The aim of this study is to propose a new feature selection method based on the class-based contribution of Shapley values. For this purpose, a clinical decision support system was developed to assist doctors in their diagnosis of lung diseases from lung sounds. The developed systems, which are based on the Decision Tree Algorithm (DTA), create a classification for five different cases: healthy and disease (URTI, COPD, Pneumonia, and Bronchiolitis) states. The most important reason for using a Decision Tree Classifier instead of other high-performance classifiers such as CNN and RNN is that the class contributions of Shapley values can be seen with this classifier. The systems developed consist of either a single DTA classifier or five parallel DTA classifiers each of which is optimized to make a binary classification such as healthy vs. others, COPD vs. Others, etc. Feature sets based on Power Spectral Density (PSD), Mel Frequency Cepstral Coefficients (MFCC), and statistical characteristics extracted from lung sound recordings were used in these classifications. The results indicate that employing features selected based on the class-based contribution of Shapley values, along with utilizing an ensemble (parallel) system, leads to improved classification performance compared to performances using either raw features alone or traditional use of Shapley values.
Background and ObjectiveAccurate diagnosis of respiratory diseases via auscultation is crucial but often subjective. This study investigates the effectiveness of TabPFN, a novel transformer-based deep learning model designed for tabular data, for the automated diagnosis of respiratory diseases from sound recordings. To our knowledge, this work represents the first evaluation of TabPFN for respiratory sound classification, comparing its performance against widely used machine learning models.MethodsThe study classified seven categories: bronchiectasis, bronchiolitis, chronic obstructive pulmonary disease, healthy, lower respiratory tract infection, pneumonia, and upper respiratory tract infection. Fifty-two distinct features were extracted from raw audio samples, including Mel-frequency cepstral coefficients, zero-crossing rate, root mean square energy, spectral centroid, bandwidth, rolloff, and crackle/wheeze indicators. While traditional classifiers utilized data augmentation and synthetic minority over-sampling technique to address data imbalance, TabPFN was evaluated on the original dataset to leverage its optimization for small-to-moderate-sized tabular data.ResultsTabPFN achieved superior performance compared to all baseline classifiers. The model attained an overall accuracy of 0.9913, a macro-averaged F1-score of 0.9625, and a Matthews Correlation Coefficient of 0.9708. The results indicate that the model effectively captures complex, nonlinear relationships in high-dimensional tabular data without the need for extensive data augmentation.ConclusionsTabPFN demonstrates significant potential as a quick and reliable tool for automated respiratory sound analysis. Its high data efficiency and robust performance suggest it is a promising candidate for clinical decision support systems, particularly in scenarios where data sources are limited.