Basic structure of the epidemiological model – a “horizontal” component

The experts who developed this model agreed that, in order to understand the potential effectiveness of the new diagnostic tests on neonatal sepsis mortality, we need to follow all neonates born within one year in the world “horizontally”, as they progress through the first 28 days of life. We also need to “branch” this progress and separate them into different ‘pathways’ every time a change in their risk of dying from neonatal infection occurs because of their personal circumstances. This would result in a horizontally laid “decision tree” in which different neonates follow different “branches” (pathways), each one of them carrying different risk of mortality from neonatal infection. However, the key challenge would be to understand how many neonates pass through each those branches each year globally, and what could be done at different levels of this “decision tree” to divert them from one pathway to another using information provided by a point-of-care diagnostic test. Application of the test should ensure that more neonates are diverted into branches with lower risk of dying, and less of them are remaining on pathways to higher mortality risk.

However, it has been noted that we also need to know the distribution of neonatal sepsis deaths by place in health system where they occur: ie, how many deaths are untreated and occur at home because neonates either do not seek care, do not have access to care, or their parents do not accept care; then, how many deaths occur after treatment is provided by a community health worker (CHW), at the level of primary health care facility (PHC), or within secondary health care facility (SHC). Once an infection occurs, the position of neonates within health systems of their countries will significantly affect their risk of dying, based on the ability of health system to reach them and provide effective treatment.

All the experts eventually agreed that a useful model should incorporate both a “horizontal” and a “vertical” component. The ‘horizontal’ one would track the progress of all neonates globally through time, ie, their first 28 days of life, and separate them into different “pathways” every time a change in their risk of dying from neonatal infection occurs because of their personal circumstances. The ‘vertical’ one would track their spatial circumstances, ie, their position within the health systems of their countries, and separate them into different categories based on the ability of health system to intervene and help them.

Figure 1 shows the expert’s consensus on the basic structure of the epidemiological model. The logic of horizontal branching is reflecting any change in mortality risk of the neonates. The ‘tree’ starts, on the top of the left hand side, with 136 million neonates who were born globally in 2010. The first branching occurs when there is a sign or a symptom of suspected infection. Clearly, neonates with no signs or symptoms (the bottom branch following the second node) will not be under risk of dying from infection, because we assumed that each death would always be preceded with at least some sign or symptom of suspected infection. Those who develop signs/symptoms of suspected infection during neonatal period should be separated in the upper branch of the tree, following the second node. This pathway then branches further, depending whether the parents choose to seek care or not. This is because, all other things being equal, those who seek care are expected to reduce the risk of dying of their neonate in comparison to those who do not. Having agreed on this, we also recognized that the assumption of all other things being equal between non-care seekers and care seekers is a major assumption. Those with more severe presentation, and those with more highly educated parents, wealthier parents, and who live closer to health care facility will be more likely to seek care. This is further discussed in assumptions of the model (see next subheading).

Figure 1.  Basic epidemiological model showing possible pathways to deaths from neonatal infections globally and four points of possible interventions with diagnostic tests. CHW – community health worker; PHC – primary health care physician; SHC – secondary health care.

Figure 1 further shows that those who do not seek care may survive or die, while those who seek care may eventually reach different level of health system, depending on their access to care, the advice given by CHWs and primary care physicians, their compliance with this advice, their resources to pay for travel to higher-level facility, and the severity of their child’s illness. Eventually, care-seekers may end up self-treating the child, receiving care from CHWs, PHC physicians, or reach the hospital (Figure 1). The model is only interested in the highest level of care obtained. Self-treatment or treatment provided by CHWs have no further branching – the outcome is either survival or death from infection.

In an attempt to simplify the model, some experts questioned the reasons for splitting the pathway of those who self-treat, or receive treatment from CHWs, from those who do not seek any care. We have two justifications: first, we may assume some limited effectiveness of treatment attempts by non-health professionals, which reduces the risk slightly in comparison to those neonates to whom no care was offered; second, if point-of-care diagnostic tests became available, the parents who seek care either from CHWs or through self-treatment would be considerably more likely to accept testing and comply with its results, in comparison to those parents who do not seek care at all. The lack of care seeking can be caused by a number of reasons, including (but not limited to) lack of access, failure to recognize symptoms, insufficient funding and other context-specific and cultural reasons.

