Artificial Intelligence, Lung; Disease, Interstitial, With Fibrosis, Pulmonary Disease, Chronic Obstructive
Conditions
Keywords
Lung; Disease, Interstitial, with Fibrosis, Idiopathic Pulmonary Fibrosis, Idiopathic Interstitial Pneumonias, Pulmonary Disease, Chronic Obstructive, Artificial Intelligence, Deep Learning, Respiratory Sounds; Auscultation, Ultrasonography, Diagnostic imaging
Brief summary
Idiopathic pulmonary fibrosis (IPF), non-specific interstitial pneumonia (NSIP), and chronic obstructive pulmonary disease (COPD) are severe, progressive, irreversibly incapacitating pulmonary disorders with modest response to therapeutic interventions and poor prognosis. Prompt and accurate diagnosis is important to enable patients to receive appropriate care at the earliest possible stage to delay disease progression and prolong survival. Artificial intelligence (AI)-assisted digital lung auscultation could constitute an alternative to conventional subjective operator-related auscultation to accurately and earlier diagnose these diseases. Moreover, lung ultrasound (LUS), a relevant gold standard for lung pathology, could also benefit from automation by deep learning.
Detailed description
Aim: To develop and determine the predictive power of an AI (deep learning) algorithm in identifying the acoustic and LUS signatures of IPF, NSIP and COPD in an adult population and discriminating them from age-matched, never smoker, control subjects with normal lung function. Methodology: A single-center, prospective, population-based case-control study that will be carried out in subjects with IPF, NSIP and COPD. A total of 120 consecutive patients aged ≥ 18 years and meeting IPF, NSIP or COPD international criteria, and 40 age-matched controls, will be recruited in a Swiss pulmonology outpatient clinic with a total of approximately 7000 specialized consultations per year, starting from August 2022. At inclusion, demographic and clinical data will be collected. Additionally, lung auscultation will be recorded with a digital stethoscope and LUS performed. A deep learning algorithm (DeepBreath) using various deep learning networks with aggregation strategies will be trained on these audio recordings and lung images to derive an automated prediction of diagnostic (i.e., positive vs negative) and risk stratification categories (mild to severe). Secondary outcomes will be to measures the association of analysed lung sounds with clinical, functional and radiological characteristics of IPF, NSIP and COPD diagnosis. Patients' quality of life will be measured with the standardized dedicated King's Brief Interstitial Lung Disease (K-BILD) and the COPD assessment test (CAT) questionnaires. Expected results: This study seeks to explore the synergistic value of several point-of-care-tests for the detection and differential diagnosis of ILD and COPD as well as estimate severity to better guide care management in adults
Interventions
DEVICELung auscultation
Digital lung auscultation with the Eko core digital stethoscope (Eko Devices, Inc., CA, USA).
DEVICELung ultrasound
Lung ultrasonography
OTHERQuality of Life's questionnaires
Impact of the diseases on subjects' health-related quality of life measured with standardized questionnaires (K-BILD, CAT)
DIAGNOSTIC_TESTPulmonary functional tests
Spirometry, body-plethysmographic parameters and lung diffusion capacity for carbon monoxide will be measured.
Sponsors
University Hospital, Geneva
Swiss Federal Institute of Technology
Hôpital du Valais
Pediatric Clinical Research Platform
Study design
Observational model
CASE_CONTROL
Time perspective
PROSPECTIVE
Eligibility
Sex/Gender
ALL
Age
18 Years to No maximum
Healthy volunteers
No
Inclusion criteria
* Written informed consent * age \> 18 years old. * patients with already-diagnosed IPF (group 1) prior to the consultation (index) date. * patients with already-diagnosed NSIP (group 2) prior to the consultation (index) date. * patients with already-diagnosed COPD (group 3) prior to the consultation (index) date. * Control subjects must be followed-up at the pulmonology outpatient clinic for: 1. obstructive sleep apnoea. 2. occupational lung diseases (miners, chemical workers, etc.). 3. pulmonary nodules (considered benign after 2 years).
Exclusion criteria
* patients who cannot be mobilized for posterior auscultation. * patients known for severe cardiovascular disease with pulmonary repercussion. * patients known for a concurrent, acute, infectious pulmonary disease (e.g., pneumonia, bronchitis). * patients known for asthma. * patients known or suspected of immunodeficiency, alpha-1-antitrypsin deficit, and or under immunotherapy. * patients with physical inability to follow procedures. * patients with inability to give informed consent.
