What question did this study set out to answer?

The research aims to develop machine learning models that predict gastrointestinal injury in sepsis patients using electronic health record (EHR) data.

March 26, 2026

1768: Development of Machine Learning Models to Predict Sepsis-Induced Gastrointestinal Injury by Ehr Data

Key Points

The research aims to develop machine learning models that predict gastrointestinal injury in sepsis patients using electronic health record (EHR) data.
Conducted a retrospective cohort study of sepsis patients in the ICU.
Performed K-M survival analysis and ROC curve evaluations.
Divided patient data into training and testing sets.
Created feature groups using Lasso, Boruta, and Ridge selection methods.
Applied 9 machine learning classifiers and assessed model performance using AUC and other metrics.
AGI and GIDS grades were significantly linked to short- and long-term mortality.
AGI grading outperformed GIDS grading in predictive performance (AUC 0.691 vs. 0.645, p< 0.001).
The CatBoost classifier achieved the highest performance (AUC 0.877).
All 36 ML models demonstrated good discriminative performance (AUC > 0.75).
SHAP analysis highlighted key factors influencing GI injury onset, including fluid input and intra-abdominal pressure.

Abstract

Introduction: The gastrointestinal (GI) tract plays a pivotal role in the initiation and progression of sepsis. Early identification and protection of the GI injury are critical for improving outcomes. Although machine learning (ML) is increasingly applied in sepsis research, predictive models specifically targeting GI dysfunction in sepsis remain limited. Methods: We conducted a retrospective cohort study of sepsis patients admitted to the ICU of the First Affiliated Hospital of Sun Yat-sen University between 2018-2023. K-M survival analysis was performed to assess 1-year mortality across AGI and GIDS grades, and ROC curves were generated to evaluate their prognostic value. To develop predictive models, clinical features from the day of infection were preprocessed. The cohort was randomly divided into training and testing sets (80:20 split). 4 feature groups were created using three selection methods (Lasso, Boruta, Ridge) and the original feature set. 9 ML classifiers were applied: logistic regression, Lasso, Ridge, SVM, MLP, RF, XGBoost, CatBoost, and LightGBM. A total of 36 model combinations were evaluated using AUC, accuracy, precision, sensitivity, and specificity. The best-performing model was further interpreted using feature importance rankings and SHAP analysis. Results: Among 4,167 sepsis patients, both AGI and GIDS grades were significantly associated with short- and long-term mortality. Both AGI grading and GIDS grading can improve the prognostic predictive performance of SOFA score, and the predictive performance of AGI grading is better than GIDS grading (0.691 vs. 0.645, p 0.75). The best-performing model was the CatBoost classifier based on Ridge-selected features (AUC 0.877). SHAP analysis showed fluid input, PEEP, intra-abdominal pressure, gastric residual volume, and blood gas pH are important influencing factors for onset of GI injury. Conclusions: ML-based models enable early identification of sepsis-induced GI injury with strong performance and clinical interpretability. These findings offer a foundation for incorporating predictive modeling into sepsis management strategies to guide early intervention and improve patient outcomes.

Bookmark

Cite This Study

Changjie Cai (Sun,) studied this question.

synapsesocial.com/papers/69c4ccf7fdc3bde448918aff https://doi.org/https://doi.org/10.1097/01.ccm.0001189068.74426.17

Bookmark