Towards a decision support tool for intensive care discharge: machine learning algorithm development using electronic...

Area of Science:

• Medical Informatics
• Clinical Decision Support

Background:

• Assessing patient readiness for intensive care unit (ICU) discharge is complex.
• Existing discharge criteria may not fully capture all relevant factors.
• Automated methods could enhance efficiency and accuracy.

Purpose of the Study:

• To develop and validate an automated method for identifying ICU patients ready for discharge.
• To improve upon previously proposed discharge criteria using machine learning.

Main Methods:

• Utilized two large historical patient datasets (Bristol GICU and MIMIC-III).
• Developed and trained random forest and logistic regression classifiers.
• Employed multiple-source cross-validation for robust performance assessment.

Main Results:

• Classifiers demonstrated improved performance over original discharge criteria.
• Models generalized well across different patient cohorts.
• Identified key predictive features consistent with clinical expertise.

Conclusions:

• The proposed automated approach for discharge readiness classification is feasible.
• This method can complement existing risk models in decision support systems.
• Further improvements are identified for clinical utility.