Author: Sumanth Swaminathan

Coauthor(s): Klajdi Qirko, PhD

Status: Work In Progress

COPD (Chronic Obstructive Pulmonary Disease) is a lung condition that imposes a significant burden on patients’ daily lives. Flare-ups of this condition (exacerbations) are a frequent trigger of physician and hospital visits, which are both costly and distressing to patients. Moreover, exacerbations are associated with long-term decline in lung function. The need for novel solutions that limit the impact of exacerbations on global health is abundantly apparent. One emerging approach to addressing COPD exacerbation is early detection by way of mobile app technology. Many of these apps, however, utilize rule-based decision frameworks, which struggle to capture the size of the variable space and complexity of variable interactions involved in triage and diagnosis. In this study, we discuss a machine learning-based strategy for early detection of COPD exacerbations and subsequent triage. Our application uses physician assisted data generation and triage to train a supervised machine learning (ML) algorithm. The accuracy of the model is assessed against 9 physicians triaging 101 identical patient cases. Deployment of the ML algorithm behind a simple to use, on demand mobile app yields a product with the potential to reduce severe exacerbations in patients with COPD.

The most relevant variables for COPD management and triage are identified through a combination of literature review and physician consultation. The algorithm is trained on the opinion of 6 pulmonologists each triaging 750 computer generated patient cases that form a training dataset. Each case is simulated as a distinct combination of variables, optimized to cover a statistically comprehensive set of plausible clinical scenarios. The algorithm outputs 1) an exacerbation diagnosis, and 2) A triage recommendation from four choices (no action, continue usual treatment and check back in 1-2 days, call MD, and go to ER). The decision-making within the app contains a variety of supervised machine learning classifiers that are cross validated on 5 folds of the training data set.

We tested the algorithm by comparing its assessments of exacerbation and triage recommendations with that of 9 physicians on 100 validation cases. 90 of these cases were statistically generated and clinically validated while 10 high complexity cases were added a priori by physicians. On this set of validation patients, the algorithm correctly assigned the exacerbation and triage classes with accuracy of 97% and 89% respectively using consensus majority opinion as the gold standard. This score was better than all 9 MDs participating individually in the same test in all classification categories. The algorithm further demonstrated superior accuracy and sensitivity in diagnosing emergency events. The final model has been deployed behind Revon Systems’ Smart Symptom Tracker, which is currently undergoing a clinical trial on anxiety reduction, quality of life, and exacerbation reduction at The Vancouver Clinic in Vancouver, Washington.