Speakers
Description
Mechanical ventilation (MV) is a life-saving therapy used in the intensive care unit (ICU). However, improper settings can lead to lung injury and organ damage. Determining the ideal ventilation settings is challenging due to the large number of variables involved, making it difficult to provide clear guidelines.
The IntelliLung project addresses this issue with a reinforcement learning (RL) algorithm that was trained on 19 groups of variables, such as blood gas analysis, circulatory function, demographics, and gas exchange, to recommend optimal settings for up to 10 key ventilator parameters. The goal was to maximize ventilator free days during a patient treatment while maintaining safe vital signs. Factored action spaces were used to overcome the challenge of the large number of discrete action combinations. The algorithm has been trained on diverse datasets, including MIMIC IV, eICU, HiRID, and TUD with the cohort consisting of adults on invasive MV. A use case specific preprocessing pipeline has been developed to convert raw medical measurements into a usable state vector using techniques like data imputation, encoding, and measures of central tendency. In safety critical evaluation such as ours, online evaluations are not feasible. Policy evaluation was done offline using Fitted Q-Evaluation (also with factored action space). This evaluation showed that the algorithm is able to achieve higher returns than clinicians across all datasets. In addition domain experts validated the policy action distributions from a medical perspective.
