Controlling surfactant therapy for acute respiratory distress syndrome with two-phase flow and machine learning – INHALE

Surfactant replacement therapy (SRT) consists of instilling a liquid-surfactant mixture into the trachea of premature newborns, whose high alveolar surface tension makes their lungs stiff and difficult to inflate. Despite indisputable efficacy, the dose and administration method have been primarily directed by trial and error, while poor insight into the underlying physical mechanisms blurs the clinical benefits expected for other respiratory diseases and age groups. In this respect, developing new, systematical capabilities to support decision-making and assessment of treatment options by medical doctors is of tremendous interest.

Meanwhile, the rapid expansion of machine learning to multiple domains has yielded important progress in the domain of decision-making techniques, by the coupling of deep neural networks with reinforcement learning algorithms (called deep reinforcement learning, or DRL). DRL has consistently, albeit slowly, spread to fluid mechanics, and is now widely believed to have matured enough to breakthrough into real-life applications. Indeed, the ability of neural networks to leverage past data into actionable information, to enhance future decision-making while taking full advantage of actuation possibilities, offers a viable alternative to model-based control of strongly nonlinear flows. For all that, DRL has never been applied to pulmonary simulations. This project is thus an attempt at pushing forward the development of the method while expanding its scope of real-life applications via targeted control of surfactant delivery in human airways.

The proposal aims then at devising enhanced surfactant therapy concepts from a computational framework combining state-of-the-art numerical methods (to characterize delivery efficiency and homogeneity from high-fidelity, two-phase flow simulations of the air-surfactant system) and deep reinforcement learning algorithms (to optimize the delivery by dynamical adjustment of functional parameters: dose, flow rate, surfactant properties, patient posture). Realistic surfactant behavior (surfactant rheology and surface tension of the air-surfactant interface) will be embedded in the multiphase solver via experimentally validated constitutive laws. Particular attention will be paid to testing and validating the project results by conducting experimental pilot studies in idealized (symmetric, asymmetric) models of human airways, all the way up to patient-specific geometries reconstructed from medical imaging data of actual subjects. This has never been done before in this context and should open both new theoretical and numerical opportunities.

INHALE relies on a strong consortium bringing together the complementary skills and expertise of high-impact scientists in the key disciplines: computational mechanics coupled with artificial intelligence, transport and delivery in the pulmonary airway system, and surfactant rheology transport. The project brings to the next level novel computational and optimization frameworks capable of simulating and optimizing SRT throughout idealized and patient-specific models of lung airway trees. This will allow reducing the subjectivity in clinical evaluations of SRT efficacy, supporting the decision of treatment options by doctors and finally providing guidance in the development of new therapeutical routes. Such capabilities can save lives, improve the life quality of patients, and eliminate the lifelong side-effects and increased health care costs associated with poor patient outcome.

We envision the proposed work as a first step towards enabling personalized care of respiratory diseases. The proposed research is highly multidisciplinary, and the methods proposed and developed in the course of the project can be quickly adapted to a wide range of engineering and bio-medical applications.