October 20, 2026 interva.net Editorial

Eclampsia, Twenty-Six Times Over

A new Maputo study needed blood, placental histology and a three-person panel to tie 63 perinatal deaths to preeclampsia. Six years earlier, an autopsy series at the same hospital showed an InterVA model calling 26 maternal deaths eclampsia when four were.

verbal autopsymaternal mortalityInterVAvalidation

At Maputo Central Hospital, a quaternary referral centre that records 400 to 500 perinatal deaths a year, a research team recruited 100 women who delivered a stillborn baby or a newborn that died within twelve hours of birth. Between March 2021 and April 2022 they did not settle the question of what killed those babies with a single test. They drew maternal blood at delivery, took fetal and placental blood during minimally invasive tissue sampling, sent the tissue to Barcelona for histology and microbiology, weighed and measured the placentas, and gave the file to a panel of a pathologist, an obstetrician and a microbiologist, who assigned causes under the WHO classification of deaths during the perinatal period without seeing the biomarker results.

That panel attributed 63 of the 100 deaths to preeclampsia, as the main maternal condition in 50 and as a contributor in 13. The assay, published in PLOS One on 15 October, then performed respectably against this benchmark. The maternal ratio of soluble fms-like tyrosine kinase-1 to placental growth factor reached an area under the curve of 82%; at the established cut-off of 85 it identified 84% of preeclampsia-related deaths, mis-flagged 27% of the rest, and carried a positive predictive value of 84%. Placental blood did worse (AUC 75%) and fetal blood worse again (61%).

The study is a cross-sectional diagnostic accuracy study at one referral hospital. Its cut-offs were chosen post hoc from the same data, no correction was applied for testing several markers and sample types, and 36 of the 98 stillbirths were macerated to grade 3, which the authors say may have degraded both the placental histology and the fetal blood. Their conclusion is that post-mortem angiogenic markers “could significantly enhance cause-of-death determination in low-resource, high-burden settings”. That is a reasonable claim about a laboratory test, in places where the test does not exist.

The version of the question a questionnaire has to answer

Almost every perinatal death in the world is investigated with something cheaper than a panel in Maputo. Where a death is counted at all, the instrument is an interview: a trained interviewer, a relative who was present, and a structured set of questions about symptoms, timing and care. The 2022 WHO verbal autopsy instrument is the current standard form of that conversation, and the InterVA model, developed at Umeå University, is one of the ways its answers are turned into a probable cause.

The awkward evidence about how that works for preeclampsia comes from the same city. In 2021, a group including several of the authors behind the new biomarker paper published a validation of InterVA against complete diagnostic autopsy — the full post-mortem examination, not a panel reading tissue samples — in 316 deaths at the same tertiary hospital between 2013 and 2015. Agreement at the level of the individual death was poor: kappa ran from −0.030 to 0.232, lowest in stillbirths and highest in adults, and chance-corrected accuracy for cause-specific mortality fractions ran from −1.00 to 0.62.

Eclampsia was the instructive case. The model’s sensitivity for it was 75%, meaning it rarely missed a maternal death that autopsy confirmed as eclampsia. Yet 26 maternal deaths were assigned eclampsia as the most probable cause while four had actually died of it. By autopsy, eclampsia accounted for 4.4% of maternal deaths; by the model, 27.5%. Most of the deaths it mislabelled were infections.

Why a high sensitivity still produced a wrong number

Sensitivity and a cause fraction answer different questions. A model that leans toward eclampsia will seldom miss a true case, and will manufacture cases wherever the symptoms fit something else. In a population where puerperal sepsis and other infections are common, fever, convulsions and loss of consciousness are shared signatures, and a probabilistic model weighing symptom patterns against a cause hierarchy can score them as obstetric.

The caveats on that study are real and worth carrying into any citation of it. Its inputs were not gathered by interviewing families: the attending physician extracted them from the clinical record and arranged them in the WHO 2012 verbal autopsy format, and 43 of the model’s 245 input indicators could not be extracted at all. The model was InterVA-4.04, not the version now in use. In the journal’s open peer review, reviewers objected that InterVA-5 had been available since April 2019 — before the analysis was written up — and that the 2013–2015 data cannot have been collected with the WHO instrument then in force. So the study measured how the model reasons over hospital-recorded symptoms, not over a bereaved relative’s account. That distinction cuts in both directions, and it is why the eclampsia result is best read as a warning about a failure mode rather than as an accuracy figure for verbal autopsy as practised.

What physician-coded verbal autopsy returns at population scale

The clearest picture of the alternative — clinical judgement applied to interview data — comes from Preah Vihear province in rural Cambodia, where community health workers ran pregnancy surveillance for four years. Of 522 deaths recorded between 2018 and 2022, 229 stillbirths and 293 neonatal deaths, 79.1% had a verbal autopsy, and two physicians interpreted each one before causes were classified with ICD-PM, adapted locally to handle stillbirths whose timing of death was unknown.

A primary cause was identified for 95.0% of the neonatal deaths. For stillbirths the figure was 36.6%, and a further 12.0% could not be assigned a timing of death at all. Where causes could be named, they were the expected ones: hypoxia for 78.3% of intrapartum stillbirths, low birth weight and prematurity for 40.9% of early neonatal deaths, infection for 51.4% of late neonatal deaths. That neonatal column can carry a cause profile. The stillbirth column cannot, and the study’s contribution is to publish the size of the gap rather than to close it.

Set beside the Maputo work, the two arrangements look less like rivals than like instruments with different reach. A questionnaire reaches the household where the death happened and costs a conversation. An assay or a tissue panel reaches a hundred deaths in thirteen months at a referral hospital and names preeclampsia with something approaching precision. Between them sits the number most health systems actually publish, and its meaning depends entirely on which instrument stood behind it and how much of the denominator that instrument declined to classify. Our explainer on how verbal autopsy works sets out the chain from interview to cause; the harder question, as these three studies show, is what the last link is entitled to claim.

Sources

← Back to the news digest