Digital auscultation in PERCH: Associations with chest radiography and pneumonia mortality in children

Abstract Background Whether digitally recorded lung sounds are associated with radiographic pneumonia or clinical outcomes among children in low‐income and middle‐income countries is unknown. We sought to address these knowledge gaps. Methods We enrolled 1 to 59monthold children hospitalized with pneumonia at eight African and Asian Pneumonia Etiology Research for Child Health sites in six countries, recorded digital stethoscope lung sounds, obtained chest radiographs, and collected clinical outcomes. Recordings were processed and classified into binary categories positive or negative for adventitial lung sounds. Listening and reading panels classified recordings and radiographs. Recording classification associations with chest radiographs with World Health Organization (WHO)‐defined primary endpoint pneumonia (radiographic pneumonia) or mortality were evaluated. We also examined case fatality among risk strata. Results Among children without WHO danger signs, wheezing (without crackles) had a lower adjusted odds ratio (aOR) for radiographic pneumonia (0.35, 95% confidence interval (CI): 0.15, 0.82), compared to children with normal recordings. Neither crackle only (no wheeze) (aOR: 2.13, 95% CI: 0.91, 4.96) or any wheeze (with or without crackle) (aOR: 0.63, 95% CI: 0.34, 1.15) were associated with radiographic pneumonia. Among children with WHO danger signs no lung recording classification was independently associated with radiographic pneumonia, although trends toward greater odds of radiographic pneumonia were observed among children classified with crackle only (no wheeze) or any wheeze (with or without crackle). Among children without WHO danger signs, those with recorded wheezing had a lower case fatality than those without wheezing (3.8% vs. 9.1%, p = .03). Conclusions Among lower risk children without WHO danger signs digitally recorded wheezing is associated with a lower odds for radiographic pneumonia and with lower mortality. Although further research is needed, these data indicate that with further development digital auscultation may eventually contribute to child pneumonia care.

binary categories positive or negative for adventitial lung sounds. Listening and reading panels classified recordings and radiographs. Recording classification associations with chest radiographs with World Health Organization (WHO)defined primary endpoint pneumonia (radiographic pneumonia) or mortality were evaluated. We also examined case fatality among risk strata.
Neither crackle only (no wheeze) (aOR: 2.13, 95% CI: 0.91, 4.96) or any wheeze (with or without crackle) (aOR: 0.63, 95% CI: 0.34, 1.15) were associated with radiographic pneumonia. Among children with WHO danger signs no lung recording classification was independently associated with radiographic pneumonia, although trends toward greater odds of radiographic pneumonia were observed among children classified with crackle only (no wheeze) or any wheeze (with or without crackle). Among children without WHO danger signs, those with recorded wheezing had a lower case fatality than those without wheezing (3.8% vs. 9.1%, p = .03).
Conclusions: Among lower risk children without WHO danger signs digitally recorded wheezing is associated with a lower odds for radiographic pneumonia and with lower mortality. Although further research is needed, these data indicate that with further development digital auscultation may eventually contribute to child pneumonia care. conjugate vaccines in these regions. 5,6 This epidemiologic transition, along with rising rates of antimicrobial resistance, has important implications for application of the WHO guidelines. 7 Both issues potentially escalate the need for the guidelines to reduce misdiagnosis and antibiotic overtreatment. Ancillary diagnostics that are feasible for low-income and middle-income countries (LMICs), are needed.
The acoustic stethoscope is an important diagnostic tool, its origins dating back to the early 1800s. 8 While many consider chest auscultation with a stethoscope synonymous with medical care, it is not included as a diagnostic in the WHO pneumonia primary care guidelines where most children first access care. This is likely because teaching lung auscultation with acoustic stethoscopes requires medical expertise and time, both of which are not readily available in often overburdened primary care clinics in LMICs. Furthermore, lung sound interpretation with acoustic stethoscopes is largely considered subjective, achieving modest agreement between experienced physicians. 9,10 Children pose additional challenges given their breathing patterns can vary, as can their cooperation, contributing to breath-to-breath lung sound variation.
Digital stethoscopes may overcome these challenges. They are portable electronic devices that can noninvasively transmit, filter, and amplify lung sounds for real-time human interpretation. 11 Digital stethoscopes can also record lung sounds to allow post-processing of sound recordings, more rigorous human interpretation, and computerized automated analysis. 12,13 In LMICs with limited resources, digitally recorded lung sounds have the potential for use in research and, with further advancements, as a point-of-care respiratory diagnostic during clinical care or in the emerging field of telemedicine.
Currently little is known whether digitally recorded lung sounds, when interpreted acoustically by humans, are associated with clinical outcomes or radiographic disease among children in LMICs.
During PERCH we used a digital stethoscope to record lung sounds from children 1-59 months of age hospitalized with WHO-defined clinical pneumonia in six sub-Saharan African and South Asian countries. 14 Our objectives for this research were twofold. First, we aimed to evaluate the association of digitally recorded lung sounds with WHO-defined radiographic primary endpoint pneumonia, and, second, we sought to determine whether digitally recorded lung sounds are associated with mortality among PERCH children with WHO-defined clinical pneumonia.

