Fig. 1

Schematic representation of the study methodology. a We built an initial dataset from all sources of information available in the first 3 h of the patient’s presentation in the ICU for CAP. We matched these presenting cases with their final identified causal respiratory pathogen. The initial dataset was randomly split into a work dataset, used for the machine learning and training the ICU experts on how the data were presented, and an external validation dataset used to assess the prediction performances of the artificial intelligence (AI) algorithm and the panel of experts. b Data flow to engineer the data-driven algorithm