Artificial intelligence assisted nutritional risk evaluation model for critically ill patients: Integration of explainable machine learning in intensive care nutrition.

Abstract

Background and objectives: Critically ill patients require individualized nutrition support, with assessment tools like Nutrition Risk Screening 2002 and Nutrition Risk in the Critically Ill scores. Challenges in continu-ous nutrition care prompt the need for innovative solutions. This study develops an artificial intelligence assisted nutrition risk evaluation model using explainable machine learning to support intensive care unit dietitians.

Methods and study design: Ethical approval was obtained for a retrospective analysis of 2,122 pa-tients. Nutrition risk assessment involved six dietitians, with 1,994 patients assessed comprehensively. Artificial intelligence models and shapley additive explanations analysis were used to predict and understand nutrition risk.

Results: High nutrition risk (35.2%) correlated with elder age, lower body weight, BMI, albumin, and higher disease severity. The AUROC scores achieved by XGBoost (0.921), CatBoost (0.926), and LightGBM (0.923) were superior to those of Logistic Regression. Key features influencing nutrition risk included Acute Physiology and Chronic Health Evaluation II score, albumin, age, BMI, and haemoglobin.

Conclusions: The study introduces an artificial intelligence assisted nutrition risk evaluation model, offering a promising avenue for continuous and timely nutrition support in critically ill patients. External validation and exploration of feature relationships are needed.