An ML-powered meta-analysis framework incorporating transfer learning techniques was developed to integrate cross-laboratory evidence. Studies investigating lipid emulsion stability in PN were systematically identified, screened, and subjected to original data. This process established a multidimensional ML database. Domain adaptation techniques were applied to address heterogeneous data distributions across laboratories. Subsequently, ML algorithms were employed to construct a lipid emulsion stability prediction model, with performance validated through cross-validation.

A systematic literature search was conducted across PubMed, Embase, and Web of Science to identify relevant studies on PN stability. The search was restricted to English-language publications from 2000 to 2024. The following search syntax was employed: (“parenteral nutrition” OR “total parenteral nutrition” OR “total nutrition admixture” OR “all in one nutrition” OR “lipid emulsion”) AND (“stability/instability” OR “compatibility/incompatibility” OR “safety/unsafety”).

Studies were included if they met the following criteria:

Experimental investigations of PN stability relevant to clinical practice.

Studies were excluded if they:

Assessed only non-physical stability aspects (e.g., chemical degradation, precipitation).

Flow diagram shown in Figure 1, summarizes the search strategy following PRISMA (5).

The features were extracted from the experimental studies reported in the literature including Environmental variables, PN composition characteristics and stability assessment results. Two researchers independently extracted the features data of experimental studies based on the following features definitions and formed the original dataset after cross-verification.

Environmental variables included:

Missing value handling: An automated protocol classified missingness as: Low (<5%), Moderate (5%–15%), High (15%–30%), and Extreme (>30%). Missing data mechanisms were evaluated using Little’s Missing Completely at Random (MCAR) test (α = 0.05). For variables with 5%–20% missingness, Multiple Imputation by Chained Equations (MICE) was implemented (10 iterations generating 5 imputed datasets).

Data were expressed as frequencies (percentages) for categorical variables and as Mean ± SD or median (IQR) based on the test of normality for continuous variables. The Kolmogorov–Smirnov test was applied to test normality. Differences between stability and instability groups were analyzed using Student’s t-test for continuous variables and Fisher’s exact test for categorical variables. A correlation heatmap was used to test potential multicollinearity between features. Statistical analyses and modeling were conducted using IBM SPSS Statistics (v25.0) and scikit-learn (v1.1) within Python 3.7 (Python Software Foundation; accessed January 1, 2025).

Algorithmic diversity: Tree-based models (random forest, decision tree, XGBoost, gradient boosting), Non-tree models (logistic regression, support vector machine, K-Nearest neighbors, Naïve Bayes) and Ensemble/Neural (Multilayer Perceptron and AdaBoost).

We employed SHAP values to quantify feature importance and interpret model predictions. Grounded in cooperative game theory’s Shapley values, SHAP provides consistent, theoretically-grounded attribution of feature contributions to individual predictions. Predictions. Model-specific implementations were utilized: Trees SHAP for tree models and Kernels HAP for non-tree models. Additivity was verified (tolerance <0.01), with ensemble SHAP values computed as weighted averages of base models. Global importance was ranked by mean absolute SHAP values, identifying top features.

The final analytical dataset comprised 1,518 PN samples representing 872 unique clinical prescriptions, extracted from 19 studies published between 2000 and 2024. These studies involve six different types of emulsions (three kinds of LCTs, two kinds of MCT /LCTs and one kind of SMOF lipid), storage temperature from 4 to 37 °C and storage time from 0 h to 14 days, the concentration of monovalent cation (Na⁺/K⁺) from 0 to 230 mmol/L, divalent cations (Ca²⁺/Mg²⁺) from 0 to 92.5 mmol/L. The concentration of macronutrient (mean ± SD): Amino (3.4 ± 2.0 g/100 mL), Dextrose (11.2 ± 6.2 g/100 mL), Lipid (2.8 ± 1.0 g/100 mL), Complete study characteristics are detailed in Table 1.