The next branch in Figure 1 shows parents who reach primary health care facility, where we assumed that at least some antibiotic treatment would be available. However, the attending physician may decide that clinical algorithm for the diagnosis of neonatal infection does not provide sufficient indication for antibiotic treatment, thus sending the parents and the child back to their home (Figure 1, lower branch). There, the child may get well and survive, or they may die if physicians made an incorrect diagnosis. In cases where physicians provide antibiotic treatment (Figure 1, upper branch), the parents may either comply or not comply. Although the lower branch (‘non-compliance’) is unlikely among parents who initially sought care, we decided to leave it in the model to account for this theoretical possibility in cases where PHC physicians or nurses provide active community-based surveillance for neonatal infections, and where non-compliance may become an issue. The final branch at the primary health care level follows the fate of those neonates who received antibiotic treatment. They may be released home following the treatment, or be told by a concerned physician to proceed and seek care at the nearest hospital. Those who comply with this advice and can afford the trip will join other parents who went to hospital directly (Figure 1).

This creates a mixture of as many as four different groups of neonates seen at the hospital level (Figure 1): (i) some will present there regardless of the severity of their symptoms, simply because the access to hospital care is easy for their parents; (ii) others will be brought directly to hospital by those parents who were the most concerned, usually because of very severe symptoms, and these children bypassed all other levels of care; (iii) the third group will consist of children who already received antibiotic treatment at the PHC level, but the physicians were sufficiently concerned to refer them to SHC facility, and the parents complied with the suggestion; (iv) finally, the fourth group is also defined as the neonates born in hospital who were never released home after birth because of a suspected infection (see rationale for this later). The first group will have the lowest risk of dying, the second group the highest, the risk in the third group will probably somewhat lower than in the second group (because treatment has been initiated early), and the fourth group will have intermediate risk (early initiated treatment will reduce the risk, but the etiological spectrum of sepsis in the first 3-5 days of life is more life-threatening). The diagnostic procedures at the hospital will lead to the decision to either treat or release without treatment, and in both branches there will be neonates who will either survive, or eventually die from neonatal infection (Figure 1).

Points for intervention using point-of-care diagnostic tools – a “vertical” component

The previous sub-heading explained the basic structure of the epidemiological model in terms of its ‘horizontal’ component. However, in order to understand the possible points of intervention using point-of-care diagnostic tool we also need to consider a ‘vertical’ component – ie, where do infected neonates find themselves within the health system once they develop symptoms of infection, and what is the highest level of care accessible to them. Figure 1 shows the four main points of possible intervention from health systems’ perspective. At those four points, designated as ‘no screening’ in Figure 1 (implying that under the current ‘status quo’ situation no point-of-care diagnostics are routinely available), information obtained from a highly sensitive and specific diagnostic tool could change the fate of the neonates who present themselves to that particular point of care. Based on the information obtained through the use of diagnostic tool those neonates could be ‘re-directed’ into another pathway, in which their risk of mortality would be lower. In this way, the point-of-care diagnostic tool would reduce mortality from neonatal infections, albeit indirectly, and this reduction could be quantified.

Figure 1 shows that, conceptually, the four key points where diagnostics could be introduced are: (i) diagnostic testing (‘screening’) of seemingly healthy neonates for signs of invasive bacterial infection immediately after birth, and then subsequently throughout the neonatal period; (ii) diagnostic testing of those neonates who developed symptoms or signs of neonatal infection at the community level, either by parents or community health workers; (iii) diagnostic testing of neonates with suspected infection who sought care at the level of primary health care; and (iv) diagnostic testing of neonates with suspected infection who sought care at the level of secondary health care. Clearly, each of these four points for potential diagnostic testing would have its own characteristics in terms of skill and capacity of implementers of the test, feasible complexity and affordable cost of the test, and supporting laboratory requirements. The four points present a spectrum, in which community-based diagnostic screening would require very low cost and complexity of the test, very low level of skill of the implementers, and no laboratory requirements. At the other end of the spectrum, hospital-based tests could potentially have somewhat higher cost, greater complexity, and allow for greater skill required from the implementers and some laboratory requirements.

Finally, once that ‘horizontal’ and ‘vertical’ component of the model has been defined, we still needed to identify the potentially useful characteristics of a diagnostic test itself. The key question posed to the experts was: what information should a diagnostic test ideally provide at each of the four different points of care defined above to reduce the risk of mortality among the tested neonates? The experts agreed that, conceptually, there would be three general tiers of diagnostic tests of potential interest, and only some would be useful at each of the four points of care.

