Abstract
OBJECTIVES: Predicting optimal coagulation control using heparin in intensive care units (ICUs) remains a significant challenge. This study aimed to develop a machine learning (ML) model to predict activated partial thromboplastin time (aPTT) in ICU patients receiving unfractionated heparin for anticoagulation and to identify key predictive factors. METHODS: Data were obtained from the Tokushukai Medical Database, covering six hospitals with ICUs in Japan, collected between 2018 and 2022. The study included 945 ICU patients who received unfractionated heparin. The dataset comprised both static and dynamic features, which were used to construct and train ML models. Models were developed to predict aPTT following initial and multiple heparin doses. Model performance was evaluated using the area under the receiver operating characteristic curve (ROC AUC), area under the precision-recall curve (PR AUC), precision, recall, F1 score, and accuracy. SHAP analysis was conducted to determine key predictive factors. RESULTS: The random forest model demonstrated the highest predictive performance, with ROC AUC values of 0.707 for the first infusion and 0.732 for multiple infusions. Corresponding PR AUC values were 0.539 and 0.551. Despite moderate overall predictive performance, the model exhibited high precision (0.585 for the first infusion and 0.589 for multiple infusions), indicating effectiveness in correctly identifying true positive cases. However, recall and F1 scores were lower, suggesting that some cases, particularly in sub-therapeutic and supra-therapeutic ranges, may have been missed. Incorporating time-series data, such as vital signs, provided only marginal improvements in performance. CONCLUSIONS: ML models demonstrated moderate performance in predicting aPTT following heparin infusion in ICU patients, with the random forest model achieving the highest classification accuracy. Although the models effectively identified true positive cases, their overall predictive performance remained limited, necessitating further refinement. The inclusion of static and dynamic features did not significantly enhance model accuracy. Future studies should explore additional factors to improve predictive models for optimizing individualized anticoagulation management in ICUs.