Community use of digital auscultation to improve diagnosis of childhood pneumonia in Sylhet, Bangladesh

Abstract

BACKGROUND: Pneumonia is a major cause of death in children younger than five years, especially in low- and middle-income countries (LMICs). The World Health Organization (WHO) Integrated Management of Childhood Illnesses (IMCI) tool primarily relies on manually counting the respiratory rate and observing for clinical signs of respiratory distress to diagnose childhood pneumonia. This tool intentionally prioritised sensitivity over specificity to ensure children with possible bacterial disease received antibiotics. Community-based health workers play a vital role in the early identification and treatment of pneumonia cases using the IMCI tool in LMICs. The IMCI tool does not include lung auscultation in the diagnostic algorithm for pneumonia because of its high interobserver variability and subjectivity and the related difficulty in training healthcare workers with limited auscultation experience to effectively use a conventional stethoscope. Digital auscultation augmented by an artificial intelligence algorithm for classification of lung sounds has the potential to be both a feasible and accurate diagnostic tool for use by healthcare workers in LMICs. Operationally, the inclusion of digital auscultation findings in the current IMCI guidelines could increase the specificity of the IMCI algorithm to diagnose childhood pneumonia and help to reduce unnecessary use of antibiotics in LMICs.

AIM AND OBJECTIVES: My PhD aims to evaluate the feasibility of digital auscultation use by frontline health workers in Bangladesh within the context of IMCI among 2 – 59 months old children, as well as its accuracy compared to a reference panel of paediatricians. The specific objectives are:

1. To synthesise the evidence on the performance of digital auscultation compared to human listeners in identifying adventitious lung sounds among children with pneumonia.

2. To assess whether lung sounds from children recorded by community health care providers (CHCPs) at community-level health facilities using a prototype digital stethoscope meet pre-defined quality thresholds established by experts. (‘pre-defined quality threshold’ was defined over 50% of the patients would have ‘quality’ lung sound recordings, ‘quality’ was defined as having at least 75% interpretable lung sound segments per patient according to the human listening panel (i.e., ≥ 3 out of 4 chest positions)).

3. To determine the diagnostic accuracy of an automated lung sound classification algorithm for identifying adventitious lung sound, compared with a reference panel of paediatricians.

4. To explore the acceptability of a prototype digital stethoscope (Feelix Smartscope) among potential end-users (CHCPs, carers and community leaders).

METHODS AND RESULTS: 1. Digital auscultation as a diagnostic aid to detect childhood pneumonia – systematic review I conducted a systematic review by searching eight bibliographic databases and citation indices (MEDLINE, Embase, CINAHL Plus, Web of Science, Global Health, IEEExplore database, Scopus, and ClinicalTrial.gov) and reference lists of included studies. Reported methodologies/approaches and performance metrics for classifying adventitious lung sounds varied widely across the included ten studies. All included studies except one reported the overall diagnostic performance of the digital auscultation/computerised lung sound analysis to distinguish adventitious lung sounds, irrespective of the disease condition or age of the participants. The reported accuracy for classifying adventitious lung sounds in the included studies varied from 66.3% to 100%. However, it remained unclear to what extent these results would be applicable for classifying adventitious lung sounds in children with pneumonia. This systematic review found very limited evidence on the diagnostic performance of digital auscultation to diagnose pneumonia in children.

