Early prediction of ICU readmissions using classification algorithms.

Area of Science:

• Critical Care Medicine
• Health Informatics
• Predictive Analytics

Background:

• Intensive Care Unit (ICU) readmissions are linked to increased mortality, longer hospital stays, and higher costs.
• Accurate risk stratification is essential for improving outcomes in critically ill patients.
• Current prediction models often rely on discharge data, limiting their utility for early intervention.

Purpose of the Study:

• To test the hypothesis that patient characteristics and data available at ICU admission can accurately predict ICU readmission.
• To develop an early warning system for patients at high risk of clinical deterioration post-discharge.

Main Methods:

• Analysis of an anonymized dataset of 9,926 adult patients from three ICUs in Brazil.
• Modeling ICU readmission as a binary classification task using eight different algorithms.
• Evaluation of prediction accuracy using multiple metrics, including AUROC.

Main Results:

• Predictions based solely on admission data achieved high accuracy, comparable to using the full dataset.
• The developed model achieved an AUROC score of 0.91, outperforming existing literature.
• Feature selection methods did not diminish prediction accuracy when using admission data.

Conclusions:

• Patient data at ICU admission is sufficient for accurate readmission prediction.
• Early identification of high-risk patients is feasible using admission data.
• Findings support the use of early markers to anticipate clinical deterioration after ICU discharge.