“Nothing is certain except death and taxes” – Benjamin Franklin, 1789

Unfortunately, Franklin was correct in excluding cause of death from the realm of certainty. Whatever method is used to determine the cause of a particular death, there is some degree of uncertainty. Even if a body is put in front of several pathologists, there can well be multiple opinions about cause of death. But for more than half of the world’s deaths today, no expert opinion is even sought or recorded around cause of death. Most of these unrecorded deaths occur in Africa and Asia, where, for the foreseeable future, individual deaths are unlikely to be certified by physicians.

So, enter “verbal autopsy” and computer models for determining cause of death. This is the only sensible medium-term fix for the current lack of knowledge about cause of death in Africa and Asia. Verbal autopsy shares two important characteristics in common with every other method related to cause of death: it is not going to achieve certainty, and it cannot get away from using an inevitable mix of evidence and expertise.

Umeå Centre for Global Health Research has been working on verbal autopsy methods over the last decade, developing the InterVA series of models for determining cause of death from verbal autopsies. These are freely available tools which are now aligned with the WHO 2012 verbal autopsy standard. These tools facilitate interpreting verbal autopsy data into causes of death cheaply, consistently, and reasonably accurately given the general uncertainties around causes of death.

In my recent BMC Medicine paper , I looked at a large-scale verbal autopsy study from the Population Health Metrics Research Consortium (PHMRC). This study captured around 12,000 well-documented hospital deaths, recording cause of death and subsequently undertaking a verbal autopsy. In principle this was a great way to build the evidence base between causes of death and responses to specific questions in verbal autopsy interviews. This evidence base is potentially important, because it can be used to build better models for interpreting verbal autopsy material.

Although the PHMRC database is an important new asset in the quest for better cause of death ascertainment, it also has some limitations. All the cases were admitted to high-level hospitals before they died, and so may differ in some respects from the more usual scenario of deaths in the community in Africa and Asia.

Inevitably, some of the verbal autopsy responses were anomalous – for example pregnancy-related deaths recorded in the hospital for women who were not said to be pregnant in the subsequent verbal autopsy interviews. On the one hand, that represents part of the intrinsic uncertainty around verbal autopsy. On the other hand, if such kinds of anomalous data are used to build models for cause of death assignment, then the models will incorporate the same anomalies.

Unfortunately, perhaps because of the lack of high-quality verbal autopsy data in general, the PHMRC group built new models and then tested them within the same dataset. Thus, in a methodological comparison of these internally validated models against other independent approaches, not surprisingly the internally derived models performed best.

Understanding mortality patterns depends critically on having clear and standardised definitions of causes of death. WHO has led this process for many decades via the International Classification of Diseases (ICD) system. Although PHMRC carefully defined the causes of death used in their study, these did not match ICD-10 criteria in all cases. In my paper I explored the effect of this in relation to premature infant deaths – where relatively small differences in definition led to major differences in cause of death patterns.

In conclusion there are four important points emerging for the future development of verbal autopsy methods:

1. When using empirical data to build verbal autopsy models, it is important to exclude cases which human experts can clearly identify as anomalous or contradictory
2. If a dataset is used to build a verbal autopsy model, that model must be evaluated on other, independent data
3. Ideally – though this may be the hardest to achieve – empirical evidence for verbal autopsy systems should be community-based rather than coming from hospitalised cases
4. Common WHO international standards must be used universally to avoid confusion from differences of definition

The ultimate aim for verbal autopsy is to move out of specific research settings and controlled environments, towards routine implementation on the millions of deaths currently passing unnoticed across Africa and Asia. UN Secretary-General Ban Ki-Moon has noted that “To be counted is to become visible”, calling on decision makers to “make each and every person count”. Verbal autopsy is a critical tool for this moving this process forward.

This post is the third of  a series about verbal autopsy, the subject of a selection of articles just published in BMC Medicine’s Medicine for Global Health collection.