Verbal Autopsy

Modestly, and that is by design. Inferring a precise cause from a relative's recollection of symptoms is inherently uncertain, and accuracy varies sharply by cause — injuries and maternal deaths are easier than causes with vague, overlapping symptoms, and neonatal deaths are particularly hard. For this reason verbal autopsy is evaluated and used primarily at the population level: the target is the cause-specific mortality fraction, where errors in individual assignments partially cancel. A single verbal-autopsy result should be read as a probable cause, not a clinical diagnosis, and never used for medico-legal purposes.

Both turn a symptom interview into a cause, but by different logic. InterVA is a Bayesian probabilistic model: it combines prior cause prevalences with physician-derived probabilities of each symptom given each cause to compute, for each death, posterior probabilities across candidate causes, and InterVA-4 aligns with the WHO instrument. The Tariff method, behind SmartVA, is data-driven: using a gold-standard reference dataset it assigns each symptom-cause pair a 'tariff' reflecting how distinctively the symptom marks the cause, scores each death by summing tariffs, and ranks causes. Murray and colleagues found Tariff performed as well or better than alternatives, though the methods can produce somewhat different cause fractions.

Physician-coded verbal autopsy is slow, costly, and inconsistent — different doctors assign different causes to the same interview, and the workload is impractical at scale. Automated tools like InterVA and Tariff apply identical, transparent logic to every death, are fast and inexpensive, and require no physician time, making routine, large-scale mortality surveillance feasible. The PHMRC validation work demonstrated that these automated methods can match or exceed physician review in reproducing true causes, which is why the field, and WHO guidance, have moved toward standardized instruments paired with automated cause assignment.