Learning to Breathe: Physics v. ML-based Lung Simulations for Control of Medical Ventilators

Abstract

We compare two different physics-based lung simulations to an ML-based data-driven lung simulation, using two evaluation metrics. The first metric quantifies how well the simulation mimics the pressure state of a human lung. The second metric quantifies how well the simulation serves as a training ground for a PID controller targeting a specific pressure waveform. We find that the physics-based simulation performs worse than the data-driven simulation according to the first metric, but on one of the 9 lung settings, the physics based simulation performs better according to the second metric. This paper contributes to an ongoing project on ventilator control conducted by Princeton University and Google AI, aimed towards building more effective and robust mechanical ventilators for use in clinical settings.