(T0930-05-27) Toward Disease-Specific Digital Twins of Ex Vivo Human Lungs for Restrictive and Obstructive Lung Diseases Using Machine Learning

Purpose: Obstructive and interstitial lung diseases can progress to end-stage lung failure, for which organ transplantation becomes the only available treatment option. Unfortunately, developing effective therapeutics to slow or reverse disease progression remains elusive. Existing preclinical animal models fail to fully replicate human pathophysiology. Therefore, despite promising preclinical results, numerous candidate therapies fail in late-stage clinical trials due to lack of efficacy or safety in humans. Ex vivo lung perfusion (EVLP) is a biomedical platform that preserves human lungs outside the body and enables comprehensive evaluation of lung function by surgeons to assess suitability of lungs for transplant. More recently, EVLP has been used to conduct preclinical evaluation of novel therapeutics in human lungs under near-physiological conditions. Leveraging longitudinal EVLP data from disease-free donor lungs, we developed digital twin models capable of emulating the physiological behaviour of human lungs on EVLP. The digital twin models facilitate lung-specific treatment effect analysis by enabling direct comparison between model predictions and observations from the lung undergoing treatment on EVLP. In this study, we extend the ex vivo digital twin approach to create disease-specific digital twins of chronic obstructive pulmonary disease (COPD) and idiopathic pulmonary fibrosis (IPF) lungs by: (1) generating synthetic EVLP data with clinically realistic characteristics of emphysema, chronic bronchitis, and pulmonary fibrosis, and (2) evaluating the predictive performance of our digital twin models on these diseased cases.

Methods: Data Collection and Preprocessing: During n=53 clinical EVLPs performed at our centre, high-resolution (100 Hz) flow, pressure, and volume data were extracted from an ICU-grade ventilator and segmented into individual breaths, each with inspiratory and expiratory phases. Respiratory mechanics parameters were calculated per breath: dynamic compliance, peak and mean airway pressures, and expiratory volume. Synthetic Data Generation: EVLP data was transformed to simulate diseased lungs by modifying the elastance and resistance of each breath to predefined disease-specific ranges. For emphysema, resistance was increased and elastance was decreased; for chronic bronchitis, resistance was increased; and for IPF, elastance was increased. During inspiration, elastance and resistance were estimated via ridge regression using the equation of motion and the measured flow, pressure, and volume. The elastance and resistance were then adjusted and used to update pressure. For expiration, an exponential decay model was fit to estimate the time constant, which was adjusted to update pressure and volume, and flow was computed as the derivative of the updated volume. The transformed breath phases were merged, and new respiratory parameters were computed from the modified signals. Digital Twin Inference: Next, we applied our previously trained gated recurrent unit (GRU)-based digital twin models to forecast the respiratory mechanics parameters of the diseased lung profiles in two steps: (a) using 1st hour data to predict the 2nd hour, and (b) using 1st and 2nd hour data to predict the 3rd hour.

Results: Generated Synthetic Data: Analysis of flow, pressure, and volume waveforms, as well as dynamic pressure-volume and flow-volume loops for the synthetic diseased and normal lungs reveals that changes in elastance and resistance resulted in disease-like characteristics (Figure 1 a-c). Increased elastance led to higher peak pressures, faster expiration, and elevated peak expiratory flow, while increased resistance caused higher peak pressures, slower expiration, and reduced peak expiratory flow. The synthetic emphysema and chronic bronchitis lungs often showed air trapping, unlike the normal and synthetic IPF lungs. Dynamic compliance was lower in the synthetic chronic bronchitis and IPF lungs and higher in the synthetic emphysema lungs (Figure 1 d). Digital Twin Forecasting Performance: Mean absolute error (MAE) and mean absolute percentage error (MAPE) are reported for the hourly GRU forecasts of the respiratory parameters. The MAPEs ranged from approximately 2-10% across the four parameters and two time points, demonstrating that the digital twins effectively captured trends in respiratory parameter behaviour (Figure 2, Table 1). The models forecasted dynamic compliance and peak pressure of the diseased lungs with MAEs 3.9-6.5 mL/cmH2O and 0.6-1.2 cmH2O respectively.

Conclusion: There is a critical need for more appropriate preclinical models for translational research on chronic lung diseases. EVLP combined with digital twin technology offers a promising solution by enabling direct comparison between treated diseased lungs and corresponding digital controls. We generated physiological-based synthetic EVLP data for COPD and IPF by transforming clinical EVLP cases using disease-specific changes in elastance and resistance. Varying the magnitude of these changes enabled simulation of different disease severity. The digital twin models show potential to forecast the behaviour of synthetic diseased lungs. Ongoing experiments include fine-tuning these models with both synthetic and real lung disease data to create specialized digital twins for evaluating targeted therapies.

References: Cypel M, Yeung JC, Liu M, Anraku M, Chen F, Karolak W, Sato M, Laratta J, Azad S, Madonik M, Chow CW, Chaparro C, Hutcheon M, Singer LG, Slutsky AS, Yasufuku K, de Perrot M, Pierre AF, Waddell TK, Keshavjee S. Normothermic ex vivo lung perfusion in clinical lung transplantation. N Engl J Med. 2011 Apr 14;364(15):1431-40. doi: 10.1056/NEJMoa1014597. PMID: 21488765.
Drummond D, Gonsard A. Definitions and Characteristics of Patient Digital Twins Being Developed for Clinical Use: Scoping Review. J Med Internet Res. 2024 Nov 13;26:e58504. doi: 10.2196/58504. PMID: 39536311; PMCID: PMC11602770.

Acknowledgements: Zhou, X., Wang, B., Cypel, M., Sage, A., Keshavjee, S. have patent filings related to the current work. Cypel, M., Keshavjee, S. are consultants for Lung Bioengineering, Traferox Technologies Inc. All others authors have no relationships to disclose. This work is funded by a Project Grant (PJT 185944) from the Canadian Institutes of Health Research (CIHR) and the J.P. Bickell Foundation.


Figure 1: Synthetic data for diseased lungs generated by transforming each breath from a normal EVLP case. ​(a) Flow, pressure, and volume over time for one breath; (b) dynamic pressure-volume (PV) loop; (c) dynamic flow-volume (FV) loop; (d) calculated respiratory mechanics parameters: dynamic compliance, peak and mean airway pressure, and expiratory volume.​


Figure 2: Gated recurrent unit (GRU) model forecasts on a generated synthetic IPF case. Two steps are shown: (a) use 1st hour EVLP data to forecast the 2nd hour, and (b) use 1st and 2nd hour EVLP data to forecast the 3rd hour. 50 breaths are used for each hour. ​


Table 1: Gated recurrent unit (GRU) model forecasting performance on normal lung data and on generated synthetic emphysema, chronic bronchitis, and IPF data averaged over all 53 original or transformed cases. Two steps are shown: (a) use 1st hour EVLP data to forecast 2nd hour, and (b) use 1st and 2nd hour EVLP data to forecast 3rd hour.​