B29-20 A Novel Method of Continuous, Automated Wheeze Detection Using a Remote Biosensor in Patients With Asthma and COPD in Both In-patient and Ambulatory Settings

Rationale Wheezing is a marker of exacerbations and inadequate control of obstructive lung diseases such as asthma, COPD and bronchiectasis. Ambulatory wheeze monitoring may enable early detection of exacerbations and remote monitoring of disease control. Currently, assessing the character and frequency of wheezes requires bedside auscultation and patient recall, which are subjective and intermittent in nature. In contrast, wearable remote lung sound monitors can enable automated longitudinal assessment during ambulatory conditions. We report here the performance in wheeze detection by a machine learning algorithm applied to continuous ambulatory lung sound recordings. To our knowledge, in contrast to previously published reports, this is the first report of automated wheeze detection from ambulatory data with varying background noises, patient positioning and motion. Methods The WheezeCheckTM detection algorithm was trained to identify the presence of wheezing in 9,283 minutes (4,699 expert-annotated wheeze sounds) of respiratory sound recordings collected using a dedicated wearable biosensor (Strados Labs RESP®) in 107 subjects with obstructive lung diseases. Biosensor recordings were obtained during a variety of in-patient and ambulatory conditions, including sleep, sedentary, and active time. Additional supplemental training data (228 minutes, 709 annotated wheeze sounds) were included from the open-source ICBHI dataset. Recordings from 12 separate subjects (2,409 minutes, 1,069 annotated individual wheezes) were used to compare the algorithm’s performance against human-annotated wheezes. The classification of 1-minute audio recordings as “wheeze-containing” vs. “non-wheeze-containing” by WheezeCheck was assessed via receiver operating characteristic (ROC) curve, generated by plotting sensitivity vs. false positive rate (FPR) for varying thresholds of positive-class detection. Results The measured area under the ROC curve (AUC-ROC) was 0.870. Corresponding false positive rates of 8.4%, 24.1%, and 53.9% were measured at sensitivities of 60%, 80%, and 95%, respectively. Conclusions The WheezeCheck algorithm identified wheezing in 1-minute audio segments with satisfactory agreement vs. human auscultation in real-world in-patient and ambulatory lung sound recordings. Performance measured by sensitivity vs. FPR is comparable to metrics reported for other automated wheeze presence classifiers which assess lung sound recordings obtained during clinical exam conditions with non-ambulatory lung sound monitors. This breakthrough in performance in real-world, “noisy” ambulatory condition warrants further evaluation of WheezeCheck for use in remote patient monitoring and decentralized clinical trials. Future work will evaluate WheezeCheck on broader ambulatory patient datasets and examine additional metrics e.g. time spent wheezing and wheeze-free time, as well as relevant contextual measures e.g. body orientation and activity levels.