Systematic Evaluation of Handcrafted Features and Classical Machine Learning for Respiratory Sound Analysis

The classification of respiratory diseases is an important problem that most researchers have tried to solve directly by using deep learning. Traditional machine learning with handcrafted features has been left behind and is less explored in this context, although it may offer efficiency and interpretability. In this paper, we performed a comparison between multiple machine learning algorithms. We applied the algorithms on respiratory sound data from the ICBHI Dataset. We extracted various features from the data, features that are more suitable for signals. We used the features both individually and together. The task was a classification one, both binary and multiclass. We ran each algorithm separately with the feature sets. For each algorithm, we performed more than 1200 runs using different parameters to optimize their learning and overall performance. Random Forest performed best, showing very promising results with accuracy and a F1 score of almost 80%, closely followed by k Nearest Neighbors. Other algorithms stood out only with certain features, while others returned suboptimal results. This experiment showed that a good choice of features, preprocessing and hyperpa rameter