The Chart Said Tuberculosis. The Autopsy Said Histoplasmosis.
In 101 autopsies of adults living with HIV in Manaus, the clinical record and the pathologist agreed on the underlying cause of death in just over a third of cases — a reminder that every verbal autopsy validation is scored against a reference standard with an error rate of its own.
The Fundação de Medicina Tropical Doutor Heitor Vieira Dourado sits in Manaus, a referral centre for tropical disease in the middle of the Brazilian Amazon. Between January 2014 and December 2019, according to municipal mortality records, 1,318 adults living with HIV died there. Autopsy was offered routinely to their families — not reserved for the puzzling cases — and 130 were carried out. A hundred and one had records complete enough to analyse.
That hundred and one is the basis of a paper published on 26 October in PLOS One, and it does something the cause-of-death literature rarely does so plainly: it prices the clinical diagnosis. Attending physicians had written an underlying cause of death in the chart. Pathologists had opened the body. Two reviewers, one of them a senior pathologist, then classified each pair of diagnoses using the Goldman criteria, with a third pathologist called in when they disagreed.
In 48 of the 101 cases — 47.6% — the discrepancy fell into Goldman class I, the category reserved for a major missed diagnosis that was directly related to death and that, had it been recognised before death, could have led to treatment capable of prolonging survival or curing the condition. Thirty-seven of those 48 involved untreated infectious disease. Eight were undiagnosed cancers. Three were something else again.
A classification built around the question “would it have mattered?”
The Goldman scheme is not a similarity score. It ranks disagreements by clinical consequence: class I as above; class II for a major missed diagnosis that would probably not have changed management; class III for a minor missed diagnosis whose treatment could have influenced prognosis; class IV for findings of epidemiological or genetic interest with no bearing on the outcome; class V for no discrepancy at all, in which case the autopsy diagnosis is taken as the reference standard.
Against that scale, 36.7% of the series fell into class V. Put the other way round, 63.3% of cases carried some discrepancy, but only 47.6 percentage points of that were major and management-relevant — the rest were class II (7.9%), III (5.9%) and IV (1.9%). Both figures appear in the paper, and they measure different things. The one worth carrying around is the class I share, because it is the closest thing an autopsy series can offer to a count of deaths that a better-resourced diagnostic pathway might have postponed.
Errors that run in both directions
The disease-level tables are where the study earns its keep. Tuberculosis was recorded clinically in 31 patients and found at autopsy in 16, with 13 concordant cases and a kappa of 0.43. Bacterial pneumonia or bronchopneumonia was called in 20 and found in 16, eight of them concordant, kappa 0.32. In both instances the clinicians were reaching for a familiar diagnosis more often than the tissue supported.
Neoplasms went the other way: seven diagnoses in life, fourteen at autopsy, six concordant, kappa 0.52. So did the fungal infections, which is the part that matters clinically. Histoplasmosis appeared in three charts and eleven autopsies, kappa 0.40. Pneumocystis was diagnosed three times and found ten, with a single concordant case and a kappa of 0.11 — the weakest agreement in the set. Cryptococcosis travelled better: eight at autopsy, six in the chart, six concordant, kappa 0.52.
The errors do not cancel. A hospital that over-diagnoses tuberculosis on clinical grounds and under-diagnoses disseminated fungal disease will produce a chart-based cause distribution that is wrong in two directions at once, and a verbal autopsy validated against those charts inherits both.
Twenty years earlier, in the same building
The authors note that an earlier FMT-HVD series covering deaths from 1996 to 2003 found 51.9% concordance — worse than the current figure, in an era before widespread antiretroviral therapy. They also set their result against a recent United States tertiary-hospital autopsy study in which class I discrepancies occurred in 9.9% of cases.
What sits between those numbers is capacity and timing. In this cohort the median CD4 count was 42 cells per microlitre and the median detectable viral load was just over 110,000 copies per millilitre; 26.7% of patients were receiving no antiretroviral therapy at all and another 30.7% were on it irregularly. The median hospital stay was eleven days. The authors attribute the high share of missed major diagnoses to advanced disease at presentation, to opportunistic infections whose clinical pictures overlap, and to limits on the specific diagnostic tests available during the study period — a candid set of explanations, and one that implicitly says the problem is the diagnostic pathway rather than the diligence of the physicians.
Only one variable was statistically associated with the Goldman class: antiretroviral status (Fisher’s exact test, p = 0.016), with the association driven by more class I discrepancies among people not on treatment. Age, sex, comorbidity, CD4 count and viral load showed no significant differences in a sample this size, and the correlation with length of stay was weak (Spearman’s rho = 0.20).
What this has to do with verbal autopsy
Nothing directly — and that is the point. The study contains no verbal autopsy, no algorithm and no coding comparison. It measures clinical diagnosis against autopsy, full stop.
Indirectly, it speaks to the ground every verbal autopsy validation stands on. When a study reports that a questionnaire-based method agreed with the reference standard in 80% of deaths, the reference standard is doing an enormous amount of the work. Sometimes that standard is physician review of the same interview. Sometimes it is the hospital record. The abridged instrument piloted in Kolkata, for instance, was scored against the medical records of 270 hospital deaths — records of exactly the kind this Manaus series examined, and found wanting in roughly half the cases where a missed diagnosis could have changed the outcome.
Biologically anchored standards are stronger, but they are still judgements. In the CHAMPS mortality surveillance network, cause-of-death determinations are made by expert panels reading histology, multiplex polymerase chain reaction results, placental sections, antenatal records, photographs and the findings of a WHO-instrument verbal autopsy interview, then coding to ICD-10 and ICD-PM. That is a far richer evidence base than a death certificate. It is not a truth machine.
And autopsy is not one either. Post-mortem sampling is incomplete, tissue changes after death, and interpretation varies between pathologists. The 2022 WHO verbal autopsy instrument exists precisely because most deaths in the world will never be opened, and the field’s working assumption has always been that a structured interview with a bereaved family is a better guide to the population distribution of causes than the alternative, which is usually no information at all.
The limits of this particular ledger
The study is retrospective, single-centre and small, and it analyses 101 of 1,318 deaths — an autopsy coverage of roughly one in ten, with families consenting case by case, which is its own kind of selection. Autopsy was offered without institutional criteria favouring diagnostic uncertainty, which helps, but consent is not random. The statistical tests are bivariate and unadjusted, several disease-specific kappa values rest on a handful of positive cases, and the paper does not describe whether the reviewers classifying discrepancies were blinded to the clinical record — an omission that matters for a judgement about what a physician could plausibly have diagnosed.
None of that undermines the headline. A referral hospital in the Amazon, in the years just before the pandemic, missed a management-relevant diagnosis in nearly half of the HIV deaths it opened. Whatever the number is elsewhere, it is not zero, and every cause-of-death measurement built on clinical records — including the ones used to judge verbal autopsy — is built on top of it.
Sources
- Martins YF, Almeida TVR, Reis MFD, Corrêa VVP, Marinho EPM, Pinto RAA, Bastos MS, de Lima Ferreira LC. Clinicopathological diagnostic concordance among autopsied adults living with HIV in the Brazilian Amazon: a retrospective cross-sectional study. PLOS One 2026;21(10):e0361097. pmc.ncbi.nlm.nih.gov
- World Health Organization. 2022 WHO verbal autopsy instrument. who.int
- Rahman A, Lee KH, El Arifeen S, et al. Causes of stillbirth in sub-Saharan Africa and South Asia: findings from Child Health and Mortality Prevention Surveillance, 2016-2023. PLOS Global Public Health 2026;6(10):e0005235. pmc.ncbi.nlm.nih.gov