A risk-adjusted length of stay to evaluate severe Community-Acquired Pneumonia (sCAP) outcomes: A machine learning analysis of 16,985 ICU admissions

Background

Severe community-acquired pneumonia (sCAP) strains ICU resources, demands efficient allocation and robust performance metrics for value-based care. Current ICU assessment methods often lack the necessary detail and risk adjustment for effective resource management, especially for complex patient populations such as those with sCAP.

Research question

This study investigated whether the machine learning-based Standardized Length of Stay Ratio (SLOSR) could reliably measure risk-adjusted LOS for sCAP patients, enabling benchmarking. SLOSR is calculated as the sum of observed length of stay (LOS) divided by the sum of predicted LOS, enabling ICU benchmarking for resource use.

Study design and methods

We conducted a multicenter retrospective cohort study of 16,985 adult sCAP admissions across 220 ICUs in 57 Brazilian hospitals (January–December 2023). Data included demographics, comorbidities, SAPS 3, and ventilatory support. The machine learning model predicted LOS and calculated SLOSR. Rigorous validation included cross-validation, calibration plots, funnel plot analysis, RMSE, MAE, and R².

Results

Hospital mortality was 9.3 %, ICU mortality 6.4 %, median ICU LOS 4 days, mean SAPS 3 score 50; 28.1 % received ventilatory support. The SLOSR demonstrated a robust grouped R² of 0.89. The model achieved RMSE = 4.57 and MAE = 3.10, with excellent calibration. Funnel plot analysis revealed a median SLOSR of 1.13 (Q1 = 0.9; Q3 = 1.34), underscoring its potential for benchmarking.

Conclusion

SLOSR shows promise as tool for assessing adjusted LOS as a surrogate of resource use in sCAP patients in the context of Brazilian ICUs. Further research is needed to validate its performance in other settings.