The first type of point-of-care diagnostic test would need to be able to identify the presence of invasive bacterial infection in a neonate with suspected infection. This test would require high degree of sensitivity and specificity, and its usefulness would be in establishing an indication for immediate antibiotic treatment. This test would be useful at all four levels of health systems – community level (as a screening among healthy neonates or intervention among possibly infected neonates), primary health care level and hospital care. This test would also empower community health workers to apply antibiotic treatment, which could dramatically expand coverage of community case management with oral antibiotics – proven to be effective even when implemented by CHWs [9,10].

The second type of diagnostic test should be able to identify the exact bacterial (or viral) pathogen that caused the infection. This information would prevent treatment with antibiotic which would not be effective against that specific pathogen, or which may fail to cure an infected neonate due to antibiotic resistance. Clearly, this type of test would not be particularly useful for screening of healthy neonates, or even for testing the neonates with suspected infections at the community level. This is because it is unlikely that CHWs would have the spectrum of low-cost and higher-cost antibiotics available to them in the present global context, and also necessary skill to apply them properly based on test results. However, this test may be very useful to a PHC-level physician, who may have a wider choice of antibiotics, and whose rapid and effective response to neonatal infection with appropriate and effective antibiotic may indeed save lives. Similarly, this test could also be useful at the hospital level, especially among severe cases who require intensive care and where a rapid and effective treatment is critical.

The third type of diagnostic test should imply that the episode is likely to become severe and require hospital care. This information would be expected to not only indicate immediate antibiotic treatment, but also prompt care seeking directly at the level of hospital care. The experts debated whether this test would really be useful, or whether any invasive bacterial infection among neonates should be considered life threatening and require hospital referral. Eventually, they agreed that it would be useful to have a test that implies development of a severe or very severe episode, because those children are most likely to die and rapid transfer to hospital-level intensive care unit could perhaps avert many of those deaths. With respect to the level of health system at which this test could be useful, the experts agreed that this test would not be feasible for screening of healthy neonates because they would be unlikely to appear healthy if they had biochemically detectable signs of severe or very severe infection. The test could be useful for those with suspected infection at the community level, because it could make a difference whether a CHW would suggest referral to PHC or SHC. It would also be useful to PHC physicians, to help them decide whether to suggest further hospital referral. It could be used in the hospitals to assist triage procedures when assigning sick neonates to intensive care units.

Definitions and assumptions underlying the proposed epidemiological model

The PATHS tool is an epidemiological model based on decision trees that describes pathways to deaths in a defined population. In this particular example, the population are world's neonates in one year and the deaths are caused by neonatal infection. Therefore, we first need to define what do we mean by ‘neonatal infection’. The experts agreed that these should actually be termed more precisely – as “suspected potentially life threatening invasive bacterial infections” (SPLTIBI). This definition obviously excludes all skin infections, tetanus, diarrhoea, malaria, HIV/AIDS and other non-invasive or non-life threatening infections in neonates. For the ease of presentation, however, we will simply continue to refer to SPLTIBI as ‘neonatal infections’ in further text.

An important suggestion from the experts was that, given that the etiological spectrum and severity of neonatal infections differ considerably between the first 3 days of life and the period between 4-28 days of life, and that the number of neonatal deaths globally are split almost equally between those two categories, perhaps we need two separate epidemiological models – for “early” vs “late” neonatal infections? After discussing this issue at some length, the experts agreed that fundamentally the severity of infection and reaction of the parents/carers will have similar consequences in both time periods, and that the pathogens that can be responsible for infection are not exclusively found in only one period – they are just more likely. Therefore, we agreed that early and late infections are not clearly separate entities from the etiological perspective and we agreed not to develop two separate models. Finally, as it will become apparent in later text, the information available for different model parameters was very sparse. Therefore, introducing an additional separation in the model would produce less reliable estimates. For similar reasons (lack of specific information), we chose not to separate neonates according to their place of birth, birth weight and other risk factors that might affect the development, progression and outcome of neonatal infections.

For similar reasons, we made no distinction between “early vs. late” sepsis, “advanced vs. non-advanced” sepsis, or “sepsis vs. severe sepsis”. In the early versions of the model we had two defined steps in the sepsis progression: sepsis and severe sepsis. We assumed that all sepsis deaths will derive from the severe sepsis phase. However, the experts at the third TAG meeting advised that the severe sepsis step is not necessary because sepsis is a very severe disease and that a high proportion of sepsis cases will lead to death if left untreated. This is why we considered all cases of neonatal sepsis as a dangerous and life-threatening condition which requires treatment and hospital care. Also, we used ‘no action’ abundantly in the model, although this term may not be strictly correct, because there’s usually always some action at some stage, although it may be too late to change the course of the disease.

