Pneumonia is a leading cause of mortality worldwide in children under five years of age. In 2019, there were approximately 740 000 child deaths due to pneumonia globally . There were an estimated 68 million pneumonia episodes in 2016, equivalent to 0.11 cases per child-year. A notable discrepancy has been observed in the incidence of pneumonia between high-income and low- and middle-income countries (LMICs) . Pneumonia is a key reason for hospital admissions of children and is a substantial burden on health systems .
In LMICs, child pneumonia is usually poorly understood by caregivers, and care-seeking for treatment is not adequate . Pneumonia is further underdiagnosed and undertreated due to the low doctor-to-population ratio . Access to doctors and hospitals is difficult [6,7] and treatment costs are often not affordable . Consequently, a large proportion of pneumonia cases are diagnosed and treated out of hospitals by non-physician health workers . These health workers apply pragmatic case management algorithms to diagnose, treat, and refer children suspected to have pneumonia during household visits or in community-level health facilities [10,11]. The role of health workers in community-based pneumonia case management has had a significant impact on lowering child mortality .
The World Health Organisation (WHO) Integrated Management of Childhood Illness (IMCI) guidelines primarily use fast breathing and chest indrawing to diagnose pneumonia in children. The health worker observes the child’s chest to identify fast breathing and chest indrawing [13,14]. Evidence shows that health workers can identify fast breathing with moderate accuracy . A child is identified to have chest indrawing if the tissue below the lower chest wall moves inward when the child inspires (Figure 1) . This clinical sign can occur when the lungs are inflamed from an infectious process and have poor lung compliance . Although chest indrawing is insufficient for diagnosing pneumonia, this signifies the severity of pneumonia and might be useful in detecting children at risk of hypoxaemia . Identifying chest indrawing can be challenging for health workers . The possible reason could be the low prevalence of chest indrawing cases in the population [20,21]. The characteristics of the training received by the health workers have an effect on their performance , often leading to misdiagnosis of pneumonia and incorrect treatment .
Figure 1. A child with chest indrawing. Reproduced with permission from World Health Organization.
The diagnosis and management of pneumonia in LMICs depends on health workers’ capability to accurately identify chest indrawing. Regardless of existing literature assessing the performance of health workers in identifying chest indrawing, to the best of our knowledge, the evidence has never been systematically collated. Most of the existing literature involves studies with small samples. Therefore, a systematic review could provide more powerful evidence impacting clinical practice and health care policy. In this review, we summarized the evidence on how accurately non-physician health workers can identify chest indrawing in under-five children with suspected pneumonia in LMICs.
We conducted this systematic review and meta-analysis in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) 2020  and the Preferred Reporting Items for Systematic Reviews and Meta-analyses of Diagnostic Test Accuracy Studies (PRISMA-DTA) guidelines  following the guidance provided in the Handbook for Diagnostic Test Accuracy (DTA) Reviews of Cochrane . We established the review methods prior to conducting it and prospectively registered it with PROSPERO (registration number CRD42022306954; January 26, 2022) .
Population, index test, and reference standard
The target participants were children younger than five years assessed for chest indrawing in the community or a health facility. The index test was chest indrawing assessment by a non-physician health worker (e.g. community health worker, medical assistant, nurse, nursing assistant). The reference standard was assessment by a human expert, defined as a pediatrician, general physician, or an IMCI expert assessor.
We developed a search strategy comprising medical subject headings (MeSH) terms and keywords. We systematically searched the following electronic databases for relevant studies published from January 1, 1990, until January 20, 2022: MEDLINE (via Ovid), Embase (via Ovid), Web of Science, and Scopus. We did not search for older studies considering that the IMCI strategy was launched in the 1990s. The detailed search strategy for each database is available in Table S1 in the Online Supplementary Document. We searched the identified studies’ reference lists to avoid missing relevant studies. The search strategy did not include any filters or limitations and we included studies published in any language. An expert librarian reviewed the search strategy.
