key: cord-0283727-qohntipf
authors: Porter, P.; Brisbane, J.; Abeyratne, U.; Bear, N.; Wood, J.; Peltonen, V.; Della, P.; Purdie, F.; Smith, C.; Claxton, S.
title: Rapid, point of care detection of Chronic Obstructive Pulmonary Disease using a cough-centred algorithm in acute care settings.
date: 2020-09-08
journal: nan
DOI: 10.1101/2020.09.05.20164731
sha: 36cc3f7ad68efdb54638e936e35acaee8359ef73
doc_id: 283727
cord_uid: qohntipf

Rapid and accurate diagnosis of Chronic Obstructive Pulmonary Disease (COPD) is problematic in acute-care settings, particularly in the presence of infective comorbidities. The aim of this study was to develop a rapid, smartphone-based algorithm for the detection of COPD, in the presence or absence of acute respiratory infection, and then evaluate diagnostic accuracy on an independent validation set. Subjects aged 40-75 years with or without symptoms of respiratory disease who had no chronic respiratory condition apart from COPD, chronic bronchitis or emphysema, were recruited into the study. The algorithm analysed five cough sounds and four patient-reported clinical symptoms providing a diagnosis in less than one minute. Clinical diagnoses were determined by a specialist physician using all available case notes, including spirometry where available. The algorithm demonstrated high percent agreement (PA) with reference clinical diagnosis for COPD in the total cohort (n=252, Positive PA=93.8%, Negative PA=77.0%, AUC=0.95); in subjects with pneumonia or infective exacerbations of COPD (n=117, PPA=86.7%, NPA=80.5%, AUC=0.93) and in subjects without an infective comorbidity (n=135, PPA=100.0%, NPA=74.0%, AUC=0.97.) In those who had their COPD confirmed by spirometry (n=229), PPA = 100.0% and NPA = 77.0%, AUC=0.97. The algorithm demonstrates high agreement with clinical diagnosis and rapidly detects COPD in subjects presenting with or without other infective lung illnesses. The algorithm can be installed on a smartphone to provide bedside diagnosis of COPD in acute care settings, inform treatment regimens and identify those at increased risk of mortality due to seasonal or other respiratory ailments.

Chronic obstructive pulmonary disease (COPD) is the fourth leading cause of mortality, affecting more than 384 million individuals worldwide [1] , and is characterised by airflow limitation and a progressive decline in lung function [2] . The population prevalence of COPD, via spirometry screening, is reported to be 9-26% in those greater than 40 years old [3] . It is estimated that 80% of people with COPD are undiagnosed [4] and up to 60% of those with a diagnosis of COPD were found to be misdiagnosed upon subsequent spirometry [5, 6] . 30-60% of patients who have been diagnosed by a physician with COPD have not undergone spirometry testing [7] . In a study of 533 COPD patients, 15% of those with spirometry tests did not show obstruction and 45% did not fulfil quality criteria [8] .

COPD should be considered in patients who present with dyspnoea, chronic cough, sputum production or recurrent lower respiratory tract infections and who have been exposed to tobacco or air pollution. Airflow limitation, demonstrated by a FEV 1 /FVC ratio of < 0.7 on post-bronchodilator spirometry is considered diagnostic of COPD according to criteria stipulated by the Global Initiative for Chronic Obstructive Lung Disease (GOLD) [2] . Severity of airflow limitation in COPD can be classified by the degree of reduction in FEV 1 as a percentage of the predicted value [2] . However, spirometry is not routinely used in emergency departments or primary care settings due to inexperience, time constraints and availability of equipment [9] .

Early and accurate diagnosis of COPD is imperative to ensure initiation of correct treatment, particularly as evidence suggests the incipient stages represent a period of rapid decline in lung function where cessation of smoking and intervention may be of value [10] . Rapid identification and management of COPD is important in acute care settings as there is a heightened risk of mortality from respiratory infections such as seasonal influenza [11] . SARS-CoV-2 has a reported case fatality rate of 1.4% for patients without comorbid conditions vs 8.0% for those with chronic respiratory conditions [12] .

Screening for COPD in primary care settings using spirometry in asymptomatic patients has not been found to be efficient as high numbers need to be screened to detect any cases [13] . Screening questionnaires such as the COPD diagnostic questionnaire (CDQ) have performed poorly in an asymptomatic cohort in the primary care setting [14] . We propose that the best use of an algorithm for screening is in a scenario where patients present to a healthcare facility with symptoms, where there is a higher pre-test probability of case detection.