Correlational analyses between extracted features and emulsion stability are presented in the correlation heatmap (Supplementary Figure S1). MDD analysis performed on 1,058 bags (69.7%), with instability (MDD ≥ 0.5 μm) detected in 32 samples (3.0% 32/1058). PFAT5 were determined 971 samples (64.0%), instability (PFAT5 ≥ 0.05%) observed in 173 samples (17.8% 173/971). Overall instability defined as violation of either USP threshold (MDD ≥ 0.5 μm or PFAT5 ≥ 0.05%), occurring in 199 samples (13.1% 199/1518). Comparative analysis between stable (n = 1,319) and unstable (n = 199) groups demonstrated significant differences (p < 0.05) in most variables except the concentration of Calcium, Magnesium, Lipid, Lipid composition and Trace elements. Complete statistical comparisons are detailed in Table 2.

We compared the performance of ML models to predict PN lipid stability using ROC analysis. The AUC-ROC of the XGBoost model had the highest predictive value (AUC-ROC: 0.968, Accuracy: 0.980), followed by a random forest model (AUC-ROC: 0.962) and Gradient Boosting (AUC-ROC: 0.961). Detailed performance measures for all models are provided in Supplementary Table S1. We demonstrated the ROC curve comparison of 10 prediction models (Figure 2). Imbalance treatment optimization (Supplementary Table S2) confirmed SMOTE significantly enhanced XGBoost performance versus alternative methods. Finally, the XGBoost model achieved a recall of 0.98, a Precision of 0.98 and F1 score of 0.98.

Detailed mechanism of XGBoost: Algorithm Principle, An ensemble learning method based on gradient-boosted decision trees, which builds a strong classifier by sequentially training multiple weak learners; Key Parameters, max_depth = 6 (controls tree depth to balance bias and variance), learning_rate = 0.1, n_estimators = 100 (number of trees); Objective Function, Obj = Σ l (yᵢ, ŷᵢ) + Σ Ω(ƒₖ), where “l” is the loss function and “Ω” is the regularization term; Feature Interaction, Capable of automatically learning non-linear interaction relationships between features without manual feature engineering.

The test set confusion matrix shows: True Positives (correctly identified unstable cases) = 55, False Negatives (unstable cases misclassified as stable) = 4. True Negatives (correctly identified stable cases) = 392, False Positives (stable cases misclassified as unstable) = 5. This results in a Sensitivity/Recall of 93.2% (55/59) and a Specificity of 98.7% (392/397). We demonstrated the confusion matrix of SMOTE+XGBoost model (Supplementary Figure S2).

To provide recommendations for practice with a straightforward understanding of key features related to predicting lipid stability, SHAP analysis of the optimal XGBoost model identified key determinants of lipid emulsion stability with or without Multicollinearity handing (Figure 3). Amino, Phosphate, Dextrose, Time storage and Lipid composition demonstrated predominant predictive influence.

This study established an XGBoost-driven multimodal data fusion framework to achieve high-accuracy prediction of lipid emulsion physical stability in PN. The model demonstrated exceptional performance on the test set, significantly outperforming conventional logistic regression and random forest. Key innovations include intelligent multisource data integration through transfer learning and domain adaptation, resolving heterogeneity from environmental conditions and prescription differences by enabling equivalent data transformation across 19 studies. Optimized Feature Engineering, 17 stability relevant features were extracted.

To address the potential impact of multicollinearity among input variables (as shown in Supplementary Figure S1), we systematically evaluated and managed feature correlations during data preprocessing. Variables were carefully reviewed, and those with lower physicochemical relevance (e.g., CAN) were excluded. A comparative sensitivity analysis of SHAP values was conducted before and after this multicollinearity handling. Although the ranking of specific features saw minor shifts, the overall importance trend remained largely consistent. This process refined the contribution weight of individual variables to the model, thereby enhancing the robustness and interpretability of our final model.

Although multiple continuous variables showed statistically significant differences between the stable and unstable PN groups (Table 2), their inherent variability and intercorrelations make it difficult to determine the stability of lipid in individualized TNA. This further highlights the advantage of using a ML approach, which can integrate multidimensional interactions among features—such as amino acids, dextrose, and electrolytes—to produce a robust predictive model. Importantly, the SHAP analysis reaffirmed the model’s ability to synthesize these interrelated influences, with key factors (e.g., amino acids, dextrose, and storage time) consistently emerging as major contributors, thereby validating the model’s interpretability and clinical relevance.

