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

Area of Science:

• Pharmaceutical Sciences
• Computational Chemistry
• Clinical Nutrition

Background:

• Physical instability of lipid emulsions in parenteral nutrition (PN) presents significant clinical safety risks.
• Factors influencing lipid stability are complex and not fully understood, necessitating further investigation.
• Accurate prediction of PN formulation stability is essential for safe clinical practice.

Purpose of the Study:

• To quantify the relative importance of various determinants on lipid emulsion stability in PN.
• To develop a machine learning (ML) model for predicting stability in individualized PN prescriptions.
• To address cross-laboratory data heterogeneity in stability studies.

Main Methods:

• A retrospective meta-analysis integrated multi-laboratory experimental data.
• A machine learning framework utilized transfer learning for data harmonization and SMOTE for imbalance mitigation.
• Model performance was assessed using AUC-ROC and accuracy metrics.

Main Results:

• The XGBoost model demonstrated high predictive performance (accuracy: 98.2%, AUC: 0.968) using 17 stability-related features from 1,518 samples.
• Key stability determinants identified include amino acid and phosphate concentrations, storage time, and lipid composition.
• The model successfully resolved cross-laboratory data heterogeneity.

Conclusions:

• This study presents the first interpretable ML framework for predicting lipid emulsion stability in PN.
• A high-accuracy prediction tool for assessing PN lipid emulsion stability has been developed.
• The methodology is generalizable for stability studies of other complex drug and nutrient formulations.