We have previously demonstrated high diagnostic agreement of a similar automated algorithm with clinical diagnoses for paediatric respiratory diseases including croup, asthma, bronchiolitis and pneumonia. The algorithm also accurately separated upper-from lower-respiratory tract conditions [15] . The technology, which has regulatory approval, is similar to that used in speech recognition software and uses cough sounds and simple patient-reported clinical symptoms to derive the diagnostic probability output [16] . Cough sounds are recorded by a standard smartphone; the 1 3 in-built diagnostic algorithm provides a rapid result without requiring clinical examination or additional diagnostic tests.

In this paper, we describe the development and evaluate the accuracy of an algorithm for diagnosing COPD from a cohort of mixed respiratory disorders including acute respiratory infections. The intended use population is those who present to health settings with suspected respiratory illness.

Between Jan 2016 and March 2019, a convenience study sample was obtained by prospectively recruiting participants from the emergency department, low-acuity ambulatory care and in-patient wards of a large, general hospital in Western Australia; and from the consulting rooms of a respiratory physician.

This diagnostic accuracy study is part of a larger development program (Breathe Easy / ANZCTR: ACTRN12618001521213). Subjects were approached if they presented to a participating site with signs or symptoms of respiratory disease or to specialist rooms for a lung function test. Subjects with no discernible symptoms of respiratory disease were also recruited. Subjects were excluded if they were on ventilatory support, had terminal disease, were medically unstable, had structural upper airway disease or had a medical contraindication to providing a voluntary cough (eg severe respiratory distress; eye, chest or abdominal surgery within 3 months; history of pneumothorax). Subjects with uncontrolled heart failure/cardiomyopathy, neuromuscular disease or lobectomy/pneumonectomy were also excluded. From this cohort, only subjects aged 40-75 years were used for the COPD development program.

Written informed consent was obtained from all participants and the study was approved by a Human Research Ethics Committee (Reference Number: 1501). There were no adverse events reported. The study did not interfere with clinical care and all treatment decisions were at the discretion of the treating physician.

The development of the mathematical techniques used to derive the algorithm have been described elsewhere [15] [16] [17] [18] . Briefly, an independent training cohort (n=564) was used to obtain clinical data and cough samples (from which mathematical features were extracted). In developing the algorithm, selected features were weighted and combined to build various continuous classifier models used to determine the probability of a COPD diagnosis (reference test). The probability output of the algorithm represents the specific, weighted combination of features used and thus the performance of individual features cannot be reported separately. The optimal model and corresponding probability decision threshold was selected using a Receiver Operating Characteristic (ROC) curve with due consideration given 1 3

to achieving a balance of PPA and NPA [16] . Different algorithms could be developed looking at different outputs such as very high specificity. Once the optimal model was developed, an independent testing set was prospectively recruited.

Subjects provided five coughs that were recorded using a smartphone (iPhone6) held approximately 50cm away from the subject at a 45-degree angle to the direction of the airflow. Recordings were undertaken in standard clinical environments; however, care was taken to ensure that other people's coughs and voices were not recorded. The cough recording was obtained within 30 minutes of the physical examination of the patient to ensure the clinical features had not changed. If the subject was unable to provide five coughs that were recognised by the coughdetection software or the cough recording became corrupted, the subject was excluded from further analysis.

The following four clinical symptoms were selected in building the model: subject age; smoking pack-years and subject-reported presence of acute cough or fever during this illness. One smoking pack-year is defined as 20 cigarettes or 20 g tobacco, smoked each day over the course of 1 year [19] . Where the clinical symptoms were unknown, the algorithm did not return a response.

A full medical assessment was performed on all participants at time of enrolment, including history and clinical examination. Diagnostic tests were ordered by the treating clinician independently of the study and results were available to researchers.

A specialist physician assigned a clinical diagnosis to each subject based on a review of their medical file including: discharge diagnosis, all outpatient and inpatient notations and radiology/laboratory results. The clinical diagnosis definitions (Table 1) were employed in both the testing set (described here) and in the training set used for algorithm development:

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted September 8, 2020. . https://doi.org/10.1101/2020.09.05.20164731 doi: medRxiv preprint Spirometry was conducted according to standard methodology [2, 20] . Where the case definition was not met or the symptoms were significantly altered by treatment, the subject was scored as "unsure" and was excluded from further analysis.

Diagnostic accuracy tests were performed for four groups using an independent, test set of subjects. The same inclusion and exclusion criteria were used for both training and test sets ( Table 2) :

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted September 8, 2020. Of GROUP 1, excluding those whose COPD has not been confirmed by spirometry . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted September 8, 2020. . https://doi.org/10.1101/2020.09.05.20164731 doi: medRxiv preprint When a clinical diagnosis had been assigned to all subjects, the database was locked and the software algorithm was run by an independent researcher to ensure blinding was maintained. Each subject's cough sound recording and clinical diagnosis were only used once in the prospective test.

