Background

Verbal autopsy (VA) is an old and well-established approach to ascertain cause of death when it is not feasible or practical to conduct medical certification and full autopsies. After a death is identified, a specially-trained fieldworker interviews the caregivers (usually family members) of the decedent. A typical VA interview includes a set of structured questions with categorical/quantitative responses and a narrative section that records the ‘story’ of the death from the respondent’s point of view. The resulting data are interpreted by various means to assign causes to the death.

The most common practice is to have clinically trained, experienced physicians read the interviews and determine causes. To address the fact that physicians frequently do not agree on causes, VA interviews are often read by two physicians, and sometimes three, and the final causes are determined through a consensus mechanism. This implicitly acknowledges two of the challenges inherent to VA:

1. without either clinical data or an autopsy it is difficult to be either specific or certain about the cause of death, and
2. each physician has unique training and experience and is therefore biased in various ways when assigning causes of death.

The first is a fundamental limitation of VA resulting from the fact that VA data contain comparatively less information than clinical records and autopsy. A clear consequence of the second is that traditional physician-assigned VA causes of death are biased and not comparable across either single physicians or groups of physicians.

An alternative to physician review is the use of an algorithmic method that processes the categorical/quantitative responses in VA interviews to identify causes of death. The algorithmic approach has three important advantages:

1. physicians are free to spend their time caring for patients,
2. VAs can be coded very quickly without having to wait for the always-lengthy physician review process, and
3. physician-associated bias is removed from the process so that cause assignments are reproducible and comparable.

I have led a small team to develop a new algorithmic method for coding verbal autopsies. Building on InterVA - 'Interpret VA', the first VA algorithm - we developed InSilicoVA. InSilicoVA assigns a cause to all deaths (no 'indeterminate' causes), provides consistent estimates of uncertainty for both individual causes and population-level cause-specific mortality fractions, utilizes both 'yes' and 'no' answers, and is fully probability-based so that results are presented as true probability distributions.

The team responsible for developing InSilicoVA includes Richard Li, Tyler McCormick, and myself.

Key papers

• The openVA Toolkit for Verbal Autopsies
• Probabilistic Cause-of-Death Assignment using Verbal Autopsies
• Bayesian Joint Spike-and-Slab Graphical Lasso
• Using Bayesian Latent Gaussian Graphical Models to Infer Symptom Associations in Verbal Autopsies
• Bayesian Factor Models for Probabilistic Cause of Death Assessment with Verbal Autopsies
• Estimating Causes of Death Where There Is No Medical Certification: Evolution and State of the Art of Verbal Autopsy

Software and the openVA Team

Software has been developed to implement InSilicoVA and all of the other commonly-used computer algorithms for cause-coding VA data, including R packages and universal executables delivered in a Docker container. This is openVA and has been developed by the openVA Team consisting of Richard Li, Jason Thomas, Peter Choi, Tyler McCormick, and myself, with increasing help from Yue Chu.

• Main openVA site and openVA Toolkit paper
• R Package
• openVA App
• Video for installation of openVA App
• Video for example analysis using openVA App

The VA Interview

Clarissa Surek-Clark has led the openVA Team's thinking around the VA interview. Drawing on her experience as a professional translator and interpreter and training as a sociolinguist, she is curious to know how VA interviews are conducted across a variety of cultural and linguistic settings. VAs are conducted in many different languages, and sometimes multiple languages are used in a single interview; however, the questionnaire is typically in English or other official language. This results in (a lot of) ad hoc translation and interpreting. We want to know how this affects the VA data and the eventual comparability of causes of death ascertained through VA.

• Verbal Autopsy Interview Standardization Study: A report from the field

Additional papers

• Cause-specific mortality rates in sub-Saharan Africa and Bangladesh
• Strengthening standardised interpretation of verbal autopsy data: the new InterVA-4 tool
• The WHO 2016 verbal autopsy instrument: An international standard suitable for automated analysis by InterVA, InSilicoVA, and Tariff 2.0
• Automated versus physician assignment of cause of death for verbal autopsies: randomized trial of 9374 deaths in 117 villages in India
• An integrated approach to processing WHO-2016 verbal autopsy data: the InterVA-5 model
• Non-confirming replication of “Performance of InSilicoVA for assigning causes of death to verbal autopsies: multisite validation study using clinical diagnostic gold standards,” by Flaxman et al.
• Direct maternal deaths attributable to HIV in the era of antiretroviral therapy: evidence from three population-based HIV cohorts with verbal autopsy