This study examined how well different computer models could predict death among patients with severe pneumonia and respiratory failure in the ICU. The researchers used data from 164 patients who were already admitted to the hospital. They compared new machine learning models against the commonly used APACHE-II scoring system, which doctors often rely on to estimate risk.
The machine learning models, including GBDT and XGBoost, achieved an accuracy score of 0.83, compared to 0.70 for the APACHE-II system. The study also found that the predictions matched actual patient outcomes well and offered a higher net benefit than simply treating everyone or no one. These models performed better at identifying which patients were at highest risk for dying during their hospital stay.
No safety issues were reported because the models are tools for calculation rather than treatments. However, because this was a small, past-look study, the results are not yet ready for widespread use. Readers should understand that this research suggests a potentially useful tool, but it does not prove that these models are ready to replace current methods in everyday practice.