We focused on studies that evaluated the performance of health workers in identifying chest indrawing against a reference standard. We included studies when all the following criteria were met:
- Identification of chest indrawing by non-physician health workers.
- Evaluation of the accuracy of identifying chest indrawing by a reference standard.
- The age of the participants was below five years.
- Carried out in LMICs .
The exclusion criteria were as follows:
- Non-human subjects or mechanically ventilated subjects.
- Lack of information on the reference standard.
- Videotaped subjects were assessed by health workers.
- Not possible to disaggregate data of chest indrawing.
- Not possible to disaggregate data of under-five children.
Study selection and data extraction
We imported the search results from different databases into Covidence Systematic Review software  and removed duplicates, after which two reviewers (AMK and SS) independently screened the retrieved studies’ titles and abstracts. Both reviewers independently read full papers of potentially relevant articles according to the eligibility criteria. The same reviewers extracted data independently using a structured form (Table S2 in the Online Supplementary Document), which included the following information: author, year, study location, study setting, sampling method, number of participants, index test, and reference standard. We also extracted data on true positive (TP), false positive (FP), false negative (FN), true negative (TN), sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) or calculated them from reported data. If relevant data were missing or not reported, we contacted the corresponding author by e-mail. We entered the data into a Microsoft Excel spreadsheet. We resolved disagreements for both the literature screening and the data extraction through discussion until we reached a consensus.
We used the Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2) tool (Table S3 in the Online Supplementary Document) to assess the methodological quality of all studies. The tool includes four risk of bias domains and three domains of applicability . Each domain has an overall judgment of “low risk of bias” if it was judged as “low” in all signaling questions. In contrast, the domain would be judged as “high risk of bias” if it was judged “high” in one or more signaling questions. We used the “unclear” category if inadequate data were reported. Again, any discrepancy between reviewers was discussed until a consensus was reached. We used Review Manager (version 5.4) to generate the figure presented in this report.
Data synthesis and analysis
We presented the sensitivity, specificity, PPV, NPV, and accuracy of each study and computed median values and interquartile ranges (IQRs). We conducted a meta-analysis following a bivariate random effects model with studies where TP, FP, FN, and TN data could be retrieved and where an adequate number of chest indrawing cases was present in the sample. We presented study diagnostic sensitivity and specificity estimates with 95% confidence intervals (CIs) in paired forest plots using a user-written command (midas) . The heterogeneity between studies was assessed from coupled forest plots and using the I2 statistics . We used Stata (version 17.0) to perform the analyses.
A PRISMA flowchart summarizing the study selection process is presented in Figure 2 . The initial search retrieved 8389 records from all databases, forty of which we reviewed in full-text. We identified three additional articles from their reference lists for full-text review. We contacted the corresponding authors of 12 studies by email for relevant data not reported in the paper. Only three responded, but no one could provide any data. We included a total of nine studies in this review and five in the meta-analysis. The list of excluded studies and the reasons for exclusion are available in Table S4 in the Online Supplementary Document.
Figure 2. PRISMA flow diagram; CI – chest indrawing, HW – health worker, RS – reference standard.
Characteristics of the included studies
The major characteristics of the included studies are presented in Table 1. Out of nine studies, six were done before the year 2000 [21,32,33,36–38] and three studies were done after [20,34,35]. Most studies were conducted in Africa [21,32,34–38], two in Asia [20,36], and one in Oceania . Only one study was conducted in a community setting  and the rest were conducted in health facility settings [21,32–38]. Two studies assessed young infants [20,32], while six assessed children aged 2-59 months [21,33–38]. The number of children per study ranged from 34 to 1405. The number of health workers ranged from 6 to 114. The workers were trained on the identification of pneumonia according to the WHO guidelines for a short duration at the beginning of the study. In seven studies, a pediatrician or a physician was the reference standard [20,21,32,34,36–38]. There was a short delay in assessment between the health worker and expert (i.e. expert assessment immediately after health worker assessment) in seven studies [21,32–34,36–38], while there was a long delay (i.e. reference standard assessed a few hours after health worker assessment) between assessments in two studies [20,35].