The performance gain achieved through the SMOTE-based resampling strategy can be attributed to the intrinsic characteristics of our dataset. SMOTE-generated synthetic samples effectively populate sparse regions within the feature space. Consequently, SMOTE’s interpolation mechanism effectively populated these sparse regions with plausible synthetic samples, which significantly enhanced the model’s ability to identify the minority class. This approach elevated the AUC from 0.95 to 0.97 and elevated the Accuracy from 0.97 to 0.98, thereby ensuring the predictive safety of the model. We add a sensitivity analysis regarding key SMOTE parameters in Supplementary Table S3.

The stability of lipid in TNA is most significantly influenced by the concentration of amino acids, which play a crucial role in regulating the pH of TNA. This study first identifies phosphate concentration as a novel predictor of PN lipid stability through ML modeling—a previously unreported association whose mechanism warrants further investigation, though we hypothesize involvement of phosphate groups in the emulsion’s hydrophilic layer. Following features such as concentration of others electrolyte and dextrose, type of lipid, CAN, or storage time that have been largely evaluated and confirmed as relevant predictors of lipid stability in PN. CAN was excluded due to multicollinearity handling. This was also discovered in all models SHAP analysis (Supplementary Figure S3). Cations can disrupt the electrostatic repulsion between lipid droplets, thereby compromising the stability of lipid emulsions, though modulated by compositional factors (6). The 100-fold higher aqueous solubility of medium-chain caprylic acid (C₈) versus palmitic acid (C₁₆) (7), reduces oil-water interfacial tension, explaining the low stability of LCT emulsions (8, 9) observed in our model. While temperature and storage duration influence stability, their limited predictive role may relate to anomalous PFAT5 reduction over time—potentially attributable to compounding-induced air bubbles (10). Our stability assessment aligned with USP and Chinese Pharmacopoeia standards using MDD/PFAT5, as enlarged droplets (>5 μm) pose embolism risks: murine studies demonstrate PFAT5 > 0.05% causes microvascular damage and reticuloendothelial injury (11), consistent with clinical reports of pulmonary fat deposition in neonates (12); critically, PFAT5 > 0.4% induces phase separation visible as surface oil droplets (1), representing immediate life-threatening risks.

As shown above, there has a critical impact of PN stability on patient safety, particularly the complex physicochemical behavior of lipid emulsions influenced by multifactorial interactions. In clinical practice, there are complex decisions in medical order, prescription review, PN preparation and PN safety monitoring. This ML-driven solution empowers pharmacists and physicians to rapidly evaluate PN compatibility by analyzing the influencing factors of specific prescriptions. ML has been increasingly used in the field of nutrition science, considering the diversity and complexity of individuals, ML can be used to predict precision nutrition support through multivariate and high-dimensional data analysis. In 2022–2027, the National Institutes of Health of the United States will provide $170 million to fund the development of algorithms for precision nutrition support (13). Our work provides a distinct contribution. Rather than merely applying XGBoost model, this study establishes a comprehensive, ML-enabled meta-analysis framework tailored specifically to PN stability. Our approach integrates a uniquely standardized clinical dataset, domain-specific feature engineering that combines environmental, electrolyte, and macronutrient variables, and an interpretable predictive tool with direct clinical utility. By adapting ML techniques to this specific challenge, we offer not only a practical system for stability prediction but also a reusable methodological framework for studying complex systems stability in pharmaceutical and nutrition sciences.

As a retrospective investigation, this study necessitates further validation through prospective multi-laboratory trials, particularly to verify phosphate’s mechanistic role in emulsion destabilization. In addition, lipid emulsion stability parameters were insufficient: pH, Zeta potential, peroxide parameter, and free fatty acid were lacking, but we followed scientific physical stability verification criteria. Subsequently, extended multi-task learning was used to predict calcium and phosphorus precipitation, vitamin degradation and other stability risks.

ML approach overcomes the clinical impracticality of experimental verification for individualized PN formulations by modeling multifactorial interactions across heterogeneous datasets. Future research will expand training data to enhance generalizability and integrate physical–chemical stability predictions, ultimately constructing a comprehensive risk assessment paradigm for PN quality assurance.