| PERCH enrollment
The PERCH study prospectively enrolled hospital cases and community controls over a 2-year period at each site in seven countries in Africa and Asia. 5 As previously described, from December 2012 to January 2014 hospitalized children 1-59 months of age who were eligible for PERCH in Bangladesh, The Gambia, Kenya, South Africa, Thailand, and Zambia could have their lung sounds recorded during enrollment; the Mali site did not participate. 14 Cases were eligible if 1-59 months old and they met pre-2013 WHO severe or very severe pneumonia criteria (Table 1). If the child with chest indrawing in the absence of danger signs was found to be wheezing during enrollment screening they received bronchodilator treatment, and if chest indrawing was present and subsequently resolved after treatment they were excluded. 5 Antero-posterior chest radiographs were obtained on cases at admission and interpreted by a panel of physicians standardized to interpret chest radiographs per the WHO research methodology. 15,16 See Table 1 for WHO chest radiograph classifications. Discharge status and hospital outcome were recorded, and children discharged alive were followed up 30 days after hospital admission to obtain vital status. PERCH clinical study staff, which included a mix of nurses, nonphysician clinicians, and physicians, which varied by study site, received intensive clinical training on respiratory assessments and laboratory and radiographic procedures before study commencement and throughout the study at regular frequencies. 17

| Lung sound recordings
All study staff were trained to record lung sounds according to a protocol using a commercial digital stethoscope (ThinkLabs ds32a). 14   using an innovative automated multiband denoising filter. 12,14 Lung sounds were classified according to a previously described methodology. 14 In brief, each lung sound was randomly assigned to two members of an expert listening panel of eight pediatricians and pediatric-experienced physicians who were standardized to interpret lung sounds according to a reference panel of previously recorded lung sounds. 14 After adjudicating interpretation discrepancies, the listening panel assigned each PERCH case 1 prespecified summary lung sound classification. 14  Radiographic pneumonia a • An opacity that includes a portion or whole of a lobe, or the entire lung, that is dense or fluffy in appearance and may or may not contain air bronchograms. • An opacity of any size or density that creates a silhouette sign with the mediastinal border.
• An opacity of any size or density associated with a pleural effusion in the lateral pleural space at the costophrenic angle or adjacent lateral chest wall. May not be associated with an opacity if the effusion occludes a majority of the hemithorax (opacity assumed). Pleural effusion does not include fluid in the horizontal or oblique fissures.
Other infiltrate Densities in both lungs that appear linear, patchy, and lacy (interstitial infiltrate) includes peribronchial thickening and atelectasis; can also be smaller patchy infiltrates or atelectasis that does not meet the criteria of radiographic pneumonia.