Table 1. Characteristics of the included studies
CHW – community health worker, wks – weeks, d – days
Methodological quality of included studies
Figure 3 presents the risk of bias and concerns regarding the applicability of the selected studies. For patient selection, we rated one study as having high risk of bias because of convenient sample selection . Two studies had unclear information on the sampling method, so we judged them as having unclear risk of bias [34,37]. For the index test, we rated all studies as low risk of bias, as the health workers were blinded to the finding of the reference standard [20,21,32–38]. For the reference standard, we rated two studies as high risk of bias as the reference standard was not blinded [35,38] and three studies as unclear risk of bias due to unclear reporting on blinding [20,32,37]. For patient flow and timing, we evaluated two studies as having a high risk of bias because of a prolonged delay between the index test and reference standard [20,35]. Overall, the studies had low concerns regarding applicability for all domains [20,21,32–38]. The concern in one study was related to inclusion criteria for patient selection .
Figure 3. Risk of bias and applicability concerns summary: review authors’ judgements about each domain for each included study.
Accuracy in chest indrawing identification by health workers compared to reference standard
The summary results of all included studies are presented in Table 2. The median sensitivity, specificity, PPV, NPV, and accuracy are reported in Table 3. The median sensitivity and specificity were 44% and 97%, respectively.
Table 2. Studies reporting health worker identification of chest indrawing compared to reference standards
CI – confidence interval
Table 3. Health worker identification of chest indrawing compared to reference standards
IQR – interquartile range
Results of meta-analysis
Individual and summary estimates of sensitivity and specificity for the studies included in the meta-analysis are shown in Figure 4. The pooled sensitivity was 46% (95% CI = 37-56), the pooled specificity was 95% (95% CI = 91-97) and there was considerable heterogeneity (I2 = 80%).
Figure 4. Accuracy of health workers’ chest indrawing identification compared to reference standards. Forest plots of individual and summary estimates of sensitivity and specificity.
Recognizing chest indrawing is a necessary skill for health workers in LMICs to diagnose and classify childhood pneumonia [13,14]. This systematic review demonstrated the performance of the non-physician health workers in identifying chest indrawing varied across the studies. The median sensitivity was 44%, with an IQR of 33%-61%; the pooled estimate of sensitivity was 46%. This low sensitivity implies that the health workers failed to identify chest indrawing among a substantial proportion of children who actually had chest indrawing. These children might have been diagnosed with pneumonia if they had other signs like fast breathing. Failure to identify chest indrawing may lead to underdiagnosis and inappropriate treatment. Sometimes the health workers were good at identifying chest indrawing. For example, in one study, the health workers identified chest indrawing with a sensitivity of 68% , and another study reported a sensitivity of 57% . The median specificity and IQR were 97% and 91%-98%, respectively, with a meta-estimate of 95%. This high specificity indicates that most children who did not have chest indrawing were correctly identified. However, the high specificity does not necessarily mean that health workers’ ability was excellent in excluding non-chest indrawing accurately. The possible reason might be the low prevalence of chest indrawing cases in the study population [39,40].
In studies conducted among children aged 2-59 months, the sensitivities ranged from 33% to 68%, and the specificity ranged from 86% to 98%. Brady et al.  conducted a study with infants aged 0-2 months, reporting a 38% sensitivity and a 89% specificity. We identified another study with neonates that fulfilled our eligibility criteria for this review. However, this study was not selected for the meta-analysis because of the insufficient number of chest indrawing cases . Accurate identification of chest indrawing in young infants is often challenging for health workers. Mild chest indrawing is considered normal, often occurring in healthy infants, as the chest wall is not yet ossified and is more compliant . However, severe chest indrawing is usually thought to be a very deep inward movement of the subcostal tissue and should be easier to identify. This is considered a danger sign for young infants . Further studies are needed to evaluate the ability of health workers to identify this sign in young infants.
