(C) PLOS One
This story was originally published by PLOS One and is unaltered.
. . . . . . . . . .



Real-time counting of wheezing events from lung sounds using deep learning algorithms: Implications for disease prediction and early intervention [1]

['Sunghoon Im', 'Department Of Mechanical Engineering', 'Ajou University', 'Suwon-Si', 'Gyeonggi-Do', 'Republic Of Korea', 'Taewi Kim', 'Choongki Min', 'Waycen', 'Inc.']

Date: 2024-02

Abstract This pioneering study aims to revolutionize self-symptom management and telemedicine-based remote monitoring through the development of a real-time wheeze counting algorithm. Leveraging a novel approach that includes the detailed labeling of one breathing cycle into three types: break, normal, and wheeze, this study not only identifies abnormal sounds within each breath but also captures comprehensive data on their location, duration, and relationships within entire respiratory cycles, including atypical patterns. This innovative strategy is based on a combination of a one-dimensional convolutional neural network (1D-CNN) and a long short-term memory (LSTM) network model, enabling real-time analysis of respiratory sounds. Notably, it stands out for its capacity to handle continuous data, distinguishing it from conventional lung sound classification algorithms. The study utilizes a substantial dataset consisting of 535 respiration cycles from diverse sources, including the Child Sim Lung Sound Simulator, the EMTprep Open-Source Database, Clinical Patient Records, and the ICBHI 2017 Challenge Database. Achieving a classification accuracy of 90%, the exceptional result metrics encompass the identification of each breath cycle and simultaneous detection of the abnormal sound, enabling the real-time wheeze counting of all respirations. This innovative wheeze counter holds the promise of revolutionizing research on predicting lung diseases based on long-term breathing patterns and offers applicability in clinical and non-clinical settings for on-the-go detection and remote intervention of exacerbated respiratory symptoms.

Citation: Im S, Kim T, Min C, Kang S, Roh Y, Kim C, et al. (2023) Real-time counting of wheezing events from lung sounds using deep learning algorithms: Implications for disease prediction and early intervention. PLoS ONE 18(11): e0294447. https://doi.org/10.1371/journal.pone.0294447 Editor: Mohammad Amin Fraiwan, Jordan University of Science and Technology, JORDAN Received: May 17, 2023; Accepted: October 23, 2023; Published: November 20, 2023 Copyright: © 2023 Im et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability: Codes and utilized data are available in our open repository (https://github.com/sunghoon-most/Wheeze_Counter). Funding: This work was supported by the Korea Environment Industry & Technology Institute (KEITI) through the Environmental Health Digital Program funded by the Korea Ministry of Environment (MOE) (2021003330010, 2021003330009) The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Competing interests: The authors have declared that no competing interests exist.

