57.4) < 0.001 Mechanical prophylaxis22 (0.55)11 (1.33) No prophylaxis2571 (63.9)341 (41.3)VTE score Caprini score4 (3–6)5 (4–8) < 0.001 Padua score3 (2–4)5 (3–6) < 0.001 Wells score1 (0–1)1 (1–1) < 0.001 Geneva score2 (1–2)2 (1–3) < 0.001 Autar score10 (7–14)14 (11–17) < 0.001 IMPROVE score1 (0–2)2 (1–3) < 0.001BMI body mass index, CRP C- reactive protein, WBC white blood cell, Hb hemoglobin, PLT platelet count, Cr creatinine, FIB fibrinogen, VTE venous thromboembolism Baseline patient characteristics, stratified by venous thromboembolism event BMI body mass index, CRP C- reactive protein, WBC white blood cell, Hb hemoglobin, PLT platelet count, Cr creatinine, FIB fibrinogen, VTE venous thromboembolism Risk stratification Based on the Caprini score, VTE incidence was 17.5% (95% CI 16.4%–18.7%) in high-risk patients (712 of 4062) and 14.4% (95% CI 12.1–17.1%) in low-risk patients (114 of 789).The Padua score indicated a VTE risk of 32.9% (95% CI 30.7–35.2%) for the high-risk group (572 of 1739) and 8.2% (95% CI 7.2–9.2%) for the low-risk group (254 of 3112).The Wells score indicated an 80.0% incidence (95% CI 72.8–86.0%) in the high-risk group (124 of 155) and a 14.9% incidence (95% CI 13.9%–16.0%) in the low-risk group (702 of 4696). Using the Geneva Score, the VTE risk was 26.6% (95% CI 24.2–29.1%) in the high-risk group (349 of 1312) and 13.5% (95% CI 12.4–14.6%) in the low-risk group (477 of 3539). The Autar score showed a high-risk group incidence of 24.0% (95% CI 22.4–25.7%) (630 of 2627) and a low-risk group incidence of 8.8% (95% CI 7.7–10.1%) (196 of 2224). The IMPROVE score showed a high incidence of 30.8% (95% CI 28.6–33.1%) in the high-risk group (522 of 1694) and 9.6% (95% CI 8.6–10.7%) in the low-risk group (304 of 3157) (e-Table 2). RAMs performance High-risk patients based on the Caprini score do not have a statistically significant increased VTE risk compared to the low-risk group (adjusted OR, 0.80 [95% CI 0.64–1.01]; P = 0.059), whereas significant differences were noted for the other five RAMs (Table 2). The sensitivity was highest for the Autar score at 70.5% (95% CI 67.3–73.6%) and lowest for the Wells score at 15.0% (95% CI 12.6–17.4%). Specificity was highest for the Wells score at 99.2% (95% CI 99.0–99.5%) and lowest for the Autar score at 59.9% (95% CI 58.4–61.4%). The Wells score also showed the highest positive predictive values (80.0%; 95% CI 73.7–86.3%) and positive likelihood ratios (19.50; 95% CI 13.25–28.69), significantly higher than the other five RAMs, while the negative predictive values and negative likelihood ratios were similar across all models (Table 3). The Padua score demonstrated the highest discriminative performance with an AUC of 0.713 (95% CI 0.693–0.733), while the Geneva score had the lowest with an AUC of 0.594 (95% CI 0.572–0.615). Overall, all RAMs exhibited poor performance (Fig. 2 and e-Table 3). Calibration belts (Fig. 3) further illustrate that, aside from the IMPROVE score, most models exhibit significant miscalibration across the prediction range (e-Table 3). Table 2Risk of VTE in High-Risk vs. Low-Risk groups based on each risk assessment modelRisk assessment modelUnadjusted OR (95% CI)P-valueAdjusted OR (95% CI)aP-valueCaprini score1.26 (1.02–1.56)0.0360.80 (0.64–1.01)0.059Padua score5.52 (4.70–6.50) < 0.0015.70 (4.68–6.96) < 0.001Wells score22.76 (15.44–34.57) < 0.00117.42 (11.74–26.59) < 0.001Geneva score2.33 (1.99–2.72) < 0.0011.82 (1.54–2.15) < 0.001Autar score3.26 (2.78–3.88) < 0.0012.55 (2.09–3.12) < 0.001IMPROVE score4.18 (3.58–4.89) < 0.0013.96 (3.24–4.83) < 0.001IMPROVE International Medical Prevention Registry on Venous Thromboembolism, OR Odd RatioaAdjusted for site and use of pharmacologic thromboprophylaxis Risk of VTE in High-Risk vs. Low-Risk groups based on each risk assessment model IMPROVE International Medical Prevention Registry on Venous Thromboembolism, OR Odd Ratio aAdjusted for site and use of pharmacologic thromboprophylaxis Table 3Predictive Accuracy of Each Risk Assessment Model for VTE% (95%CI)LHR (95%CI)Risk assessment modelSensitivitySpecificityPPVNPVPositiveNegativeCaprini Score49.3 (45.9–52.7)73.3 (72.0–74.7)27.5 (25.2–29.8)87.6 (86.5–88.7)1.85 (1.70–2.02)0.69 (0.65–0.74)Padua Score69.2 (66.1–72.4)71.0 (69.6–72.4)32.9 (30.7–35.1)91.8 (90.9–92.8)2.39 (2.24–2.55)0.43 (0.39–0.48)Wells Score15.0 (12.6–17.4)99.2 (99.0–99.5)80.0 (73.7–86.3)85.1 (84.0–86.1)19.50 (13.25–28.69)0.86 (0.83–0.88)Geneva Score42.3 (38.9–45.6)76.1 (74.8–77.4)26.6 (24.2–29.0)86.5 (85.4–87.6)1.77 (1.60–1.95)0.76 (0.71–0.81)Autar Score70.5 (67.3–73.6)59.9 (58.4–61.4)26.5 (24.7–28.4)90.8 (89.7–91.9)1.76 (1.66–1.86)0.49 (0.44–0.55)IMPROVE Score63.2 (59.9–66.5)70.9 (69.5–72.3)30.8 (28.6–33.0)90.4 (89.3–91.4)2.17 (2.02–2.33)0.52 (0.47–0.57)IMPROVE International Medical Prevention Registry on Venous Thromboembolism, LHR likelihood ratio, NPV negative predictive value, PPV positive predictive value Predictive Accuracy of Each Risk Assessment Model for VTE IMPROVE International Medical Prevention Registry on Venous Thromboembolism, LHR likelihood ratio, NPV negative predictive value, PPV positive predictive value Receiver Operating Characteristic (ROC) Curves for Each Risk Assessment Model. The area under the ROC curve was 0.615 (95% CI 0.593–0.637) for Caprini score, 0.713 (95% CI 0.693–0.733) for the Padua score, 0.605 (95% CI 0.586–0.623) for the Wells score, 0.594 (95% CI 0.572–0.615) for the Geneva score, 0.700 (95% CI 0.681–0.720) for the Autar score, and 0.711 (95% CI 0.692–0.730) for the IMPROVE score GiViTI calibration belt for Each Risk Assessment Model Net benefit of the RAMs The DCA curve was used to obtain the net benefit of RAMs. The Autar and IMPROVE scores have a relatively high threshold interval for positive net benefit compared with the other four RAMs (Fig. 4). DCA indicates that the IMPROVE and Autar scores yield a positive net benefit rate for VTE prophylaxis between risk probabilities of 0.1 and 0.6. The DCA curve of RAMs in prediction of VTE Subgroup and sensitivity analysis In our subgroup analysis of patients without pharmacologic thromboprophylaxis, which included 2945 individuals, 352 developed VTE (12.0%) (e-Table 4). The RAMs’ accuracy did not improve relative to the overall population results (e-Table 5 and e-Table 6). Similar results were observed in subgroups undergoing orthopedic surgery (e-Table 7, e-Table 8, and e-Table 9), abdominal surgery (e-Table 10, e-Table 11, and e-Table 12), and neurosurgery (e-Table 13, e-Table 14, and e-Table 15). In a sensitivity analysis, the performance of all RAMs experienced a further decline when considering PE as the outcome. The Autar score demonstrated the highest discriminatory ability with an AUC of 0.653 (95% CI 0.578–0.728). Discussion In this comprehensive, multicenter evaluation of RAMs for predicting VTE in surgical inpatients, we found that the Caprini, Padua, Wells, Geneva, Autar, and IMPROVE scores all demonstrated suboptimal performance. Despite their widespread adoption, these models failed to accurately discriminate between patients who did and did not develop VTE following surgery. Our findings underscore the limitations of current VTE risk prediction strategies in surgical populations and suggest that existing RAMs may not be sufficiently reliable for clinical decision-making. Our study is the first to perform a head-to-head comparison of these six RAMs in a retrospective manner within a surgical cohort. The poor discriminative performance of all models, as indicated by the AUC values, is concerning. An AUC value of 0.5 suggests no better than chance prediction, and while our results showed slightly higher values, they were well below the threshold considered to represent good discrimination (AUC > 0.7), except for the IMPROVE score and Padua score which are slightly above 0.7. In addition, except for the IMPROVE score, the calibration of other RAMs is very unsatisfactory. This suggests that the model has limited ability to identify patients with VTE, which may lead to underuse or overuse of thromboprophylaxis. However, the near-perfect calibration of the IMPROVE score (p = 0.905) may be context-specific and warrants validation in larger external cohorts. In a systematic review by Stuck et al.