P431 - PREDICTING RESTING ENERGY EXPENDITURE USING MACHINE LEARNING WITH PHYSIOLOGICAL AND CLINICAL DATA: A RETROSPECTIVE STUDY

P431

"PREDICTING RESTING ENERGY EXPENDITURE USING MACHINE LEARNING WITH PHYSIOLOGICAL AND CLINICAL DATA: A RETROSPECTIVE STUDY"

O. Raphaeli^1,*, L. Statlender², I. Kagan², E. Robinson², M. Hellerman-Itshaki ², P. Singer²

¹Industrial Engineering and Management, Ariel University, Ariel, ²Intensive Care Unit, Beilinson Hospital, Petah Tikva, Israel

Rationale: Indirect calorimetry (IC) is the gold standard for measuring resting energy expenditure (REE), to avoid overfeeding when prescribing parenteral nutrition in the ICU. However, its routine use is limited. This study aims to develop a machine learning (ML) based regression model for REE prediction using physiological and clinical variables.

Methods: The cohort included 955 adult ICU patients hospitalized for over 48 hours (2011-2018) with IC measurements. Linear models (Linear, Ridge, Lasso) and non-linear regression (Random Forest (RF), Gradient Boosting (GB)) were assessed for REE prediction using physiological and clinical data. Five-fold cross-validation evaluated performance using mean average error (MAE). Ethical approval was obtained.

Results: RF and GB achieved MAEs of 303 and 307 calories, respectively, while the linear model had the poorest performance (MAE = 326). Including BMI improved MAE across all models. The RF model highlighted BMI as the most influential feature: higher BMI categories increased REE, while lower BMI reduced it. Temperature significantly impacted REE, heart rate and urine output showed moderate effects and arterial parameters had minimal influence. 

Image:

Conclusion: Non-linear ML models outperformed linear models in REE estimation, effectively capturing complex variable interactions. Weight and temperature had the strongest impact, while heart rate and urine output added value despite variability. ML may aid energy estimation and reduce overfeeding risk when prescribing parenteral nutrition. 

References: [1] Spolidoro, G. C. I., D'Oria, V., De Cosmi, V., Milani, G. P., Mazzocchi, A., Akhondi-Asl, A., Mehta, N. M., Agostoni, C., Calderini, E., & Grossi, E. (2021). Artificial Neural Network Algorithms to Predict Resting Energy Expenditure in Critically Ill Children. Nutrients, 13(11), 3797. https://doi.org/10.3390/nu13113797

Disclosure of Interest: O. Raphaeli: None declared, L. Statlender: None declared, I. Kagan: None declared, E. Robinson: None declared, M. Hellerman-Itshaki : None declared, P. Singer Grant / Research Support from: Baxter