Shi Yuxin
Medical College, Shantou University
Abstract：
This systematic review summarizes research progress in the field of machine learning-based prediction of feeding intolerance (FI) in critically ill patients from 2021 to 2026. It summarizes the development logic, clinical efficacy, and existing limitations of current models to provide evidence-based guidance for critical care nutrition practice and future research. Using original research on machine learning models for predicting FI in critically ill patients and related systematic reviews published between 2021 and 2026 as core literature, this study provides a systematic review of the research trajectory, key directions, model performance, and existing issues in this field. Results: From 2021 to 2026, this field has evolved from early, isolated proof-of-concept studies into a specialized subfield with a comprehensive research framework. Research has centered on the development of direct FI risk prediction models, extending to models for specific patient subgroups, prediction of feeding-related outcomes, assessment of adjacent complication risks, and dynamic precision nutritional support. Most models demonstrated good discriminatory power (AUROC 0.79–0.94), but key issues remain, including inconsistent definitions of FI outcomes, a predominance of single-center retrospective studies, insufficient external validation, and a high overall risk of bias. Machine learning holds great promise for the early prediction of FI in critically ill patients; however, current research remains in the early stages of integration and has not yet met the standards for routine clinical deployment. Future efforts should focus on standardizing outcome definitions, conducting multicenter external validation, and pursuing prospective intervention studies to facilitate the translation of these models into clinical practice and advance the precision of critical care nutrition.
Key Words：
machine learning; critical illness; enteral nutrition; feeding intolerance; risk prediction; nursing
