Machine learning tool helps evaluate ICU resource efficiency in severe pneumonia care

A study published in the Journal of Critical Care, conducted with the participation of the D'Or Institute for Research and Education (IDOR), investigated how to measure efficiency in the use of resources for patients with severe community-acquired pneumonia (CAP), an illness contracted outside hospital settings and most common among older adults.

Severe CAP represents one of the greatest challenges for ICUs. It requires complex resources, ranging from prolonged hospitalizations to respiratory support, directly affecting hospitals' ability to deliver quality care. Despite its relevance, traditional methods of evaluating hospital performance do not always take patient severity into account, which undermines fair comparisons between institutions and hinders more effective management strategies.

Risk-adjusted care

To address this problem, researchers tested the Standardized Length of Stay Ratio (SLOSR), a tool developed with machine learning techniques, a branch of Artificial Intelligence. The aim was to determine whether SLOSR could predict, in a patient risk–adjusted way, the appropriate length of ICU stay. This would allow for more accurate comparisons across hospitals, highlighting both overuse and underuse of resources.

The study was retrospective and multicenter, analyzing 16,985 adult CAP admissions in 220 ICUs across 57 Brazilian hospitals during 2023. Variables such as age, comorbidities, need for mechanical ventilation, and disease severity were taken into account.

A machine learning model was applied to predict expected length of stay, allowing researchers to calculate the SLOSR as the ratio between observed and predicted times. To ensure robustness, they performed strict statistical validation, including calibration plots, cross-validation, and error metrics, confirming the model's alignment with clinical reality.

Key findings

Median length of stay was four days, and approximately 28% of patients required ventilatory support. The model showed strong explanatory power with low prediction errors, reinforcing SLOSR's potential as a reliable indicator of resource efficiency across ICUs.

The study demonstrates that SLOSR could be a valuable tool for hospitals and healthcare managers, enabling ICU performance evaluation adjusted for patient severity. This approach helps identify where resources are being used efficiently and where waste may be occurring. Researchers, however, note that further investigations are needed to test the method's applicability in other contexts, such as different countries and healthcare systems.