October 2, 2026 interva.net Editorial

A 30-Minute Verbal Autopsy: A Kolkata Team Cuts the Questionnaire to 61 Questions

An abridged version of the WHO verbal autopsy instrument — 61 questions or fewer per age group instead of hundreds — agreed with hospital diagnoses in 80% of 270 deaths, in a single-city pilot validation in Kolkata.

verbal autopsyvalidationcause of deathfield studies

Every verbal autopsy is a second conversation with a family that has already lost someone. The interview is long on purpose: the WHO’s latest instrument runs to hundreds of questions across age-specific modules, because the signs and symptoms that separate one cause of death from another are numerous, and each one has to be asked before it can be coded. The cost is time — and time is exactly what mortality surveillance teams lack when they are working through a backlog of deaths in a slum, a displacement camp or a district with one functioning hospital.

A research team based at ICMR’s National Institute for Research in Bacterial Infections in Kolkata has now published a pilot validation of a much shorter alternative: an abridged questionnaire of 54 to 61 closed-ended questions per age group, derived from the 2022 WHO instrument, tested against the medical records of 270 hospital deaths. The results, reported in BMJ Global Health, put overall agreement with clinical diagnoses at 80%.

Cutting the instrument down to essentials

The logic of the abridged tool is subtractive. Its authors began from the 2022 WHO VA instrument and kept the questions tied to the causes that dominate in each age group, dropping duplicated and low-yield items while preserving the domains that make cause assignment possible: signs and symptoms before death, how long the illness lasted, and relevant medical history. The result is three age-specific versions — 61 questions for neonates, 60 for children aged one month to 11 years, and 54 for adults aged 12 to 49.

That compression shows up at the doorstep. Interviews in the study lasted roughly 30 to 35 minutes, against the 45 to 60 minutes or more the authors attribute to field implementations of the full questionnaire. Interviewers were given three days of training, worked in Bengali and Hindi, and conducted the interviews one-on-one after explaining the study and building rapport with the respondent, usually a parent or closest caregiver. Sessions were audio-recorded and later translated into English, and interviewees could go back to the family by phone to confirm details — the kind of workflow that adds cost to every case, and that a shorter instrument trims.

A test with a hospital record as the yardstick

The design is a record-based validation rather than a field survey. The team drew 90 deaths in each age group from three state-run tertiary general hospitals and one tertiary paediatric hospital in Kolkata, for deaths recorded in 2021 and 2022, then interviewed families between July 2024 and January 2025. Each death was coded twice: once from the underlying diagnosis written in the medical record, and once by a hierarchical computer-based expert algorithm applied to the interview. The medical cause was treated as the reference standard, and the two were compared using sensitivity, specificity, positive and negative predictive value.

A VA diagnosis was reached for 95.5% of the 270 deaths. Agreement on the cause-specific mortality fraction — how well the algorithm reproduced the distribution of causes, not just individual cases — was 81.1% for neonates, 71.1% for children and 87.8% for adults.

Age group	Overall agreement	Leading cause (medical record)	VA sensitivity / specificity
Neonates	81.1%	Sepsis (34.4%)	83.8% / 94.9%
Neonates	—	Intra-partum-related events (25.6%)	69.6% / 97.0%
Children 1 mo–11 y	71.1%	Pneumonia (53.3%)	100% / 50.0%
Children 1 mo–11 y	—	Other infections incl. sepsis, unknown source (25.6%)	21.7% / 98.5%
Adults 12–49 y	87.8%	Acute febrile illness (38.5%)	80.9% / 98.6%
Adults 12–49 y	—	Cardiovascular disease (38.5%)	100% / 98.6%

The table also shows where the shortened tool struggled. Childhood pneumonia came out with perfect sensitivity but only 50% specificity, because the algorithm counted possible pneumonia and acute lower respiratory infection in the same category — a deliberately inclusive rule that pulls in deaths caused by something else. Sepsis without a known source, grouped among “other infections,” was largely missed. In adults the algorithm had no code for a slice of hospital-diagnosed non-communicable diseases, left 6.7% of deaths unclassified, and returned five false positives for hepatitis that the authors trace to overlap with dengue. Overall, 3.3% to 6.7% of deaths stayed unspecified — a range the paper notes is close to the 5% to 10% typical of large VA studies.

What the pilot does not settle

The reference standard is the main caveat, and the authors say so plainly. Hospital records can be incomplete or wrong, they reflect only the people who reached a tertiary hospital, and the disease mix at that level of care may not match the community — a point the paper illustrates by contrasting neonatal cause patterns at tertiary hospitals with those at subdistrict level. Ninety deaths per age group is a small base, several causes the instrument is designed to detect never appeared in the sample at all, recall bias cannot be excluded, and the study is set in one city’s urban poor population. The authors’ own next step is validation in crisis-affected and displaced populations, not extrapolation from Kolkata.

There is a second, quieter limitation for anyone who reads this as a statement about VA coding generally. The abridged questionnaire was built to work with the expert-algorithm method used in the study; the authors write that adaptations of InterVA, InSilicoVA and physician coding “may also perform well” with it, but that was not tested here, and they describe that assessment as beyond the study’s scope. Whether a shorter questionnaire travels well across different coding methods — and different disease mixes — remains open.

Not the first attempt to shorten the interview

The trade-off is old. In 2015, researchers behind the multi-site PHMRC validation published a shortened VA instrument built for routine mortality surveillance, and a follow-up study across three sites in Bangladesh, the Philippines and Sri Lanka timed it: a median of about 25 minutes per interview, of which the free-text open narrative accounted for only about four, with 60% of respondents mentioning at least one keyword in that section. The Kolkata team’s bet is similar in spirit but different in method — a computer-coded questionnaire with the 2022 WHO instrument as its parent, shortened for settings where a full interview is a luxury.

Taken together, the two strands point at the same unresolved question: how much of the questionnaire can be removed before cause-of-death estimates stop carrying information, and which setting — a routine surveillance round, a humanitarian response, a national survey — each trade-off actually suits. The Kolkata pilot supplies one data point on that curve, from 270 hospital deaths, with an 80% overall agreement and a clear-eyed list of what it could not see. The 2022 WHO instrument remains the reference the field works from; the shortened versions are experiments around its edges.

interva.net is an independent digest of verbal autopsy and cause-of-death research. It is not affiliated with, endorsed by, or speaking for the InterVA project, Umeå University, or the World Health Organization.

Sources

← Back to the news digest