Diagnostic accuracy of asthma severity grading using machine learning features and lung sounds

Introduction: Recent medical technological advances in acoustics, digital sound analysis and classification give new possibilities of increase in accuracy of diagnosis of respiratory diseases like COPD. Aim and Objectives: We aimed to determine the accuracy of diagnostic classification of COPD severity using machine learning features and auscultated lung sounds. Methods: We auscultated lung sounds of 53 COPD patients bilaterally from 11 sites on anterior, posterior and lateral thoracic sites, including trachea using modified microphone based stethoscope. In addition, these lung sounds were analyzed electronically using a spectrogram. Accuracy was assessed by measuring the area under the curve (AUC value) of the machine learning ROC curve. Results: The accuracy of diagnostic classification between moderate COPD Vs severe COPD using Spectrogram features alone was [0.74(0.65-0.82)], using type of clinical lung sound feature alone was [0.61(0.57-0.70)] and highest when combined [0.79(0.69-0.81)]. The accuracy of diagnostic classification between moderate COPD Vs very severe COPD using Spectrogram features alone was [0.77(0.65-0.85)], using type of clinical lung sound feature alone it was [0.51(0.50-0.53)] and highest when combined [0.88(0.80-0.92)]. The accuracy of diagnostic classification between severe COPD Vs very severe COPD using Spectrogram features alone it was [0.88(0.70-0.94)], using type of clinical lung sound feature alone it was [0.50 (0.50-0.50)] and highest when combined [0.88(0.72-0.88)]. Conclusion: Spectrogram features can be used to improve the accuracy of diagnosis and disease severity grading among patients with COPD when added to traditional clinical lung sound identification.