All hospital births after which sepsis was suspected at hospital and that were not released home for this reason were presented as “seeking care at the SHC level”, because this was much simpler to present and easier to understand. This approach essentially defines parameters 6-9 in Table 1 as ‘the highest level of received care’. Also, the experts agreed that parameters 14-21 in Table 1 are all needed, because diagnosis at home by parents, by CHW, or by doctors at the PHC or SHC level all have their different sensitivity and specificity. Similarly, the treatments that they provide will have different effectiveness, as captured by parameters 10-13 in Table 1. The experts argued whether sensitivity at the PHC level may increase due to multiple visits, but this showed not to be of sufficient importance to the model's output to justify a special consideration.

Also, we treated each episode of suspected sepsis as an individual one, with no “backward loops” of multiple episodes per neonate. This is because a possible increased incidence risk and/or case fatality ratio (CFR) of neonates with multiple sepsis episodes is reflected in the overall CFRs. Essentially, multiple episodes don’t matter because the affected neonates must survive all previous episodes to experience the last one that occurred within the first 28 days of their life. We also assumed that there would be no increase in the effectiveness of interventions for neonates that visited more than one type of point of care. Therefore, the effectiveness of intervention for each neonate was the one provided at the highest point of care. This is particularly relevant to neonates who received antibiotic treatment at the PHC level but still proceeded to seek care at the CHC level. We assumed that PHC doctors referred all the cases of sepsis that they recognised based on their clinical algorithm to a SHC facility after one course of treatment.

Experts suggested that one arm should perhaps follow the neonates treated with oral, and another one with parenteral antibiotics. However, others stated that trials show very high effectiveness with either one of those two treatment types, albeit this is actually a comparison between an earlier oral treatment of less severe sepsis with a later treatment of more severe sepsis with parenteral antibiotic, which may not be directly comparable. However, for simplicity of messages the difference in administration of antibiotics was not taken into account.

Experts noticed that there would be neonates who would seek care, but wouldn't actually be able to access the level of care that they wanted to seek. Some of them are treated in the PATHS tool as those who did not get care. The experts questioned whether they should have been termed as “obtaining / not obtaining care” rather than “seeking / not seeking”? Given that this difference doesn’t mean anything to their outcome, we kept the terminology as “not seeking” for clarity, although the point is valid. Similarly, we termed some parents who sought care at the PHC level and were referred to SHC (but could not comply, although they probably wanted to) as “did not comply”, although the more correct term would be “could not comply”.

Experts commented that care seeking is likely to be greater in cases that are apparently more severe and asked whether this is somehow factored in the model. However, this notion would then need to be attributed to parents’ own diagnostic algorithms, which is captured in sensitivity and specificity of diagnosis at the home care level. The same is true at the CHW and PHC level.

Finally, experts suggested that there may be some hidden correlations or interactions between the parameters used in the model and wondered if eg, a 50% change in care seeking would need to be corrected for actual access, etc.? However, since all parameters in the tool are related to specific branches of the decision tree and each one can be modified, the users can account for these possible correlations wherever they use the tool and think that there is a reasonable case to do this.

Agreed values for model parameters based on evidence and expert opinion

In this part of the paper, we will explain how we decided on the most likely value of the key model parameters, using both evidence and expert opinion, and being constrained by the global estimate of 907 000 neonatal deaths in 2008.

1. Number of neonates. Data for this model parameter were obtained from the World Health Organization’s (WHO) website [15]. The annual total number of neonates was estimated to be 136 241 000 and this estimate was based on the 2008 WHO data (Figure 2, Table 1).

Figure 2.  Parameters of the model relevant to the number of neonates each year globally, proportion of those who develop symptoms of suspected infection, care-seeking behaviour and case-fatality rates (CFRs) in different pathways. (CFR for “neonates” represents the absolute risk of dying during the first month of life; all other CFRs are case-fatality rates among neonates with suspected infections or real sepsis).

2. Proportion of neonates with symptoms of suspected infection. The only estimate of larger geographic relevance that we could find was based on a review by Thaver and Zaidi published in 2009 [16]. Their paper reviewed studies published in the early 2000s and this parameter was estimated to equal about 0.17 [16]. However, most technical experts thought that this estimate would be too high globally for the year 2008. Therefore, we performed Delphi exercise and took the median of what the total of 12 experts reported (see their list in acknowledgement). The final value in our model was set to 0.12 (Figure 2, Table 1).

