Predicting and simulating effects of PEEP changes with machine learning

Abstract

Background/Objectives Choosing ventilator settings, especially positive end-expiratory pressure (PEEP), is a very common and non-trivial task in intensive care units (ICUs). Established solutions to this problem are either poorly individualised or come with high costs in terms of used material or time. We propose a novel method relying on machine learning utilising only already routinely measured data. Methods Using the MIMIC-III (with over 60000 ICU stays) and eICU databases (with over 200000 ICU stays) we built a deep learning model that predicts relevant success parameters of ventilation (oxygenation, carbon dioxide elimination and respiratory mechanics). We compare a random forest, individual neural networks and a multi-tasking neural network. Our final model also allows to simulate the expected effects of PEEP changes. Results The model predicts arterial partial pressures of oxygen and carbon dioxide and respiratory system compliance 30 minutes into the future with mean absolute percentage errors of about 22 %, 10 % and 11 %, respectively. Conclusions The deep learning approach to ventilation optimisation is promising and comes with low cost compared to other approaches.