Модели прогнозирования риска венозного тромбоэмболизма у пациентов с множественной миеломой: систематический обзор и метаанализ.

Introduction

Multiple myeloma (MM) is a hematologic malignancy caused by clonal abnormal proliferation of plasma cells, which accounts for approximately 2% of all malignancies and 13% of all hematologic malignancies ( McCurdy et al., 2023 ; Elbahoty, Papineni & Samant, 2024 ). According to the global epidemiological data in 2022, there will be about 188,000 new cases of MM and 121,000 deaths ( Mafra et al., 2024 ). Patients with MM are often accompanied by highly disabling complications, such as bone destruction, renal insufficiency, and anemia, which seriously affect the quality of life and survival rate of patients ( Long et al., 2025 ).

Novel therapeutic agents such as proteasome inhibitors, immunomodulatory drugs (IMiDs), and anti-CD38 monoclonal antibodies have significantly improved survival outcomes in MM patients. Nevertheless, certain treatment regimens increase the risk of venous thromboembolism (VTE), with IMiDs being a major contributing factor, while proteasome inhibitors are typically associated with lower thrombotic risk ( Fotiou, Dimopoulos & Kastritis, 2022 ). Recent prospective studies report that the incidence of VTE in MM patients receiving IMiD therapy has exceeded 10% ( Bradbury et al., 2020 ; Sborov et al., 2022 ). Even after anticoagulant therapy (including aspirin, prophylactic low molecular weight heparin, and warfarin), the VTE incidence remains above 4% ( Palumbo et al., 2011 ; Chalayer et al., 2022 ), far exceeding that of the general population (0.1–0.3%) ( Beckman et al., 2010 ). VTE not only leads to chemotherapy interruption, it also increases the risk of anticoagulation-related bleeding as well as the risk of death ( Schoen et al., 2020 ; Khorana et al., 2022 ; Escobar et al., 2022 ). Studies have shown that the risk of death within 1 year in MM patients with VTE is three times higher than that in patients without VTE ( Kristinsson et al., 2012 ). Therefore, accurately identifying patients with a high risk of VTE and optimizing prevention strategies are crucial to improve the clinical outcome of patients with MM.

The National Comprehensive Cancer Network (NCCN) ( Streiff et al., 2021 ) emphasized that risk assessment models play a key role in disease diagnosis, prevention, and prognostic evaluation. In VTE management, risk prediction models can be used for individualized risk stratification to assist clinicians in accurately assessing patients’ thrombosis risk, thereby optimizing prevention strategies and balancing the benefits of anticoagulation therapy with potential bleeding risks ( Horner et al., 2024 ; Stefano et al., 2022 ). Many scholars have studied VTE risk prediction models in MM patients ( Li et al., 2023 ; Chakraborty et al., 2022 ; Sanfilippo et al., 2019 ; Li et al., 2019 ), however, due to factors such as study populations, treatment modalities, and differences in sample sizes, there is significant heterogeneity in the predictive efficacy of the current mainstream models ( e.g ., IMPEDE-VTE, SAVED, IMWG guideline, etc .), and their applicability and accuracy in clinical practice remain controversial. Therefore, this study aimed to evaluate the predictive performance and methodological quality of the existing models through systematic review and meta-analysis of all published VTE prediction models in MM patients, to provide a reference for clinical medical staff to select or develop more practical prognostic risk assessment models.

Methods

We designed this systematic review according to the checklist provided by the preferred reporting items for systematic reviews and meta-analyses guidelines ( Supplemental Material ). The study protocol has been registered on PROSPERO (registration number: CRD420251024346).

Search strategy

Due to the project’s tight timeline, formal screening of search results against eligibility criteria began before submitting to PROSPERO. This early screening was intended to streamline the review process and meet deadlines. All steps followed pre-specified criteria, and PROSPERO registration was submitted once the process was completed.

We conducted a systematic search of four databases (PubMed, Embase, Cochrane Library, and Web of Science) using the following keywords: “Multiple myeloma”, “Plasma cell myeloma*”, “MM”; “Venous thromboembolism”, “Venous thrombosis”, “VTE”, “Deep vein thrombosis”, “DVT”, “Pulmonary embolism”, “PE”; “Risk assessment”, “Risk factors”, “Risk prediction”, “Risk score”, “Prognostic model”, “Predictive Model”, “Nomogram”, “Algorithm”, with the search updated through March 31, 2026. In the case of Pubmed, the detailed search strategy is shown in the Supplemental Material . Furthermore, we manually reviewed the reference lists of both original studies and review articles to identify additional relevant studies. All English-language predictive model studies meeting our inclusion criteria (PICOTS) were included as follows:

P (Population): Patients diagnosed with MM

I (Intervention model): All risk prediction models used to predict VTE in MM patients

C (Comparator): Not applicable

O(Outcome): Objectively diagnosed VTE (including, but not limited to, deep vein thrombosis types, pulmonary embolism, and catheter-related venous thrombosis)

T (Timing): Outcomes were measured across the entire projection period without restrictions

S (Setting): Applicable across diverse healthcare environments.

Inclusion and exclusion criteria

Two searchers independently screened the literature based on the following criteria. The inclusion criteria included: (a) studies on MM patients aged ≥18 years; (b) studies on the construction and/or validation of VTE prediction models for MM patients; (c) study type of cross-sectional, case-control, or cohort studies; and (d) in cases where studies on the same topic were repeated, preference was given to more recent publications or those with more detailed information. The following exclusion criteria were applied: (a) articles that were not published, conference abstracts, or review articles; (b) studies where the full text was unavailable or data was incomplete. Two independent researchers initially performed the literature screening, and any disagreements were resolved by a third reviewer. The basic details and data extracted from the studies were cross-verified by the two researchers.

Study selection and data extraction

Two researchers (JY, YP) independently screened the literature according to the inclusion and exclusion criteria, deleted duplicates using NoteExpress software, first screened the titles and abstracts of the literature, and then re-screened the literature by reading the full text, and finally included the literature that met the criteria. When research choices were inconsistent, the third reviewer (LPZ) had talks to reach an agreement.

