A Respiratory Digital Twin for Intervention-Specific Prediction of Lung Ventilation and Aerosol Deposition Validated against Multi-Modal Clinical Data
Please login to view abstract download link
Predicting patient-specific response of the human lung to mechanical ventilation and inhaled therapy is a central challenge in both intensive care medicine and pulmonary drug development. In mechanically ventilated patients suffering from respiratory failure, ventilator settings must be individualized to maintain gas exchange while limiting ventilation-induced lung injury. Similarly, inhaled therapies rely on efficient delivery and patient-specific targeting. In both cases, substantial pre-existing inter- and intrapatient variability is further compounded by disease-specific pathophysiological processes. While pulmonary imaging and bedside monitoring provide valuable snapshots of anatomy and function, they cannot by themselves provide the counterfactual predictions required to compare alternative ventilator settings or aerosol delivery strategies. We present a respiratory digital twin for patient-specific prediction of regional ventilation and aerosol deposition under varying ventilation and inhalation conditions. The framework integrates computed tomography to construct a reduced-dimensional representation of the lung resolving subject-specific lobar structure, airway geometry, and deformable parenchymal mechanics, coupled to transient airflow and particle transport. Inter- and intrapatient variability is inferred from imaging and clinical data, and the pipeline is designed for robust and scalable cohort-level application, leveraging machine learning for automated processing of clinical inputs. Model validity is assessed against multi-modal experimental ground truth, with a particular focus on dynamic imaging validation. For mechanically ventilated ARDS patients, regional ventilation is compared against electrical impedance tomography (EIT), as well as against gated 4DCT-derived Jacobian ventilation maps (local volumetric expansion computed from deformable image registration). Additionally, we compare predicted deposition heterogeneity against 3D SPECT/CT imaging of radiotracer deposition, following administration of a radiolabeled aerosol.
