April 2026

PULMONARY

STUDY TESTS MACHINE LEARNING MODELS FOR PREDICTING PNEUMONIA TREATMENT SUCCESS

Featuring: Catherine A. Gao, MD
​
This study analyzed data from the Successful Clinical Response in Pneumonia Therapy (SCRIPT) trial, a prospective observational cohort of mechanically ventilated patients who underwent bronchoalveolar lavage (BAL) for suspected community-acquired pneumonia (CAP).

Using ICU data from the first three days after BAL, investigators trained multiple extreme gradient boosting models to predict if patients’ pneumonia episodes would be treated successfully within seven or eight days.

Model features
  • Vital signs and physiologic variables adapted from Halm’s clinical stability criteria
  • Measures of oxygenation and ventilatory support
  • Severity of illness indicators, including vasopressor use
  • Neurologic status assessed by Glasgow Coma Scale

The models relied on early clinical features already available in routine ICU care, without requiring novel testing or delayed endpoints.
 
Key Findings
  • The best‑performing model included the worst Halm’s clinical features and worst ventilator features in the first three days
  • The best-performing model demonstrated:
    • Area under the receiver operating characteristic curve: 0.79 on the holdout test set
    • 79% sensitivity and 78% specificity for identifying likely treatment failure within the next few days.
  • Important features factored into prediction included:
    • Neurologic status
    • Norepinephrine use
    • BMI
  • Adding lab values or medication data (including antibiotic spectrum and steroid dose) did not significantly improve model performance.
 
Implications for Clinical Practice
While preliminary, this study demonstrates that short‑term outcome prediction in severe CAP is feasible using explainable machine learning models and routinely available ICU data.

​This study is an important first step toward real‑time, decision‑support tools for critically ill patients with pneumonia. Future work will focus on prospective validation and evaluation of clinical impact across broader care settings.
Importantly, this model is investigational and not intended for standalone clinical decision‑making.
Catherine A. Gao, MD, Assistant Professor of Pulmonary and Critical Care

Dr. Gao was a co-author of this research study.

Refer a Patient

Northwestern Medicine welcomes the opportunity to partner with you in caring for your patients. ​

You May Also Like