Related Experiment Video
Updated: Jun 5, 2026
07:31
Implementation of a Real-Time Psychosis Risk Detection and Alerting System Based on Electronic Health Records using CogStack
Published on: May 15, 2020
Development and Interpretability Analysis of a Stacking Ensemble Model for Early Prediction of Nutritional Risk in
1Institute of Medical Information/Medical Library, Chinese Academy of Medical Sciences and Peking Union Medical College, No. 3 Yabao Road, Chaoyang District, Beijing, 100020, China, 86 01052328782.
JMIR Medical Informatics
|June 3, 2026
Summary
A new machine learning model, E-NUTRIC, accurately predicts malnutrition risk in ICU patients within 24 hours. This data-driven approach outperforms traditional scores, enabling earlier intervention for critically ill patients.
Area of Science:
- Critical care medicine
- Nutritional science
- Machine learning in healthcare
- Electronic health records analysis
Background:
- Malnutrition in critically ill patients increases morbidity and mortality.
- Traditional screening tools like the modified NUTRIC (mNUTRIC) score have limitations in the dynamic ICU environment due to subjective data and delays.
- Electronic health records (EHRs) offer potential for real-time, data-driven risk stratification.
Purpose of the Study:
- To develop and validate a machine learning model, E-NUTRIC, for early malnutrition risk prediction within 24 hours of ICU admission.
- To improve predictive performance over traditional scoring systems and individual machine learning models using ensemble learning.
- To maintain clinical interpretability of the developed model.
Main Methods:
- Retrospective cohort study using MIMIC-IV database (version 3.1).
- Developed a stacking ensemble model (E-NUTRIC) with Logistic Regression, Random Forest, XGBoost, and LightGBM as base learners and a logistic metalearner.
- Assessed performance using AUROC, precision-recall curves, and calibration curves, comparing against the mNUTRIC score.
Main Results:
- The E-NUTRIC model achieved an AUROC of 0.875, significantly outperforming the mNUTRIC score (AUROC=0.635).
- E-NUTRIC demonstrated superior performance compared to individual base learners, including XGBoost (AUROC=0.871) and LightGBM (AUROC=0.866).
- SHAP analysis identified key predictors such as minimum serum albumin, admission weight, and early hypokalemia.
Conclusions:
- The E-NUTRIC stacking ensemble model offers an interpretable and accurate method for early nutritional risk screening in ICUs.
- Utilizes routinely available EHR data for automated and objective risk stratification.
- Demonstrates superior discrimination to the mNUTRIC score, enabling better identification of high-risk patients.
