Machine learning predicts lipid emulsion stability in parenteral nutrition using multi-laboratory literature data.

Objective: Physical instability of lipid in parenteral nutrition (PN) poses significant clinical safety risks. As lipid stability is influenced by multiple complex factors and remains incompletely characterized, this study aimed to quantify the relative importance of stability determinants and to develop a machine learning (ML) model for predicting stability in individualized PN prescriptions. Methods: A retrospective meta-analysis integrated experimental data from multi-laboratory studies. The ML framework employed transfer learning for cross-laboratory data harmonization and Synthetic Minority Over-sampling Technique (SMOTE) for class imbalance mitigation. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC-ROC) and accuracy. Results: The datasets comprised 17 stability-related features (electrolytes, macronutrients, and storage conditions) extracted from 1,518 samples representing 872 unique PN formulations across 19 studies (2000 and 2024). The XGBoost model achieved exceptional predictive performance (accuracy: 98.2%, AUC 0.968). SHAP-based feature importance analysis identified the concentrations of Amino and phosphate, storage time and lipid composition as key stability determinants. Conclusion: This study establishes the first interpretable ML framework for predicting lipid emulsions stability in PN, resolving cross-laboratory data heterogeneity. We have provided a high-accuracy prediction tool for assessing lipid emulsion stability in PN, while the methodology demonstrates generalizability for stability studies of complex drugs and nutrients formulations.