Validation of a Method for Automatic Detection of Lung Sounds in Fibrotic Interstitial Lung Disease

BACKGROUND Delayed diagnosis of fibrotic interstitial lung disease (ILD) is characterized by significant repercussions for management and survival. A timely assessment of “velcro-type” crackles at lung auscultation might prompt a proper diagnostic process in these patients. The accuracy of objective, computerized methods in detecting ILD from respiratory sounds has not been systematically assessed in a clinical setting. We aimed to validate automatic detection of “velcro-type” crackles by comparing the results of lung sounds analysis with chest high resolution computed tomography (HRCT) scans. METHODS Lung sounds were recorded using an electronic stethoscope (Littmann 3200) from anatomically defined sites in 56 subjects (derivation cohort) undergoing a chest HRCT scan in Modena (Italy). Three-hundred anonymized, single-layer HRCT images corresponding to the sites of sound recording were reviewed by two radiologists with expertise in ILD and scored for the signs indicating lung fibrosis. All audio recordings were analyzed by extracting a set of global acoustic features from each file and different classification algorithms were employed on a subset of the best ranked features to classify the data into two groups. The gold standard used to label the ground truth for each sound file was based on the evidence of fibrosis in the corresponding HRCT image. Two ILD physicians were also asked to independently assess the sound files for the presence of “velcro-type” crackles. The results were validated in a further cohort of 59 patients (validation cohort) undergoing chest HRCT in Parma (Italy). Three-hundred nineteen HRCT images and corresponding sound files were obtained. RESULTS In the derivation cohort, accuracy of 74.4% was achieved in automatic detection of lung sounds associated with fibrotic HRCT images. However, although specificity was high (87.3%), sensitivity was lower (48.5%). The two respiratory physicians showed comparable performance, with accuracy 71.6%, specificity 82.3% and sensitivity 49.9%. The results of the automated detection were substantially reproduced in the validation cohort (accuracy 70.2%, specificity 84.7%, sensitivity 45.8%). CONCLUSION An automated method can identify lung sounds associated with pulmonary fibrosis at HRCT, and replicates the performance of experienced clinicians on the same data set. This suggests that there are substantial grounds for developing an automated method for clinical detection of fibrotic ILD.