Multi-input deep learning to predict ventilator-associated pneumonia in patients treated for major trauma

Purpose: Ventilator-associated pneumonia (VAP) is a frequent complication among critically ill patients with major trauma. We aimed to develop and evaluate a multi-input deep learning model that integrates early clinical data and chest radiographs to predict VAP in mechanically ventilated trauma patients.Methods: We retrospectively analyzed patients aged ≥16 years with an Injury Severity Score ≥16 who were admitted to a level I trauma center intensive care unit and required mechanical ventilation for more than 2 days. VAP was defined as the presence of new or progressive pulmonary infiltrates within the first week after admission together with compatible microbiological findings. Selected clinical variables and early chest radiographs were processed using a ResNet50-based convolutional neural network with long short-term memory layers. Clinical and imaging features were integrated into a multi-input neural network and evaluated using a held-out test dataset.Results: Among 491 patients, 147 (29.9%) developed VAP. In the test dataset, the integrated model achieved a sensitivity of 0.92, a negative predictive value of 0.97, and an area under the receiver operating characteristic curve of 0.84.Conclusion: A multi-input deep learning model that combined early clinical variables with chest radiographs demonstrated high sensitivity and negative predictive value for predicting VAP in patients with major trauma.