[14], the predictive performance of RAMs was found to be inconsistent, with sensitivities and specificities ranging widely. Similarly, a study by Greene et al.[15] highlighted the challenges in validating VTE RAMs, noting that their predictive accuracy was often lower than anticipated in clinical practice. Our results echo these findings, suggesting that the predictive performance of these models may be less robust than required for effective clinical utility. The DCA suggests that IMPROVE and Autar may guide prophylaxis in moderate-risk patients, warranting further validation in prospective cohorts. In our study, these models exhibited poor to fair discrimination (AUC 0.594–0.713), with Brier scores of 0.126–0.138 indicating limited precision. Calibration was suboptimal for most models (p < 0.001), except IMPROVE (p = 0.905), highlighting a need for improved predictive tools. These findings underscore the challenge of reliable individual risk prediction in surgical inpatients. The Caprini score has been reported to have good sensitivity but poor specificity in certain studies, leading to potential overtreatment of patients [16]. The Padua score has shown promise in identifying patients at risk, but its performance in surgical populations has been less consistent [17]. The Geneva and IMPROVE scores, while validated in medical inpatients, have shown limited generalizability to surgical patients [9]. The Wells score, known for its simplicity, has demonstrated variable predictive performance in different settings [18]. In our analysis, although the RAMs successfully distinguished between high- and low-risk groups for VTE, their predictive accuracy for individual patient outcomes was found to be weak. This discrepancy suggests that while the models are capable of identifying broad patterns of risk, they fall short in precisely forecasting VTE in individual cases. The significant differences observed between the risk groups indicate that the models have some utility in stratifying patients, yet the weak predictive power implies that there are likely other influential factors not captured by the current models. This limitation underscores the necessity for ongoing refinement of the RAMs. The ability of the RAMs to differentiate between high- and low-risk groups for VTE, despite their weak predictive accuracy, indicates that they serve as a starting point for risk stratification. However, to enhance their clinical utility. Future efforts should focus on incorporating a more comprehensive set of risk factors, reassessing the weighting of existing factors, and updating the models to reflect the latest clinical evidence and changes in practice [19]. Our findings highlight the need for more accurate and tailored risk prediction tools for surgical patients. The current models, developed using data from various patient populations and settings, may not be generalizable to all surgical patients. There is a clear opportunity for the development of new models or the refinement of existing ones using modern statistical techniques, such as machine learning, which can incorporate a broader range of variables and interactions [19, 20]. Additionally, our study raises questions about the clinical implications of using RAMs to guide thromboprophylaxis decisions [9]. While the intention is to target prophylactic measures to those at highest risk, our results suggest that the current models may not effectively identify these patients. This could lead to unnecessary treatment and associated risks for some patients, while others at risk may be missed [20]. The suboptimal performance of RAMs (AUC 0.594–0.713, poor calibration except IMPROVE) poses risks of overtreatment, with associated bleeding hazards, or undertreatment, potentially missing preventable VTE events. Their utility for broad stratification is supported by significant ORs (e.g., Padua OR 5.70), but poor individual prediction suggests caution. Future models incorporating dynamic variables (e.g., D-dimer trends) or machine learning could improve accuracy, while current use should be supplemented with clinical judgment and regular reassessment. Although DCA suggests that IMPROVE and Autar may guide prophylaxis in moderate-risk