Based on the critical appraisal and data extraction for systematic reviews of prediction modeling studies (CHARMS) proposed by Moons et al. (2014) . A data extraction form was developed. Two investigators extracted data independently, including Author(s)/year, Country, Study design, Participants, Treatment plan, Follow-up time, VTE (median time), VTE type, Sample size/VTE cases, Missing data handling, Variable selection, Final predictors, Model development method, Validation method, Model performance, Model presentation, and so on. Any differences of opinion among the authors were settled through discussion or by consulting a third author for arbitration (LPZ). For studies involving multiple predictive models, we extracted the relevant data for each model separately. For models that were externally validated in multiple studies, we included only externally validated studies based on non-overlapping populations to avoid replication effects.

Quality assessment

In this study, two investigators assessed the risk of bias and applicability of the included prediction models using PROBAST (prediction model risk of bias assessment tool) ( Moons et al., 2019 ). When assessing the risk of bias, the analysis is mainly carried out from four aspects: study subjects, predictors, outcomes, and analysis; The evaluation of applicability focuses on the suitability of study subjects, predictors, and outcomes. The assessment results are divided into three categories: “low risk”, “high risk” and “unclear”. During the evaluation process, the two researchers independently evaluated the quality, and when they encountered differences, they reached an agreement through discussion and invited third-party experts to make a ruling if necessary.

Statistical analysis

In this study, R version 4.4. 0 software was used for data analysis ( R Core Team, 2024 ), and meta-analysis was completed using the “meta” and “metafor” software packages. When no less than two original studies were included, a fixed-effect or random-effect model was used to perform a Meta-analysis of the incidence of VTE in MM patients and the discriminant efficacy of the risk prediction model. For the pooled incidence of VTE, if there is high heterogeneity and the number of studies is not less than two, further subgroup analyses will be conducted to explore the source of heterogeneity. For external validation studies of predictive models, if two or more studies were included, the area under the receiver operating characteristic curve (AUC) was pooled and analyzed to assess the overall predictive performance of the models and their applicability in different clinical settings. AUC = 0.5 indicated no predictive ability, AUC > 0.5 and ≤0.7 indicated weak predictive ability, AUC > 0.7 and ≤0.9 indicated good predictive ability, and AUC = 1 indicates perfect predictive performance ( de Hond, Steyerberg & van Calster, 2022 ). Heterogeneity was assessed by the I 2 statistic, where I 2 = 25%, 50%, and 75% represent low, moderate, and high levels of heterogeneity, respectively ( Huedo-Medina et al., 2006 ). If ≥10 studies are included, funnel plots and Egger’s test can be used to assess publication bias; p > 0.05 indicates no significant publication bias ( Egger et al., 1997 ). In addition, to ensure the robustness of the results, sensitivity analyses were conducted by excluding studies one at a time to assess their impact on the overall findings.

Results

Literature screening

A total of 2,578 relevant studies were retrieved from the database, and two articles were included in the manual search. After removing duplicates ( n = 156) and excluding irrelevant titles and abstracts ( n = 2,357), 67 studies were selected for full-text review, of which 53 were excluded due to study subject in compliance, non-predictive models, conference abstracts, duplicates, and finally 14 studies were included. The study screening process is detailed in Fig. 1 .

Basic information about the studies

A total of 14 studies were included, published between 2019 and 2025. Of these, 13 were retrospective cohort studies ( Chalayer et al., 2022 ; Li et al., 2023 ; Chakraborty et al., 2022 ; Sanfilippo et al., 2019 ; Li et al., 2019 ; Gursoy et al., 2025 ; Fernandez-Vargas et al., 2024 ; Bao et al., 2024 ; da Costa et al., 2023 ; Dima et al., 2023 ; Sanfilippo et al., 2022 ; Bravo-Perez et al., 2021 ; Covut et al., 2021 ), one was a retrospective case-control study ( Sanfilippo et al., 2023 ), and two data ( Li et al., 2023 ; Gursoy et al., 2025 ) were derived from multicenters. Studies were mainly from Europe and the United States, including the United States ( n = 7) ( Chakraborty et al., 2022 ; Sanfilippo et al., 2019 ; Li et al., 2019 ; Dima et al., 2023 ; Sanfilippo et al., 2022 ; Covut et al., 2021 ; Sanfilippo et al., 2023 ), China ( n = 2) ( Li et al., 2023 ; Bao et al., 2024 ), and one each from Brazil ( da Costa et al., 2023 ), France ( Chalayer et al., 2022 ), Mexico ( Fernandez-Vargas et al., 2024 ), Spain ( Bravo-Perez et al., 2021 ), and Turkey ( Gursoy et al., 2025 ). Sample sizes ranged from 131 to 4,446 cases, with four studies ( Li et al., 2023 ; Sanfilippo et al., 2019 ; Li et al., 2019 ; Sanfilippo et al., 2022 ) having sample sizes of more than 1,000 cases. The duration of follow-up varied across studies, with VTE follow-up dates ranging from 1 month to 12 months, and six studies ( Chalayer et al., 2022 ; Chakraborty et al., 2022 ; Li et al., 2019 ; Gursoy et al., 2025 ; Bravo-Perez et al., 2021 ; Sanfilippo et al., 2023 ) described a median time to VTE appearance, ranging from 1.7 to 4.8 months. A total of 13 studies ( Chalayer et al., 2022 ; Li et al., 2023 ; Chakraborty et al., 2022 ; Sanfilippo et al., 2019 ; Li et al., 2019 ; Gursoy et al., 2025 ; Fernandez-Vargas et al., 2024 ; Bao et al., 2024 ; da Costa et al., 2023 ; Dima et al., 2023 ; Sanfilippo et al., 2022 ; Bravo-Perez et al., 2021 ; Covut et al., 2021 ) reported on the percentage of IMiD use, covering a range of 14.3% to 100%, with five studies ( Chalayer et al., 2022 ; Li et al., 2023 , 2019 ; da Costa et al., 2023 ; Dima et al., 2023 ) reaching 100% IMiD use (see Table 1 for details).

