Using machine learning to predict venous thromboembolism in patients with uterine cancer (081)

Objectives: Gynecologic cancers are traditionally associated with a high risk of venous thromboembolism (VTE). Recent ASCO clinical practice guidelines have expanded recommendations for pharmacological thromboprophylaxis in high-risk patients; however, high risk is not well defined for the endometrial cancer population. Both the Caprini and the Khorana scores are clinically validated to assess the risk of VTE - Caprini developed for surgical patients and Khorana for cancer patients undergoing chemotherapy. We sought to develop a better tool to assess endometrial cancer patients at high risk for VTE. Methods: A single-institution retrospective cohort study was performed from January 2016 through January 2020. All patients with uterine cancer were screened, and the patients with complete clinical information were included. VTE was evaluated from diagnosis through 90 days post-diagnosis. Both the Caprini and the Khorana scores were calculated for each patient. High-risk Caprini score was defined as ≥ 9 based on prior studies in cancer patients, whereas high-risk Khorana score was defined as ≥ 3. In addition to variables included in the risk assessment tools, additional variables included race/ethnicity, tumor grade, stage, and histology. Microsoft Azure Machine Learning Studio was utilized to develop a machine learning (ML) algorithm to predict VTE. Five different ML models were assessed, including random forest, extreme random trees (ERT), logistic regression, XGBoost classifier, and support vector machine. Data were split into 70% for ML training and 30% for validation. Models were analyzed based on classification accuracy, AUC, and sensitivity and specificity. Results: A total of 1000 patients were included. The majority of patients were White (93.5%), obese (mean BMI of 37.8 kg/m2) with an average age of 61.2 years. Most tumors were grade 1 (62.7%), endometrioid histology (79.3%), and stage I (76.4%). Most patients underwent a hysterectomy (95.4%). A minority received neoadjuvant (4.0%) and adjuvant chemotherapy (13.1%), whereas 269 patients received radiation (26.9%). Only 18 patients (1.8%) were diagnosed with a VTE within 90 days of diagnosis. The Khorana score had a sensitivity of 38.9% and a specificity of 80.1% for VTE. The Caprini score had a sensitivity of 66.7% and a specificity of 61.0%. Combining the features of the Khorana and Caprini scores, the best-performing ML algorithm had an accuracy of 67.3%, AUC of 83.8%, and the sensitivity improved to 100%, retaining specificity at 66.8% (ERT). To further improve the model, histology and grade were included, which performed better as a random forest model with an accuracy of 75.7%, AUC of 84.8%, sensitivity 100%, and specificity of 75.3%. Lastly, we added stage and race/ethnicity to the model (random forest model), which improved the accuracy to 83.7%, AUC 89.4%, sensitivity 100%, and specificity to 83.4%. In the final model, tumor grade and stage were the most weighted for importance, followed by age, platelet count, histology, and hemoglobin. Conclusions: The use of ML to enhance VTE prediction in endometrial cancer is promising. Using the same parameters as clinically validated tools, we were able to greatly improve prediction, and by adding clinically relevant variables (histology, grade, stage, race/ethnicity), we were able to further improve the model with 100% sensitivity and 83.4% specificity for VTE within 90 days of diagnosis. Using ML, VTE thromboprophylaxis could be more individualized for patients at the highest risk for VTE and minimize the potential risks associated with unnecessary anticoagulation. Objectives: Gynecologic cancers are traditionally associated with a high risk of venous thromboembolism (VTE). Recent ASCO clinical practice guidelines have expanded recommendations for pharmacological thromboprophylaxis in high-risk patients; however, high risk is not well defined for the endometrial cancer population. Both the Caprini and the Khorana scores are clinically validated to assess the risk of VTE - Caprini developed for surgical patients and Khorana for cancer patients undergoing chemotherapy. We sought to develop a better tool to assess endometrial cancer patients at high risk for VTE. Methods: A single-institution retrospective cohort study was performed from January 2016 through January 2020. All patients with uterine cancer were screened, and the patients with complete clinical information were included. VTE was evaluated from diagnosis through 90 days post-diagnosis. Both the Caprini and the Khorana scores were calculated for each patient. High-risk Caprini score was defined as ≥ 9 based on prior studies in cancer patients, whereas high-risk Khorana score was defined as ≥ 3. In addition to variables included in the risk assessment tools, additional variables included race/ethnicity, tumor grade, stage, and histology. Microsoft Azure Machine Learning Studio was utilized to develop a machine learning (ML) algorithm to predict VTE. Five different ML models were assessed, including random forest, extreme random trees (ERT), logistic regression, XGBoost classifier, and support vector machine. Data were split into 70% for ML training and 30% for validation. Models were analyzed based on classification accuracy, AUC, and sensitivity and specificity. Results: A total of 1000 patients were included. The majority of patients were White (93.5%), obese (mean BMI of 37.8 kg/m2) with an average age of 61.2 years. Most tumors were grade 1 (62.7%), endometrioid histology (79.3%), and stage I (76.4%). Most patients underwent a hysterectomy (95.4%). A minority received neoadjuvant (4.0%) and adjuvant chemotherapy (13.1%), whereas 269 patients received radiation (26.9%). Only 18 patients (1.8%) were diagnosed with a VTE within 90 days of diagnosis. The Khorana score had a sensitivity of 38.9% and a specificity of 80.1% for VTE. The Caprini score had a sensitivity of 66.7% and a specificity of 61.0%. Combining the features of the Khorana and Caprini scores, the best-performing ML algorithm had an accuracy of 67.3%, AUC of 83.8%, and the sensitivity improved to 100%, retaining specificity at 66.8% (ERT). To further improve the model, histology and grade were included, which performed better as a random forest model with an accuracy of 75.7%, AUC of 84.8%, sensitivity 100%, and specificity of 75.3%. Lastly, we added stage and race/ethnicity to the model (random forest model), which improved the accuracy to 83.7%, AUC 89.4%, sensitivity 100%, and specificity to 83.4%. In the final model, tumor grade and stage were the most weighted for importance, followed by age, platelet count, histology, and hemoglobin. Conclusions: The use of ML to enhance VTE prediction in endometrial cancer is promising. Using the same parameters as clinically validated tools, we were able to greatly improve prediction, and by adding clinically relevant variables (histology, grade, stage, race/ethnicity), we were able to further improve the model with 100% sensitivity and 83.4% specificity for VTE within 90 days of diagnosis. Using ML, VTE thromboprophylaxis could be more individualized for patients at the highest risk for VTE and minimize the potential risks associated with unnecessary anticoagulation.