Verbal autopsy as a tool for identifying children dying of sickle cell disease: a validation study conducted in Kilifi district, Kenya

Sickle cell disease (SCD) [1, 2] represents a growing health problem both in Africa and in populations of African origin [3]. In a recent study, we estimated that, currently, more than 300,000 children are born with SCD worldwide every year [3]. Approximately three-quarters of these births occur in sub-Saharan Africa (SSA), where facilities for the diagnosis and treatment of SCD are few. As a result, between 50 and 90% of children born with SCD on the continent die undiagnosed in the first 5 years of life [4]. The combination of these high birth rate and mortality figures mean that SCD currently accounts for more than 6% of all deaths among children younger than 5 years in SSA [5].

Poor awareness of SCD among health leaders, combined with a lack of facilities for proper diagnosis and care, means that few data are available regarding such basic questions as the modes and age profile of deaths among children with SCD throughout much of SSA. Recognizing this knowledge gap, we investigated whether verbal autopsy (VA), in which a structured questionnaire is administered to the relatives of deceased people, with a view to ascertaining the probable causes of death, might be one potential approach to investigating the contribution of SCD to childhood deaths in areas where facilities for diagnosis are suboptimal. Here we present data from a study in which we investigated the utility of VA as a tool for diagnosing SCD among children dying in Kilifi County on the coast of Kenya.

Although common, SCD remains neglected in many parts of SSA [18]. Few countries have programs for the early diagnosis or specialist treatment of SCD, and the majority of those affected therefore continue to die undiagnosed during early childhood [4]. Although infections, including bacterial diseases [19] and malaria [7, 20, 21], are widely regarded as probable causes of this early mortality, to a large extent the natural history of SCD in African populations remains poorly documented [18].

In this study, we investigated the question of whether VA might prove a useful approach to the identification of deaths due to SCD in populations for which more robust data on COD are not available. If VA does prove reliable then it could provide a useful tool for improving knowledge on the epidemiology of SCD in resource-poor settings, which could also be helpful both for advocacy and for public health planning.

Consistent with previous studies [22–28], we found that PCVA performed better for some diagnoses than for others. A single previous study was conducted in the same population, which aimed to validate VA against COD data from the hospital wards [25]. In that study, VA was associated with the following sensitivities, specificities and positive predictive values with regard to the attribution of CODs in children older than 1 month, which were broadly similar to those in the current study: acute respiratory infection 28%, 91%, and 29%, respectively; meningitis 38%, 94%, and 20%; and malnutrition 89%, 96%, and 87%. Conversely, the sensitivity of VA for the attribution of malaria was somewhat higher in the current (83%) than in the previous study (46%), although specificity and positive predictive values were similar in both (90% versus 89% and 50% versus 57%, respectively). Measles, a common COD in the earlier study, did not feature among the top 10 diagnoses in the current study, whereas HIV did. Some of these differences might be explained by changes that have been made to the VA tools since the earlier study was conducted, but there have also been significant changes in the patterns of morbidity and mortality in the community in the intervening period. Malaria, previously the single most common cause of admission and death at KDH, has declined significantly in recent years [29], universal childhood vaccination against measles has been introduced, and the HIV epidemic has become well established.

The main focus of the current manuscript, however, is SCD. We found that COD attribution by VA for this diagnosis was consistently good from a number of perspectives. First, agreement between coders was considerably higher for SCD (97.3%) than for all other diagnoses combined (54.6%) (Table 1). Second, agreement between the InterVA-4 model and hospital COD was considerably higher for SCD (96.9%) than for other diagnoses combined (42.5%) (Table 1). Third, agreement between PCVA data and hospital COD was considerably higher for SCD (95.5%) than for other diagnoses combined (50%) (Table 1). Finally, agreement between VA coded data and laboratory-proven SCD was also high at 87.5% (Table 1). Moreover, the proportion of all deaths that were attributed to SCD by VA (6%) was in broad agreement with that predicted from other studies in Africa [5], and is also consistent with that we would predict within the Kilifi population. During the period covered by this study, overall mortality among children aged 1 month to 15 years was in the region of 40/1000, while around 0.8% of all births are affected by SCD [19]. Assuming that 50% of these children die during the first 5 years of life, and that most of these deaths occur after the first month of life [4], SCD should account for around 10% of all deaths among children of this age. Taken together, therefore, these data suggest that the VA method can deliver a plausible estimate of the burden of SCD in our study population.