3a. Proportion of neonates with symptoms of suspected infection that sought care. Data for this indicator were reviewed and summarised by Chandran et al. [11] and Herbert et al. [12]. In total, 22 studies (18 from South-East Asia and 4 from Africa) were identified that reported care seeking behaviour for newborn infections. Care seeking outside of the home ranged from a lower bound of 0.10, an upper bound of 1 and a median of 0.69 (Figure 2, Table 1).

3b. Proportion of neonates with symptoms of suspected infection that didn't seek care. We reviewed the information for this indicator separately. However, the data assembled was of a much lesser quality compared to the data that were assembled for the indicator of proportion of neonates with symptoms of suspected infection that sought care. Therefore this indicator was set to be equal to 1 minus the proportion of neonates with symptoms of suspected infection that sought care (1 – 0.69 = 0.31) (Figure 2, Table 1).

4. Proportion of neonates with symptoms of suspected infection that developed sepsis. According to Thaver and Zaidi [16], 27.5% of neonates with suspected infection developed laboratory confirmed sepsis. However, the experts in the third TAG meeting thought that this estimate was too high and did not reflect the current situation. Based on the pneumonia model by Rudan et al. [13,14], 8.7% of pneumonia cases have severe disease and therefore we decided to use this estimate to reflect the amount of neonates with suspected infection that will develop sepsis (Figure 2, Table 1).

5. Case fatality ratio of untreated sepsis. There is no evidence of what the CFR of untreated neonatal sepsis is. Neonatal sepsis is a very severe disease and the experts in all three TAG meetings agreed that a CFR of 80% will reflect the current situation in low income settings (Figure 2, Table 1).

6. Proportion of neonates with symptoms of suspected infection that sought care received treatment from their carers at home (self treatment). Data for this indicator were reviewed and summarised by Chandran et al. [11] and Herbert et al. [12]. In total, 3 studies (2 studies from SE. Asia, 1 study from Africa) were identified that reported the proportion of neonates of those that sought care that received home treatment for newborn infections with a lower bound of 0.06, an upper bound of 0.37 and a median of 0.16 (Figure 3, Table 1).

Figure 3.  Parameters of the model relevant to the highest level of health system to which access has been realized for the neonates with suspected infection who sought care. CFR – case-fatality rate in each pathway.

7. Proportion of neonates with symptoms of suspected infection that sought care and received treatment from community health workers (CHW treatment). In total 4 studies, all from SE. Asia identified reported the proportion of caregivers who accepted care from CHWs for their ill neonates with a lower bound of 0.01 and an upper bound of 0.72. However, since these studies addressed care acceptance and not care seeking, this indicator was set to be equal to 1 minus the proportion of neonates that sought care at home, at a primary health care (PHC) facility and at a secondary health care (SHC) facility (1-0.16-0.37-0.32 = 0.15) (Figure 3, Table 1).

8. Proportion of neonates with symptoms of suspected infection that sought care and received treatment in a primary health care facility (PHC treatment) or from a medically trained provider. Data for this indicator were reviewed and summarised by Chandran et al. [11] and Herbert et al. [12]. In total 11 studies (9 studies from SE. Asia, 2 studies from Africa) were identified that reported the proportion of neonates of those that sought care that received treatment for newborn infections in a PHC facility with a lower bound of 0.04 and an upper bound of 0.83 and a median of 0.37 (Figure 3, Table 1).

9. Proportion of neonates with symptoms of suspected infection that got treatment in a SHC treatment. Data for this indicator were reviewed and summarised by Chandran et al. [11] and Herbert et al. [12]. In total 6 studies (5 studies from SE. Asia, 1 study from Africa) were identified that reported the proportion of neonates of those that sought care that received treatment for newborn infections in a SHC facility with a lower bound of 0.04 and an upper bound of 0.66 and a median of 0.32 (Figure 3, Table 1).

10. Effectiveness of home care interventions. Although it is unlikely that the home treatment interventions will have a major effect on neonatal sepsis (especially if they are based on traditional methods), the experts thought that the effectiveness of home treatment should be set equal to 0.10 as they may use antibiotics prescribed to their children previously that are still effective (Figure 4, Table 1).

Figure 4.  Parameters of the model relevant to the outcome of neonates with suspected infection whose parents applied self-treatment. CFR – case-fatality rate in each pathway.