patients, warranting further validation in prospective cohorts. Future research should focus on the development and validation of more accurate prediction tools and the evaluation of their impact on clinical outcomes. Our study has some limitations. First, the retrospective nature of the data collection may introduce biases, such as incomplete documentation or selective reporting of outcomes [14]. Second, the generalizability of our findings may be constrained by the specific characteristics of the patient population studied, which was drawn from four healthcare systems. Third, the study’s reliance on administrative data for determining VTE outcomes may have resulted in misclassification, as clinical diagnoses were not independently verified [11]. Fourth, our study did not account for all potential risk factors for VTE, such as genetic predispositions, detailed aspects of postoperative care, and prophylactic use of anticoagulants, which could influence the models' predictive performance. Additionally, the high VTE rate in small subgroups (e.g., Wells high risk, n = 155) may be unstable due to limited sample size, necessitating cautious interpretation and further validation in larger cohorts. Despite its limitations, our study offers valuable insights into the current state of VTE risk assessment in surgical inpatients. The findings highlight the need for further research and development of more accurate and clinically relevant prediction tools. In conclusion, this study provides a critical evaluation of the predictive performance of six established RAMs in surgical inpatients. The results indicate that these models have limited accuracy and may not be suitable for guiding clinical decisions regarding thromboprophylaxis. The IMPROVE score emerged as the top performer, with excellent calibration, a moderate AUC, and notable net benefit in DCA. Nonetheless, its discriminative ability remains suboptimal (AUC < 0.75), suggesting that while it may guide risk stratification, enhanced models are needed for precise individual predictions. There is an urgent need for more reliable and effective VTE risk prediction strategies in surgical care. Future research should prioritize the development and validation of novel models that can better inform clinical decision-making and improve patient outcomes. Conclusions Our study reveals that the Caprini, Padua, Wells, Geneva, Autar, and IMPROVE risk assessment models demonstrate limited accuracy in predicting VTE in surgical inpatients, suggesting the need for more reliable prediction strategies. Further research is warranted to develop and validate improved models that can enhance clinical decision-making and patient outcomes in this setting. Supplementary Information Supplementary material 1. Author contributions M.X., Z.X., Z.Z, F.F., J.W., and Z.W. designed the study and drafted the manuscript; Y.D., C.L., W.J. and F.C. helped interpret the results; All authors contributed to the enrolment of patients and sample collection; C.L. and L.W. helped in the statistical analysis. Z.X, C.C. and Z.W. take responsibility for the paper as a whole. All authors critically reviewed and approved the final manuscript. Funding This research was supported by Zhejiang Provincial Natural Science Foundation of China under Grant No. ZCLY24H0101, and Zhejiang Medicines Health Science and Technology Program (2023RC216, 2023KY904, and 2023KY893). Data availability The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request. Declarations Ethics approval and consent to participate The Ethics Review Committee of the Second Affiliated Hospital of Wenzhou Medical University approved this retrospective study and waived the requirement for patient informed consent. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Mengying Xie, Zhiliang Zhou and Fan Fei have contributed equally to this work. Contributor Information Zhiyi Wang, Email: wzy1063@126.com. Jie Weng, Email: wengjie198811@163.com. Zhe Xu, Email: fgsw10@163.com. References Associated Data Supplementary Materials Supplementary material 1. Data Availability Statement The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.