A55-50 A Biomarker-based AI Tool That Predicts Corticosteroid Treatment Benefit or Harm in Pneumonia

Importance

Community-acquired pneumonia (CAP) is a leading cause of infectious mortality. Corticosteroids have mixed efficacy in CAP. A precision medicine approach that identifies patients most likely to benefit from or be harmed by steroids could optimize treatment, leading to improved outcomes in CAP.

Objective

To develop and validate an artificial intelligence (AI) algorithm, cortiCAP, which integrates clinical features and plasma biomarkers to identify patients with predicted clinical benefit or harm from corticosteroid treatment in CAP.

Design

Secondary analysis of a prospective observational, multicenter cohort study. CortiCAP is a multivariable algorithm that uses patient characteristics at hospital admission for CAP to predict benefit or harm from acute treatment with corticosteroids. This study describes the derivation and validation of cortiCAP, created using supervised machine learning techniques and validated using logistic regression and propensity-matched analyses based on known steroids receipt and mortality.

Setting

Seven hospitals in the United States.

Participants

Included were patients aged 18+ presenting to the hospital February 2018-September 2023 with a blood culture order within 6 hours and an ICD-9 or ICD-10 discharge diagnosis of CAP. Exclusions included COVID-19, asthma, COPD, late or subtherapeutic corticosteroid exposure, and missing >30% of data. Of 1,025 eligible patients, 656 were assigned to the derivation cohort and 369 to the validation cohort.

Exposure

Receipt of systemic corticosteroids (hydrocortisone-equivalent ≥100 mg within 24 hours of index). Use of steroids was a clinical decision.

Main Outcomes and Measures

The primary outcome was 30-day all-cause mortality. Treatment effect estimates compared mortality between steroid and nonsteroid groups overall and stratified by cortiCAP classification as steroid responsive positive (SR+) or negative (SR−).

Results

Corticosteroid treatment showed no difference in mortality in the overall population (absolute risk difference [ARD], +1.3%; p = 0.726). In the validation cohort, cortiCAP classified 46.3% of patients as SR+, and steroids were associated with reduced 30-day mortality (ARD, −9.8%; p < 0.042). Among SR− patients, steroids were associated with increased mortality (ARD, +10.9%; p < 0.047). CortiCAP outperformed CRP, SOFA, and CURB-65 severity measures in predicting steroid benefit.

Conclusions and Relevance

In this study of patients with non-COVID CAP, an AI-based clinical and biomarker algorithm, cortiCAP, stratified patients into one subgroup (46%) predicted to benefit from corticosteroids and another subgroup (54%) predicted to not benefit. Corticosteroid treatment was associated with reduced 30-day mortality (benefit) in the subgroup predicted to benefit, and increased 30-day mortality (harm) in the subgroup predicted to not benefit.

This abstract is funded by: Prenosis, NIH