Incidence of VTE in patients with MM

One case-control study was excluded, and 13 cohort studies were finally included for meta-analysis of VTE incidence. The incidence of VTE reported in each study ranged from 4.8% to 16.4%. A pooled analysis using a random effects model showed a combined VTE incidence of 7.9% (95% CI [6.2–10.1%]; I 2 = 93.1%, p < 0.001) ( Fig. 2 ). Sensitivity analysis was performed by sequentially excluding each individual study ( Fig. 3 ). The pooled incidence of VTE remained stable, ranging from 7.4% to 8.3%, and the heterogeneity remained high (I 2 = 89.8~93.6%). Further subgroup analysis by follow-up time (6 months vs . 12 months) showed a combined incidence rate of 6.6% (95% CI [4.8–8.9%]; I 2 = 90%, p < 0.001) in the 6-month follow-up group ( Chalayer et al., 2022 ; Sanfilippo et al., 2019 ; Gursoy et al., 2025 ; Bao et al., 2024 ; Sanfilippo et al., 2022 ; Covut et al., 2021 ); The combined incidence in the 12-month follow-up group ( Li et al., 2023 ; Chakraborty et al., 2022 ; Li et al., 2019 ; da Costa et al., 2023 ; Dima et al., 2023 ) was 9.5% (95% CI [6.6–13.5%]; I² = 91.1%, p < 0.001). The results of the random effects model analysis showed that there was no statistically significant difference in the incidence of VTE between groups at different follow-up times (χ² = 2.33, p = 0.127). The results of the publication bias assessment showed that the overall funnel plot distribution was approximately symmetrical, with no obvious asymmetry ( Fig. 4 ). The Egger test also did not indicate significant publication bias (t = 1.28, p = 0.228).

Basic information of the models

Fourteen studies reported clinical practice with one or more of seven risk assessment tools, including IMPEDE VTE score, SAVED score, IMWG guideline, PRISM score, IMPEDED VTE score, RAM score, and Khorana score. In terms of missing data processing, seven studies ( Chakraborty et al., 2022 ; Sanfilippo et al., 2019 ; Gursoy et al., 2025 ; Fernandez-Vargas et al., 2024 ; da Costa et al., 2023 ; Sanfilippo et al., 2022 ; Covut et al., 2021 ) excluded patients with missing data, two studies ( Li et al., 2023 ; Dima et al., 2023 ) used complete datasets, and the remaining five studies ( Chalayer et al., 2022 ; Li et al., 2019 ; Bao et al., 2024 ; Bravo-Perez et al., 2021 ; Sanfilippo et al., 2023 ) mentioned missing data processing methods. In terms of predictor selection, five studies ( Li et al., 2023 ; Chakraborty et al., 2022 ; Sanfilippo et al., 2019 ; Li et al., 2019 ; Sanfilippo et al., 2023 ) employed univariate analysis to screen candidate variables during the model construction phase. For model development and validation, four studies ( Li et al., 2023 ; Chakraborty et al., 2022 ; Sanfilippo et al., 2019 ; Li et al., 2019 ) conducted model development and adopted both internal validation (Bootstrapping) and external validation, nine studies ( Chalayer et al., 2022 ; Gursoy et al., 2025 ; Fernandez-Vargas et al., 2024 ; Bao et al., 2024 ; da Costa et al., 2023 ; Dima et al., 2023 ; Sanfilippo et al., 2022 ; Bravo-Perez et al., 2021 ; Covut et al., 2021 ) conducted independent external validation, and one study ( Sanfilippo et al., 2023 ) did not report validation methods. During model development, the most commonly used methods were Cox regression ( n = 4) ( Li et al., 2023 ; Chakraborty et al., 2022 ; Sanfilippo et al., 2019 ; Li et al., 2019 ) and logistic regression ( n = 1) ( Sanfilippo et al., 2023 ) (See Table 2 for details).

Model performance index and presentation

We performed a pooled analysis of AUC for seven risk assessment tools (IMPEDE VTE score, SAVED score, IMWG guideline, PRISM score, IMPEDED VTE score, RAM score, Khorana score). The pooled AUC values ranged from 0.57 to 0.68. Estimates for I 2 range from 0% to 93%. The best predictive performance was achieved with the IMPEDE VTE score and the IMPEDED VTE score. The predictive efficacy of the IMPEDE VTE score was evaluated in nine studies ( Chalayer et al., 2022 ; Sanfilippo et al., 2019 ; Gursoy et al., 2025 ; Fernandez-Vargas et al., 2024 ; Bao et al., 2024 ; da Costa et al., 2023 ; Bravo-Perez et al., 2021 ; Covut et al., 2021 ; Sanfilippo et al., 2023 ), of which data from seven studies ( Gursoy et al., 2025 ; Fernandez-Vargas et al., 2024 ; Bao et al., 2024 ; da Costa et al., 2023 ; Bravo-Perez et al., 2021 ; Covut et al., 2021 ; Sanfilippo et al., 2023 ) were available for pooled analysis with a pooled AUC of 0.68 (95% CI [0.63–0.72]; I 2 = 28.7%, p = 0.163). The IMPEDED VTE score was assessed in two studies ( Gursoy et al., 2025 ; Sanfilippo et al., 2023 ) with a combined AUC of 0.68 (95% CI [0.62–0.74]; I 2 = 0.0%, p = 0.445). RAM score ( Li et al., 2023 ) (AUC = 0.62) and Khorana score ( Sanfilippo et al., 2022 ) (AUC = 0.53) were assessed in only one study, and no pooled analysis was performed. In terms of calibration analysis, only two studies ( Sanfilippo et al., 2019 , 2023 ) used the Hosmer-Lemeshow test to assess model calibration, and meta-analyses could not be performed due to inadequate reporting in external validation studies. All models are presented as formulas. For the models that can be analyzed in combination, their AUC results are summarized and presented in a single multi-panel plot to visually compare the prediction performance of each model ( Table 3 , Fig. 5 ). A limited number of studies (2–7) were included in each AUC pooled analysis, so no formal publication bias assessment was performed.

