A Machine Learning Approach to Predicting Dyspnea with Noninvasive Biomarkers

Rationale: Dyspnea is the subjective sensation of breathing discomfort. This symptom is highly prevalent in patients with chronic and critical illness, and its presence is associated with poor clinical outcomes and long-term psychological trauma. The multidimensional nature of the neurophysiological mechanisms underlying dyspnea, paired with individual variation in its presentation, makes identifying and monitoring this symptom difficult, particularly in non-communicative patients. Objective: Undetected and untreated dyspnea in critically ill patients is a significant problem contributing to patient suffering. Therefore, the objective of this study was to investigate the feasibility of machine learning methods for assessing and continuously monitoring dyspnea using easily obtained noninvasive biomarkers. Methods: We recruited healthy participants (N = 60, 35 women) and stimulated dyspnea using a forced end-tidal semi-rebreathing circuit to modulate arterial oxygen and carbon dioxide levels while collecting non-invasive biomarker data and continuous self-reported dyspnea severity scores. This data was used to train machine-learning models to predict the presence or absence of significant dyspnea (Numeric Rating Scale ≥ 3). We then compared the performance of our final model to observational estimates by trained healthcare providers. Measurements and Main Results: The final model (Random Forest) performed well (PR-AUC=0.832) and exceeded the accuracy of observation estimates made on the same participants using the Respiratory Distress Observational Scale (RDOS) (accuracy=54%). Conclusions: These results indicate that machine learning models can utilize non-invasive biomarker inputs to accurately predict carbon dioxide- and hypoxia-induced dyspnea in a healthy population during spontaneous breathing.