Development and validation of a deep learning model to predict heparin response in critically ill patients with deep vein thrombosis receiving continuous intravenous heparin infusion

Continuous intravenous heparin infusion is widely used for deep vein thrombosis (DVT) in the intensive care unit (ICU), but accurate prediction of activated partial thromboplastin time (APTT) remains challenging due to patient heterogeneity and complex drug responses. A deep learning model was developed using heparin administration data, laboratory tests, and patient history to predict future APTT values in ICU patients with DVT. Data from 796 patients receiving continuous heparin infusion were collected, including demographics, comorbidities, treatment, and laboratory information, with external validation performed on 514 patients from the Medical Information Mart for Intensive Care (MIMIC) database. A 2-layer Long Short-Term Memory model with dropout was trained and evaluated. Optimal performance was achieved with the [ t − 12, t − 2] window, yielding a mean absolute error of 1.885, root mean square error of 4.732, R ² of 0.945, and mean absolute percentage error of 4.433. Subgroup analyses demonstrated robust accuracy across critical APTT zones and therapeutic categories (subtherapeutic, therapeutic, supratherapeutic), with area under the curve values of 0.92, 0.81, and 0.92. The model predicted abnormal APTT values a mean of 6 hours in advance and could improve guidance over physician-led management in approximately 70% of cases. External validation confirmed good generalizability. Feature importance identified the difference between 45 s and APTT, cumulative heparin dose, and infusion rate as the leading predictors. Collectively, this deep learning model accurately forecasts future APTT values in critically ill patients with DVT receiving intravenous heparin, supporting a shift from reactive to proactive, data-driven heparin management in the ICU with potential benefits for patient safety and treatment efficacy.