A Novel Machine-Learning Algorithm to Predict the Early Termination of Nutrition Support Team Follow-Up in Hospitalized Adults: A Retrospective Cohort Study

Abstract

Background: For hospitalized adults, it is important to initiate the early reintroduction of oral food in accordance with nutrition support team guidelines. The aim of this study was to develop and validate a machine learning-based algorithm that predicts the early termination of medical nutritional therapy (the transition to oral feeding).

Methods: This retrospective cohort study included consecutive adult patients admitted to the Hacettepe hospital (from 1 January 2018 to 31 December 2022). The outcome of the study was the prediction of an early transition to adequate oral feeding before discharge. The dataset was randomly (70/30) divided into training and test datasets. We used six ML algorithms with multiple features to construct prediction models. ML model performance was measured according to the accuracy, area under the receiver operating characteristic curve, and F1 score. We used the Boruta Method to determine the important features and interpret the selected features.

Results: A total of 2298 adult inpatients who were followed by a nutrition support team for medical nutritional therapy were included. Patients received parenteral nutrition (1471/2298, 64.01%), enteral nutrition (717/2298, 31.2%), or supplemental parenteral nutrition (110/2298, 4.79%). The median (interquartile range) Nutritional Risk Screening (NRS-2002) score was 5 (1). Six prediction algorithms were used, and the artificial neural network and elastic net models achieved the greatest area under the ROC in all outcomes (AUC = 0.770). Ranked by z-value, the 10 most important features in predicting an early transition to oral feeding in the artificial neural network and elastic net algorithms were parenteral nutrition, surgical wards, surgical outcomes, enteral nutrition, age, supplemental parenteral nutrition, digestive system diseases, gastrointestinal complications, NRS-2002, and impaired consciousness.

Conclusions: We developed machine learning models for the prediction of an early transition to oral feeding before discharge. Overall, there was no discernible superiority among the models. Nevertheless, the artificial neural network and elastic net methods provided the highest AUC values. Since the machine learning model is interpretable, it can enable clinicians to better comprehend the features underlying the outcomes. Our study could support personalized treatment and nutritional follow-up strategies in clinical decision making for the prediction of an early transition to oral feeding in hospitalized adult patients.

Keywords: clinical nutrition; hospitalization; machine learning prediction; medical nutrition therapy; nutrition support team; oral intake.

Figure 1

Flow chart of ML procedure.

Figure 2

Feature importance according to the Boruta method. Blue boxes represent shadow attributes, green color signifies significant attributes, and red boxes denote attributes considered unimportant. CAD: coronary artery disease, COPD: chronic obstructive pulmonary disease, CKD: chronic kidney disease, DM: diabetes mellitus, CHF: congestive heart failure, ICUs: intensive care units, BMI: body mass index, NRS-2002: Nutrition Risk Screening Score-2002, GIS: gastrointestinal system, EN: enteral nutrition, PN: parenteral nutrition. The dashed vertical line represents the discrimination between variables that were and were not important according to Boruta algorithm.

Figure 3

ROC curves from six models. ANN: artificial neural network, EN: elastic net, RF: random forest, XGBoost: extreme gradient boosting, SVM: support vector machine.