Automatic identification of respiratory diseases from stethoscopic lung sound signals using ensemble classifiers

Abstract

This paper investigates the application of different homogeneous ensemble learning methods to perform multi-class classification of respiratory diseases. The case sample involved a total of 215 subjects and consisted of 308 clinically acquired lung sound recordings and 1176 recordings obtained from the ICBHI Challenge database. These recordings corresponded to a wide range of conditions including healthy, asthma, pneumonia, heart failure, bronchiectasis or bronchitis, and chronic obstructive pulmonary disease. Feature representation of the lung sound signals was based on Shannon entropy, logarithmic energy entropy, and spectrogram-based spectral entropy. Decision trees and discriminant classifiers were employed as base learners to build bootstrap aggregation and adaptive boosting ensembles. The optimal structure of the investigated ensemble models was identified through Bayesian hyperparameter optimization and was then compared to typical classifiers in literature. Experimental results showed that boosted decision trees provided the best overall accuracy, sensitivity, specificity, F1-score, and Cohen's kappa coefficient of 98.27%, 95.28%, 98.9%, 93.61%, and 92.28%, respectively. Among the baseline methods, SVM provided the best yet a slightly poorer performance, as demonstrated by its average accuracy (98.20%), sensitivity (91.5%), and specificity (98.55%). Despite their simplicity, the investigated ensemble classification methods exhibited a promising performance for detecting a wide range of respiratory disease conditions. The data fusion approach provides a promising insight into an alternative and more suitable solution to reduce the effect of imbalanced data for clinical applications in general and respiratory sound analysis studies in specific.

Introduction

According to a recent report published by the Forum of International Respiratory Societies (FIRS), respiratory diseases (RDs) are among the leading causes of severe illness worldwide, with a death toll exceeding 4 million lives annually [1]. In 2017, the World Health Organization (WHO) declared that chronic RDs accounted for more than 10% of the global disease burden, second only to cardiovascular diseases [2].

The diagnostic process of RD involves auscultation, which is a clinical examination that involves listening to the internal sounds of moving air inside and outside the lungs. Pulmonary sounds are commonly auscultated through the anterior and posterior chest walls or the trachea using a stethoscope [3]. During auscultation, a medical practitioner examines a patient to identify adventitious (atypical) lung sounds superimposing regular breathing patterns. Examples of commonly heard adventitious lung sounds (ALS) include coarse or fine crackles, pleural rubs, wheezes, and stridors [3], [4]. Many RDs causing obstructed or restricted respiratory pathways are characterized by the existence of ALS while breathing. These sound types can be distinctly identified based on their characteristic frequency, pitch, intensity, and energy. For example, both wheezes and stridors are continuous high-pitched sounds occurring at a frequency range of 400 and 500 Hz, respectively. Lower pitched wheeze sounds are also known as rhonchi. High-pitched wheeze sounds may be present as a result of inflamed or narrowed bronchial tubules and are thus an indication of asthma or chronic obstructive pulmonary disease [5]. Stridors usually occur due to tracheal or laryngeal edema [6]. On the other hand, crackles are discontinuous high-pitched (fine) or low-pitched (coarse) waves associated with pneumonia, bronchitis, or heart failure conditions [7]. Similarly, pleural rubs are low-pitched rhythmic sounds associated with inflamed lung lining due to pleural effusion.

Regardless of the type of stethoscopic system used, auscultation is considered one of the safest, easiest, and cheapest examination procedures. Moreover, it provides a patient-friendly and non-invasive solution to monitoring the function of the lungs and other respiratory organs [4]. These qualities are of great value in resource-constrained primary care settings where advanced diagnostic tools are technologies, such as spirometry and radiography, are inaccessible. However, despite being routinely used in healthcare settings, it is a consensus among clinicians that standard pulmonary auscultation has some notable limitations. Firstly, acquiring good auscultatory skills requires extensive training and expertise. Moreover, efficient detection of adventitious sounds is sensitive to the level of experience and auditory acuity of the healthcare professional. Even if auscultation is performed by an expert practitioner, abnormal patterns are sometimes overlooked or misinterpreted during the examination [8]. Thus, subjectivity and inter-variability in observations and interpretations may limit the diagnostic effectiveness of such an approach. These challenges have given rise to the significance of computer-aided auscultation systems that can perform automated identification of ALS and RDs.

