Decision tree methodology for electronic health record (EHR) clinical data endpoints
Date
2026
Authors
Kasdorf, Kale
Journal Title
Journal ISSN
Volume Title
Publisher
University of Victoria
Abstract
This study explores the use of decision tree models on electronic health record (EHR) data to support clinical decision-making in intensive care units (ICUs). The publicly available Massachusetts Institute of Technology (MIT) MIMIC-IV eICU dataset, spanning 20 hospitals and over 35,000 deidentified patient encounters, was analyzed to identify relevant tables and variables. Decision trees were developed in KNIME, a data science software, starting with simple algorithms, which were iteratively refined into more complex and realistic predictive models. Initially, the simple models achieved high predictive performance, although overfitting and unrealistic splits were observed. Further development of these models produced slightly worse predictive performance, but addressed the overfitting and generated more realistic and interpretable predictions. This study establishes a preliminary pipeline for applying decision trees to ICU EHR data and highlights the need for ongoing refinement to improve predictive reliability.
Description
Keywords
decision trees, machine learning, healthcare, clinical decision support, prediction, HER, Jamie Cassels Undergraduate Research Awards (JCURA)