Adventitious Pulmonary Sound Detection: Leveraging SHAP Explanations and Gradient Boosting Insights

Pulmonary illnesses are annually reported as highly prevalent. Patient outcomes can, however, be improved through the aid of automated processes for early diagnoses. This study aims to develop an automated method with a comprehensive explanation for diagnosing underlying adventitious sounds in respiratory diseases. For this purpose, stethoscope audio signal recordings are initially segmented to merely include either normal or adventitious sounds such as wheezing and crackling. A comprehensive feature set representing the temporal and spectral dynamics in the respiratory events is extracted to identify the respiratory condition. An extreme gradient-based boosting (XGBoost) model is used and evaluated on the ICBHI 2017 dataset recordings based on a five-fold cross-validation approach. To further increase the model explainability, shapley values are calculated and analyzed. Our predictive method suggests specificity, sensitivity, and ICBHI scores of 94.57%, 77.96%, and 86.27%, respectively, demonstrating superior results, outperforming the state-of-the-art techniques. It is also concluded that Mel-frequency cepstral coefficients (MFCC), spectral centroid, zero crossing rate, and signal intensity are the most consistent discriminating features within the adventitious sounds. Clinical relevance- This work contributes to the development of advanced smart digital stethoscopes and respiratory monitoring systems that can be used in clinical, telemedicine and personalized healthcare settings for early detection of breathing disorders or pulmonary conditions.