Introduction Lung diseases are a major cause of global morbidity and mortality, including asthma, COPD, lung infections like pneumonia, lung cancer, bronchitis, and other breathing problems [1, 2]. Lung sounds can be indicative of most lung and respiratory diseases [3]. When there is no respiratory disorder, normal breathing sounds are heard, whereas abnormal breathing sounds such as wheezing or crackling are detected when there is a lung disease [4, 5]. For this reason, regular or routine monitoring of breathing sounds is essential for symptom prevention and alleviation, as well as for the early detection of various respiratory diseases [6, 7]. Typically, respiratory abnormalities are diagnosed by spirometry and auscultation [8]. While spirometry is impossible for certain groups, such as children, and is difficult to use practically to monitor a long-term pattern of patient condition in non-clinical settings [9, 10], auscultation is non-invasive, inexpensive, and easy to use [11, 12]. Medical professionals listen to these sounds to evaluate and diagnose patients [13]; however, conventional auscultation requires considerable training and expertise, and its quality depends on the doctor’s experience and hearing [14]. The misunderstanding of breathing sounds and making incorrect diagnoses is not rare among medical students [15, 16]. To overcome the limitation of conventional auscultation, various methods such as neural networks [17], classifiers [18, 19], and NMF [20] are suggested in many cases in order to assist in the automatic detection and classification of adventitious lung sounds [21]. Among them, deep learning algorithms train the machine to automatically learn the characteristics of the signals or waveforms of lung sounds to recognize abnormal lung or breathing sounds (wheezing, crackling) [22]. The most common deep learning algorithm used for lung sound classification is a convolutional neural network (CNN) [11, 15, 23, 24] or recurrent neural network (RNN) model [25, 26] that extracts breathing sound features from a two-dimensional spectrogram image, or a combination of the two, a convolutional-recurrent neural network (CRNN) [22, 27]. The accuracy of the models ranges from 63% [11] to 99% [28], and in general, the CNN-based model has the highest accuracy [5]. Incorporating AI-based lung sound analysis into automated diagnosis systems has been suggested to determine the degree of airway inflammation [29] or the risk of a number of lung diseases [30]. Recently, efforts have been made to collect breathing sounds from smartphones or real-time lung sounds from wearable devices to develop automated AI-based solutions for lung sound analysis and classification [31–34]. Through this technological advancement, abnormal respiratory and asthmatic symptoms could be detected or diagnosed at an early stage via real-time self-monitoring or telemedicine [35, 36]. However, most existing models focus on the automatic diagnosis of single recorded data, and applications to real-time monitoring data are still limited [21, 37]. They tended to be developed based on the learning data collected by auscultation for a short period of 10 to 70 s and labeled by clinicians [38, 39]. Much of the previous work focused on addressing methodological challenges associated with noise cancellation or reduction [40, 41], detection of the breathing section, or binary classification of an individual cycle of respiration [11, 22, 23, 42, 43]. Due to a lack of adaptability for real-time, continuous long-term signals, most lung sound classification algorithms have not been widely implemented in practice, with limited applicability in self-symptom management or telemedicine [2, 44]. Considering that respiratory patterns represent the holistic physical and psychological state of humans, not only the presence of abnormal sounds but also the location, duration, and relationships of a sequence of respiration cycles, including atypical breathing activities, could serve as important reference data for clinicians and patients to diagnose and monitor lung diseases [45]. To provide comprehensive information about the lung’s breathing functionality, which may not be well noticed or recognized in a clinical setting, the pattern and frequency of abnormal lung sounds within a relatively long time must be analyzed rather than most of the existing models for determining the presence or absence of abnormalities at each respiratory unit [46, 47]. The real-time data collection and automated pre-processing system would be critical for long-term monitoring and intervention [48]. We have summarized the relevant papers in a table and included them in the S1 Table. Considering this loophole, in this exploratory study, we have developed a real-time event counting algorithm to identify abnormal breathing sounds, especially wheezing, and record their frequency to determine the pattern over a certain period and present this information in real-time. We utilize a unique method that involves the meticulous categorization of a single breathing cycle into three types: break, normal, and wheeze. The algorithm not only detects abnormal sounds in each breath but also collects extensive data on their location, duration, and connections within the entire respiratory cycle, including unusual patterns. This counting algorithm may improve existing studies that aim to predict lung diseases based on long-term breathing patterns [49–51], going beyond simply classifying respiratory units. In addition, when integrated with wearable devices that are being actively developed, its utility will be maximized [52, 53]. Using three types of labeled lung sound data, we trained a one-dimensional convolutional neural network and a long short-term memory (1D-CNN-LSTM) network model for discriminating three breathing statuses (break, normal, and wheezing) and then developed a “real-time wheezing counter” as a pilot; we suggested the possibility of its application for early diagnosis or the remote treatment of respiratory diseases. Our research demonstrates the potential of AI-based technology for diagnosing and monitoring lung diseases in real-time, offering the prospect of earlier detection and improved treatment outcomes. Existing research gaps include limitations in real-time applications and a focus on short-term data. We address these gaps with a real-time event counting algorithm designed for continuous, long-term signals, emphasizing the pattern and frequency of abnormal lung sounds over time, rather than just detecting their presence or absence at individual respiratory units. This advancement holds promise for enhancing the diagnosis and monitoring of lung diseases.

