Warfarin adjustment in pediatric cardiology is particularly challenging and requires frequent, individualized modifications based on international normalized ratio (INR) monitoring. While some adolescent patients are autonomous, most pediatric patients depend on caregivers and frequently contact their cardiologist to report INR values and contextual factors. Although rule-based dosing algorithms are available, clinicians often rely on experience, as these tools insufficiently account for inter- and intra-patient variability, including dietary changes, inflammatory status, and behavioral factors. We hypothesized that a machine learning (ML) model could learn from individual longitudinal INR monitoring data to personalize and fine-tune rule-based dose adjustment strategies at the patient level, thereby approximating clinician-driven decision-making in pediatric anticoagulation management.

Monitoring data, including patient age, INR target, INR values, warfarin doses, and clinician dose adjustment decisions, from patients under 18 years of age receiving oral anticoagulation and followed at La Timone Hospital (Marseille) were analyzed. Clinician dose adjustment decisions served as the reference standard. The existing rule-based algorithm was evaluated. Four instead ML models were developed, trained, and evaluated to predict dose adjustments using longitudinal patient data. We calculated the Mean Squared Error (MSE) and the coefficient of determination ( R ² ) for each prediction of the model.

29 patients were included (mean age was 7.5 years ± 4.3), with a mean follow-up of 36 months and a total number of INR and clinician response of 2470 lines. The performance of the rule-based model to predict the clinician decision was MSE: 0.12, R ² : 0.90 whereas the Light Gradient Boosting (LGB) ML models performed better with a MSE: 0.09, and a R ² : 0.92.

LGB model achieved the best performance in predicting the clinician advice about INR adaptation. An interactive clinical decision-support application is planned to assist clinicians and, following evaluation and validation, to support patient self-monitoring and self-management.