Machine learning based prediction of platelet concentration from ROTEM measurements

Viscoelastic testing is an established standard in modern coagulation medicine, reducing transfusion needs. However, it does not provide direct information on platelet count. Here, we present a machine learning approach to predict platelet count and to detect low platelet count based on standard rotational thromboelastometry (ROTEM) parameters. We analyzed 2,333 anonymised datasets collected from multiple centres between 2014 and 2023. Six machine learning methods (Regression, Gradient Boosting Methods, extreme Gradient Boosting, Random Forests, Neural Networks, and Stacked Ensemble Learning) were trained to predict platelet count and to classify thrombocytopenia below thresholds of <100 x109/L and <50 x109/L. Mean absolute error and root mean square error judged model quality for regression models; area under the curve, balanced accuracy, and F1-score for classification models. The best model for predicting platelet count was stacked ensemble learning (RMSE 57.8 ±9.1; R2 0.68 ±0.1). The best classification models were random forest classification for a platelet count below 100 x109/L (AUC 0.90 ±0.03) and stacked ensemble learning for a platelet count below 50 x109/L (AUC 0.95 ±0.03). Machine learning methods showed excellent discriminative performance for detecting low platelet count whereas only moderate predictive accuracy for predicting platelet level. This novel approach may contribute to rule-out thrombocytopenia and ensure safe and timely platelet transfusion.