Deep learning diagnostic and severity-stratification for interstitial lung diseases and chronic obstructive pulmonary disease in digital lung auscultations and ultrasonography: clinical protocol for an observational case–control study

Interstitial lung disease (ILD) encompasses several pulmonary conditions defined by an alteration of the pulmonary interstitium, a restrictive pattern of lung function, and fibrotic scarring on chest computed tomography (CT). Approximately one-third of these disorders have known endogenous or exogenous causes, including environmental or occupational factors, infections, drugs and radiation. Two-thirds are of idiopathic aetiology [1] and comprise a range of subcategories, the most common of which is idiopathic interstitial pneumonia (IIP). In turn, IIP comprises a range of sub-pathologies, such as idiopathic pulmonary fibrosis (IPF) and non-specific interstitial pneumonia (NSIP) [2]. Identifying patients with IIP at the earliest possible stage is essential for care management as treatment is aimed at slowing the irreversibly debilitating and ultimately fatal progression. Delay in specialist referral is associated with a higher mortality, irrespective of disease severity [3]. With a mean delay of 2.2 years between the onset of symptoms and specialist referral, the investigation of competing diagnoses by non-specialist providers can be costly for both patients and healthcare providers [3]. However, given the initial non-specific symptomatic presentation, the need for advanced diagnostic tools, such as high-resolution chest CT (HRCT), and an expertise in the early-stage diagnosis of IPF and NSIP remain desirable and achievable objectives [4]. Distinguishing between IPF and NSIP raises considerable diagnostic challenges as their clinical presentations share many overlapping features. However, the distinction is useful as their response to treatment differs markedly [5]. Until now, with limited treatment options benefiting mostly patients in the early stages of the disease [6, 7], many patients will progress towards disability or death [8]. Despite research and advances in therapy, ILDs remain a worldwide health challenge affecting millions of people each year [9], emphasizing the need to make progress in diagnosis and prevention.

Different measures have been proposed to improve the early diagnosis of ILDs [10, 11]. In particular, the identification of the so-called “velcro”-like crackles on lung auscultation by primary care doctors has been suggested as an early and strongly predictive sign of IPF or fibrotic NSIP [4, 12]. For instance, while IPF and NSIP typically have fine velcro-like crackles audible on the mid-to-late inspiratory cycle, chronic obstructive pulmonary disease (COPD) tends to have coarse crackles occurring during the early inspiratory cycle [13]. As stethoscopes are readily available, inexpensive and non-invasive, they constitute an adequate tool to detect velcro-like crackles in the early stages of IPF or fibrotic NSIP to shorten the diagnostic delay and allow the prompt referral to specialised care. However, conventional auscultation is a highly subjective skill limited by inter-listener variability and human perceptual ability to distinguish between lung sounds and their temporal occurrence in the respiratory cycle. Inherent heterogeneity in stethoscope quality, background noise and patient-related factors, such as obesity or chest deformities, are other limiting factors. To overcome these drawbacks, research efforts have been devoted to improve computerised respiratory sound recording with electronic stethoscopes and an objective analysis based on advanced digital acoustic signal processing [14‐17]. The advent of deep learning in recent years took the analysis of auscultation signals one step further by allowing an enhanced detection of abnormal lung sounds in patients with respiratory diseases [18‐20].

Several studies have assessed the broad adoption and impact of deep learning to help diagnose COPD [21‐23], the third leading cause of death worldwide [24]. The vast majority developed predictive models to cover a wide range of objectives, the main ones being the diagnosis and severity classification of the disease [25, 26]. A 2022 review of artificial intelligence (AI) techniques in COPD yielded 156 articles relevant to the application of AI in COPD research, including 56 concerning diagnosis, 65 on its prognosis, 54 on COPD severity classification, and 17 on the management of the disease [27]. Most studies have used a variety of features, including patient physiological characteristics, comorbidities, symptoms, vital signs, biomarkers, genomic information, pulmonary function tests, CT images, hospitalization information, and/or breath sounds [28, 29]. Regardless of the method(s) chosen, COPD remains an incurable and progressive disease and diagnosis at the early risk stage is important. In this sense, the work of Altan et al. is innovative. The deep learning algorithms they used on analysing multiple lung auscultation points for the early diagnosis of COPD achieved high classification performance rates [30, 31]. Achieving this with a method as conventional as lung auscultation can reduce the need for additional, more extensive, time-consuming, expensive or invasive diagnostic tests.

