The majority of Verbal Autopsy (VA) work over the past 20 years has relied on lay interviewers to gather the relevant information, which has then been fed to one or more physicians for interpretation and cause of death determination. However, the process of interpretation by physicians is time consuming, diverts highly qualified personnel from other tasks, and gives rise to considerable inter- and intra- observer variation in results.

More recently a probabilistic model for interpreting VA findings has been developed, which uses Bayes’ theorem to relate probabilities of various circumstances, signs and symptoms to likely causes of death. This approach has both advantages and disadvantages – a mathematical model may lack some of the subtle nuances that physicians can offer, but on the other hand can offer 100% consistency of interpretation.

The initial development of the model was described in Byass et al. 2003 (Scandinavian Journal of Public Health 31, Supplement 62, p32-37). This included an initial evaluation against physician- interpreted VA material from Vietnam. Subsequently this InterVA model was refined by an international panel of experts, leading to an improvement in its performance (Byass et al.(2005), Scandinavian Journal of Public Health). More recently the model has been evaluated using data from the Butajira DSS in Ethiopia, the results of which of which will shortly be available.

The current version of the model exists in both a PC- based version and also an implementation for a Windows-   based PDA, which can be used for direct data capture at the VA interview stage, with immediate interpretation of the likely cause(s) of death.

Further information and opportunities to download the computer models are available at www.interva.net