Parenteral nutrition is an essential therapeutic option for patients who are unable to meet their nutritional needs through oral or enteral routes, but its complex composition and potential complications require specialized management. Pharmacists in ...
Parenteral nutrition is an essential therapeutic option for patients who are unable to meet their nutritional needs through oral or enteral routes, but its complex composition and potential complications require specialized management. Pharmacists in the nutrition support team (NST) are responsible for ensuring the safety and appropriateness of parenteral nutrition prescriptions; however, limited personnel and time often hinder the provision of comprehensive consultations. Although research on automated parenteral nutrition systems using artificial intelligence has recently progressed, evidence remains scarce in Korea. Therefore, this study aimed to pilot-develop a ‘machine learning-based parenteral nutrition consultation decision-support algorithm’ applicable in hospital settings, thereby establishing a methodological foundation for streamlining pharmacist workflows and standardizing decision-making.
This study collected de-identified NST consultation replies written by pharmacists at Seoul National University Hospital from January 1, 2022, to December 31, 2024, via the hospital’s clinical data warehouse (SUPREME 2.0). Free-text fields were segmented into semantic units (chunks) based on delimiters and keywords, and structured variables were constructed using natural language processing, resulting in a total of 74 analytical variables. Subsequently, a three-step algorithm was designed. Step 1 (Model 1) is a regression model predicting calorie and protein requirements. Its performance was evaluated using mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and the coefficient of determination (R²) by comparing extreme gradient boosting (XGB), multilayer perceptron (MLP), and ensemble models. Step 2 (Model 2) is a Random Forest–based classification model predicting commercial product suitability (is_special) and central venous administration (is_central). Accuracy, F1 score, AUROC, and AUPRC were calculated, and feature importance was analyzed. Step 3 (Model 3) targeted patients eligible for commercial PN-only regimens and consisted of two components: (1) predicting required calorie, protein, lipid, and dextrose amounts using CatBoost-based out-of-fold estimates, and (2) converting these predictions into feasible combinations of commercial PN products and infusion rates by applying clinical guideline–based constraints and practice-derived rules (e.g., infusion rate 10–80 mL/hr, GIR ≤ 4 mg/kg/min, full-bag administration preference). This study was conducted after approval by the Institutional Review Board (IRB No. 2407-164-1556).
Out of 54,020 total cases, 8,688 PN consultations were analyzed after applying exclusion criteria. Calorie and protein requirements were presented as single values in 66.8% and 60.9% of cases, respectively; when minimum–maximum ranges were provided, the mean absolute ranges were 240.7 kcal/day for calories and 16.4 g/day for protein. In Model 1, XGBoost demonstrated the best performance, with MAE 69.63 kcal/day and R² 0.832 for calorie prediction, and MAE 6.87 g/day and R² 0.697 for protein prediction. Model 2 showed high discriminative power, with AUROC 0.906 for predicting commercial PN suitability and 0.987 for central venous administration. Key predictors included ICU status, renal function indicators (e.g., eGFR_effective, BUN), anthropometric and metabolic variables, and the presence of a central venous catheter. Model 3 structured actual clinical rules (infusion rate 10–80 mL/hr, GIR ≤ 4 mg/kg/min, preference for full-bag administration) to generate optimal formulation combinations. Model 3 showed significant positive correlations with actual pharmacist prescriptions, with Pearson correlation coefficients of 0.75 for calories, 0.76 for protein, 0.68 for lipids, and 0.77 for dextrose (all p < 0.001). Bland–Altman analysis indicated mean differences close to zero, although the 95% limits of agreement were relatively wide due to limitations in available input information.
This is a pilot study that quantified and structured the pharmacist PN consultation process using large-scale real-world data. The developed algorithm was confirmed to predict requirements within safe clinical ranges and support efficient patient triage and standardized prescription combination searches. These findings provide key foundational evidence for the construction of advanced intelligent clinical decision support systems and are expected to expand clinical applicability through future multi-center validation and prospective studies.