Personalization of Mechanical Ventilation Treatment using Deep Conservative Reinforcement Learning

Anonymous

Keywords: reinforcement learning, mechanical ventilation, conservative Q-learning, healthcare

TL;DR: We propose DeepVent, a deep reinforcement learning model to personalize mechanical ventilation treatment settings.

Abstract: Mechanical ventilation is a key form of life support for patients with pulmonary impairment. Nonetheless, the optimal treatment regime is often unknown, leading to sub-optimal care and increased risks of complications. This work aims to develop a decision support tool to personalize mechanical ventilation. We present DeepVent, an off-policy deep reinforcement learning model that determines the best ventilator settings throughout a patient's stay. We evaluate our model using Fitted Q Evaluation, and show that it is predicted to outperform physicians. Moreover, we address the challenge of policy value overestimation in out-of-distribution settings using Conservative Q-Learning and show that it leads to safe recommendations for patients. We also design an intermediate reward based on the Apache II score to further improve our model's performance.