(366) Hourly Heparin Dosage Recommendation Strategy to Predict Anticoagulant for Cardiac Surgery Patients

Introduction: Anticoagulant therapy for patients who underwent cardiac surgery is a challenge. Both inadequate and excessive anticoagulation can cause fatal complications. Previous studies have failed to provide real-time guidance for heparin pump speed adjustment. This study intended to provide an hourly dosage recommendation model to predict heparin treatment for cardiac surgery patients based on a machine learning method.

Methods: Patients who underwent cardiac surgery and were admitted to an intensive care unit in Peking Union Medical College Hospital (PUMCH) from Jan 2013 to Dec 2021 were retrospectively analyzed. Target activated partial thromboplastin time (aPTT) was defined as 35-45 s. Linear regression, SVM, XGBoost, LSTM, GRU, FNN (feedforward neural network), and FNN + self-attention models were used to establish an hourly adjustment recommendation for heparin pump speed. Mean absolute, square, and absolute percentage errors were used to evaluate the reliability of the models. Shapley additive explanations (SHAP) was used to interpret the important features of the FNN + self-attention model.

Results: A total of 862 patients admitted to the PUMCH were included in this study, and 88267 heparin pump administrations were recorded. Among seven candidate models, the FNN + self-attention model yielded the lowest mean absolute error of 0.9133 and the lowest mean absolute percentage error of 0.1908. Linear regression, SVM, XGBoost, GRU, LSTM, and FNN models yielded mean absolute percentage errors of 0.476, 0.365, 0.276, 0.386, 0.393, and 0.259, respectively. Gap to target aPTT, history of coronary heart disease and atrial fibrillation, prothrombin activity, international normalized ratio, and concentration of factor IV (Ca2+) were identified as important features affecting heparin adjustment. Compared with actual clinical diagnosis and treatment, the model based on FNN+self attention can adjust the heparin pump speed earlier according to the aPTT value, with a median time of 2.8 hours. It can reduce the time duration for patients under over anticoagulation or insufficient anticoagulation state.

Conclusions: Our model recommending hourly heparin pump speed was potentially applicable for reducing physicians’ decision errors and providing personalized and reliable heparin adjustment strategies for patients.