Recently, researchers have proposed various artificial intelligence solutions for the identification of adventitious lung sounds. Generally, proposed approaches were based on feature extraction paired with different classification models. At the feature extraction stage, breathing sounds were commonly characterized using several signal processing techniques, including higher-order statistics [9], spectrograms or scalograms [10], [11], wavelet transform coefficients [12], [13], Hilbert–Huang transform [14], and mel-frequency cepstral coefficients (MFCC) [15]. These feature extraction methods have been utilized in conjunction with the standard machine or deep learning methods such as naive Bayes classifiers [9], k-nearest neighbors [14], support vector machines [16], artificial neural networks (ANN) [13], convolutional neural networks (CNN) [17], and recurrent neural networks (RNN) [18]. Generally, obtained accuracy results varied between 97% and 70.2% for wheeze [19], [20], [21], [22], 97.5% and 86% for crackle [23], [24], and 99% for normal sound types. In [9], discriminating between normal, fine crackles, coarse crackles, mono-wheezes, and ploy-wheezes sound using a tree-based classifier provided an overall accuracy of 94%.

Despite the various efforts to develop automated adventitious lung sound detection algorithms, their usefulness in identifying RDs is still limited. Recent studies have shown that the presence of abnormal respiratory sounds is not a distinctive characteristic of impaired respiratory functions [3], [25]. For example, atypical respiratory sounds might not reflect impaired breathing patterns, and abnormalities do not always translate into audible sounds. These findings necessitate the need for first-hand computer-aided tools that are capable of identifying RDs directly from lung sound signals regardless of the existence of adventitious sounds. In this regard, a recent subclass of studies in the literature focused on exploring different machines and deep learning techniques to perform binary (normal vs. pathological) [26], [27], ternary (normal vs. chronic vs. non-chronic) [28], or multi-class [28], [29] classification of RDs. Investigated disease conditions included respiratory tract infections, pneumonia, bronchiectasis, bronchiolitis, asthma, and COPD. These studies reported accuracies up to 93.3%, 99%, and 98% for binary, ternary, and multi-class classification, respectively.

It is worth noting that most of the achieved satisfactory results in the context of multi-class classification were based on hybrid deep learning approaches. Despite exhibiting a highly promising performance without the need to incorporate sophisticated feature engineering techniques, training a reliable deep network architecture can be time-consuming and requires significant computational resources. Besides, the training process is iterative, and it involves multiple model parameters and enormous datasets. In clinical contexts, the lack of sufficient high-quality, diverse, and annotated training data is considered among the main limitations. To address the issue of data available, previous studies have employed several data augmentation techniques to over-sample minority classes such as variational autoencoder, adaptive synthetic sampling, and synthetic minority oversampling. However, minority oversampling techniques can introduce data leakage during the validation process, and thus, obtained results might be biased towards high fake accuracies. In fact, the majority of data augmentation approaches are mainly employed to improve the classification performance, without addressing the imperative requirement of using fine-grained datasets that represent the studied population.

In this study, we carried out a multi-class RD classification task considering six different conditions, namely normal, asthma, pneumonia, heart failure, bronchiectasis and bronchitis (BRON disorders), and chronic obstructive pulmonary disease (COPD). To this end, a novel stethoscopic lung sound dataset was collected locally at King Abdullah University Hospital, Jordan University of Science and Technology, Irbid, Jordan. This dataset was complemented by the publicly available ICBHI Challenge database to obtain a more balanced distribution among the respiratory disease classes. In terms of validity in the context of clinical applications, we believe that this data fusion approach provides a better alternative to data augmentation techniques employed in the literature. We propose to tackle the classification problem through a simple yet effective framework utilizing entropy features along with homogeneous ensemble classification methods. Subsequently, a comparative investigation is carried out to compare the proposed ensemble models to several baseline machine learning classifiers that were repeatedly employed in previous works.

The rest of this paper is organized as follows. Section 2 describes the methods used to acquire the lung sound signals along with the mathematical formulation of the entropy features. A brief overview touching on the mathematical groundwork and the implementation of the classification models is also provided. Section 3 presents the experimental results, and Section 4 provides a discussion of these results. Finally, Section 5 concludes this paper.