Addressing Disparities Between Physician Rated and Self-Reported Dyspnea Using Machine Learning

Nearly 50% of intensive care and 25% of tertiary care patients experience dyspnea, or the sensation of unsatisfied breathing. Dyspnea decreases quality of life, impacts clinical outcomes, and can have long standing psychological consequences for some patients. Current methods of identifying dyspnea severity (such as the respiratory distress observation scale or RDOS) lack suffcient quantifiable measurements of dyspnea to accurately predict its severity. In our testing, while the RDOS does have predictive power, it frequently produces underestimates compared to self-reported dyspnea severity (r = 0.32; p = 0.0081). Furthermore, dyspnea severity can change throughout the day, requiring healthcare professionals to conduct RDOS evaluations multiple times to attempt to meet patient needs. Therefore, there is an urgent need for continuous, accurate monitoring of dyspnea severity in the hospital setting. To address this need, we are using non-invasive data collected during induced bouts of dyspnea to train a machine learning (ML) algorithm in dyspnea prediction. We employ a steady-state end-tidal forcing breathing system and adjust end-tidal gas tensions manually while collecting P O­­­ ­ 2 , P CO2 , SpO 2 , and ventilation data. Preliminary findings show several physiological and demographic factors independently correlate to dyspnea severity, including BMI (r = 0.49; p = 0.007), end-tidal P CO ­­ 2 (r = 0.33; p = 0.008), and inspired P O2 (r = -0.25; p = 0.043). We achieved an accuracy of 84%, 83.3%, and 82% in-dataset predictions using k-nearest neighbor, random forest, and logistic regression approaches respectively. Several physiological biomarkers and demographic factors were used as inputs and self-reported dyspnea was used as the predicted output. These findings show promise for the use of ML in dyspnea prediction. With further testing and use of a broader participant population, it may be possible to use ML models such as these to improve patient outcomes and overall quality of life. University of California, Riverside. This is the full abstract presented at the American Physiology Summit 2024 meeting and is only available in HTML format. There are no additional versions or additional content available for this abstract. Physiology was not involved in the peer review process.