In LMICs, pneumonia signs are usually poorly recognized by parents and care-seeking at the facilities is low . Community-based health workers play a key role in identifying pneumonia during household visits. Out of the nine included studies, only one was based in the community and had a single case of chest indrawing . Therefore, the performance of health workers in identifying this sign at the community level could not be evaluated, necessitating large community-based studies.
In all of our included studies, the health workers’ ability to identify pneumonia signs was assessed using actual sick children, which should be ideal. However, children with pneumonia signs are often not readily available, and sick children may need to be treated immediately to safeguard their well-being. Therefore, some studies used videotaped subjects that may not have adequately depicted the signs [42–44]. The findings of those studies might have been different from those with actual children, and we excluded those studies from our review.
A human expert assessing chest indrawing in the same setting is usually considered the reference standard for health worker performance assessment. Seven of the nine included studies used a pediatrician or physician as a reference standard [20,21,32,34,36–38], and the other two did not report the expert’s qualification [33,35]. An expert is thought to be more accurate, but can over-assess or under-assess pneumonia signs. Hence, using expert assessment as a reference standard itself raises questions due to doubtful precision. Child assessment could be videotaped, and this video could be interpreted systematically by a panel of experts  or by a video-based automated system . This could be an ideal reference standard for evaluating health workers in future studies and further research is needed in this area.
This review has several limitations. First, most of the selected studies were conducted in Africa, while two were in Asia and one in Oceania. This may limit the generalization of our findings to other LMICs. Second, most studies were conducted in the 1990s. We cannot determine anything about the current generation of health workers from our findings. We had identified some recent studies but could not include them in this review, as disaggregated chest indrawing data was irretrievable, even after we contacted the corresponding authors. Third, the health workers were trained before their assessment, which could affect the review results . Their performance may change over time from training. The performance in the study might be better due to the direct observation by the evaluator  and might not reflect the health workers’ day-to-day performance. Lastly, health workers often assessed chest indrawing as a part of a larger study. Those studies may not have provided enough information on chest indrawing for being selected in this review or included in the meta-analysis.
We provide evidence on the necessity of improving the performance of health workers in identifying chest indrawing pneumonia. The health system constraints in LMICs include a lack of mentoring, supervision, and continuing development program for health workers. Additionally, there are limited auditing and quality improvement processes to evaluate the program . The health workers’ performance could be improved by training combined with job aides, supportive supervision, regular performance evaluation, and feedback for those who have a poor ability to recognize chest indrawing . A well-functioning monitoring process can identify health system constraints and can improve their performance . The development of a video-based automated method for chest indrawing assessment  for health workers might be useful for identifying pneumonia cases in LMICs. An appropriate non-biased reference standard should be applied to assess health workers’ performance in identifying chest indrawing pneumonia.
Through this review, we found that the performance of non-physician health workers in LMICs was relatively poor in identifying chest indrawing pneumonia. They could identify chest indrawing with poor sensitivity and reasonable specificity, showing a need for improvement. However, all the studies were conducted quite some time ago. New studies should be conducted to assess a new generation of health workers and to investigate possible reasons behind the challenges in identifying chest indrawing encountered by health workers. Appropriate measures should be taken to improve their performance for accurate diagnosis of pneumonia and appropriate treatment.
Online Supplementary Document
We are grateful to Ruth Jenkins, Academic Librarian, and Bohee Lee, Systematic Review Tutor, at the University of Edinburgh, for their help in developing the search strategy. We also thank RESPIRE collaboration for their contribution in bringing the manuscript to its final shape. The RESPIRE collaboration comprises the UK Grant holders, Partners, and research teams as listed on the RESPIRE website (www.ed.ac.uk/usher/respire).
Data availability: Input data used for the meta-analysis is available in Table S5 in the Online Supplementary Document.