Risk of bias assessment

The quality assessment was performed according to the PROBAST tool, and the 14 studies included all showed a high risk of bias, mainly focused on the results domain and the data analysis domain. Involving the following: In the predictor domain, two studies ( Fernandez-Vargas et al., 2024 ; da Costa et al., 2023 ) did not report whether the measurement method of the predictor was uniform, and one study ( Sanfilippo et al., 2023 ) assessed the predictor with knowledge of the outcome due to study design issues. In the outcome domain, seven studies ( Chalayer et al., 2022 ; Chakraborty et al., 2022 ; Gursoy et al., 2025 ; Fernandez-Vargas et al., 2024 ; da Costa et al., 2023 ; Sanfilippo et al., 2022 ; Covut et al., 2021 ) had non-standardized outcome definitions, and three studies ( Chalayer et al., 2022 ; Gursoy et al., 2025 ; Sanfilippo et al., 2023 ) failed to exclude factors overlapping with predictors. In addition, one study ( Sanfilippo et al., 2023 ) did not provide information on the blinded assessment of outcome-predictor relationships and failed to clarify whether there was an appropriate time interval between predictor assessment and outcome determination. In the area of data analysis, three studies ( Fernandez-Vargas et al., 2024 ; da Costa et al., 2023 ; Sanfilippo et al., 2023 ) failed to meet the recommendation of “events per variable” (EPV) exceeding 20, and seven studies ( Chakraborty et al., 2022 ; Sanfilippo et al., 2019 ; Gursoy et al., 2025 ; Fernandez-Vargas et al., 2024 ; da Costa et al., 2023 ; Sanfilippo et al., 2022 ; Covut et al., 2021 ) directly eliminated missing data, and five studies ( Li et al., 2023 ; Chakraborty et al., 2022 ; Sanfilippo et al., 2019 ; Li et al., 2019 ; Sanfilippo et al., 2023 ) used univariate analysis to screen variables. Only two studies ( Sanfilippo et al., 2019 ; Sanfilippo et al., 2023 ) evaluated the discrimination and calibration of the model simultaneously, and none of them evaluated the clinical utility. In terms of applicability risk assessment, six studies ( Chalayer et al., 2022 ; Gursoy et al., 2025 ; Fernandez-Vargas et al., 2024 ; da Costa et al., 2023 ; Sanfilippo et al., 2022 , 2023 ) were rated as at high risk of bias, mainly because the definition of predictors was not clearly stated in the text, and some studies used subjectively defined outcome measures (See Table 4 for details).

Discussion

This systematic review provides a comprehensive analysis of the predictive performance of existing VTE risk prediction models in patients with MM, involving the development and/or validation of seven risk assessment tools. The results showed that the overall predictive efficacy of the existing VTE risk prediction models for MM patients was weak, and the comprehensive AUC values were less than 0.7. In addition, all included studies had a high risk of bias. Therefore, this study highlights the need to optimize and update VTE prediction models in MM patients to improve their clinical applicability and predictive accuracy.

In this study, the combined incidence of VTE in MM patients was 7.9% (95% CI [6.2–10.1%]), with significant heterogeneity in the included studies (I² = 93.1%), and the heterogeneity could not be reduced by subgroup analyses (6-month vs . 12-month follow-up). This could be attributed to inconsistencies in population characteristics, treatment modalities, or outcome definitions. Regarding the study population, some studies included only patients with newly diagnosed MM, whereas others included patients with both relapsed and refractory MM, and the heterogeneity of the study population may have been exacerbated by differences in disease severity, treatment options, and risk of complications by disease stage. In addition, the proportion of IMiDs used varied greatly (14.3–100%). Although IMiDs have antitumor activity, their antiangiogenic effects also significantly increase the risk of thrombosis ( Epstein & Morrison, 2022 ; Raghupathy et al., 2017 ). Differences in anticoagulation prevention strategies are also one of the influencing factors, as in Chalayer et al. (2022) , more than 90% of patients received VTE prophylaxis (mainly low molecular heparin), whereas the studies by Fernandez-Vargas et al. (2024) and Bao et al. (2024) excluded patients who received anticoagulation. The definition of outcome measures is also inconsistent. Most studies have deep vein thrombosis and pulmonary embolism as the primary endpoints, while some studies ( Sanfilippo et al., 2019 ; Gursoy et al., 2025 ; Fernandez-Vargas et al., 2024 ; Bao et al., 2024 ; Bravo-Perez et al., 2021 ; Sanfilippo et al., 2023 ) include superficial venous thrombosis and central venous thrombosis. Studies by Bao et al. (2024) and Sanfilippo et al. (2023) both found that the placement of central venous catheters significantly increased the risk of VTE, and the incidence rate was nearly three times higher than that of patients without catheterization (21.2% vs . 6.0%; 32% vs . 10%).

The literature included in this study covers the development and/or validation of seven risk assessment tools including IMPEDE VTE score, SAVED score, IMWG guidelines, PRISM score, IMPEDED VTE score, RAM score, and Khorana score. Due to the strong heterogeneity among the included studies, the AUC of all models was not analyzed overall, but five of the models (IMPEDE VTE score ( Gursoy et al., 2025 ; Fernandez-Vargas et al., 2024 ; Bao et al., 2024 ; da Costa et al., 2023 ; Bravo-Perez et al., 2021 ; Covut et al., 2021 ; Sanfilippo et al., 2023 ), SAVED score ( Li et al., 2023 , 2019 ; Gursoy et al., 2025 ; Fernandez-Vargas et al., 2024 ; da Costa et al., 2023 ; Dima et al., 2023 ), IMWG guidelines ( Li et al., 2023 ; da Costa et al., 2023 ), PRISM score ( Chakraborty et al., 2022 ; Fernandez-Vargas et al., 2024 ), and IMPEDED VTE score ( Gursoy et al., 2025 ; Sanfilippo et al., 2023 ) were analyzed separately. The results showed that the combined AUC of these models ranged from 0.57 to 0.68, all of which were lower than 0.7, suggesting their limited discrimination ability. Furthermore, for the RAM score ( Li et al., 2023 ) and Khorana score ( Sanfilippo et al., 2022 ), which had a small number of original studies, the reported AUC alone was also less than 0.7, indicating that its predictive power was also weak. Therefore, existing VTE risk prediction tools for MM patients should be used with caution in clinical practice.