11. Effectiveness of CHW interventions. The experts thought that the effectiveness of CHW treatment interventions will depend mainly on whether CHWs have access to antibiotics or not. However, they thought that access to antibiotics of CHWs is limited in the developing countries and therefore it is more likely that they will use traditional based methods. The consensus was that the effectiveness of the CHW interventions should be set to about 0.15 (Figure 5, Table 1).

Figure 5.  Parameters of the model relevant to the outcome of neonates with suspected infection who received some care from the local community health worker (CHW). CFR – case-fatality rate in each pathway.

12. Effectiveness of PHC interventions. In a review by Theodoratou et al. [17] we found that the effectiveness of antibiotics on childhood pneumonia in a community setting was 35% and after accounting for compliance this estimate was scaled up to 70%. We assumed, however, that the effectiveness of antibiotics for neonatal sepsis will be lower since: (i) neonatal sepsis is a more severe disease and (ii) the administration of antibiotics to neonates is more challenging, which could compromise their effectiveness. There are also issues related to access to correct antibiotics and antibiotic formulations and their availability in low-resource settings, that we reviewed separately in the paper by Lee et al. [18]. Therefore, after discussion with the experts at the third TAG meeting we set the effectiveness of PHC interventions to 0.35 (Table 1).

13. Effectiveness of SHC interventions. The experts thought that the effectiveness of the interventions used in a SHC facility will be a mixture of first- and second-line antibiotics depending on the setting and country. However, they thought that the effectiveness of the interventions in SHC facilities should be closer to the effectiveness of community administered antibiotics for pneumonia after accounting for compliance and this parameter was set equal to 0.60. Then, it was further revised to 0.58 in order to “fit” the total number of deaths in this model to the “envelope” of 907 000 defined by Black et al. [1]. This was the only parameter that was slightly altered in order to achieve the perfect fit of the number of deaths to the envelope (Table 1).

14-15. Sensitivity and specificity of clinical algorithm of carer to detect sepsis at home care treatment. The experts at the third TAG meeting decided that there was no evidence to suggest that the ability of the parents to recognise the symptoms of true sepsis was better than a random guess and therefore the sensitivity and specificity of the clinical algorithm that the parents used to decide whether to self treat the newborn or not were set equal to 0.50 and 0.50 respectively (Figure 4, Table 1).

16-17. Sensitivity and specificity of clinical algorithm of CHW to detect sepsis. The experts at the third TAG meeting decided that the ability of a CHW to diagnose sepsis was quite low, but still higher than the ability of the parents. Therefore, sensitivity and specificity of the clinical algorithm that the CHWs used to decide whether to treat the newborn or not were set equal to 0.60 and 0.60, respectively (Figure 5, Table 1).

18-19. Sensitivity and specificity of clinical algorithm of medical doctor at PHC facility to detect sepsis. The experts at the third TAG meeting agreed that the ability of a medical doctor to diagnose sepsis at PHC facility is better than the ability of CHW or parents and they thought that a sensitivity and specificity of 0.70 and 0.70 reflects the situation in 2010 (Figure 6, Table 1).

Figures 6.  Parameters of the model relevant to the outcome of neonates with suspected infection who received care from a primary health care physician (PHC). CFR – case-fatality rate in each pathway; SHC – secondary health care.

20-21. Sensitivity and specificity of clinical algorithm at SHC to detect sepsis. The experts at the third TAG meeting decided that the ability of a medical doctor to diagnose sepsis at a SHC facility is better than the ability of a medical doctor in a PHC (based on experience and number of severe cases that are reviewed in SHC facilities compared to PHC facilities) and they thought that a sensitivity and specificity of 0.90 and 0.70 reflects the situation in 2010 (Figure 7, Table 1).

22. Compliance of parents with antibiotics prescription from PHC doctor. The experts at the third TAG meeting suggested that the compliance of the parents with antibiotic prescription at a PHC facility will be high and the consensus was that this would be 0.90 (Figure 6, Table 1).

23. Proportion of parents that comply with referral to SHC from PHC doctor. The proportion of parents that will comply with a referral to SHC facility by a PHC doctor will depend on several factors including: (i) proximity or even existence of a SHC facility and (ii) ability (mainly financial) and willingness (determined mainly by social factors) of the parents to travel the distance between the PHC and SHC facilities. There was no hard evidence for this indicator and our estimate of 0.60 was based on the experts opinions mainly based on their experiences in the field (Figure 6, Table 1).