Machine Learning-Based Prediction of Heparin Levels in Pediatric Patients

Introduction: Heparin infusion requires regular laboratory monitoring, a challenge in pediatric care due to risk of iatrogenic anemia. We hypothesized that machine learning models could predict anticoagulation parameters for heparinized children using a patient’s own prior clinical data, with the ultimate hypothesis that this could reduce lab frequency without increasing complications.


Methods: All heparin infusions ≥6 hours for inpatients within an academic health system, regardless of patient age, between 2011-2024 were used for modeling of a variational recurrent neural network (vRNN). Input variables included heparin dosage (and prior adjustments, medication holds, etc), selected medications, demographics, and prior coagulation parameters. We used one-hot encoding for categorical features and converted numerical features to percentiles. Three classes were defined for the patient’s coagulation parameters given clinical standards: low (anti-XA level < 0.3 IU/mL or PTT level < 60sec), normal [anti-XA 0.3-0.7 or PTT 60-99s), and high (anti-XA >0.7 or PTT level > 99s); a separate classifier was used for each. We used a “nowcast” approach, predicting the class label at collection time. Patients were partitioned into training and validation cohorts stratified by age and complication rates.


Results: Heparin infusions occurred 28,789 times for 15,079 patients (1,125 infusions in children 1-18 years and 1,006 in infants < 1 year). Median age was 62 years (IQR 50-74) and median heparin duration was 36.5 hours (IQR 19-67), with longer durations in those < 18 years than adults (p < 0.01). Anti-Xa was the most common lab test (93% of all heparin infusions) followed by PTT (26%). Coagulation parameters were measured a median 2.2 times per day (IQR 1.4-2.9) across the heparin infusion period. Bleeding complications were documented in 81 children (11.2%) during heparin infusions while 153 (21.1%) had clotting complications. The vRNN achieved AUROC of 0.77, 0.75, 0.79 for predicting low, normal, and high coagulation parameters, respectively, in the pediatric validation cohort.


Conclusions: This vRNN predicts coagulation parameters for heparinized children with acceptable AUROC. These data suggest that computer modeling could potentially be used to reduce lab monitoring frequency for heparinized pediatric patients.