Leukemia is a group of hematologic malignancies, primarily classified into acute leukemia (AL) and chronic leukemia. Acute leukemia includes acute lymphoblastic leukemia (ALL) and acute myeloid leukemia (AML). Among these, acute promyelocytic leukemia (APL) represents a distinct and rare subtype of AML. Thrombosis is one of the critical complications requiring vigilance in leukemia patients, leading to multiple negative impacts. These include chemotherapy interruption and increased bleeding risk due to anticoagulant therapy, post-thrombotic syndrome and various sequelae, as well as shortened overall survival and progression-free survival (1, 2). From the perspective of thrombosis risk in the overall cancer population, thrombosis ranks as the second leading cause of mortality and morbidity, surpassed only by the cancer itself (3–5). Data indicate that cancer patients face a four to ninefold increased risk of thrombosis compared to the general population (6–8), and cancer is associated with 20-30% of initial venous thrombotic events (9). Moreover, the risk of thrombosis in hematologic malignancies is similar to or even higher than that in solid tumor patients (10, 11). Specifically, regarding leukemia subtypes, the thrombosis incidence rate in ALL is 1.7-16% (12–22), in AML (excluding APL) it is 1.6-14.6% (23–27), while APL, due to the potential coexistence of thrombotic and hemorrhagic manifestations (28), has a thrombosis incidence rate of 5.2-20.6% (29–32). Thrombotic events in leukemia are predominantly venous thromboembolism (VTE), including deep vein thrombosis, pulmonary embolism, and central venous catheter-related thrombosis, while arterial thromboembolism (ATE) has been less studied (33, 34). Therefore, early identification of leukemia patients at high thrombosis risk and implementing targeted prevention are key to reducing the thrombotic burden and improving clinical outcomes.

Currently, the commonly used risk prediction models for cancer-associated thrombosis are the Khorana score (35) and the Vienna Cancer and Thrombosis Study (CATS) score (36). The Khorana score includes five predictive factors: cancer site, platelet count ≥350×10⁹/L, hemoglobin <100 g/L, and/or use of erythropoiesis-stimulating agents, white blood cell count >11×10⁹/L, and body mass index ≥35 kg/m². The CATS score adds two biomarkers, soluble P-selectin and D-dimer, to the factors in the Khorana score. However, the development populations for both models did not include leukemia patients. Furthermore, the universal features of leukemia, such as anemia and leukocytosis, may limit the risk assessment performance of these models (37). Consequently, developing thrombosis risk prediction models specific to leukemia is crucial.

In recent years, although risk prediction models for thrombosis in leukemia patients have been successively proposed, providing new tools for clinical risk stratification, a comprehensive assessment of their quality and applicability has been lacking. This study aims to systematically review the development methods, predictive indicators, and validation results of thrombosis risk prediction models in leukemia patients, grade the evidence, and propose directions for optimization. The findings will inform clinical decision-making and guide the design of subsequent research initiatives.

The study protocol was registered on the International Prospective Register of Systematic Reviews (PROSPERO; Registration number: CRD420251146253).

This systematic review evaluated 14 studies comprising 16 risk prediction models for thrombosis in leukemia patients. By systematically examining model characteristics, performance reporting, risk of bias, and clinical relevance, we found that although several models reported moderate to good apparent discrimination and identified clinically plausible predictors, all included studies were judged to be at high risk of bias, and the overall evidence base is insufficient to support routine clinical implementation at this stage. The main methodological limitations were widespread, including small effective sample sizes with low EPV, suboptimal handling of continuous predictors, incomplete handling and reporting of missing data, and limited validation. Overall, these issues increase the possibility of overfitting and optimism, undermine calibration and transportability, and limit the real-world applicability of existing tools. Therefore, current models should be regarded as exploratory and should undergo rigorous prospective external validation before being used to guide thromboprophylaxis or other clinical decision-making. In the following sections, we discuss in detail the main sources of bias and limited applicability observed across studies.

The development of thrombosis risk prediction models in leukemia patients falls under prognostic modeling research. Prospective cohort studies are widely regarded as the optimal design for minimizing retrospective bias. However, among the 14 studies reviewed, only two adopted a prospective design, with the remaining 12 relying on retrospective cohort or case-control methodologies. A key distinction lies in the blinding process: the two prospective studies inherently ensured that predictors were assessed prior to outcome occurrence, whereas none of the retrospective studies explicitly described whether blinding was implemented—either in predictor assessment or in outcome evaluation. Proper blinding is essential to prevent information bias. If assessors are aware of thrombosis outcomes, they may overinterpret borderline predictor values, leading to inflated associations. Conversely, if clinicians know predictor values such as D-dimer levels, they might be more inclined to diagnose equivocal symptoms or imaging findings as thrombosis, increasing the likelihood of false-positive outcomes. Therefore, future efforts in model development should emphasize standardized implementation of blinding procedures.

