A cough sound-based deep learning algorithm for accessible prompt detection of chronic obstructive pulmonary disease with smartphones.

Early COPD diagnosis is vital for effective management, yet conventional tools such as professional spirometers are often inaccessible in resource-limited settings. We present Cough Search, a smartphone-based deep learning algorithm that uses voluntary cough sounds to detect COPD, offering a cost-efficient and accessible diagnostic approach. The presented COPD detection algorithm (Cough Search) employs a transformer-based neural network model. It was trained on a training cohort (406 COPD and 1631 non-COPD) with hyperparameters tuned on the balanced internal validation cohort (151 COPD and 225 non-COPD participants). The algorithm was finally validated on the external validation cohort (105 COPD and 617 non-COPD participants from four hospitals). Participants were classified as COPD or non-COPD based on spirometry and clinical diagnoses. Cough Search achieved an area under the receiver operating characteristic curve (AUC) of 0.92 and 0.94 in the internal and external validation cohorts, respectively. In the external validation cohort study, the model demonstrated high sensitivity (92%) and specificity (86%) in distinguishing COPD from non-COPD cases. Performance remained robust across all COPD stages, with a sensitivity exceeding 93% for severe stages (GOLD 3-4) and above 91% for moderate stages (GOLD 1-2). The algorithm maintained its accuracy across non-COPD respiratory conditions and smartphone models. Cough Search shows promise as a scalable, accessible tool for COPD detection, particularly in underserved areas, potentially transforming early COPD diagnosis and management. Trial registration: ClinicalTrials.gov Identifier: NCT06082791.