Development of a deep learning-based classification method for distinguishing wet and dry cough sounds in older adults: a retrospective study

Cough classification can assist in the assessment of respiratory conditions in older adults. Distinguishing wet and dry coughs is clinically important, and expert auditory assessment is commonly used for this classification. This study developed an automated system to reproduce pulmonologists’ auditory classification of wet and dry cough sounds in older adults. A total of 122 older adults with ages ranging from 65 to 98 years were included in the study. Coughs were induced via inhalation of a citric acid-physiological saline solution using an ultrasonic nebulizer, and 631 cough sounds were recorded. Two expert pulmonologists annotated the cough sounds as either dry or wet based on auditory judgment. Four acoustic features—voiced/unvoiced flag, band aperiodicity index, continuous logarithmic fundamental frequency, and mel-cepstrum—were extracted from the audio waveforms. Image maps of cough sounds were created for convolutional neural network training and classification. The pulmonologists demonstrated a 90.81% (κ = 0.679) agreement rate. Coughs were classified into three classes: 58 unclassifiable (disagreed annotation), 494 dry, and 79 wet coughs. Overall, the system demonstrated high discriminative performance across the five folds, with mean AUCs of 0.985 ± 0.016 for dry cough, 0.976 ± 0.040 for unclassifiable cough, and 0.993 ± 0.006 for wet cough. Based on the annotations of two expert pulmonologists, the developed system demonstrated high discriminative performance for distinguishing dry cough from wet and unclassifiable coughs in an internal validation setting.