Simultaneously, methodological deficiencies are also evident in sample size and variable handling. According to PROBAST guidance, prognostic model studies should ensure an “Events Per Variable (EPV) ≥ 20” to reduce overfitting (39). However, among the models included in this review, only one met this standard; the EPV in the remaining studies mostly ranged between 0.5 and 11.33, potentially leading to reduced generalizability of the models in external populations. Furthermore, 9 out of the 14 studies categorized continuous variables, which can lead to loss of information, reduced statistical power, and even mask non-linear relationships between predictors and the thrombotic outcome. To reduce categorization bias, we recommend that future studies use continuous variables in their original form during statistical analysis or consider non-linear regression techniques to fit the models.

Regarding missing data handling and model validation, among the 14 studies, 9 used complete-case analysis (direct deletion of cases with missing data), and 5 did not report the handling method. We recommend employing multiple imputation methods in future research to handle missing data, thereby enhancing the accuracy of reported results. Concerning model validation, only 9 studies performed internal validation, and only 1 conducted external validation, making it impossible to effectively assess the model’s overfitting risk and generalizability. Future studies should perform both internal and external validation, and consider temporal validation where necessary to evaluate the long-term stability of the models.

In the PROBAST assessment of concerns regarding applicability, 5 out of the 14 leukemia thrombosis risk prediction models raised high concerns. The core issues focused on insufficient participant representativeness and flaws in the temporal sequence of predictors. Regarding participants, 1 model was limited to patients on specific chemotherapy regimens, and 4 were confined to populations with PICC insertions. These limitations prevent the models from covering the heterogeneity of leukemia patients with different chemotherapy regimens and venous access choices in clinical practice, directly restricting their generalizability to the broader population. Regarding predictors, 3 studies potentially collected predictors after the thrombotic event had occurred, violating the core logic that predictors must precede the outcome, thereby rendering the models clinically non-predictive. Future research needs to include heterogeneous leukemia populations and strictly ensure the “predictor-outcome” temporal sequence to enhance the clinical promotion value of the models.

Another key consideration is generalizability. Nearly half of the included studies were conducted in China, and several reports were published in Chinese, including two dissertations. Differences in patient genetics, contemporary treatment protocols, supportive care standards, and thrombosis screening practices may limit the transportability of these models to other settings. To address concerns regarding evidence-source quality and indexing, we conducted restricted analyses excluding articles not published in English and excluding non–peer-reviewed dissertations. These analyses yielded consistent overarching conclusions but further highlighted the paucity of robust external validation across diverse populations.

From the analysis of predictors, the high-frequency factors identified across the 14 studies can be categorized into three core dimensions: patient characteristics (e.g., prior thrombosis history, comorbidities, age), treatment intervention factors (central venous catheter placement, chemotherapy/radiotherapy, use of hemostatic drugs), and disease-related indicators (e.g., D-dimer levels, platelet count, white blood cell count, ECOG score, leukemia type).

A prior history of thrombosis is a significant warning sign for recurrent thrombotic diseases. Leukemia patients with a history of thrombosis have approximately a 6-fold higher risk of subsequent venous or arterial thrombosis compared to those without such a history (33, 53). Comorbidities such as diabetes and hypertension indirectly increase thrombosis risk by damaging vascular endothelial function, promoting a hypercoagulable state, and activating platelets (54–56). Studies show that leukemia patients with ≥2 underlying chronic diseases have a significantly higher incidence of thrombotic events, by 1 to 3 times, compared to those with a single disease (57). Advancing age is closely related to alterations in coagulation mechanisms. Individuals over 65 years old exhibit a significantly increased thrombosis risk due to decreased vascular elasticity, increased activity of certain coagulation factors, and a chronic inflammatory state (58–60).

Central venous catheter placement is widely used in leukemia patients to administer chemotherapy drugs, provide parenteral nutrition, transfuse blood products, and facilitate blood sampling, while minimizing the discomfort of frequent venipuncture. Central venous catheters (CVCs) can be classified into four types: non-tunneled catheters, tunneled central catheters, fully implantable catheters, and peripherally inserted central catheters (PICCs) (61). CVCs can promote thrombosis due to mechanical injury to the vascular endothelium and adsorption of coagulation factors on the catheter surface, with PICC-related thrombosis incidence potentially being higher than other catheter types (62–64). Chemotherapy exacerbates abnormalities in the coagulation system of leukemia patients, increasing thrombotic risk. For instance, L-asparaginase, a common drug for ALL treatment, causes deficiencies in antithrombin, fibrinogen, protein C, and protein S, disrupting the physiological balance between anticoagulation and hemostatic pathways (65–68). Radiotherapy is important in specific scenarios for leukemia treatment, such as central nervous system leukemia, pre-transplant conditioning, or for tumor masses formed by leukemic cells. In a sub-analysis of the COMPASS-CAT study, radiotherapy was associated with an increased risk of venous thromboembolism (HR 2.47, 95% CI: 1.47-4.12), more pronounced in women than men (10.8% vs. 2.7%, p=0.03) (69). Some leukemia patients experience bleeding during chemotherapy. While hemostatic drugs can effectively control bleeding, their overuse can lead to a hypercoagulable state, increasing thrombosis risk (43).