We compared the performance of multiple prediction models and found that IMPEDED VTE score and IMPEDE VTE score had the best predictive performance, with pooled AUCs of AUC = 0.68 (95% CI [0.62–0.74]) and AUC = 0.68 (95% CI [0.63–0.72]), respectively, and the heterogeneity of the two models across studies was low (I² < 30%), suggesting their stability in different populations. The IMPEDED VTE score was constructed based on the IMPEDE VTE score by integrating D-dimer metrics, and the predictive power of the model (AUC = 0.65 vs . AUC = 0.61) was improved ( Sanfilippo et al., 2023 ). D-dimer has a high sensitivity in the diagnosis of suspected VTE, normal levels help to exclude VTE and guide anticoagulation therapy, and its dynamic changes better reflect the evolution of thrombotic risk ( Yang et al., 2020 ; Posch et al., 2020 ). Posch et al. (2020) demonstrated that for every doubling of the trajectory level of D-dimer, the risk of VTE was increased by 2.8 times (95% CI [1.69–4.58]). Gibson et al. (2017) also found that the inclusion of D-dimer in the IMPROVE score improved the ability to discriminate VTE risk in medical patients (Δ AUC: 0.06 (95% CI [0.02–0.09]), p = 0.0006). In addition to D-dimer, other novel biomarkers, such as platelet activation markers ( e.g ., P-selectin), coagulation factor levels, and inflammatory markers ( e.g ., IL-6, fibrinogen), also show potential predictive value and can provide supplementary information for future VTE risk models ( Joy, Venkatachalam & Binesh, 2026 ; Riondino et al., 2019 ). Combining these biomarkers for longitudinal monitoring and integrating dynamic indicators with clinical features is expected to further optimize the performance of existing predictive models. Additionally, using machine learning or multi-omics methods to construct joint multi-marker prediction models may improve the ability to identify individualized risk for different patient groups. It should be noted that the IMPEDED VTE score has been validated only in a small-sample retrospective study ( Gursoy et al., 2025 ), and its broad applicability remains to be confirmed. Future studies should verify the model’s performance in large samples and across different populations, and explore the gain effect of novel biomarkers in dynamic monitoring and model optimization, to improve the accuracy and clinical applicability of VTE risk prediction.

PROBAST tool evaluation showed that all included studies had a high risk of bias, mainly reflected in insufficient sample size, improper handling of missing data, non-standard independent variable screening methods, and incomplete evaluation of model performance. Although da Costa et al. (2023) reported a high AUC value (0.80), their small sample size ( n = 130) may have led to the overfitting of the model in the training set, thus limiting its ability to generalize to a wider population ( Wynants et al., 2015 ). PROBAST recommends multiple imputations for handling missing data. However, most studies in this study directly excluded missing values ( n = 7), which may lead to information loss and selective bias ( Xin et al., 2025 ). Although Sanfilippo et al. (2019) and Li et al. (2019) had relatively large sample sizes, variable screening relied solely on univariate analysis and failed to fully consider interactions between variables, potentially missing key predictors and affecting model stability and predictive performance ( Chowdhury & Turin, 2020 ). It is suggested that more advanced variable selection methods such as LASSO regression ( McNeish, 2015 ) should be adopted in future research to effectively deal with multicollinearity problems and improve the stability and generalization ability of the model.

Currently, the traditional Cox regression ( Li et al., 2023 ; Chakraborty et al., 2022 ; Sanfilippo et al., 2019 ; Li et al., 2019 ) and logistic regression ( Sanfilippo et al., 2023 ) are still used for model construction, which has limitations in dealing with nonlinear relationships and confounding factors and may limit the improvement of model performance ( Dalicandro et al., 2021 ). In the future, diversified modeling techniques such as machine learning and deep learning can be considered to enhance the accuracy and adaptability of prediction models ( Koppe, Meyer-Lindenberg & Durstewitz, 2021 ). Furthermore, although all studies reported the discriminative ability of the models ( e.g ., AUC), only two ( Sanfilippo et al., 2019 , 2023 ) further evaluated their calibration performance. The degree of calibration measures the consistency between the predicted probability and the actual incidence rate and is an important indicator to evaluate the clinical practicability of the model, which is usually presented by a calibration plot ( Steyerberg & Vergouwe, 2014 ). The lack of systematic calibration assessment may lead to one-sided interpretations of model efficacy and increase the risk of reporting bias. Another key issue is that none of the available studies conducted clinical utility assessments. Although statistical indicators such as discrimination and calibration can reflect model performance, they cannot directly evaluate their value in clinical decision-making. Future research should introduce methods such as decision curve analysis to clarify whether the model truly “benefits outweigh disadvantages” in practical application.

Although the IMPEDE score has been recommended by the National Comprehensive Cancer Network (NCCN) ( Streiff et al., 2021 ) to assess the risk of VTE in patients with MM, where those with a score ≥8 are classified as high risk and recommended for anticoagulation prophylaxis, this systematic review shows that this score failed to effectively identify high-risk patients in some studies, suggesting that it may have limitations in specific patient groups. This result is consistent with the International Myeloma Working Group (IMWG) and European Multiple Myeloma Network (EMN) guidelines: the guidelines recommend standard risk stratification for MM patients receiving IMiDs, emphasize that high-risk patients should implement anticoagulation prophylaxis, and point out that existing scoring tools still need to be externally validated in different patient populations ( Yang et al., 2023 ). These findings suggest that existing risk scores may not capture all high-risk patients; therefore, external validation through large-sample, multicenter, prospective studies is still required before routine clinical application. In the future, optimizing risk assessment models in combination with dynamic clinical features or novel biomarkers may improve their applicability across different clinical settings and patient populations.

This study still has the following limitations: (a) Most of the included studies are based on European and American populations, which may limit the generalization of the study results in other populations; (b) only English literature is included, which may omit important studies in other languages and affect the comprehensiveness of the results; (c) all included studies are retrospective cohort or case-control studies, which are susceptible to selection bias and information bias, and the completeness and accuracy of data collection are difficult to guarantee, which may reduce the reliability and external validity of the results; (d) since 95% CI was not reported for AUC in some studies, this study was excluded, which may cause some bias to the quantitative synthesis results.

Conclusion

A total of 14 studies on risk prediction models for VTE in MM patients were included in this systematic review, involving the development and/or validation of seven risk assessment tools. Among them, the IMPEDE VTE score and IMPEDED VTE score had the best predictive performance, but both lacked the effective identification ability of high-risk VTE patients (pooled AUC < 0.7). In addition, all included studies had a high risk of bias as assessed by PROBAST. Therefore, it is recommended that future studies strictly follow the TRIPOD statement for study design and standardized reporting, and optimize or update the existing models in multicenter, large-sample prospective cohorts by optimizing variable selection, improving modeling methods, and introducing new biomarkers to improve their predictive efficacy and clinical applicability.

Supplemental Information

PRISMA checklist.

PubMed searchable.

Raw data.

Additional Information and Declarations

Competing Interests

The authors declare that they have no competing interests.

Author Contributions

Jin Yang conceived and designed the experiments, performed the experiments, analyzed the data, authored or reviewed drafts of the article, and approved the final draft.

You Pu performed the experiments, analyzed the data, prepared figures and/or tables, and approved the final draft.

Yuqin Li performed the experiments, analyzed the data, authored or reviewed drafts of the article, and approved the final draft.

