Venous thromboembolism (VTE), encompassing deep vein thrombosis (DVT) and pulmonary embolism (PE), is a leading cause of morbidity and mortality in hospitalized cancer patients. Among cancer types, lymphoma carries a particularly high VTE burden because patients require prolonged inpatient chemotherapy cycles and frequently use central venous catheters (CVCs), both of which independently elevate clot risk. A large multicenter cohort study covering 1995-2012 found that VTE occurred in as many as 90% of hospitalized lymphoma admissions in the most severe reporting context, with an associated mortality rate of approximately 16%.
Contemporary incidence: Improved VTE prevention guidelines from academic organizations have reduced overall rates substantially. Current estimates place the VTE incidence in lymphoma patients at roughly 7.9%, but diffuse large B-cell lymphoma (DLBCL), the most common and aggressive NHL subtype, carries rates as high as 12.8%. These figures matter clinically because lymphoma patients spend extended periods in the hospital across multiple chemotherapy cycles, compounding the cumulative risk compared with solid tumor patients.
The gap in existing tools: The Khorana Score (KS) is the most widely used clinical VTE prediction tool in oncology. However, it was developed and validated primarily for outpatient solid tumor populations. It does not account for the number of chemotherapy cycles a patient has received, nor does it incorporate CVC use. It also requires extranodal localization data that can take 3 to 7 days to confirm, limiting rapid bedside application. A 2016 lymphoma-specific model by Thorly and colleagues introduced the ThroLy score, but it similarly omits chemotherapy-cycle count and CVC use, and was not designed for hospitalized patients actively undergoing treatment.
This 2025 multicenter study from Chongqing, China, directly addresses these gaps by developing a machine learning-based early warning system (VTE-EWS) built on 1,141 hospitalized lymphoma patients across four academic centers, training six distinct ML algorithms on 12 routinely collected clinical variables, and visualizing the top-performing variable set via a web-based nomogram tool.
The study retrospectively enrolled 1,141 lymphoma patients hospitalized for chemotherapy at four academic medical centers in Chongqing, China between February 2020 and February 2024: Chongqing University Cancer Hospital (CQUCH), Yongchuan Hospital of Chongqing Medical University, Third Affiliated Hospital of Chongqing Medical University, and People's Hospital of Rongchang District. Patients were included if they were over 18 years of age, had a histopathological diagnosis of lymphoma, and had been hospitalized for chemotherapy at least once. Patients with unknown VTE status, unknown histological type, unknown Ann Arbor stage, or missing other required variables were excluded.
Dataset split: The training set comprised 799 patients from CQUCH (70% of the total), and the external validation set comprised 342 patients drawn from the three remaining centers (30%). Statistical comparison of the two cohorts found no significant differences in baseline characteristics (p greater than 0.05), supporting the assumption that the external set represents a genuinely independent population rather than a random split of a single institutional sample.
Feature selection via LASSO: A candidate pool of clinical variables was initially assembled from routine admission data. LASSO (Least Absolute Shrinkage and Selection Operator) regularization was applied to identify features with non-zero coefficients, effectively penalizing redundant or low-impact variables. Variables with high collinearity, excessive missing data (more than 10%), or limited clinical interpretability were deprioritized. This process yielded 12 final variables: age, sex, body mass index (BMI, scored per WHO criteria), CVC use, ECOG performance status, histological lymphoma subtype, Ann Arbor stage, white blood cell (WBC) count, hemoglobin (HB) level, D-dimer level, platelet count (PLT), and number of chemotherapy cycles. A correlation heatmap confirmed weak pairwise associations between retained variables (Pearson coefficients below 0.4), indicating that multicollinearity was not a concern.
Outcome definition and class imbalance handling: The primary outcome was VTE occurrence during the chemotherapy hospitalization, confirmed by Doppler ultrasound or venography for DVT and by CT pulmonary arteriography (CTPA) or nuclear lung ventilation/perfusion imaging for PE. VTE incidence was 10.89% in the training set (87/799) and 10.82% in the validation set (37/342), creating a class imbalance that could bias models toward predicting the majority non-VTE class. The study applied three resampling strategies (undersampling, oversampling with SMOTE, and mixed sampling using ROSE), determining that mixed sampling produced the best performance and was used for all final models.
