Machine learning and sensitivity analysis for predicting nasal drug delivery for targeted deposition

Abstract

Targeted nasal drug delivery can provide improved efficacy for drug formulations to be delivered at high efficacy rates. Some parameters that influence drug delivery have a dependency on the patient's technique of administration and the spray device itself. When the different parameters, each having a specific range of values are combined, the combinatory permutations for studying its effects on particle deposition become large. In this study, we combine six input spray parameters (the spray half-cone angle, the mean spray exit velocity, the breakup length from the nozzle exit, the diameter of the nozzle spray device, the particle size, and the sagittal angle of the spray) with a range of values to produce 384 combinations of spray characteristics. This was repeated for three inhalation flow rates of 20, 40, and 60 L/min. To reduce the computational costs of a full transient Large Eddy Simulation flow field, we create a time-averaged frozen field and perform the time integration of particle trajectories through the flow field to determine the particle deposition in four anatomical regions of the nasal cavity (anterior, middle, olfactory and posterior) for each of the 384 spray field. A sensitivity analysis determined the significance of each input variable on the deposition. It was found the particle size distribution significantly affected deposition in the olfactory and posterior regions, while the spray device insertion angle was significant for deposition in the anterior and middle regions. Five machine learning models were evaluated based on 384 cases and it was found that despite the small sample dataset the simulation data was sufficient to provide accurate machine-learning predictions.

Keywords: CFD; Drug delivery; Machine learning; Nasal spray; Sensitivity analysis.