A76-23 Increased Accuracy of the CoughCheck Automated Cough Detection Algorithm in Real World Settings via Predictive Modeling of Error-Prone Segments

Rationale Manual cough annotation of respiratory recordings remains the gold standard for objective cough monitoring in clinical trials. Automated cough detection via machine learning can accelerate and reduce the costs of cough monitoring, but its performance can be impaired in real-world conditions (e.g. high background noise, throat clearing). To improve performance of automated cough detection in real world respiratory recordings, we developed an error prediction model (EPM) to automatically identify specific error-prone subsets for manual labeling during an otherwise automated approach. Methods The CoughCheckTM cough detection algorithm was trained to identify cough events in 3,099h (233,626 annotated coughs) of respiratory sound recordings collected in real-world conditions using a dedicated biosensor (Strados Labs RESP®) in 249 subjects with subacute (3 to 8 weeks) or chronic (≥8 weeks) cough. Recordings from 74 separate subjects (62,079 annotated coughs) were used to train the EPM to predict the number of CoughCheck-generated cough labels discordant with human-annotated cough labels for a given 1-minute recording segment based on audio and non-audio features. Results Automated cough detection was conducted in: 1) recordings from 36 subjects with subacute cough (mean duration 23.5h ([22.5:23.66] min:max), geometric mean cough frequency 8.53 ± 3.34 coughs/h per subject), and 2) recordings from 40 subjects with chronic cough (mean duration 23.46h ([22.12:23.65] min:max), geometric mean cough frequency 20.1 ± 2.90 coughs/h). Ground truth cough labels were obtained via manual annotation. The EPM identified the subset of 1-minute predicted cough-containing audio segments corresponding to the 10th percentile highest-predicted discordance between manual and machine learning annotation (2% and 2.4% of subacute and chronic cough datasets, respectively). Using the EPM, substitution of automated labels by manual labels for these segments led to increases in precision at 90% sensitivity for cough detection from 78.8% to 85.7% in the subacute cough dataset and 85.3% to 89.5% in the chronic cough dataset. Accompanying generalized linear model analysis indicated that the mean cough peak sound intensity, cough sound power, and median RMS amplitude per 1-minute audio segment were factors most-associated with discordant cough predictions. Conclusions Error prediction modeling can reliably identify subsets of real-world respiratory sound recordings which will benefit most from targeted manual human annotation in an otherwise automated cough labeling approach. This enables increased cough detection accuracy with a limited impact on costs or process efficiency through a minimum amount of human overread, as well as objective analysis of error-contributing factors to help guide clinical or research decisions. This abstract is funded by: Strados Labs