The aim of this study was to assess whether physicians' access to free-texts explains the large discrepancy with the InterVA-3 model in malaria diagnoses in this holo-endemic malaria region, not to judge whether one method or the other is superior in its accuracy of malaria diagnosis. Overall we found that access to free-text does not explain why physicians diagnose more malaria than does the model. Nonetheless, information from free-text was found to provide additional information not available in the structured portion of the questionnaire, which resulted in non-negligible changes in the model output.
Not surprisingly, the highest proportion of malaria diagnoses by both the physicians and the model was amongst under five year olds and during the wet season. However, both methods diagnosed malaria in some age and sex groups in ways contradicting the standard knowledge on malaria. Physicians diagnosed malaria within the first four weeks of life, generally accepted not to be possible , while such diagnoses did not occur in model output. Also, it was notable that amongst physicians' malaria diagnoses there was a higher proportion aged 65 and above, and amongst the model diagnoses there was a higher proportion of women of reproductive age. Very little literature exists on malaria deaths amongst the elderly, likely due to lower sample sizes in this age category and other causes of death with greater priority. However, perhaps physicians are justified in diagnosing more malaria amongst the elderly, whose weakened immune systems may be less able to combat the disease than those aged below 65. A review of malaria mortality rates in sub-Saharan Africa and Bangladesh found that malaria death rates drop in late childhood and young adulthood, and then steadily increase with age in West and East Africa . A study in India also suggests malaria mortality rises in older ages, though the epidemiology of malaria mortality may be different in India as compared to Burkina Faso . This trend was not present in the "Global Burden of Diseases and Risk Factors" report, which showed consistently decreasing malaria rates with increasing age, for sub-Saharan Africa as a whole .
The model estimates higher levels of mortality due to malaria amongst women of reproductive age than physicians. Findings by Lemma et al suggest that women of reproductive age are more likely to die of malaria compared with other adults, supporting the trends indicated in malaria mortality in model output .
Although the InterVA has recently been validated with varying types of clinical data, it remains difficult to judge whether physicians or the model may be closer to the real causes of death, as contradicting results were seen under different conditions [14, 22, 23]. Unfortunately, this study suffers from a lack of a gold standard, as do all verbal autopsy-related studies based in developing country communities without full access to health services, and thus the true causes of death are impossible to be obtained . This problem is particularly challenging for malaria, a disease that is difficult to diagnose accurately without parasitic evidence, and from which fatality is much reduced when health services are accessed, rendering in-person validation by a physician near impossible [19, 25–27].
Had the free-text hypothesis been correct, one would have expected a closing of the gap in malaria diagnoses between physicians and the model. Though absolute changes between models were small in both malaria-specific sub-groups, trends tended toward an increase in model malaria diagnosis in the physicians malaria sub-group and a decrease in the model malaria sub-group. This trend was not evident when comparing absolute differences in proportions at population level for the representative sample, though separate analyses by age group revealed increases in malaria diagnoses in the youngest and oldest age groups, but a decrease in 5-49 year olds. However, additional shifts in malaria diagnoses at the individual level may have cancelled each other out at the population level. A more important closing of the malaria diagnosis gap was achieved when setting the model malaria indicator to "yes" when any fever was present, without regard to duration or treatment . This suggests that physicians are highly influenced by presence of any fever when diagnosing malaria, but also shows the sensitivity of the model to the malaria indicator being set to "yes", despite difficulties in obtaining clinical diagnosis of this disease in developing countries.
The most consistent outcome in all three sub-groups in the free-text analysis was a rise in malnutrition/diseases of the digestive system. This is a result of the availability of information on jaundice (a principal symptom of liver disease) in the free-text, an item not captured in the questionnaire. In a further sub-group analysis of the effects of free-text on HIV diagnosis, the sample was too small to see any HIV-related impact, but even in the smaller sub-group, the greatest change was an increase in malnutrition/diseases of the digestive system (data not shown).
A notable pattern was visible in the interplay of three disease categories with similar symptoms to malaria: pneumonia/sepsis, diarrhoea, and meningitis. These infectious diseases all display fever, some cough, and some convulsions (in severe cases [17, 28]). While physicians may struggle to distinguish between these causes without having seen the individual, obtaining laboratory tests, or, in many cases, having full information on the circumstances of death, the model also shows inter-dependent fluctuations, which are visibly sensitive to additional free-text information. In particular, meningitis and malaria appear to have overlapped in diagnosis, the model diagnosing about twice as much meningitis as physicians did in the original model. Indeed, Burkina Faso is one of the core countries of the meningitis belt, where more than 90% of all meningitis cases occur during calendar weeks 1-20, corresponding to the meningitis season, with the typical bell shaped incidence curve peaking by end of March and seen every year at district level . As malaria is endemic throughout the year, during the meningitis season both diseases are present, with the potential for misdiagnosis of the one or the other. In some individuals a bacterial infection and malaria parasitaemia may even occur simultaneously (as seen in children) [30, 31]. The model's season indicator could thus result in too few malaria diagnoses in the dry season.
Though the over-all change in the model outputs was not monumental, consulting the free-text did reduce the number of unallocatable indicators from 37 to 24, in the three sub-groups combined. Additional free-text information seemed to be evenly distributed amongst all age groups, without regard to type of sub-group, indicating no particular benefit of free-text inclusion for any particular sub-population.
Ideally the VA questionnaire used would have been tailored precisely to the indicators such that there would be no unallocatable indicators. However, if a questionnaire does not have information for certain indicators, this analysis shows that it is possible to search in the free-texts for information on missing indicators (assuming free-texts of reliable consistency). A reduction in indeterminate causes also indicates fuller information as a result of the free-text information. The Nouna INDEPTH site is in the process of switching to an improved questionnaire, universally employed across the INDEPTH sites, and hence consulting free-text may have less of an impact for this site. However, with a plethora of VA questionnaires still in use in developing countries, it is undoubtedly useful to observe the impact free-text may have when searching for information not gathered in the structured portion of the questionnaire. Consulting free-text may be especially relevant for studies using historical data, in which data may have been gathered in the free-text not known at the time to be relevant for the structured portion. Thus, even if electronic coding should become the norm for VA analysis, free-texts should be maintained as part of VA questionnaires and interviewers trained on documenting narratives in a consistent manner.