Paper Type
Complete
Abstract
Hospital-acquired infections (HAIs) are a pertinent health challenge in the United States, particularly within intensive care units (ICUs). This study evaluates the efficacy of two automated machine learning (AutoML) frameworks, PyCaret and H2O, for the early prediction of HAIs using data from 20,682 pneumonia patients in the eICU Collaborative Research Database. Results show that PyCaret’s Random Forest classifier consistently outperformed H2O’s stacked ensemble across all metrics, achieving a peak precision of 87% and recall of 85% when clinical factors were integrated with baseline admission characteristics. Feature importance analysis identified length of stay (ICU and hospital) and hematocrit levels as the most significant clinical predictors. While baseline demographic models provided reliable triage assessments, the inclusion of dynamic clinical variables significantly enhanced predictive accuracy. These findings suggest that end-to-end AutoML frameworks can provide real-time surveillance tools to reduce the burden of nosocomial infections in critical care settings.
Paper Number
1370
Recommended Citation
Oladokun, Blessing Elijah and Olufadewa, Isaac, "Evaluating Automated Machine Learning Approaches for Early ICU Infection Surveillance" (2026). AMCIS 2026 Proceedings. 2.
https://aisel.aisnet.org/amcis2026/sig_odis/sig_odis/2
Evaluating Automated Machine Learning Approaches for Early ICU Infection Surveillance
Hospital-acquired infections (HAIs) are a pertinent health challenge in the United States, particularly within intensive care units (ICUs). This study evaluates the efficacy of two automated machine learning (AutoML) frameworks, PyCaret and H2O, for the early prediction of HAIs using data from 20,682 pneumonia patients in the eICU Collaborative Research Database. Results show that PyCaret’s Random Forest classifier consistently outperformed H2O’s stacked ensemble across all metrics, achieving a peak precision of 87% and recall of 85% when clinical factors were integrated with baseline admission characteristics. Feature importance analysis identified length of stay (ICU and hospital) and hematocrit levels as the most significant clinical predictors. While baseline demographic models provided reliable triage assessments, the inclusion of dynamic clinical variables significantly enhanced predictive accuracy. These findings suggest that end-to-end AutoML frameworks can provide real-time surveillance tools to reduce the burden of nosocomial infections in critical care settings.
When commenting on articles, please be friendly, welcoming, respectful and abide by the AIS eLibrary Discussion Thread Code of Conduct posted here.
Comments
SIGODIS