Digital twins in mechanical ventilation : models, identification, and prediction of patient-specific response to care.

Abstract

Mechanical ventilation (MV) is a fundamental, core treatment for both daily and critical care in hospitals. Patient-specific lung condition varies significantly due to patient-specific response to the wide range of lung disease and dysfunction. In response, ventilator settings are currently set primarily by clinician experience around broad guideline approaches. They are thus not standardized, and often non-optimal. Technologically, the response has been to design a far broader array of MV modes promising better treatment, but leading to more uncertainty in care and no big change in patient outcomes. All of these uncertainties result in higher morbidity and mortality than if care was personalized.

Personalizing care requires greater insight into time-varying, patient-specific lung condition and response to MV care. Mathematical models offer the opportunity, in combination with clinical measured data, to enhance the understanding of patient-specific lung mechanics. A predictive model would allow optimization of MV settings, while minimizing risk to the patient. However, such models and the data available from the ventilator are relatively limited.

This thesis studies a well-validated lung mechanics model and extends it with physiological-relevant basis functions to enable prediction over MV care settings. It yields accurate predictions for lung mechanics responses to changes in ventilator settings and thus enhances understanding of patient-specific lung mechanics.

The thesis also addresses over-distension, a leading cause of ventilator induced lung injury, which increases length of MV, length of stay, morbidity, mortality, and thus cost. The mechanics relevant digital-twin model and methods are extended to account for the potential appearance of distension in measured pressure-volume (P-V) loops. The new model with basis functions accurately detects and predicts excessive airway pressure resulting by over-assistance from ventilators. It also accurately captures and predicts the retained volume, denoted the dynamic functional residual volume (Vfrc), as positive end expiratory pressure (PEEP) changes. This model-based approach is further extended to create a non-invasive, predictive over-distension index to assess lung condition and MV settings breath-to-breath, which is also much more intuitively understandable than the only current validated metric.

Finally, spontaneous breathing effort and assisted breathing MV modes are considered. The digital-twin model and estimation methods are explored and combined for these partial support MV modes. The results show the model fits well and has the ability to identify physiological features from patient effort during breathing. Thus, the model is taken from fully supported ventilation to include assisted and partial support modes.

Overall, accurate predictions in patient-specific lung mechanics are important and promising in MV care. Each area studied in this thesis is validated using clinical data and shows model efficacy in optimizing care, while minimizing risk. The combination between well-validated lung mechanics models and physiological-relevant basis functions provides novel insights for future research advancement for safe and optimal clinical care. The models and methods presented offer capabilities of capturing and predicting lung mechanics response under multiple MV modes no other research has yet realized, and provide a platform for future personalized care and improved outcomes at reduced total cost.