Predicting individualized treatment effects of corticosteroids in community-acquired-pneumonia: a data-driven analysis of randomized controlled trials

Abstract

AbstractBackgroundCorticosteroids could improve outcomes in patients with community-acquired pneumonia (CAP). However, we hypothesize that corticosteroid effectiveness varies among individual patients, resulting in inconsistent outcomes and unclear clinical indication. Therefore, we developed and validated a predictive, causal model based on baseline characteristics to predict individualized treatment effects (ITEs) of corticosteroids on mortality in patients with CAP.MethodsWe obtained individual patient data from six randomized controlled trials comparing corticosteroid therapy to placebo in 1,869 adult CAP patients. The study endpoint was 30-day mortality. We performed effect modelling through logistic regression and evaluated the predicted ITEs in terms of discrimination and calibration for benefit. Our modelling procedure involved variable selection, missing value imputation, data normalization, encoding treatment variables, creating interaction terms, optimizing penalization strength, and training logistic regression models. We evaluated discriminative performance using the newly proposed ‘AUC-benefit’.FindingsThe model identified high levels of CRP and glucose, at baseline, as main predictors for benefit of corticosteroid treatment. Using a decision threshold of ITE=0, the model predicted harm in 1,004 patient and benefit in 864 patients. We observed benefit in patients where the model predicted benefit, with an odds ratio of 0.5 (95% CI: 0.3 to 0.9) and a mortality reduction of 3.2% (95% CI: 0.7 to 5.6), and no statistically significant benefit in the patients where the model predicted harm, with an odds ratio of 1.1 (95% CI: 0.7 to 1.8) and a negative mortality reduction (hence, increase) of −0.3% (95% CI: −2.6 to 1.8). The model yielded an AUC-benefit of 184.9 (28.6 to 347.6, 95% CI), underestimated ITEs in the lower ITE region and slightly overestimated ITEs in the higher ITE region.InterpretationOur model has potential to identify patients with CAP who benefit from corticosteroid treatment, and aid in the design of personalized clinical trials. We will prospectively validate the model in two recent CAP trials.