Xuelian Li performed the experiments, prepared figures and/or tables, and approved the final draft.

Xia Jiang analyzed the data, authored or reviewed drafts of the article, and approved the final draft.

Liping Zou conceived and designed the experiments, authored or reviewed drafts of the article, and approved the final draft.

Data Availability

The following information was supplied regarding data availability:

The data is available in the Supplemental File .

References

Bao et al. (2024) Bao L Fang LJ Xiao MY Lu MQ Chu B Shi L Gao S Xiang QQ Wang YT Liu X Ding YH Xin Z Chen Y Wang MZ Hu WK Guo CY Chen LY Sun K Validation of the IMPEDE VTE score for prediction of venous thromboembolism in Chinese patients with multiple myeloma: a single-center retrospective cohort study Thrombosis Research 2024 236 130 135 10.1016/j.thromres.2024.02.011 38430904

Beckman et al. (2010) Beckman MG Hooper WC Critchley SE Ortel TL Venous thromboembolism: a public health concern American Journal of Preventive Medicine 2010 38 4 S495 S501 10.1016/j.amepre.2009.12.017 20331949

Bradbury et al. (2020) Bradbury CA Craig Z Cook G Pawlyn C Cairns DA Hockaday A Paterson A Jenner MW Jones JR Drayson MT Owen RG Kaiser MF Gregory WM Davies FE Child JA Morgan GJ Jackson GH Thrombosis in patients with myeloma treated in the myeloma IX and myeloma XI phase 3 randomized controlled trials Blood 2020 136 9 1091 1104 10.1182/blood.2020005125 32438407 PMC7453153

Bravo-Perez et al. (2021) Bravo-Perez C Fernández-Caballero M Soler-Espejo E Garcia-Torralba E Sorigue M García-Malo MD Jerez A Vicente V Roldán V Arriba F Heparin versus aspirin thromboprophylaxis adds independent value to IMPEDE-VTE score for venous thrombosis prediction in multiple myeloma Journal of Thrombosis and Thrombolysis 2021 52 848 853 10.1007/s11239-021-02407-5 33649980

Chakraborty et al. (2022) Chakraborty R Rybicki L Wei W Valent J Faiman BM Samaras CJ Anwer F Khorana AA Abnormal metaphase cytogenetics predicts venous thromboembolism in myeloma: derivation and validation of the PRISM score Blood 2022 140 23 2443 2450 10.1182/blood.2022015727 35772005

Chalayer et al. (2022) Chalayer E Talbot A Frenzel L Karlin L Collet P Guyotat D Attal M Leleu X Tardy B Prediction of venous thromboembolism in patients with multiple myeloma treated with lenalidomide, bortezomib, dexamethasone, and transplantation: lessons from the substudy of IFM/DFCI, 2009 cohort Journal of Thrombosis and Haemostasis 2022 20 8 1859 1867 10.1111/jth.15758 35557490

Chowdhury & Turin (2020) Chowdhury MZI Turin TC Variable selection strategies and its importance in clinical prediction modelling Family Medicine and Community Health 2020 8 e000262 10.1136/fmch-2019-000262 32148735 PMC7032893

Covut et al. (2021) Covut F Ahmed R Chawla S Ricaurte F Samaras CJ Anwer F Garcia AVM Angelini DE Mazzoni S Faiman B Valent J Khouri J Validation of the IMPEDE VTE score for prediction of venous thromboembolism in multiple myeloma: a retrospective cohort study British Journal of Haematology 2021 193 6 1213 1219 10.1111/bjh.17505 33997961

da Costa et al. (2023) da Costa IHF de Pádua CAM de Miranda Drummond PL Silveira LP Malta JS Santos RMMD Reis AMM Comparison of three risk assessment models for thromboembolism in multiple myeloma patients receiving immunomodulators: a Brazilian historical cohort Journal of Thrombosis and Thrombolysis 2023 56 147 155 10.1007/s11239-023-02817-7 37133703

Dalicandro et al. (2021) Dalicandro L Harder JA Mazmanian D Weaver B How prevalent is overfitting of regression models? A survey of recent articles in three psychology journals The Quantitative Methods for Psychology 2021 17 1 6 10.20982/tqmp.17.1.p001

de Hond, Steyerberg & van Calster (2022) de Hond AAH Steyerberg EW van Calster B Interpreting area under the receiver operating characteristic curve The Lancet Digital Health 2022 4 12 e853 e855 10.1016/S2589-7500(22)00188-1 36270955

Dima et al. (2023) Dima D Li A Granat LM Dhillon P Chamseddine F Yalamanchali A Mirzai S Wei W Samaras CJ Valent J Anwer F Khouri J External validation of the SAVED score for venous thromboembolism risk stratification in patients with multiple myeloma receiving immunomodulatory drugs British Journal of Haematology 2023 201 2 280 284 10.1111/bjh.18630 36604838

Egger et al. (1997) Egger M Smith GD Schneider M Minder C Bias in meta-analysis detected by a simple, graphical test BMJ 1997 315 629 634 10.1136/bmj.315.7109.629 9310563 PMC2127453

Elbahoty, Papineni & Samant (2024) Elbahoty MH Papineni B Samant RS Multiple myeloma: clinical characteristics, current therapies and emerging innovative treatments targeting ribosome biogenesis dynamics Clinical & Experimental Metastasis 2024 41 829 842 10.1007/s10585-024-10305-2 39162964 PMC11607061

Epstein & Morrison (2022) Epstein M Morrison C Practical guidance for new multiple myeloma treatment regimens: a nursing perspective Seminars in Oncology 2022 49 1 103 117 10.1053/j.seminoncol.2022.01.010 35197198 PMC9149030

Escobar et al. (2022) Escobar A Salem AM Dickson K Johnson TN Burk KJ Bashoura L Faiz SA Anticoagulation and bleeding in the cancer patient Support Care Cancer 2022 30 8547 8557 10.1007/s00520-022-07136-w 35579752 PMC9529787

Fernandez-Vargas et al. (2024) Fernandez-Vargas OE Amezcua I Cabello B Martinez AQ Espinoza R Cesarman-Maus G Multiple myeloma: retrospective assessment of routine thromboprophylaxis and utility of thrombotic risk scores Research and Practice in Thrombosis and Haemostasis 2024 8 7 102571 10.1016/j.rpth.2024.102571 39552772 PMC11564956