It is noteworthy that SCD did not feature as a diagnosis either in the previous validation study conducted in Kilifi [25] or in other similar studies [24, 27]. The most likely explanation is that no specific code for SCD was included in the shortlist of codes used in these previous VA studies. This is certainly true for the Kilifi study, where deaths from SCD will most likely have been distributed between the classifications 'anemia' and 'others'. Similarly, no specific code for SCD was included even in much larger validation studies such as those published by Setel and colleagues [30] or by Murray and colleagues [31]. Of note, in the latter study, the authors stated that their cause list was constructed based on the WHO Global Burden of Disease estimates of the leading causes of death, potential to identify unique signs and symptoms, and the likely existence of sufficient medical technology to ascertain gold standard cases. This is of particular concern because it reflects the low visibility that SCD currently holds in the eyes of the international health community, despite accounting for rising proportion of declining deaths in children under 5 years of age. In the current study, we used the longer ICD-10 category list for our COD assignment by both PCVA and by hospital physicians, and the WHO 2012 cause list for our COD assignment by InterVA-4. While it is obvious that the absence of a code for SCD in previous code lists will have led to a systematic underestimation of SCD as a COD (to the point of absence) in previous studies, we cannot exclude the possibility that the use of these longer cause lists might have led to an overestimation of SCD as a COD in the current study; further work will be required to investigate this possibility in the future.

During the current study we noted that no specific questions relating to SCD were included in the 2007 WHO VA forms, and that the attribution of SCD as the underlying COD both by PCVA and by the InterVA-4 model therefore leant heavily on the free-text element of the questionnaires. Increasing awareness of the importance of SCD as a major COD in the African context from a number of recent studies [4, 5, 19, 32], and provisional data from this study has now led to the inclusion within the 2012 version of the WHO Verbal Autopsy Instrument of a specific question regarding SCD [11]. Similarly, a specific code for SCD has also now been included in the list of CODs formulated by WHO for VA use, and can also be assigned by the InterVA-4 model. The full process by which this decision, and the decision to introduce or to drop other specific CODs from the 2012 list, were made have been described recently by Leitao and colleagues [33].

Although our study provides new evidence that VA has potential as a tool for identifying deaths due to SCD, it is open to a number of criticisms. Perhaps most importantly, SCD constitutes a major research interest at our program [7, 13, 19, 20] and, as a result, it is possible that our coding clinicians are unusually well versed in the signs and symptoms of SCD. We aimed to mitigate against this bias by applying the InterVA-4 model, which at least guarantees consistency and comparability with possible future studies elsewhere. Similarly, in order to validate our COD assignment by VA, we used as our gold standard COD assignment by a clinician for the subset of deaths that occurred in hospital. We cannot be sure that the distribution of CODs among deaths that occur in hospital is fully representative of CODs overall, and this could potentially lead to an unrecognized and unmeasurable bias. Nevertheless, we suggest that our study would benefit from replication and validation at other sites.

We have previously estimated that approximately 240,000 children are born with SCD annually in Africa alone [3] and that 50-90% of these children die undiagnosed in their first few years of life [4]. Others have estimated that this means that SCD accounts for around 6% of all child deaths in SSA [5], placing SCD firmly on the list of leading causes of death, at least in countries where the gene frequency is high. This is recognized by the welcome inclusion of the hemoglobinopathies in the latest round of the Global Burden of Disease Study [34]. Nevertheless, much more remains to be done to raise the profile of SCD on the global health agenda.

In summary, we have presented data suggesting that VA might provide a useful approach to investigating the contribution of SCD to childhood deaths in one part of SSA. We hope that our study will prompt similar work in other parts of Africa, including the further exploration of automated approaches to VA coding that might be generalized to other settings.

CN reviewed the literature, analyzed and interpreted data, and wrote the paper. EB contributed to the study design and helped to edit the final version of the paper. VN helped in interpreting the data and editing of the paper. GM helped to code the VAs, and helped in interpretation of the results and editing of the paper. AMA, PK and SK helped with VA data coding and editing of the paper. GN helped with data management and editing of the paper. AM helped with laboratory data management and editing of the paper. PB helped with interpreting the data and editing the final version of the paper. TNW conceived the study and edited the final version of the paper. All authors read and approved the final version of the manuscript.