Comparative machine learning modeling of resting energy expenditure estimation in mechanically ventilated children after cardiac surgery

Abstract

Nutrition status is vital for children’s recovery following cardiac surgery, with substantial inter-individual variability in metabolic demands. We aimed to develop machine learning (ML) models of postoperative resting energy expenditure (REE) by analysing the influential factors. We retrospectively analyzed children who underwent indirect calorimetry (IC) for valid REE measurements within 4 to 24 h after cardiac surgery between January 2021 and December 2022 at our center. Nine ML algorithms were trained to predict REE. Bland–Altman analysis assessed agreement with measured REE, and SHAP was used for population-level interpretation and a representative patient case. A total of 278 mechanically ventilated children were analyzed. REE measured by IC ranged from 387 to 2642 kcal/d (715 [550, 964]). The root mean square error (RMSE) of the ML models ranged from 226 (95% CI: 184–267) to 281 (95% CI: 225–344) kcal/d, while the range for conventional equations was 249 (95% CI: 215–287) to 282 (95% CI: 252–315) kcal/d. Regularized linear models achieved the highest R2 of 0.64 (95% CI: 0.49–0.76). The top 10 most important variables associated with REE in the optimal model are weight, age, height, preoperative serum albumin, preoperative CK-MB, vasoactive inotropic score (VIS), CPB time, postoperative LVEDD Z-score, gender, and postoperative NT-proBNP. This study developed ML models to predict REE in children after cardiac surgery, with some models outperforming conventional equations. These findings highlight the potential of machine learning to optimize postoperative nutritional management by accurately capturing the non-linear metabolic response to surgical stress.