Abstract

Background: Objective cough measurements based on audio recordings using deep learning techniques are being developed. Open data is available, but their value has not been studied.

Aim: To evaluate the value of open sound data for training deep learning models in detecting cough from chronic cough patients? audio recordings.

Methods: For training data, we collected audio recordings from chronic cough patients through a patient-worn mp3 recorder (35h from 14 patients: 1200 coughs; 829 negatives). Also, open data were collected (COUGHVID: 1200 coughs; YouTube sounds: 10,000 negatives; NST Swedish ASR Database: 10,000 negatives). Three datasets were assembled: 1) Own data only; 2) COUGHVID+YouTube+NST; 3) Own data + COUGHVID + YouTube + NST. Each dataset was then used to learn models, balanced to minimize the total amount of false positive and negative findings.

For the validation data, recordings performed in three chronic cough patients (in total 3.5h, 72 coughs) were analyzed by the pre-trained models, and then compared with manual scoring by trained staff, to estimate false positives and false negatives.

Results: Using model 1 (own data), total false findings were 8.5/h (Sensitivity 75%, Precision 89%).

Using model 2 (open data), total false findings were 8.5/h (Sensitivity 88%, Precision 75%).

Using model 3 (all data), total false findings were 6.5/h (Sensitivity 92%, Precision 80%).

Conclusions: A small sample from actual patient recordings gave higher precision despite having roughly 1/20 negatives compared with open data. However, open data gave higher sensitivity in detecting cough from audio. A combination of actual patient recordings andopen data seems optimal to create effective models for measuring cough.