D-dimer is a degradation product of fibrin broken down by plasmin. Abnormally elevated levels in the blood are a key marker reflecting secondary hyperfibrinolysis and a hypercoagulable state. Studies indicate that ALL patients with high D-dimer levels (≥4 μg/mL) have a 100-day cumulative incidence of venous or arterial thrombosis of 52.9%, compared to 13.8% in patients with low-to-moderate D-dimer levels (<4 μg/mL) (70). A platelet count exceeding 350×10⁹/L during hospitalization is associated with a 2.5-fold increased risk of VTE (71). In the outpatient setting, patients with a pre-chemotherapy platelet count ≥350×10⁹/L had a higher VTE incidence compared to those with a count <350×10⁹/L (3.98% vs. 1.37%) (72). The tumor microenvironment can induce thrombocytosis and platelet activation. Platelet activation leads to the release of prothrombotic molecules from alpha-granules and dense granules into the blood microenvironment and exposes activated sites for thrombin, correlating with increased thrombosis risk (73). When the peripheral blood WBC count in leukemia patients exceeds 50×10⁹/L, the risk of thrombosis increases, potentially related to blood stasis, stimulation of vascular endothelium by toxic metabolic oxidative products from leukemic cells, expression of procoagulant substances by leukemic cells, and release of inflammatory factors (48). The ECOG score is a tool to assess the functional status and ability to perform daily activities in cancer patients; typically, a score ≤2 is required for considering active treatment like chemotherapy. An ECOG score ≥2 indicates the patient is incapable of work; if combined with prolonged bed rest due to reduced mobility, slow blood flow, and poor functional status, it increases thrombosis risk (46). The risk of thrombosis in leukemia patients is closely related to the specific subtype, with APL having the highest risk, where abnormal promyelocytes simultaneously activate both the coagulation and fibrinolytic systems, predisposing to life-threatening disseminated intravascular coagulation (DIC) (74).

These predictive factors provide a clear basis for candidate variable selection in future model development. In clinical practice, dynamic monitoring based on these quantifiable indicators could potentially help healthcare staff identify high-risk individuals early and reduce thrombosis risk through preemptive interventions.

Notably, within our included evidence base, chronic leukemia was represented only by chronic myeloid leukemia (CML) in a single mixed-subtype cohort. Thrombosis in leukemia patients follows the classic Virchow triad—hemodynamic abnormalities, endothelial injury, and a hypercoagulable state, but the dominant contributing factors and management differ between acute and chronic forms. In acute leukemia, thrombosis is initiated primarily by the rapid release of tissue factor, microvesicles, and other procoagulant substances from leukemic cells, leading to the coagulation cascade and are often accompanied by disseminated intravascular coagulation (DIC) (75, 76). This creates a paradoxical risk of both bleeding and thrombosis, and management focuses on treating the underlying leukemia and controlling DIC with careful consideration of anticoagulation given high bleeding risk. In contrast, chronic leukemia develops over a prolonged course, where alterations in blood flow, persistent inflammation, and excessive or dysfunctional blood cells collectively induce a hypercoagulable state, predominantly resulting in macrovascular thrombosis (77, 78). Treatment therefore emphasizes thromboprophylaxis or therapeutic anticoagulation in conjunction with disease-directed therapy.

This review has limitations. Although we implemented a comprehensive search across English and Chinese databases, relevant studies in other languages or unpublished literature may have been missed, which could contribute to reporting and publication bias. In addition, substantial clinical and methodological heterogeneity (e.g., leukemia subtype, outcome definition, predictors, and validation) precluded meaningful quantitative pooling. We therefore focused on a structured descriptive and comparative synthesis complemented by sensitivity analyses as robustness checks.

Future research should prioritize prospective, multicenter cohorts with clearly defined time horizons and standardized outcome adjudication, ensuring that predictors are measured before outcome occurrence and that thrombosis screening practices are transparently reported. Rigorous model development should be supported by adequate sample size planning (including EPV considerations), principled handling of missing data (e.g., multiple imputation), and appropriate modeling of continuous predictors (avoiding unnecessary dichotomization and considering non-linear effects when justified).

External validation across diverse populations and healthcare systems should be considered a minimum requirement before any clinical implementation. When validation reveals calibration drift or reduced discrimination, model updating and recalibration strategies should be undertaken, and comparative studies that evaluate multiple candidate models within the same contemporary cohort are needed to identify the best-performing approach for specific clinical contexts.

Given subtype-specific differences in thrombosis mechanisms and treatment pathways, future models should either be developed and validated within specific leukemia subtypes or incorporate subtype in a way that is clinically interpretable and supported by sufficient data. In addition, models targeting catheter-related thrombosis should explicitly account for catheter type, insertion technique, catheter-related complications, and prophylaxis practices.

Current thrombosis risk prediction models for leukemia patients appear promising but remain insufficiently validated for routine clinical use. They should be considered exploratory and not used as stand-alone tools to guide thromboprophylaxis. Prospective multicenter external validation, transparent reporting, and model updating are essential to develop robust and transportable prediction models.