Power calculations were derived as follows. Based on expected positive and negative percent agreement greater than 85% from the training program, to obtain a superiority end-point of 75% (lower bound 95% CI of maximum width ±0.10) a minimum of 48 cases were required.

Positive percent agreement (PPA) is defined as the percentage of subjects with a positive index test result for a specified condition who also have a positive reference standard for the same condition. Negative percent agreement (NPA) is the percentage of subjects who returned negative results for both tests.

The primary study endpoint was defined as PPA and NPA of the index test with the reference standard, with 95% confidence intervals calculated using the method of Clopper-Pearson. The probability of positive clinical diagnosis was calculated for each subject by the final classifier model and was used as the decision thresholds in the derived ROC curve.

From the prospective testing set, 270 participants met inclusion criteria for, and were enrolled in the COPD diagnostic study. Of these 153 were from the hospital emergency department or inpatient wards, and 117 were respiratory outpatients or from the ambulatory acute care unit.

Two hundred and fifty-two participants provided a valid index and reference test (figure 1). Two were excluded as the clinical diagnosis was recorded as unsure. The mean age of participants was 59.7 ± 9.2 years, 58.7% were female. Those with COPD were older than those without (65.5 vs 57.8 years, p<0.0001), although the sex proportion did not differ with diagnosis. 85.3% of the entire cohort had at least one of the following respiratory symptoms: acute, chronic or productive cough; fever; rhinorrhoea; SOB; wheeze; or hoarse voice. Subject characteristics are shown in Table 3 including spirometry results where available.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted September 8, 2020. . https://doi.org/10.1101/2020.09.05.20164731 doi: medRxiv preprint CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted September 8, 2020. For cases where spirometry (n=123) was used to confirm the presence or absence of COPD, the mean age was 60.0 ± 8.7 years and 65.0% were female with FEV 1 measurements as shown in Table 4 . The COPD negative group includes six chronic, fixed asthmatic patients with FEV 1 below 80%.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted September 8, 2020. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted September 8, 2020. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted September 8, 2020. . https://doi.org/10.1101/2020.09.05.20164731 doi: medRxiv preprint Although the algorithm was developed to discriminate based on GOLD criteria we repeated the analysis using Lower Limit of Normal (LLN) thresholds to diagnose COPD. Test performance in the "COPD confirmed by spirometry group" (n=229) returned PPA of 100.0% [90.75%, 100.0%] and NPA of 75.4% [68.65%, 81.32%].

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted September 8, 2020. . https://doi.org/10.1101/2020.09.05.20164731 doi: medRxiv preprint

We have described a simple, rapid diagnostic test for COPD which demonstrates high agreement with clinical diagnosis in the acute setting. Diagnostic agreement of the software algorithm with clinical diagnosis of COPD was PPA 93.8% and NPA 77.0%. Agreement was maintained when the patient had an acute respiratory infection (PPA 86.7% and NPA 80.5%). Importantly, the index test retains high diagnostic agreement in cases of spirometry-confirmed COPD: PPA (100.0%) and NPA (77.0%). Accurate diagnosis of COPD requires confirmation by spirometry, the gold standard tool for COPD diagnosis [2] .

We used the GOLD criteria for COPD diagnosis (FEV 1 /FVC<0.7) when developing our algorithms although COPD can also be defined using lower limit of normal (LLN). When calculated using the LLN thresholds, test performance was not significantly different from values obtained using GOLD criteria. It should be noted that as our model was developed to recognise COPD diagnosed using the GOLD criteria, we would expect a poorer performance when the diagnostic criteria were changed.

In many European countries, spirometry is available in acute and primary care settings [8] although uptake of the test is limited, leading to underdiagnosis or misdiagnosis of patients [6] . A number of barriers to using spirometry in primary and acute care settings have been reported including limitations in access, expertise and time; as well as expense [21] . Alternative testing methods have been developed. A meta-analysis of the COPD Diagnostic Questionnaire (CDQ) among ever smokers had a pooled sensitivity of 64.5% (95% CI 59.9% to 68.8%) and specificity 65.2% (52.9% to 75.8%) from four studies. Analysis of handheld flow meters showed a sensitivity of 79.9% (95% CI 74.2% to 84.7%) and specificity 84.4% (68.9% to 93.0%) from three studies [14] . In a scenario comparable to our study, when the CDQ was performed on symptomatic patients in primary care, the AUC was 0.65, sensitivity was 89.2%/65.8% and specificity 24.4%/54.0% for low risk and high risk of having COPD respectively [22] . The performance of our software algorithm exceeds that of currently available COPD screening questionnaires; outperforms the sensitivity of handheld flow meters with comparable specificity and demonstrates high agreement with the gold-standard (spirometry) in under one minute. This algorithm is intended to be used as a stand-alone device allowing for real-time diagnosis. As it is easy to operate and requires no physical patient contact, infection risk is minimised.