2. Ability of community health workers to record quality lung sounds from children younger than five years at first-level facilities – prospective observational research study A total of 990 children aged 2-59 months with possible pneumonia (cough or difficult breathing) were enrolled by nine CHCPs from selected nine community clinics in Zakiganj sub-district of Sylhet District, Bangladesh. Of them, 389 children were classified as having “any” pneumonia (pneumonia or pneumonia with respiratory danger signs), and 601 children had only cough or difficult breathing. Using a prototype digital stethoscope (Feelix Smartscope), CHCPs recorded lung sounds in four sequential chest positions – two posterior and two anterior positions – ~10 seconds in each position, allowing 3 to 4 complete breath cycles (inspiration and expiration) per recorded chest position. Each child’s recorded lung sounds were then classified by two paediatricians trained to a standardized classification protocol and who were also blinded to the clinical findings. The paediatrician classifications were compared, and if not in agreement, a third trained paediatrician reader served as an arbitrator. A quality recording was defined as a child with at least three out of four chest positions assessed as interpretable by the listening panel. An interpretable chest position was classified by the listening panel as any of the following: no wheeze and no crackle, wheeze only, crackle only, or wheeze and crackles. Among all children, CHCPs recorded three or more chest positions in more than 98% of participants. The paediatrician listening panel classified 87.6% (95% CI: 85.4%, 89.6%) of all children, 88.2% (95% CI: 84.5%, 91.2%) of those with any pneumonia and 87.2% (95% CI: 84.3%, 89.8%) without pneumonia as interpretable in at least three chest positions. Among CHCPs, we observed a linear relationship between the number of children enrolled and the percentage of successfully recorded sound files from ≥3 chest positions. On average, the rate of success increased by 1% with every ten new enrolments by CHCPs. The proportion of children who were cooperative and quiet throughout the lung sound recordings was 81.6% (95% CI: 77.3%, 85.3%) among those with any pneumonia and 78.7% (95% CI: 75.1%, 81.9%) in children with no pneumonia. In the majority of children (56.6%) lung sounds were recorded in all four positions within one minute.

3. Diagnostic accuracy of automated computerised lung sound analysis to identify adventitious lung sounds compared to the paediatrician listening panel – prospective observational research study Each child’s lung sound recordings were randomly sent to two listening panel members, and if classifications did not agree, a third reader served as the arbitrator. The listening panel classified recorded lung sounds into normal (no wheeze and no crackles), crackles, wheeze, crackles and wheeze, or uninterpretable. An automated CLSA algorithm classified recorded sounds into the same categories, which were compared with the human panel classifications. Of 990 enrolled children, 867 had ≥3 chest positions interpretable by the listening panel. Compared to the consensus listening panel, the CLSA had a moderate sensitivity (61.8%; 95% CI: 55.7%, 67.6%) and a moderate specificity (60.7%; 95% CI: 56.6%, 64.6%) for classifying lung sounds as adventitious or normal. The agreement was low between panel classifications and CLSA classifications (Cohen’s kappa = 0.20; 95% CI: 0.14, 0.26) of the recorded lung sounds.

4. Acceptability of a prototype digital stethoscope (Feelix Smartscope) among potential end-users – qualitative study Four focus group discussions (FGDs) were conducted with beneficiaries (mother, father and community leaders) and service providers (CHCPs) who used the Feelix Smartscope prototype. Verbatim transcripts were prepared, and translations were completed. Coding was executed in Microsoft Excel, and relevant quotes were extracted to ascertain the emerging themes. Two researchers coded the dataset independently to ensure validity, and inconsistencies were resolved through discussion. Mothers were more aware of the digital stethoscope than fathers. CHCPs and the female community leaders were aware of the stethoscopes and accepted the stethoscope. Most CHCPs had positive perceptions of the digital stethoscope. They appreciated stethoscope training as they learned about new technology and diagnostic approaches. The users mentioned several technical shortcomings of the prototype device.

CONCLUSIONS: This thesis demonstrated that CHCPs at rural, first-level clinics in Bangladesh could feasibly record quality lung sounds using a novel digital stethoscope prototype (Feelix Smartscope) without a substantial increase in their workload. This study also showed an initial version of an automated algorithm could classify lung sounds as either adventitious or normal with moderate sensitivity and moderate specificity when using a paediatrician listening panel as the reference. The agreement was low between panel classifications and CLSA classifications (Cohen’s kappa = 0.20) of the recorded lung sounds. CHCPs found the new technology generally acceptable. This thesis also highlights the potential for addressing current limitations that could be overcome by refining both device itself and the automated algorithms. After improving the stethoscope addressing the issues identified in this study and improving the automated sound classification algorithm, a validation study is required.