Uninterpretable
Image is not interpretable regarding the presence or absence of radiographic pneumonia.
Digitally recorded lung sound models (Adapted from McCollum ED et al. 14 ) Normal Soft sounds, not musical or popping in quality.
Crackle only Short, explosive, not musical, popping sounds; usually repetitive and occurs without wheezes.
Wheeze only Musical sounds of long duration; can be high or low pitch and occurs without crackles.
Any wheeze Musical sounds of long duration; can be high or low pitch and can be present with crackles.
Uninterpretable Persistent crying or poor quality such that no full breath sounds are heard Abbreviation: WHO, World Health Organization. a Radiographic pneumonia is termed "primary endpoint pneumonia" in the WHO methodology.
F I G U R E 1 Listening positions for sequential lung sound recordings

| Statistical analysis
To evaluate associations between lung recordings and radiographic pneumonia or death, we used the t test for continuous variables and the Pearson χ 2 or Fisher exact tests for categorical variables. We calculated unadjusted odds ratios (OR) and 95% confidence intervals (CIs) for radiographic pneumonia (vs. normal) and mortality (vs. alive), as predicted by each lung sound model (abnormal vs. normal) using simple logistic regression. Children with missing or uninterpretable lung sound recordings, or with missing or uninterpretable chest radiographs or radiographs classified as "other infiltrate" only were excluded from analyses comparing lung sounds and chest radiographs. Multiple logistic regression was used to adjust for sex, age, and study site in multivariate analyses. We also conducted a sensitivity analysis using a random-effects regression model to evaluate the association between lung sounds and radiographic pneumonia using country as the group variable. All statistical analyses were performed using SAS (version 9.4).

| RESULTS
Among 792 total PERCH cases with lung sound recordings, We described the characteristics of digital auscultation cases in Table 2, and in Table 3 and E- The distribution of lung sound classifications by chest radiograph reading varied substantially across PERCH sites (Table 3).

| Digitally recorded lung sounds and WHO-defined radiographic pneumonia
In Table 4 we report on the associations between lung sound recordings and radiographic pneumonia when using normal chest radiographs as the referent. We found a lower adjusted OR (aOR) for pneumonia, relative to normal lung sounds. However, neither of these associations reached statistical significance. The random effects sensitivity analysis in E- Table S2 suggests the fixed effect model results in Table 4 are robust to other conditions. In E- Table S3 we explored the performance of combinations of lung sounds in identifying radiographic pneumonia.

