Machine learning-based prediction of aerodynamic performance in arformoterol-lactose dry powder inhaler formulations using surface roughness features

Abstract

Dry powder inhalers (DPIs) are widely used for pulmonary drug delivery, and their aerodynamic performance is highly dependent on particle surface morphology. This paper presents a machine learning-based framework to quantitatively predict DPI aerodynamic behavior using surface roughness parameters from scanning electron microscopy (SEM) images. Ten DPI formulations were prepared with various storage conditions and compositions. Thirteen surface roughness features were extracted using ImageJ and SurfCharJ and used as inputs for predictive models, including multilayer perceptrons (MLPs), support vector machines (SVMs), logistic regression (LR), and convolutional neural networks (CNNs). The models were trained to estimate the emitted dose (ED), delivered dose (DD), fine particle fraction (FPF), and fine particle dose (FPD).

Among the models, the MLPs paired with RF-based feature selection best predicted the ED and DD. CNNs trained directly on SEM images predicted FPF and FPD better. Surface roughness parameters kurtosis (Rku) and skewness (Rsk) were the most consistent critical predictors across all performance metrics, showing nonlinear relationships with the aerodynamic behavior and specific inflection thresholds associated with marked shifts in particle adhesion and detachment. Higher Rku and Rsk values indicated sharp and asymmetric surface structures and were associated with improved dispersion performance (higher DD, FPF, and FPD), but reduced ED, reflecting more efficient deaggregation.

Combining predictive modeling and mechanistic interpretation, this integrative approach provides a data-driven and interpretable basis for a rational arformoterol-lactose DPI formulation design. The results highlighted Rku and Rsk as practical morphological indicators optimizing the performance of inhalable dry powders in pulmonary drug delivery systems.