Six machine learning algorithms were trained to predict VTE risk: (1) logistic regression, a linear baseline model; (2) random forest, an ensemble of decision trees using bootstrap aggregation; (3) backpropagation network (BP-network), a standard multi-layer neural network trained via gradient descent; (4) XGBoost (eXtreme Gradient Boosting), a gradient-boosted tree ensemble known for strong tabular data performance; (5) decision tree, a single interpretable tree classifier; and (6) support vector machines (SVM), which maximize the margin between classes in a high-dimensional feature space. Training all six algorithms on the same dataset provides a head-to-head comparison under identical conditions and helps identify which algorithm types are most suitable for this clinical prediction task.
Hyperparameter optimization: Each model underwent systematic hyperparameter tuning via 10-fold cross-validation on the training set. This approach partitions the training data into 10 equal subsets, iteratively training on 9 subsets and evaluating on the held-out 10th, cycling through all combinations. The hyperparameter configuration producing the best cross-validation performance was then used to build the final model for each algorithm. This procedure reduces the risk that any single model's reported performance reflects overfitting to a particular random split.
Evaluation metrics: Models were assessed on accuracy, precision, sensitivity (recall), specificity, F1 score, Brier score, and area under the receiver operating characteristic curve (AUC). The Brier score quantifies calibration by measuring the mean squared difference between predicted probabilities and actual binary outcomes, where lower scores indicate better-calibrated predictions. Calibration curves and decision curve analysis (DCA) were additionally applied to the top-performing model to assess clinical net benefit across a range of probability thresholds. DCA is particularly important because it translates predictive performance into clinical decision terms, showing the net benefit of acting on the model's predictions relative to treating all patients or none.
All statistical analysis and model training were conducted in RStudio version 2023.06.2-551 and R version 4.3.3. Missing data were imputed using multiple imputation via the mice package rather than simple mean or mode substitution, which preserves the uncertainty in imputed values and produces less biased parameter estimates.
Of the 1,141 patients analyzed, 124 developed VTE during chemotherapy hospitalization, all occurring after chemotherapy initiation rather than at admission. In external validation, all six models demonstrated strong and broadly comparable predictive performance. Accuracies ranged from 0.74 to 0.89 and AUC values from 0.80 to 0.83 across the six algorithms, indicating similar discriminatory ability to distinguish VTE from non-VTE patients. Sensitivities ranged from 0.62 to 0.73, specificities from 0.74 to 0.92, F1 scores from 0.36 to 0.50, and Brier scores from 0.14 to 0.40.
XGBoost as the top performer: XGBoost achieved the highest accuracy (0.89) and the lowest Brier score (0.14) in external validation, indicating both the fewest misclassifications and the best-calibrated probability estimates. In the confusion matrix, XGBoost predicted the most non-VTE patients correctly (up to 280 of 305 non-VTE cases), while identifying between 23 and 27 true VTE cases depending on the threshold. The calibration curves for XGBoost confirmed close agreement between predicted and observed VTE probabilities, and DCA showed substantial clinical net benefit across a wide range of decision thresholds.
Variable importance analysis: Two complementary approaches were used to assess which features drove model predictions. First, SHAP (SHapley Additive exPlanations) values were computed for the XGBoost model using the SHAPforxgboost package, decomposing each prediction into individual variable contributions. Second, permutation importance analysis was applied to all six models via the DALEX package, randomly shuffling each variable and measuring the consequent degradation in model performance. Results were integrated into a heatmap comparing variable importance rankings across all six models. This cross-model consensus approach identifies variables that are consistently important regardless of algorithm choice, providing a more robust signal than relying on any single model's feature rankings.
The heatmap analysis revealed that WBC count, D-dimer level, CVC use, age, number of chemotherapy cycles, and ECOG performance status were the top six variables across all models. SHAP analysis confirmed the direction of each effect: WBC count 11 x 10^9/L or higher, D-dimer above 0.5 mg/L, CVC use, advanced age, more chemotherapy cycles, and higher ECOG scores all increased predicted VTE probability.
