EURO 2025 Leeds
Abstract Submission

2869. Assessing ICU Efficiency using Machine Learning: Analysis of 16,985 Patients with Severe Community-Acquired Pneumonia (sCAP) across 220 Brazilian ICUs

Invited abstract in session TB-13: AI and Machine learning in healthcare, stream OR in Healthcare (ORAHS).

Tuesday, 10:30-12:00
Room: Clarendon SR 1.01

Authors (first author is the speaker)

1. Igor Tona Peres
Department of Industrial Engineering, Pontifical Catholic University of Rio de Janeiro
2. Amanda Quintairos e Silva
D'Or Institute for Research and Education (IDOR)
3. Guilherme Ferrari
Department of Industrial Engineering, Pontifical Catholic University of Rio de Janeiro
4. Joana da Matta Furtado Ferreira
Department of Computer Science, Pontifical Catholic University of Rio de Janeiro
5. Leonardo S. L. Bastos
Industrial Engineering, Pontifical Catholic University of Rio de Janeiro
6. Vicente Dantas
D'Or Institute for Research and Education (IDOR)
7. Jorge Salluh
D'Or Institute for Research and Education (IDOR)

Abstract

Severe community-acquired pneumonia (sCAP) strains intensive care unit (ICU) resources, demanding efficient allocation and robust performance metrics for value-based care. Traditional ICU assessments often lack granularity and risk adjustment, particularly for complex cases such as sCAP. This study evaluated the application of the Standardized Length of Stay Ratio (SLOSR), a machine learning-based metric of ICU efficiency, in sCAP patients.
We analyzed 16,985 adult sCAP admissions across 220 ICUs in 57 Brazilian hospitals (January-December 2023), the largest sCAP efficiency cohort to date. SLOSR is calculated as the sum of observed length of stay (LOS) divided by the sum of predicted LOS. The machine learning model predicting ICU LOS uses a stacking model (combining Random Forests and Linear Regression) and was updated and validated using the 2022 and 2023 data. We used cross-validation and calibration plots to evaluate the predicted LOS at ICU level. Model performance showed a root mean square error of 4.57 and R² = 0.89. Funnel plot analysis revealed a median SLOSR of 1.13 (Q1=0.9; Q3=1.34), demonstrating no bias in its use for benchmarking.
To facilitate broader adoption, an R package named SLOS was developed, enabling researchers and clinicians to calculate and analyze SLOSR efficiently. SLOSR is a valuable tool for assessing ICU efficiency in sCAP patients, aligning with value-based care principles and supporting data-driven decision-making in critical care settings.

Keywords

Status: accepted

Back to the list of papers