Machine learning predicts cancer-associated deep vein thrombosis using clinically available variables

Abstract

PurposeTo develop and validate machine learning (ML) models for cancer-associated deep vein thrombosis (DVT) and to compare the performance of these models with the Khorana score (KS). MethodsWe randomly extracted data of 2100 patients with cancer between Jan. 1, 2017, and Oct. 31, 2019, and 1035 patients who underwent Doppler ultrasonography were enrolled. Univariate analysis and Lasso regression were applied to select important predictors. Model training and hyperparameter tuning were implemented on 70% of the data using a ten-fold cross-validation method. The remaining 30% of the data were used to compare the performance with seven indicators (area under the receiver operating characteristic curve [AUC], sensitivity, specificity, accuracy, balanced accuracy, Brier score, and calibration curve), among all five ML models (linear discriminant analysis [LDA], logistic regression [LR], classification tree [CT], random forest [RF], and support vector machine [SVM]), and the KS. ResultsThe incidence of cancer-associated DVT was 22.3%. The top five predictors were D-dimer level, age, Charlson Comorbidity Index (CCI), length of stay (LOS), and previous VTE (venous thromboembolism) history according to RF. Only LDA (AUC = 0.773) and LR (AUC = 0.772) outperformed KS (AUC = 0.642), and combination with D-dimer showed improved performance in all models. A nomogram and web calculator https://webcalculatorofcancerassociateddvt.shinyapps.io/dynnomapp/ were used to visualize the best recommended LR model. ConclusionThis study developed and validated cancer-associated DVT predictive models using five ML algorithms and visualized the best recommended model using a nomogram and web calculator. The nomogram and web calculator developed in this study may assist doctors and nurses in evaluating individualized cancer-associated DVT risk and making decisions. However, other prospective cohort studies should be conducted to externally validate the recommended model.