Fotiou, Dimopoulos & Kastritis (2022) Fotiou D Dimopoulos MA Kastritis E Approach to contemporary risk assessment, prevention and management of thrombotic complications in multiple myeloma Cancers 2022 14 24 6216 10.3390/cancers14246216 36551701 PMC9777181

Gibson et al. (2017) Gibson CM Spyropoulos AC Cohen AT Hull RD Goldhaber SZ Yusen RD Hernandez AF Korjian S Daaboul Y Gold A Harrington RA Chi G The IMPROVEDD VTE risk score: incorporation of D-dimer into the IMPROVE score to improve venous thromboembolism risk stratification TH Open: Companion Journal to Thrombosis and Haemostasis 2017 1 1 e56 e65 10.1055/s-0037-1603929 31249911 PMC6524839

Gursoy et al. (2025) Gursoy V Baysal M Sadri S Hunutlu FC Ersal T Gul OO Kose E Celik E Baysal S Koca TG Cubukcu S Ergun E Yavuz S Ozkocaman V Ozkalemkas F Comparative analysis and validation of the IMPEDED VTE, IMPEDE VTE, and SAVED risk models in predicting venous thromboembolism in multiple myeloma patients: a retrospective study in Türkiye Diagnostics 2025 15 5 2025 2633 10.3390/diagnostics15050633 40075881 PMC11899139

Horner et al. (2024) Horner DE Davis S Pandor A Shulver H Goodacre S Hind D Rex S Gillett M Bursnall M Griffin X Holland M Hunt BJ de Wit K Bennett S Pierce-Williams R Evaluation of venous thromboembolism risk assessment models for hospital inpatients: the VTEAM evidence synthesis Health Technology Assessment 2024 28 20 1 166 10.3310/AWTW6200 PMC11056814 38634415

Huedo-Medina et al. (2006) Huedo-Medina TB Sánchez-Meca J Marín-Martínez F Botella J Assessing heterogeneity in meta-analysis: Q statistic or I2 index? Psychological Methods 2006 11 193 206 10.1037/1082-989X.11.2.193 16784338

Joy, Venkatachalam & Binesh (2026) Joy S Venkatachalam K Binesh A Emerging perspectives on biomarker panels in deep vein thrombosis: the combined roles of D-dimer, P-selectin, and microparticles Annals of Vascular Surgery 2026 122 781 791 10.1016/j.avsg.2025.08.050 40946911

Khorana et al. (2022) Khorana AA Mackman N Falanga A Pabinger I Noble S Ageno W Moik F Lee AYY Cancer-associated venous thromboembolism Nature Reviews Disease Primers 2022 8 11 10.1038/s41572-022-00336-y 35177631

Koppe, Meyer-Lindenberg & Durstewitz (2021) Koppe G Meyer-Lindenberg A Durstewitz D Deep learning for small and big data in psychiatry Neuropsychopharmacology 2021 46 176 190 10.1038/s41386-020-0767-z 32668442 PMC7689428

Kristinsson et al. (2012) Kristinsson SY Pfeiffer RM Björkholm M Schulman S Landgren O Thrombosis is associated with inferior survival in multiple myeloma Haematologica 2012 97 10 1603 1607 10.3324/haematol.2012.064444 22511493 PMC3487563

Li et al. (2023) Li X Sun X Fang B Leng Y Sun F Wang Y Wang Q Jin J Yang M Xu B Fang Z Chen L Chen Z Yang Q Zhang K Ye Y Geng H Sun Z Hao D Huang H Wang X Jing H Ma L Pan X Chen W Li J Development and validation of a new risk assessment model for immunomodulatory drug-associated venous thrombosis among Chinese patients with multiple myeloma Thrombosis Journal 2023 21 7 2023 2105 10.1186/s12959-023-00534-y PMC10552366 37794471

Li et al. (2019) Li A Wu Q Luo S Warnick GS Zakai NA Libby EN Gage BF Garcia DA Lyman GH Sanfilippo KM Derivation and validation of a risk assessment model for immunomodulatory drug-associated thrombosis among patients with multiple myeloma Journal of the National Comprehensive Cancer Network 2019 17 7 840 847 10.6004/jnccn.2018.7273 31319391 PMC7808759

Long et al. (2025) Long B McCurdy A Koyfman A Rosenberg H An emergency medicine review: multiple myeloma and its complications The American Journal of Emergency Medicine 2025 88 172 179 10.1016/j.ajem.2024.11.073 39643958

Mafra et al. (2024) Mafra A Laversanne M Marcos-Gragera R Chaves HVS Mcshane C Bray F Znaor A The global multiple myeloma incidence and mortality burden in 2022 and predictions for 2045 Journal of the National Cancer Institute 2024 117 5 907 914 10.1093/jnci/djae321 39658225

McCurdy et al. (2023) McCurdy A Seow H Pond GP Gayowsky A Chakraborty R Visram A Kaedbey R D’Souza A Mohyuddin GR Wildes TM Fonseca R Mian H Cancer-specific mortality in multiple myeloma: a population-based retrospective cohort study Haematologica 2023 108 12 3384 3391 10.3324/haematol.2023.282905 37439357 PMC10690919

McNeish (2015) McNeish DM Using Lasso for predictor selection and to assuage overfitting: a method long overlooked in behavioral sciences Multivariate Behavioral Research 2015 50 5 471 484 10.1080/00273171.2015.1036965 26610247

Moons et al. (2014) Moons KGM de Groot JAH Bouwmeester W Vergouwe Y Mallett S Altman DG Reitsma JB Collins GS Critical appraisal and data extraction for systematic reviews of prediction modelling studies: the CHARMS checklist PLOS Medicine 2014 11 10 e1001744 10.1371/journal.pmed.1001744 25314315 PMC4196729

Moons et al. (2019) Moons KGM Wolff RF Riley RD Whiting PF Westwood M Collins GS Reitsma JB Kleijnen J Mallett S PROBAST: a tool to assess risk of bias and applicability of prediction model studies: explanation and elaboration Annals of Internal Medicine 2019 170 W1 10.7326/M18-1377 30596876

Palumbo et al. (2011) Palumbo A Cavo M Bringhen S Zamagni E Romano A Patriarca F Rossi D Gentilini F Crippa C Galli M Nozzoli C Ria R Marasca R Montefusco V Baldini L Elice F Callea V Pulini S Carella AM Zambello R Benevolo G Magarotto V Tacchetti P Pescosta N Cellini C Polloni C Evangelista A Caravita T Morabito F Offidani M Tosi P Boccadoro M Aspirin, warfarin, or enoxaparin thromboprophylaxis in patients with multiple myeloma treated with thalidomide: a phase III, open-label, randomized trial Journal of Clinical Oncology 2011 29 986 993 10.1200/JCO.2010.31.6844 21282540