Design outcomes
Primary
| Measure | Time frame | Description |
|---|---|---|
| To differentiate ILD from control subjects based on digital lung sounds recordings and LUS. | During lung auscultation (10 minutes). Each patient will provide 10 recordings of 30 seconds. LUS images and 5 second video clips of each anatomic region (10 regions represented). | To determine the predictive performance of the AI algorithm-evaluated lung auscultation and LUS in the identification and risk stratification of ILD signatures from control subjects described in terms of descriptive statistics, area under the receiver operating characteristic curve, sensitivity, specificity, positive and negative predictive values, and likelihood ratios (95% confidence intervals). Digital lung sounds will be transformed to Mel Frequency Cepstrum Coefficients. Several data augmentation techniques will be explored. The effect of each pre-processing method will be tested. The best performing approach according to sensitivity and specificity will be reported. This dataset will then be fed into a various deep learning networks with aggregation strategies for binary classification into positive vs negative for diagnostic results for: * ILD or control subjects * ILD or COPD * (If ILD+) IPF or NSIP The same prediction will also be made using LUS images. |
| Predictive performance of the DeepBreath algorithm to stratify ILD severity based on human digital lung sounds recordings and LUS (i.e. physiological parameters) compared to grading scales. | During lung auscultation (10 minutes). Each patient will provide 10 recordings of 30 seconds. LUS images and 5 second video clips of each anatomic region (10 regions represented). | To determine the ILD clinical severity predictive performance of the DeepBreath algorithm based on human digital lung sounds recordings and LUS, risk stratification will use multiclass or regression according to grading scales obtained from: * K-BILD and CAT impact of life questionnaire. * Lung function tests (Forced Expiratory Volume in 1 sec, Forced vital capacity, Forced Expiratory Volume in 1 sec/Forced vital capacity, Total lung capacity, functional respiratory capacity, Transfer capacity for carbon monoxide, Alveolar Volume). * High-Resolution Computed Tomography (severity markers that will be used are: traction bronchiectasis, presence of honeycombing, ground glass opacities, reticulation, emphysema. Chest CT-scans will be reviewed independently by two radiologists blinded to each other). |
| Performance of the DeepBreath algorithm to subcategorize ILD by discriminating digital lung sounds recordings and LUS (i.e. physiological parameters). | During lung auscultation (10 minutes). Each patient will provide 10 recordings of 30 seconds. LUS images and 5 second video clips of each anatomic region (10 regions represented). | The performance of the DeepBreath algorithm to determine the subcategories of ILD such as IPF and NSIP based on digital lungs sounds and LUS according to gold standard diagnosis: * IPF follows the Fleischner Society Consensus criteria. * NSIP diagnosis follows the American Thoracic Society classification. |
Secondary
| Measure | Time frame | Description |
|---|---|---|
| To test whether performance of DeepBreath could be improved using clinical features (i.e., signs, respiratory symptoms, demographics, medical history and basic paraclinical tests). | During the data analysis period (i.e., after the 60-minute study intervention period) | To explore the utility of adding clinical data collected at enrolment including demographic information (age and sex), several binary clinical symptoms (respiratory symptoms), medical history and basic paraclinical tests to improve the accuracy of the DeepBreath algorithm in detecting IPF from control subjects or COPD. Clinical data will be explored for their predictive capacity in the above tasks and added to the breath sound analysis either as an Support vector machine or in conditional feature extraction upstream of the neural network. |
| Performance of human expert-identified acoustic signatures. | During the data analysis period (i.e., after the 60-minute study intervention period). | Comparison of the predictive performance of human expert-identified acoustic signatures in the predictive tasks described above in the primary outcomes (Kappa coefficient). |
| CAT | Baseline | COPD assessment test: the CAT health status questionnaire is a 8 item validated, self-completed heath status questionnaire. The total CAT score ranges from 0 to 40 where 0 represents no symptoms and 40 very bad symptoms. This questionnaire will be used to assess the Impact of COPD on subjects' health-related quality of life. It will take about 3 minutes to complete this questionnaire. |
| K-BILD | Baseline | King's brief Interstitial Lung Disease Health Status: the K-BILD health status questionnaire is a 15 item validated, self-completed heath status questionnaire. It has three domains: breathlessness and activities, psychological and chest symptoms. The K-BILD domain and total score ranges are 0-100, with the higher scores corresponding with better health-related quality of life. This questionnaire will be used to assess the Impact of ILD on subjects' health-related quality of life. It will take about 3 minutes to complete this questionnaire. |
| Agreement of human labels with objectively clustered pathological sounds by machine learning. | During the data analysis period (i.e., after the 60-minute study intervention period). | To quantify the agreement of human labels with objectively clustered pathological sounds by machine learning (ie, the DeepBreath AI algorithm). |
| Diagnostic performance of DeepBreath to detect crackles in IPF patients. | During the data analysis period (i.e., after the 60-minute study intervention period). | Diagnostic performance of the AI algorithm (DeepBreath) trained to detect crackles in IPF patients. |
Countries
Switzerland
Contacts
Primary ContactJohan N. Siebert, MD
Backup ContactPierre-Olivier Bridevaux, Prof.
Outcome results
None listed