A persistent challenge with machine learning models in clinical settings is that their internal logic is opaque and their outputs are difficult to apply at the bedside without specialized software. This study addressed that barrier by translating the ML model's key variable set into a nomogram, a visual scoring chart that allows clinicians to manually estimate a patient's VTE risk probability without executing any code. The six variables selected for the nomogram (WBC, D-dimer, CVC use, age, chemotherapy cycles, and ECOG score) were those that achieved the highest and most consistent importance rankings across all six ML models, ensuring that the simplified tool captures the most clinically meaningful information.
How the nomogram works: Each variable is assigned a score based on its value range, graphically represented as a point on a horizontal axis. Clinicians draw a vertical line from each patient's value to the "Points" axis, sum the individual scores, locate the total on the "Total Score" axis, and read off the corresponding VTE probability. This approach has a long history in clinical risk tools because it requires no computational device and generates a transparent, interpretable risk estimate that clinicians can scrutinize and question.
CQ Lymphoma VTE-EWS online tool: Going beyond the static nomogram, the authors converted it into a web-based interactive tool using the DynNom R package, accessible at https://tingtingjiang.shinyapps.io/CQ_lymphoma_VTE_EWS/. Clinicians enter six patient parameters, click "Predict," and receive an immediate risk probability estimate. Patients with a predicted probability exceeding 0.7 are classified as high-risk for VTE, triggering consideration for prophylactic anticoagulation or other targeted interventions. The tool is accessible from desktops or mobile devices, making it usable at the bedside or during pre-admission assessments.
The six variables incorporated into the VTE-EWS are all collected as part of standard pre-chemotherapy workups, meaning the tool adds no new testing burden. WBC and D-dimer are routine laboratory values; CVC use and chemotherapy cycle count are recorded in the medical record; age and ECOG score are assessed at each admission. This design minimizes implementation friction and makes the system immediately deployable in institutions with basic electronic medical records.
The study compared the VTE-EWS directly against the Khorana Score (KS), the established clinical standard for VTE risk prediction in oncology, using the same patient cohorts. The KS uses five variables (cancer site type, pre-chemotherapy platelet count, hemoglobin, leukocyte count, and BMI) and classifies patients with a score of 3 or higher as high-risk. This threshold was developed in outpatient solid tumor populations and, as the authors note, most VTE-positive lymphoma patients in their cohort scored 2 to 4 on the KS, meaning a significant share of true high-risk patients fall below the KS cutoff of 3 and are missed.
Detection rates in external validation: In the external validation set, the VTE-EWS detected 65% of VTE cases (24/37, sensitivity 0.65, 95% CI 0.49 to 0.78) in the high-risk group, compared to 54% detected by the KS (20/37, sensitivity 0.54, 95% CI 0.38 to 0.69). For the low-risk classification (non-VTE detection), the VTE-EWS correctly identified 90% of non-VTE patients (277/305, specificity 0.91, 95% CI 0.87 to 0.94), versus 77% for the KS (235/305, specificity 0.77, 95% CI 0.72 to 0.81).
AUC comparison: ROC curve analysis confirmed the VTE-EWS's substantially higher discriminatory ability. In the external validation set, the VTE-EWS achieved an AUC of 0.83 (95% CI 0.75 to 0.91) versus 0.69 (95% CI 0.61 to 0.78) for the KS. In the training set, the gap was even wider: VTE-EWS AUC 0.86 (95% CI 0.84 to 0.87) versus KS AUC 0.66 (95% CI 0.63 to 0.70).
Decision curve analysis: DCA showed that the VTE-EWS provided positive clinical net benefit across a far wider range of decision thresholds than the KS. In external validation, the VTE-EWS delivered net benefit from threshold probabilities of approximately 1% to 78%, while the KS provided benefit only in the 5% to 30% range. This means that for any reasonable clinical decision threshold for VTE prophylaxis, the VTE-EWS would generate fewer net harms from unnecessary interventions while capturing more true high-risk patients.
