Practical machine learning model for early and accurate prediction of disseminated intravascular coagulation before its progression to an overt stage

Background and aims: In patients with sepsis, anticoagulant therapy is expected to have maximal efficacy when administered before the development of sepsis-induced overt disseminated intravascular coagulation (DIC). This therapeutic strategy requires a valid method for real-time, early prediction of sepsis-induced DIC, which is likely to progress to an overt stage. We aimed to develop machine learning prediction models for overt DIC, based on the International Society on Thrombosis and Haemostasis criteria, using a large-scale electronic medical record database. Methods: This multi-center retrospective, observational study included adult patients with sepsis without overt DIC at day 1. The outcome was development of overt DIC 24 h after a certain time point. Predictive features were baseline data and time series data within 7 days post-admission. We constructed three separate models (minimum, compact, and full models), each incorporating increasingly comprehensive sets of features. Prediction models were constructed with machine learning algorithms, including eXtreme Gradient Boosting (XGBoost) and Gradient Boosting Machine (GBM), and evaluated on a randomly selected 20% test set. Results: Among 7,532 patients with sepsis, 766 developed overt DIC within 7 days. XGBoost and GBM achieved the highest prediction accuracy. The full, compact, and minimum models on the test set, exhibited area under the receiver operating characteristic curve values of 0.916, 0.884, and 0.851, respectively. The full XGBoost model achieved an area under the precision–recall curve of 0.295; at 80% recall, its precision was 14.4%. Conclusions: The machine learning model exhibited high accuracy in predicting overt DIC at a clinically reliable level.