| Digitally recorded lung sounds and mortality
We also examined the association between digitally recorded lung sounds and 30-day mortality among PERCH cases by logistic regression ( our recording techniques, ambient sound filtering, and interpretation methods were likely valid, achieving >90% interpretability, moderate between-listener agreement, and a high proportion of normal lung sound recordings among controls, compared to clinical pneumonia cases. 14 We have also developed and internally validated a fully automated lung sound processing algorithm that can identify abnormal lung sounds from PERCH recordings with nearly 90% accuracy. 13 In this research, we extend this initial body of work to show that human interpretation of digital lung recordings has important clinical relationships with radiographic pneumonia and pneumonia mortality. While these results are encouraging it is important to stress that they should be considered as only an initial step towards clinical or research application given the lack of a gold standard for pneumonia diagnosis and the inherent limitations of the WHO-defined radiographic pneumonia methodology, as discussed below. Additional research evaluating digital auscultation as a potential diagnostic tool for pediatric respiratory illnesses will be required before considering it for clinical implementation.
Although chest radiographs are considered the reference standard for pneumonia diagnosis among children, radiographic imaging exposes children to ionizing radiation. 19 Furthermore, radiographic equipment is expensive, facility-based, and there is a lack of interpretation expertise in most LMICs. 20 All of these issues pose barriers to wide-scale implementation of chest radiography in LMICs. Digital stethoscopes that incorporate an automated lung sound processing algorithm, on the other hand, circumvent these obstacles and have the potential to be a community-based, noninvasive point-of-care pneumonia diagnostic. Understanding the relationships between lung recordings and radiographic pneumonia in LMICs is, therefore, crucial, but has yet to be rigorously studied.
After controlling for demographic characteristics we found that wheezing among children with WHO severe pneumonia (i.e., no WHO danger signs) is independently associated with a lower odds of radiographic pneumonia (OR: 0.38; 95% CI: 0.15, 0.92). In contrast, we also found a higher odds of radiographic pneumonia in children with lung recordings of crackles or wheezes and WHO very severe pneumonia (i.e., with danger signs), although these results did not reach statistical significance (95% CIs crossed 1.0). We may have observed qualitatively conflicting odds of radiographic pneumonia by pneumonia severity strata because children with more severe illness (i.e., WHO very severe disease) and wheezing may have more severely narrowed airways and greater airflow obstruction, both of which could subsequently result in hyperinflation and atelectasis (i.e., collapsed areas of lung parenchyma) on chest radiographs that could be secondarily infected with bacteria. 21 A higher proportion of these more severely ill children may also have had primary bacterial pneumonia with alveolar consolidation present on imaging. Importantly, the WHO radiographic primary endpoint pneumonia does not differentiate between alveolar consolidation and atelectasis. 16,20 Overall the results from this analysis indicate that lung recordings have potential for use as a pneumonia diagnostic among children.
We previously found wheezing to be the most common abnormal recorded lung sound heard among PERCH cases, identified in about T A B L E 6 Case fatality ratio a stratified by digitally recorded lung sounds and WHO pneumonia severity   Table S1). As a result, if any bias exists, we believe our results are biased toward more conservative inferences. The second main limitation to this study is that pneumonia has no true gold standard reference. 24 Although chest radiographs are considered the best current reference standard, they are not ideal given the interpretation of radiographic abnormalities is subjective and the appearance of radiographic abnormalities can also lag behind clinical signs, potentially reducing sensitivity and delaying effective treatment. It is well known that normal chest radiographs can be present in children with signs consistent with clinical pneumonia, and this also occurred in PERCH, as 46% of children meeting WHO clinical pneumonia criteria had normal chest radiographs. 25 It is also important to note that the WHO radiographic primary endpoint pneumonia definition is not intended for clinical application, and this limits the clinical generalizability of these results. However, despite chest radiographs serving as the reference standard for pneumonia diagnosis there are few chest radiograph interpretation schemas used as widely as the WHO method. Although this interpretation approach was initially established for the evaluation of bacterial conjugate vaccine efficacy its application has been extended to epidemiologic research of child pneumonia in LMICs, 16 including as the reference standard in the PERCH Study. 5 It is important to note that not all experts agree with application of the WHO chest radiograph methodology for epidemiologic research due to its bias toward specificity rather than sensitivity, leading to underestimation of the public health burden of pneumonia. These results should be interpreted within this context. Despite the inherent limitations to chest radiographs in general and the WHO method itself, PERCH applied rigorous interpretation procedures to optimize interpretation reliability and case ascertainment, achieving 78% agreement between primary readers of radiographs (Cohen's kappa: 0.50), which was comparable to other studies using this WHO methodology. 15,26 Lastly, all children enrolled into PERCH met clinical pneumonia criteria, which does not allow us to assess digital auscultation utility among children without clinical pneumonia. Such an evaluation is an important next step.
In summary, the results of this study suggest that digital lung recordings may have a future role in pediatric respiratory research and as a point-of-care respiratory diagnostic and prognosticating tool for children in LMICs. Essential next steps include evaluating the feasibility and decision-making impact of digital stethoscope use by both formal and informally trained health workers, evaluating it against other pneumonia reference endpoints other than chest radiography, assessing agreement between standard auscultation by experts with digital recorded lung sounds interpreted by either humans or automated algorithms, and externally validating the automated lung sound processing algorithm in other similarly vulnerable pediatric populations.