Personalized prediction of aPTT responses to unfractionated heparin for ICU patients using a novel machine learning pipeline in a multi-cohort international study

Abstract Background Managing unfractionated heparin (UFH) infusions in ICU settings is challenging due to significant inter-patient variability and UFH's polydisperse molecular weight. UFH exhibits dose-dependent pharmacokinetics: at low doses, elimination occurs via saturable, receptor-mediated uptake, while higher doses shift elimination toward linear renal excretion. Its anticoagulant effects are also dose-dependent, with lower doses inhibiting Factor Xa and higher doses inhibiting both Factor Xa and thrombin (Factor IIa). These variable transitions in pharmacokinetics and pharmacodynamics complicate titration and highlight the need for predictive tools to optimize UFH therapy. Purpose This study developed and externally validated a machine learning model to predict the next activated partial thromboplastin time (aPTT) during UFH therapy in ICU patients. Methods Using 64 variables from an ICU registry, we trained an XGBoost regressor and selected the most relevant features for the model development. Based on the pharmacokinetics and pharmacodynamics of unfractionated heparin (UFH), we hypothesized that dose-specific models would outperform a general model. Accordingly, we identified an optimal global dose threshold using Optuna and trained separate models for low and high UFH doses. Given the variability in optimal thresholds across patients, we aimed to personalize this approach. Each data point was labeled with the better-performing regressor, and an XGBoost classifier was trained to determine the personalized threshold for each patient. (Figure 1) To ensure generalizability, we used data from six international databases, excluded entries with missing values, and limited aPTT measurements to 4–24 hours post-adjustment. The pipeline was trained on data from four databases and externally validated on two to ensure robust generalization. Results The training set included 68,398 observations, and the external validation set comprised 69,169 observations. Feature selection identified UFH flow rates before and after UFH adjustment, previous aPTT value, and timing-related variables (4 in total) as the most predictive inputs. The classifier divided the data into Subgroup 1 (n=6,314) and Subgroup 2 (n=62,855). The models demonstrated the following performance: Subgroup 1 achieved a root mean square error (RMSE) of 13.2 seconds, a mean absolute error (MAE) of 9.4 seconds, and R² of 0.46, while Subgroup 2 showed an RMSE of 11.5 seconds, MAE of 8.2 seconds, and R² of 0.50. (Figure 2) Conclusion We developed and externally validated a machine learning pipeline for personalized aPTT prediction, demonstrating strong performance across international datasets. Future work will focus on further personalization, model-driven optimization, and prospective clinical validation.Figure 1Figure 2