Evaluation of multi-output machine learning models for predicting inhaled particle deposition in the human upper and central airway

Abstract

Targeted drug delivery to the deep lung improves therapeutic outcomes, but respiratory system variability complicates drug spray design. Numerical simulations offer insights for individualized treatments but are computationally intensive, highlighting the need for surrogate models for real-time deposition prediction. This study comprehensively explores the multi-task predictive capability of regression models, including Linear regression (LR), Bayesian regression (BR), Back Propagation Neural Network (BPNN), Support Vector Machine (SVM), Random Forest (RF), XGBoost, and CatBoost, for predicting total and regional deposition rates of inhaled particles in airway. A training dataset is obtained from well-validated CFD simulations with realistic human airway model using Euler-Lagrangian method. The results indicate that LR, BR, and SVM yield unsatisfactory predictive accuracy, with average R² values in range of 0.21 to 0.73. Comparatively, BPNN and decisiontree-based models show great potential in predicting total deposition rate in the upper and central airway. However, for regional deposition rate prediction, BPNN did not consistently yield high accuracy, particularly for oral deposition (R² = 0.538). Comparatively, XGBoost emerges as optimal model, achieving an R² approximately close to 1 on both the training and testing datasets, with predictive errors within the range of ±0.5. The overall results demonstrate that decision-tree-based models, particularly XGBoost, have superior performance in accurately predicting both total and regional deposition rates of inhaled particles within airway. Despite limitations like geometry complexity and data quantity, the workflow developed in this study is expected to pave the way for future research integrating ML models into drug delivery device design and evaluation.