UC Riverside

Dyspnea is the subjective sensation of breathing discomfort. It is prevalent in patients with chronic and critical illness and associated with poor clinical outcomes and long-term psychological trauma. The multidimensional nature of dyspnea and individual variation in its presentation make identification difficult, particularly in non-communicative patients. To address this critical need, I aimed to (1) better understand the factors contributing to individual variation in dyspnea severity by identifying its physiological and psychological drivers, and (2) produce a novel tool for monitoring dyspnea based on noninvasive biomarker inputs. I recruited healthy adults and administered demographic and psychological evaluations to evaluate their health, anxiety levels and interoceptive awareness traits. I then induced dyspnea using a semi-rebreathing, forced end-tidal system to control arterial partial pressures of oxygen and carbon dioxide while collecting self-reported dyspnea scores during free breathing. I also collected clinically-relevant noninvasive biomarker data including respiratory air flow rates, inspiratory and end-tidal gas tensions, electrocardiogram (ECG), and oxygen saturation. Video was recorded to allow physician estimates of dyspnea during experiments. Dyspnea severity was positively associated with physiological factors including the hypercapnic ventilatory response (HCVR) (p = 0.006), respiratory rate (p < 0.001), tidal volume (p < 0.001), minute ventilation (p < 0.001), and heart rate (p = 0.006). Through mixed linear model analysis, I found that trait anxiety has a positive association with dyspnea severity (p = 0.042). “Attention regulation”, “emotional awareness” and “self regulation” also influenced dyspnea severity (p < 0.001, p = 0.004, p = 0.004 respectively). Using a combination of demographic and physiological data I trained categorical machine learning models to determine if a multivariate method would achieve higher accuracy in identification of dyspnea than observational estimates by healthcare professionals. Ensembled and classical models were tested. Random forest performed best overall with an F1 score of 0.805. We then recruited physicians and nurses to view the videos and assess participant dyspnea using the respiratory distress observation scale. They achieved a precision-recall F1 score of 0.608, meaning our model performed 32.4 percent better. This study highlights the multivariate nature of dyspnea, and the need for further evaluation of identification methods.