This prospective study investigated the reliability of identifying crackles and wheezes in breath sounds using both physicians and artificial intelligence (AI). 11,532 breath sound files were labeled by five physicians and six AI models. Doubtful labels were re-evaluated by two additional physicians. While both physicians and AI showed good sensitivity and specificity for wheezing, crackles demonstrated good sensitivity but poor specificity, indicating their unreliability for medical decision-making. Further investigation into the challenges of crackle identification is warranted.
