A Novel, Interpretable Machine Learning Model Predicts Furosemide Dosing After Congenital Cardiac Surgery.

Fluid overload is common after neonatal congenital cardiac surgery (CCS) and is frequently managed with continuous furosemide infusions requiring iterative dose titration. An interpretable prediction model could support more consistent early postoperative dosing decisions. We hypothesized that a novel, interpretable machine learning approach could accurately predict furosemide dosing decisions in neonates following CCS. We identified term neonates admitted to the Pediatric Cardiothoracic ICU at a large academic children's hospital between 8/1/2014 and 3/1/2023 following CCS with cardiopulmonary bypass. Demographic and clinical data from the first 48 postoperative hours were used to train, validate, and test a Tropical Geometry-Based Fuzzy Neural Network Regressor (TGFNN-R) tasked with predicting furosemide infusion dose changes after CCS. The TGFNN-R was primed with clinician heuristics and provides transparent explanations behind predictions. A held-out internal validation/testing cohort was drawn from the same single-center population. Data from 506 neonates were extracted; 398 received a continuous furosemide infusion. Mean age at surgery was 6.2 (± 5.1) days; 67.3% were White. The most common surgeries were Stage I (Norwood) (25.1%) and arterial switch operation (18.6%). There were 783 furosemide dose increases and 224 dose decreases. Test set performance was R²=0.515, mean absolute error = 0.119 mg/kg/hr, and false positive rate = 0.062. In this retrospective single-center cohort of neonates following CCS, an interpretable TGFNN-R model predicted and explained furosemide dose changes with good test performance. Next steps include external validation and nonclinical studies evaluating the model within clinical decision support and closed-loop paradigms to achieve prespecified fluid balance goals.