We envisage this algorithm could be positioned as an initial screening test in acute care settings for patients who present with non-specific respiratory symptoms. A positive result could be used to guide immediate care in the acute setting. Confirmatory testing by spirometry remains the gold standard test and could be performed during subsequent specialist follow up.

Population and primary care surveys have demonstrated that mild (FEV 1 ≥ 80% of percent predicted) and moderate (FEV 1 50-80% of percent predicted) airflow limitation is seldom diagnosed by clinicians [23, 24] . In our study, 48% of those with clinically-diagnosed COPD had only mild or moderate airflow limitation (Table 1) . This group represents those who would benefit most from this algorithm, both by virtue of new treatment possibilities and also because they are frequently underdiagnosed.

In this study we were able to accurately identify the presence or absence of COPD in patients with LRTI including pneumonia. In these situations, spirometry can be difficult to perform adequately, and an initial diagnostic test will help detect COPD in acutely unwell patients and identify those individuals most at risk of developing complications. Individuals with COPD are known to experience more frequent complications and mortality due to seasonal illnesses such as influenza [11] . More recently, a meta-analysis examining risk of severe outcomes from SARS-CoV-2 infection (admission to ICU, mechanical ventilation or death) showed a greater than five-fold increase in risk of severe disease in patients with coexistent COPD [25] . The authors recommend that all COPD patients with a suspected infection should be carefully monitored in view of this increased risk. The diagnosis of COPD in patients presenting with SARS-CoV-2 or similar respiratory infections, would allow more focused therapeutic pathways and usefully guide healthcare resources to this at-risk group.

There are several limitations to this study. Our study population was recruited in an urban setting with smoking-related COPD. The generalisability of these results to COPD of differing aetiologies and in other settings requires confirmation. The tests were performed by trained research personnel in controlled environments, although we would consider the device less onerous to use than spirometry. The cough recording can be affected by background noise and positioning of the device, although the program will alert the user if background noise is excessive. The population recruited reflects the intended age range of the population, however as expected, those with diagnosed COPD were slightly older than those without and it will be important to replicate this study using an older control group.

This COPD diagnostic algorithm may be used in combination with a suite of other respiratory diagnostic algorithms developed in the Breathe Easy program, including tests for asthma, pneumonia and lower respiratory tract disease [15] . The software would provide a diagnostic output for each condition simultaneously.

In conclusion, the algorithm was able to accurately identify COPD even in the presence of infection. The algorithm operates as a stand-alone tool and provides a rapid result. It may find application in the acute-care setting as a screening tool to alert clinicians to the presence of COPD and allow more rapid, targeted and appropriate management.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted September 8, 2020. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted September 8, 2020. . https://doi.org/10.1101/2020.09.05.20164731 doi: medRxiv preprint

Global Health Epidemiology Reference, Global and regional estimates of COPD prevalence: Systematic review and meta-analysis

Global Initiative for Chronic Obstructive Pulmonary Disease. Global Strategy for the Diagnosis, Management, and Prevention of Chronic Obstructive Pulmonary Disease

Implementation of a targeted screening program to detect airflow obstruction suggestive of chronic obstructive pulmonary disease within a presurgical screening clinic

Prevalence, severity and underdiagnosis of COPD in the primary care setting

Under-and over-diagnosis of COPD: a global perspective. Breathe (Sheff)

Underdiagnosis and Overdiagnosis of Chronic Obstructive Pulmonary Disease

How often is diagnosis of COPD confirmed with spirometry

Misdiagnosis of asthma and COPD and underuse of spirometry in primary care unselected patients

New developments in the assessment of COPD: early diagnosis is key

Risk of Severe Influenza Among Adults With Chronic Medical Conditions

Date last updated

Screening for chronic obstructive pulmonary disease using spirometry: summary of the evidence for the U.S. Preventive Services Task Force

Diagnostic accuracy of screening tests for COPD: a systematic review and meta-analysis

A prospective multicentre study testing the diagnostic accuracy of an automated cough sound centred analytic system for the identification of common respiratory disorders in children

Cough sound analysis can rapidly diagnose childhood pneumonia

Automatic Croup Diagnosis Using Cough Sound Recognition

Stratifying asthma severity in children using cough sound analytic technology

Quantification and chemical markers of tobacco-exposure

Standardisation of spirometry

Barriers to the use of spirometry in general practice

External validation of a COPD diagnostic questionnaire

Prevalence and underdiagnosis of COPD by disease severity and the attributable fraction of smoking Report from the Obstructive Lung Disease in Northern Sweden Studies

The prevalence of undiagnosed chronic obstructive pulmonary disease in a primary care