Conversely, research regarding IPF and NSIP are scarce and have focused mostly on datasets collected through radiological [32‐34], genomic [35, 36] or functional tests [37]. Pancaldi et al. described the use of an AI algorithm to detect the presence of velcro-like crackles in patients with rheumatoid arthritis and a suspicion of ILD [17, 38]. However, to our knowledge, no study has investigated the benefit of deep learning-aided diagnostic tools for early IPF and NSIP diagnosis using respiratory sound analysis in adults. This might allow doctors to assess acoustic signatures more objectively and thus allow a more standardised and potentially earlier diagnosis in patients presenting at primary care clinics with non-specific, chronic respiratory symptoms. On the other hand, lung ultrasound (LUS) is already the standard of care for detecting consolidations, diagnosing pneumonia and guiding pleural taps. The distinction between A (normal aeration) and B (alveolar-interstitial syndrome) lines on LUS is clinically important and forms the backbone of multiple clinical decision trees for real-time respiratory diagnoses and treatment choices [39]. As such, not only is LUS a relevant gold standard for lung pathology, but it could also benefit from automation by deep learning.

We developed a series of deep learning algorithms on digital lung auscultation (DeepBreath) and LUS to detect a range of physiological and pathological lung diseases, including (COVID-19) [40]. This study will seek to explore the synergistic value of several point-of-care-tests for the AI-aided detection and differential diagnosis of ILD and COPD, as well as estimate of severity, with the aim to better guide and improve care management in adults.

The following data will be reviewed by external experts and interpreted using standardised report forms noting the binary presence/absence of several anomalies as well as a text field for other notable observations: routine chest X-ray films (usually 2 incidences, posterior-anterior and lateral); lung auscultation audio clips (10 anatomic regions represented, 30 s recordings of each region); and LUS images (10 anatomic regions represented, 5 s video clips of each region). The analysis will be blinded and the assessor will not have any knowledge of the linked clinical data or association between the various imaging modalities (i.e., IDs are scrambled between media sources and chest X-ray images will be reviewed blinded to the patient’s LUS images, auscultation audio, clinical data, etc.). These blinding and standardisation procedures are expected to minimise performer and study management bias respectively.

For all subjects, spirometry, body-plethysmographic parameters (see details above) and lung diffusion capacity for carbon monoxide (TLco/Kco) will be measured. Participants’ lung images recorded from previous HRCT (or X-rays) during past routine visits will be used. No chest CT scans will be performed as part of this study; only lung images of participants previously recorded as part of their regular follow-up or to be performed in this context will be used. The presence on the chest CT scan of honeycombing, traction bronchiectasis, reticulations, ground-glass opacities, and emphysema will be measured for patients with IPF, NSIP or emphysema. The main chest CT scan features of IPF are reported to be basal and peripheral reticulations, traction bronchiectasis, minimal ground glass opacities, and moderate or extensive honeycombing. For NSIP, the CT scan typically demonstrates bilateral lung involvement and invariably some extent of ground-glass opacities, mainly in the lower zones with fibrotic changes, while honeycombing is not a common feature [55]. The presence on the chest CT scan of structural abnormality, such as ≥ 5% emphysema and/or ≥ 15% gas-trapping and/or airway wall thickness ≥ 2.5 mm [56], will be measured for patients with COPD. The protocol assumes normal lung parenchyma in the control group, which will not be exposed to radiation unless controls have already undergone a recent (i.e., < 5 years) chest CT scan for other reasons. This will be taken into consideration as we will investigate the association of lung sounds with radiological characteristics.

Demographics including age, sex, ethnicity, environment (smoking status, long-term exposure to occupational or environmental agents, etc.), treatments, presence of chronic respiratory symptoms or repeated lower respiratory tract infectious diseases, as well as a diagnosis of other comorbidities (obesity, immunodeficiency, alpha-1 antitrypsin deficiency, etc.) will be reported in a questionnaire (Additional file 1). Severity of functional limitations according to the New York Heart Association (NYHA) functional classification [57] will be also reported if available.

The impact of IPF and NSIP on subjects’ health-related quality of life will be measured with the standardised K-BILD questionnaire [47] (use with license agreement), which covers 15 questions exploring 3 health dimension scores (psychological, breathlessness and activities, and chest symptoms) using a 7-point Likert response scale (scores range from 0 to 100, a higher score indicating better health status). The impact of COPD will be assessed with the CAT [48] (use with license agreement) that measures eight items: cough; phlegm; chest tightness; breathlessness; limited activities; confidence leaving home; sleeplessness; and energy. Scores range from 0 to 40, with a higher score indicating a more severe impact of COPD on a patient’s life. The SF-36 will also be used to assess the impact of IPF, NSIP or COPD on patients’ quality of life. The investigators will double-check on-site that the questionnaires are fully and accurately completed. Data collection will be carried out using the online REDCap database (REDCap, Vanderbilt University, Nashville, TN, USA; https://​www.​project-redcap.​org/​resources/​citations/​). Conditions or complaints occurring after enrolment will not be considered in the statistical analyses. Current symptoms at enrolment will be registered.