Discussion Long-term monitoring of lung sounds assembled via a wearable device and AI-based diagnosis without doctor involvement would be essential to developing advanced computerized monitoring that may be used for self-symptom management or remote monitoring such as telemedicine. There have been a few research studies on these issues recently [16, 53]. However, the methods typically classify whether a signal is abnormal or normal or what kind of inadvertent sound it is. Their suggested method just distinguishes the subject’s artificially induced inadvertent lung sound as a real-life applicable demonstration, indicating a practical usage limitation. Otherwise, we present an applicable method for implementing long-term monitoring in clinical settings by counting the number of wheeze occurrences over time. Our method differs from previous research in that it counts both the number of normal breaths and the number of wheezes, that is helpful for monitoring respiratory disease patients in dynamic environments. We utilize a segment-based classification AI model, which is normally used in speech recognition [66] or rare sound detection [67, 68]. To be able to detect not only wheezing events but also the isolated breath cycle, we sliced lung sound signals into segments and utilized the predicted probability of each segment. As a result, the counting algorithm we developed could report the frequency of wheezing during entire clinical lung sounds without any additional information, such as respiration volume. Despite our contributions, several methodological limitations must be addressed. First, due to the limited availability of reference lung sound data collected in good quality, it is not sufficiently verified whether the developed algorithm would function well on the data collected from various patients in diverse recording environments. With useful methods such as assessment of lung sound quality, we would utilize the fine quality of lung sound data, resulting in a more sophisticated and accurate AI model. Furthermore, the algorithm must be enhanced and adjusted based on the clinical trials of long-term lung sound monitoring with a broad patient group in order to assure the validity, reliability, and applicability of preventive treatment in clinical and non-clinical settings. Second, this study does not provide empirical evidence on how sensitive the proposed algorithm is to different types of lung sounds. The frequency of wheezes was the focus of this study because it is known that wheezes might exacerbate asthma and COPD. A recently developed wearable stethoscope including a de-nosing function [53] could enable the widespread practical use of our counting algorithm when more adventitious sounds such as crackling, rhonchi, stridor, and pleural rub are accumulated by the device. In general, automated long-term monitoring via AI-based algorithms could assist preventative medicine by acquiring precursory information from numerous signals and images from the human body for the relevant bad health impacts. As so, long-term monitoring of wheezing occurrences and patterns may shed light on the development of various respiratory illness outcomes if combined with a patient’s clinical records, such as symptom exacerbation and response to treatment. The integration of AI-based algorithms into long-term monitoring could revolutionize preventative medicine. By acquiring precursory information from numerous signals and images from the human body, AI could potentially predict adverse health impacts. In the context of respiratory health, long-term monitoring of wheezing occurrences and patterns, combined with a patient’s clinical records, could provide invaluable insights into the development of various respiratory illness outcomes. This could lead to more personalized treatment plans and improved patient outcomes.

Conclusion This research presents a deep learning-based algorithm for counting wheezes, utilizing a 1D-CNN-LSTM model. The model is trained on a variety of reference lung sound databases to predict the probability of abnormal sounds in each segment. Our algorithm then uses this model to count wheeze instances from recorded lung sounds and validates in real-time lung sound simulation. Our wheeze counting method is straightforward yet effective, with potential for expansion into automatic symptom monitoring. This could be crucial in predicting the onset or severity of future abnormalities, as well as detecting current symptoms. Given the possible link between wheeze occurrence trends and symptom exacerbations, our approach could aid in preventing urgent emergencies like asthma attacks. Unlike traditional lung sound classification algorithms, our method can handle continuous data. With a detection accuracy of 90%, the results include identifying the number of total breath cycles and the proportion of abnormal sounds, along with real-time counting and visualization of these events throughout whole respiration. This could revolutionize research on predicting lung diseases based on long-term breathing patterns and offers utility in both clinical and non-clinical settings for immediate detection and remote intervention of worsened respiratory symptoms. Moreover, our counting algorithm can easily adapt to other bio-signals. For instance, when used with ECG (Electrocardiogram) or EMG (Electromyography) signals, it could automatically detect the intensity of heart or muscle anomaly patterns. In conclusion, our study introduces a novel and effective approach to real-time wheeze detection and counting, which has significant potential for improving self-symptom management and telemedicine-based remote monitoring. This innovative wheeze counter, with its high detection accuracy and ability to handle continuous data, could play a crucial role in predicting lung diseases based on long-term breathing patterns. Furthermore, its adaptability to other bio-signals suggests a wide range of potential applications in both clinical and non-clinical settings. Future research should focus on further refining the algorithm and exploring its potential in various healthcare contexts.

[END]
---
[1] Url: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0294447

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/

You are viewing proxied material from magical.fish. The copyright of proxied material belongs to its original authors. Any comments or complaints in relation to proxied material should be directed to the original authors of the content concerned. Please see the disclaimer for more details.