Posch et al. (2020) Posch F Riedl J Reitter E Crowther MJ Grilz E Quehenberger P Jilma B Pabinger I Ay C Dynamic assessment of venous thromboembolism risk in patients with cancer by longitudinal D-dimer analysis: a prospective study Journal of Thrombosis and Haemostasis 2020 18 6 1348 1356 10.1111/jth.14774 32073229 PMC7317804

R Core Team (2024) R Core Team R: a language and environment for statistical computing 2024 Version 4.4.0 Vienna R Foundation for Statistical Computing https://www.r-project.org

Raghupathy et al. (2017) Raghupathy R Ayyappan S Prabhakar D Campagnaro EL Mo FKF Li W Hui EP Chan ATC Lei KIK Retrospective study of the incidence and patterns of arterial and venous thrombosis in Chinese versus African American patients with multiple myeloma British Journal of Haematology 2017 176 315 317 10.1111/bjh.13942 26814167

Riondino et al. (2019) Riondino S Ferroni P Zanzotto FM Roselli M Guadagni F Predicting VTE in cancer patients: candidate biomarkers and risk assessment models Cancers 2019 11 95 10.3390/cancers11010095 30650562 PMC6356247

Sanfilippo et al. (2022) Sanfilippo KM Carson KR Wang TF Luo S Edwin N Kuderer N Keller JM Gage BF Evaluation of the khorana score for prediction of venous thromboembolism in patients with multiple myeloma Research and Practice in Thrombosis and Haemostasis 2022 6 e12634 10.1002/rth2.12634 35028491 PMC8742966

Sanfilippo et al. (2023) Sanfilippo KM Fiala MA Feinberg D Tathireddy H Girard T Vij R Paola JDi Gage BF D-dimer predicts venous thromboembolism in multiple myeloma: a nested case-control study Research and Practice in Thrombosis and Haemostasis 2023 7 102235 10.1016/j.rpth.2023.102235 38193055 PMC10772878

Sanfilippo et al. (2019) Sanfilippo KM Luo S Wang TF Fiala M Schoen M Wildes TM Mikhael J Kuderer NM Calverley DC Keller J Thomas T Carson KR Gage BF Predicting venous thromboembolism in multiple myeloma: development and validation of the IMPEDE VTE score American Journal of Hematology 2019 94 1176 1184 10.1002/ajh.25603 31379000 PMC7058359

Sborov et al. (2022) Sborov DW Baljevic M Reeves B Laubach J Efebera YA Rodriguez C Costa LJ Chari A Silbermann R Holstein SA Anderson LD Jr Kaufman JL Shah N Pei H Patel S Cortoos A Bartlett JB Vermeulen J Lin TS Voorhees PM Richardson PG Daratumumab plus lenalidomide, bortezomib and dexamethasone in newly diagnosed multiple myeloma: analysis of vascular thrombotic events in the GRIFFIN study British Journal of Haematology 2022 199 355 365 10.1111/bjh.18432 36111391 PMC9825872

Schoen et al. (2020) Schoen MW Carson KR Luo S Gage BF Li A Afzal A Sanfilippo KM Venous thromboembolism in multiple myeloma is associated with increased mortality Research and Practice in Thrombosis and Haemostasis 2020 4 1203 1210 10.1002/rth2.12411 33134785 PMC7590313

Stefano et al. (2022) Stefano VDe Larocca A Carpenedo M Cavo M Raimondo FDi Falanga A Offidani M Petrucci MT Ruggeri M Santi RM Barosi G Thrombosis in multiple myeloma: risk stratification, antithrombotic prophylaxis, and management of acute events. A consensus-based position paper from an ad hoc expert panel Haematologica 2022 107 2536 2547 10.3324/haematol.2022.280893 35861017 PMC9614522

Steyerberg & Vergouwe (2014) Steyerberg EW Vergouwe Y Towards better clinical prediction models: seven steps for development and an ABCD for validation European Heart Journal 2014 35 1925 1931 10.1093/eurheartj/ehu207 24898551 PMC4155437

Streiff et al. (2021) Streiff MB Holmstrom B Angelini D Ashrani A Elshoury A Fanikos J Fertrin KY Fogerty AE Gao S Goldhaber SZ Gundabolu K Ibrahim I Kraut E Leavitt AD Lee A Lee JT Lim M Mann J Martin K McMahon B Moriarty J Morton C Ortel TL Paschal R Schaefer J Shattil S Siddiqi T Sudheendra D Williams E Hollinger L Nguyen MQ Cancer-associated venous thromboembolic disease, Version 2.2021, NCCN clinical practice guidelines in oncology Journal of the National Comprehensive Cancer Network: JNCCN 2021 19 1181 1201 34666313 10.6004/jnccn.2021.0047

Wynants et al. (2015) Wynants L Bouwmeester W Moons KGM Moerbeek M Timmerman D Huffel SV Calster BV Vergouwe Y A simulation study of sample size demonstrated the importance of the number of events per variable to develop prediction models in clustered data Journal of Clinical Epidemiology 2015 68 1406 1414 10.1016/j.jclinepi.2015.02.002 25817942

Xin et al. (2025) Xin Y Song R Hao J Li W Wu C Zuo L Cai Y Zhang X Wu H Hui W Poor reporting quality and high proportion of missing data in economic evaluations alongside pragmatic trials: a cross-sectional survey BMC Medical Research Methodology 2025 25 61 10.1186/s12874-025-02519-z 40050714 PMC11884024

Yang et al. (2023) Yang B Liu C Lin Z Geng C Zhang Z Nursing management of treatment-related venous thromboembolism in patients with multiple myeloma Frontiers in Medicine 2023 10 1153694 10.3389/fmed.2023.1153694 37144028 PMC10151651

Yang et al. (2020) Yang M Qi J Tang Y Wu D Han Y Increased D-dimer predicts the risk of cancer-associated recurrent venous thromboembolism and venous thromboembolism: a systematic review and meta-analysis Thrombosis Research 2020 196 410 413 10.1016/j.thromres.2020.09.031 33022489