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

Area of Science:

• Pediatric Cardiac Surgery
• Neonatal Intensive Care
• Machine Learning in Medicine

Background:

• Fluid overload is a common complication following neonatal congenital cardiac surgery (CCS).
• Current management involves continuous furosemide infusions, requiring frequent dose adjustments.
• A need exists for predictive models to support consistent early postoperative dosing.

Purpose of the Study:

• To develop and validate an interpretable machine learning model for predicting furosemide dosing decisions in neonates post-CCS.
• To hypothesize that a novel Tropical Geometry-Based Fuzzy Neural Network Regressor (TGFNN-R) can accurately predict these dosing changes.

Main Methods:

• Retrospective analysis of 506 neonates undergoing CCS with cardiopulmonary bypass (2014-2023).
• Utilized demographic and clinical data from the first 48 postoperative hours.
• Trained and tested a TGFNN-R model, incorporating clinician heuristics for interpretability.

Main Results:

• The model achieved R²=0.515 and a mean absolute error of 0.119 mg/kg/hr on the test set.
• The false positive rate for predicting dose changes was 0.062.
• The interpretable TGFNN-R model demonstrated good performance in predicting furosemide dose adjustments.

Conclusions:

• An interpretable machine learning model (TGFNN-R) can effectively predict furosemide dose changes in neonates after CCS.
• This approach shows promise for enhancing clinical decision support in managing postoperative fluid balance.
• Future research includes external validation and testing in clinical decision support and closed-loop systems.