Performance Decay and Federated Learning Prospects in EHR-based AI Malnutrition Screening Models: A Systematic Review

Title: Performance Decay and Federated Learning Prospects in EHR-based AI Malnutrition Screening Models: A Systematic Review Description: [Purpose] Artificial Intelligence (AI) models leveraging Electronic Health Records (EHR) have demonstrated significant potential in automating malnutrition screening. However, the "generalization gap"—where models trained in single-center settings suffer from severe performance decay when deployed across different institutions—remains a critical bottleneck. This systematic review aims to quantify this performance decay, evaluate the methodological quality of current predictive models, and explore the necessity of Federated Learning (FL) as a privacy-preserving solution to data heterogeneity in clinical nutrition. [Methods] Following the PRISMA 2020 guidelines, this study conducts a comprehensive search across five major databases: PubMed, Web of Science, Scopus, IEEE Xplore, and Embase. We specifically target original research utilizing machine learning or deep learning for malnutrition risk prediction based on real-world EHR data. The methodological quality and risk of bias are rigorously assessed using the PROBAST (Prediction model Risk Of Bias ASsessment Tool). [Key Areas of Investigation] The review focuses on three dimensions: Performance Disparity: Comparing internal validation metrics (AUC, Sensitivity, Specificity) with external/multi-center validation results to identify the "Performance Cliff." Data Heterogeneity: Analyzing how variations in clinical documentation habits and laboratory standards contribute to model failure. The FL Roadmap: Investigating the current research gap in decentralized collaborative modeling and the potential of Federated Learning to bridge the generalization gap while maintaining data privacy. [Expected Outcomes] This study expects to reveal a widespread high risk of bias in current clinical nutrition AI models, particularly in data analysis and calibration reporting. By synthesizing evidence of performance decay, we aim to provide a strategic roadmap for transitioning from "closed-loop single-center modeling" to "open-ecosystem multi-center collaboration," establishing Federated Learning as the future standard for robust, generalizable, and privacy-compliant malnutrition screening.