Pretest Risk Estimation of Age-Adjusted D-Dimer Positivity Using Machine Learning

Pulmonary embolism evaluation in the ED often bypasses D-dimer because clinicians anticipate a positive result, leading to potentially avoidable CT pulmonary angiography. Quantifying risk of a positive result might improve patient selection for D-dimer testing. We trained a machine learning algorithm to predict age-adjusted D-dimer positivity.

Adult ED encounters from 8/2022 to 8/2024 within our integrated healthcare system with a resulted D-Dimer were included. Patient demographics, vital signs, comorbidities and lab values resulted prior to the D-dimer order were extracted from our EHRs structured data warehouse. The outcome was age-adjusted D-dimer positivity (cutoff 500 ng/mL if age 90% correlation, near zero-variance, and >90% missingness were removed. Missing values were median imputed, and 109 final features were selected via penalized LASSO regression. Classification models were trained using nested cross-validation, including ridge logistic regression, elastic net, random forest, gradient boosting, XGBoost, and neural network models. The best model was evaluated on the test set using the AUROC, AUPRC, and Brier calibration score. SHAP analysis identified feature importance.

The dataset included 45,453 encounters; 36,362 (80%) of which were used for training and 9,091 (20%) for testing. The median age was 57 years (IQR 39, 71) and 60.8% were female. In the test set, 42.5% of encounters had a positive D-dimer. The best-performing model was an XGBoost classifier using 400 trees, maximum tree depth of 5, and learning rate of 0.03. The model achieved an AUROC of 0.77 (95% CI, 0.76 – 0.77), AUPRC of 0.72 (95% CI, 0.71-0.74) and Brier score of 0.19 on the held-out test set. At a sensitivity of 0.69, the model had a PPV of 0.64, and NPV of 0.76. Most recent hematocrit, age, highest heart rate, and the most recent D-dimer value had the highest feature importance.

Machine learning can reasonably predict age-adjusted D-dimer positivity and may identify patients where a D-dimer would be reasonable to order before CT angiography. External prospective validation and evaluation of clinical utility are needed before clinical use.