Comparing Effectiveness of Machine Learning Methods for Diagnosis of Deep Vein Thrombosis

This paper presents the results of a comparative study of machine learning techniques when predicting deep vein thrombosis. We used the Ri-Schedule dataset with Electronic Health Records of suspected thrombotic patients for training and validation. A total of 1653 samples and 59 predictors were included in this study. We have compared 20 standard machine learning algorithms and identified the best-performing ones: Random Forest, XGBoost, Gradient-Boosting and HistGradientBoosting classifiers. After hyper-parameter optimization, the best overall accuracy of 0.91 was shown by Gradient-Boosting classifier using only 15 of the original variables. We have also tuned the algorithms for maximum sensitivity. The best specificity was offered by Random Forests. At maximum sensitivity of 1.0 and specificity of 0.41, the Random Forest model was able to identify 23% additional negative cases over the screening practice in use today. These results suggest that machine learning could offer practical value in real-life implementations if combined with traditional methods for ruling out deep vein thrombosis.