RATIONALE. Predicting and preventing COPD exacerbations requiring hospitalization is an important and elusive goal.Previous attempts at exacerbation prediction have been impeded by a lack of adherence when collecting data, a lack of consensus over what constitutes a COPD exacerbation, and a failure to find features across the study cohort that are sufficiently predictive. We propose a novel method of tracking individual COPD patients utilizing the machine learning paradigm of novelty detection, as a means of generating notifications that highly correlate with impending hospital visits by subjects within 0 to 7 days. Notifications are based upon departure from normality of patient physiology, and require longitudinal monitoring of respiratory force, as can be obtained from the Spire Medical Health Tag (MHT). It is hypothesized that prior to a hospitalization there is a high likelihood that breaths acquired during this period will exhibit novelty, resulting in a notification. The method has been inspired from a similar approach utilized to monitor jet engine performance.
METHODS. For this interim analysis, 31 subjects were chosen from the database. For each subject, respiratory signal data from the MHT was processed for individual breaths. Each captured breath was converted into an array of features that describe the duration and shape of the breath. A clustering algorithm was used along with Extreme Value Theory to identify sufficiently novel breaths (exceeding a defined novelty threshold) to trigger a notification.
RESULTS. From the 31 subjects, 19hospitalizations were analyzed. A novelty threshold was set that yielded notifications for 16 out of 19 hospitalizations (Mean notification warning = 4.4 days). On 2 occasions the notification occurred 0 or 1 days before hospitalization. Sensitivity andSpecificity were 84% and 93% respectively. CONCLUSIONS. By tracking respiratory force in individual subjects, continuously over time, novel features of respiratory waveforms were extracted that correlate with COPD hospitalization in the 0 to 7 daytime period. The present approach has been able to overcome problems seen previously in this area by using both longitudinal monitoring of respiratory force (via the MHT form factor) and subject-specific modeling.
Available at: https://www.atsjournals.org/doi/abs/10.1164/ajrccm-conference.2020.201.1_MeetingAbstracts.A5062