Document resource
Background Cough frequency, typically reported as 24-hour cough frequency, is used as an objective measure of cough. It may be derived by manually annotating coughs in ambulatory audio recordings collected using established cough monitoring systems such as the VitaloJAK®. However, this can be a time-consuming process. Here we present an AI-based audio trimming algorithm that removes non-cough activity to provide reduced-length (trimmed) audio recordings to assist manual cough annotation. The aim of this study was to assess its performance in sensitivity (cough retention) and reduction of audio recording length, for more efficient cough analysis.Methods A test data set was employed comprising 24-hour audio recordings from 20 adult chronic cough subjects (4M/16F) from the RaDAR database (Manchester University NHS Foundation Trust), collected using the VitaloJAK. Coughs were annotated by trained analysts totalling 14,931 coughs (range: 17–3,164; median: 469). Recordings were then trimmed by the algorithm and independent blinded analysts annotated the coughs on the trimmed recordings. Sensitivity (retention of original cough annotations) and percentage reduction in recording length post-trimming were computed. Agreement between 24-hour cough frequency derived from untrimmed versus trimmed recordings was assessed using intra-class correlation (ICC), linear regression and Bland-Altman analyses.Results The algorithm achieved a median sensitivity of 100% (95% CI: 99.9–100; range: 97.6–100) and reduction in audio recording length of 95.4% (95% CI: 93.8–96.7; range: 75.5–97.7). 24-hour cough frequency derived from untrimmed versus trimmed recordings showed excellent agreement (ICC=0.99; R 2=0.99; p<0.001), with a unity slope of 1 (95% CI: 0.99–1.01) and minimal bias (intercept=11). Median difference bias was 8 coughs (p=0.8).Conclusions The audio trimming algorithm may greatly reduce audio recording length with minimal impact on 24-hour cough frequency, showing strong potential to support more efficient cough analysis.