Untreated IPF has the worst prognosis among the different forms of ILD, with median survival ranging from 3 to 5 years from diagnosis [60]. Recent studies suggest that if novel anti-fibrotic medications (pirfenidone and nintedanib) are started early, they can slow the rate of lung function decline and prevent IPF exacerbation, thus reducing mortality [6, 7]. Unfortunately, because of the unspecific nature of the symptoms, the early stage of IPF remains underdiagnosed and many patients will progress to advanced disease and may require lung transplantation [8]. By contrast, the prognosis of NSIP is generally better than that of IPF, with a median survival time of more than 9 years. Systemic steroids and immunosuppressive therapy may be attempted to slow or reverse the course of the disease, but non-responsive individuals may also be considered for lung transplantation [8]. When left untreated, NSIP tends to progress toward fibrotic changes and persistent debilitating symptoms. COPD is also a leading cause of disability worldwide. Patients are generally unaware of their condition for years, leading to a significant delay in diagnosis, the application of preventive measures such as a smoking cessation intervention, and potential treatment [61]. Being able to recognise and diagnose these lung diseases earlier is of the utmost clinical importance.

This study will aim to collect a standardised dataset of digital lung auscultations and derive a deep leaning model able to detect the acoustic and sonographic signatures of the presence and severity of IPF, NSIP and COPD. Recent advances in deep learning are promising to support doctors in standardising the detection and interpretation of complex patterns in pulmonary diseases and AI has proven to outperform doctors in discriminating respiratory pathologies via respiratory functional explorations [62], symptoms [63, 64] and/or radiological examinations [34]. To overcome the subjectivity of human auscultation and the discrepancy in auscultation ability between doctors [16], the development of AI algorithms for the analysis of respiratory acoustic signals has been proposed [19, 20]. In order to meet this requirement, many attempts have been made to develop and apply neural networks to automate the detection and classification of various disease-related breath sounds using machine learning and deep learning-based analysis [14, 64‐66]. In particular, recent literature reviews have summarized advances in the implementation of respiratory sound-based AI algorithms in the screening, diagnosis, and classification of COPD [26, 65]. Conversely, the current state of knowledge on the computerized analysis of breath sounds in patients with ILD using AI techniques has not been assessed. Table 2 summarizes the published studies most similar to our research.

However, there are some notable differences in these studies, which justify the present work. The main ones are the frequent absence of healthy subjects as a control group and the almost unanimous lack of severity classification or joint use of LUS images. As stated by Charleston-Villalobos et al. [68], a comparison with other attempts to diagnose and classify lung sounds is difficult due to the difference in data acquisition, type of classification scheme, lack of gold standards allowing standardization between studies, and their distinct exploratory nature. In particular, a major flaw of most anterior studies aimed at building deep learning models for diagnostic classification from digital lung sounds is the use of publicly available databases, such as the R.A.L.E repository [82] or the International Conference on Biomedical Health Informatics [83]. These databases have inherent acquisition flaws due to heterogeneity in data collection and methods that create systematic biases between the predicted labels on which new algorithms are built. This is then reflected in the results of studies with an exaggerated excellent predictive performance that prevents their evaluation and comparison with each other. On the contrary, in our study, sounds will come from a cohort of patients under standardised and homogeneous recording conditions. It remains to be determined whether an AI algorithm using respiratory sounds and/or LUS analysis can be used as an initial and accurate diagnostic tool for patients with ILDs or COPD. The diagnosis of IPF, NSIP and COPD in early stages may allow practitioners to appropriately recognise exacerbations of a chronic lung disease, whereas patients may initially be diagnosed as having multiple bouts of acute disease (e.g., bronchitis) without this defined diagnosis [61]. Early diagnosis with AI may therefore allow patients to benefit from prevention measures and the allocation of appropriate treatments aiming to reduce the progression to permanent lung damage and improve the overall prognosis in patients presenting at primary care clinics for non-specific chronic respiratory symptoms. As research in this area is scarce, it is anticipated that the results generated from this study will be of great importance and may be sufficient to change and improve pulmonary primary care practice in a vulnerable population by proposing a faster diagnosis.