White blood cell count: Elevated WBC (leukocytosis) at or above 11 x 10^9/L was the most important predictor across all six models. Tumors release inflammatory cytokines that activate tissue factor on monocytes, promoting fibrin deposition and clot formation. Khorana et al. previously documented a two-fold increase in VTE risk among cancer patients with pre-chemotherapy leukocytosis and a 3% incidence in those with persistent leukocytosis after one cycle. In this cohort, 59.8% of VTE patients had WBC counts above 11 x 10^9/L, versus only 8.29% of non-VTE patients in the training set, a striking separation that explains this variable's dominant predictive weight.
D-dimer: D-dimer is a fibrin degradation product released when clot-stabilizing cross-links are broken down. Elevated D-dimer reflects active coagulation activity and strongly correlates with both DVT and PE risk, particularly in post-surgical and malignancy contexts. Among VTE patients in the training set, 80.5% had D-dimer above 0.5 mg/L, versus 48.7% of non-VTE patients. Persistent D-dimer elevation, rather than a single isolated high reading, was identified as the most clinically meaningful pattern for VTE prediction in this population.
CVC use: Central venous catheters cause direct vascular wall injury at insertion and during dwell time, trigger inflammatory responses, and impair normal blood flow. In the training set, 46% of VTE patients used CVCs, compared to only 7.16% of non-VTE patients, one of the largest absolute differences observed between groups. VTE rates among CVC users were consistently higher in both training and validation sets, confirming that CVC use is not merely a marker of disease severity but an independent mechanical contributor to clot formation.
Age, chemotherapy cycles, and ECOG score: Older age correlates with VTE across cancer types, likely reflecting accumulated cardiovascular risk factors and reduced physiological reserve. Extended chemotherapy cycles increase cumulative exposure to coagulation-activating agents and prolonged immobility, with VTE rates rising notably among patients completing 6 to 10 cycles (16.1% of VTE patients) compared to those completing 1 to 5 cycles (79.3% of VTE patients). Poor functional status as captured by ECOG score of 2 or higher predicts VTE by reflecting reduced mobility and impaired cardiovascular compensation. Together, these three variables capture the temporal and functional dimensions of VTE risk that the KS entirely omits.
Inpatient-only applicability: The VTE-EWS was developed and validated exclusively in hospitalized lymphoma patients receiving chemotherapy. The model's performance in outpatient treatment settings, where VTE risk profiles and clinical workflows differ substantially, has not been evaluated. Variables like CVC use and chemotherapy cycle count carry different distributional characteristics in outpatient contexts, and the 0.7 probability threshold for high-risk classification may not translate without recalibration for ambulatory patients.
Retrospective design without prospective validation: All study data were collected retrospectively from medical records, meaning the model has not been tested in a prospective implementation study where clinicians actually receive and act on the tool's predictions. Retrospective performance metrics, even when computed on an independent external validation set, do not capture how a tool performs when integrated into real clinical workflows, where alert fatigue, workflow friction, and clinician override behavior all influence the effective impact on patient outcomes.
Geographic concentration: All four study sites are located within Chongqing province in China, meaning the model has not been validated across diverse geographic, ethnic, or healthcare system contexts. Differences in chemotherapy protocols, CVC insertion practices, laboratory reference ranges, and VTE prophylaxis standards between institutions and countries could affect model performance. The authors acknowledge this and state that the next phase will include prospective studies across different Chinese provinces to improve generalizability before international validation is pursued.
Future directions: The authors plan to develop an intelligent prediction system integrated directly with electronic medical record (EMR) systems, enabling real-time automated risk calculation at each pre-chemotherapy admission without requiring manual data entry. EMR integration would also facilitate longitudinal risk tracking as a patient progresses through sequential chemotherapy cycles, allowing the system to capture the accumulating impact of multiple cycles on VTE probability. Prospective validation would additionally allow assessment of whether the VTE-EWS changes clinical decision-making and whether those changes translate into reduced VTE incidence or improved outcomes for identified high-risk patients.