Background

Ex vivo lung perfusion (EVLP) is a cutting-edge clinical and research platform that maintains human lungs in a physiologically stable environment outside the body. This enables detailed functional assessment of the lungs and targeted therapeutic interventions.
Our Digital Twin (DT) project leverages this platform to create a dynamic, high-fidelity digital twin of ex vivo lungs, capable of forecasting over 75 key functional parameters and serving as a powerful tool for biomedical research and preclinical therapeutic optimization.

Multi-modal Lung Function Data

Our DT is built on the world’s largest annotated dataset of ex vivo human lung function, combining high-frequency ventilator waveform recordings, hourly lung function assessment measures, biochemical and protein biomarkers, imaging data, and transcriptomics. This comprehensive multimodal resource underpins the DT’s design and supports robust model training, calibration, and validation.

DT Conceptualization and Architecture

The DT integrates physics-based models, grounded in well-established lung physiology equations, with advanced machine learning approaches. Physics-based components extract metrics such as lung compliance and airway pressures from high-resolution ventilator waveforms, while data-driven models—such as gated recurrent units (GRU) and XGBoost—capture complex temporal and nonlinear patterns. This hybrid approach ensures the DT is both interpretable and adaptable to diverse experimental conditions. The DT employs two main model architectures: (1) Gated recurrent unit neural networks for time series forecasting, capturing sequential dependencies in key lung physiology parameters over time; and (2) XGBoost for key hourly lung function parameters. These models are trained and calibrated to ensure accurate short- and long-term forecasts, with hyperparameter tuning designed to ensure model stability and generalizability.

DT Workflow

From raw data ingestion to physics-based feature extraction, model training, calibration, and deployment in real-time inference and therapeutics testing.

DT Deployment

Accessible via Docker container, Streamlit Web-based app, Google Colab, and Python, with automated data and model loading, inference, and visualization for local or cloud use.