Verbal autopsy: how the cause of death is found

A 34-year-old woman dies at home, in a village four hours from the nearest hospital. She was seven months pregnant and, according to her sister, had a severe headache, swelling in her legs and, on the last night, a seizure. No doctor examined her, and the body was buried the next day. Weeks later, a trained interviewer visits the family and asks a series of questions about the final days. The scene is fictional and was written for teaching.

This text is about the verbal autopsy, the method that uses an interview with relatives to estimate the cause of a death that no one certified. The idea is to show how it works, what studies say about its accuracy and why it matters for global health. If you play the case of the day on NEXO, you will recognize the challenge: reaching a cause from clues that are told, not seen.

Key points

• Nearly half of the world's deaths have no certified cause.
• The verbal autopsy is a structured interview with relatives or caregivers, conducted after a period of mourning.
• The World Health Organization published a standardized instrument in 2016, readable by automated programs.
• Algorithms such as InterVA, InSilicoVA and Tariff interpret the answers.
• Against the complete diagnostic autopsy, case-by-case agreement was low in Maputo.
• Questions about the circumstances of death bring useful information for planning.

Deaths that no one records

Knowing what people die from is the starting point for deciding where to invest in health. A review on the use of the verbal autopsy in health policy recalls that one in two deaths in the world is not registered (Thomas and colleagues, 2018). Another analysis estimates that fewer than a third of deaths in the world receive an assigned cause (de Savigny and colleagues, 2017). The authors of InterVA-5 describe uncertified deaths as nearly half of global mortality and as a source of much uncertainty for health planning (Byass and colleagues, 2019).

A commentary by pathologists adds the context: more than two thirds of the world's population lives and dies in countries without a working vital registration system, where most deaths happen outside health services and are neither counted nor classified by cause. In these places, the verbal autopsy has become the main method for determining the cause of death (Fligner and colleagues, 2011).

Why certification does not reach

The literature on the topic points to some reasons for the lack of a registered cause. Most deaths happen outside health services, and civil registration systems do not work in many countries (Fligner and colleagues, 2011). There is a shortage of doctors trained to certify deaths, and building a complete registration system took decades or centuries in countries such as Sweden, the United Kingdom and Australia (Hazard and colleagues, 2020).

Civil registration and vital statistics

Civil registration and vital statistics systems, known by the acronym CRVS, exist to count births and deaths and to record their causes. Where medical certification is rare, the verbal autopsy has been proposed as the best available alternative to fill the gap in causes of death (de Savigny and colleagues, 2017). This requires integrating the interview into the routine of the system, and not only into research projects, with the verbal autopsy conducted in the community and linked to the registration of each death (de Savigny and colleagues, 2017).

The same authors argue that automated methods make it possible to track causes of death within the CRVS inexpensively and almost in real time (de Savigny et al., 2017). A review of 26 studies, carried out in 15 low- and middle-income countries in Africa, Asia, the Middle East and South America, examined the use of verbal autopsy to guide policies and improvements in health systems and describes it as a pragmatic and provisional solution to the lack of a functional CRVS (Thomas et al., 2018). The text on one health shows another field in which surveillance depends on data collected far from the hospital.

The method in the routine of several countries

The SmartVA tool, which combines a shortened questionnaire and the Tariff 2.0 algorithm, has been translated and applied in countries such as Bangladesh, Brazil, China, Colombia, Papua New Guinea, the Philippines, Myanmar, Peru, Rwanda, Solomon Islands, Sri Lanka and Zambia (Hazard et al., 2020). The authors report the experience of using it as part of civil registration routine and note that it performed as well as or better than most alternative methods, including the cause assigned by physicians.

What verbal autopsy is

Pathologists note that, for most physicians, the term verbal autopsy sounds like a contradiction. Even so, it is a well-defined procedure, which classifies the cause of death from structured interviews with family members, friends and caregivers and from a review of available medical records (Fligner et al., 2011). There is no examination of the body and no tissue collection. The word autopsy appears because the goal is the same: to find out why the person died.

Verbal autopsy is a structured interview with a close relative or caregiver of the person who died, conducted after a period of mourning (Nichols et al., 2018). The answers are then interpreted, by physicians or by algorithms, to arrive at a probable cause.

Who answers and when

The interview is conducted in the community, after a culturally acceptable period of mourning (de Savigny et al., 2017). In the scene at the beginning, the sister is the informant. She saw the headache, the swelling and the seizure, and these three clues, together with the pregnancy, enter the questionnaire questions.

The questionnaire and the free narrative

The WHO questionnaires follow a common structure (Nichols et al., 2018). They begin with the date and place of the interview, the study site and the household, continue with the main informant's data and the sociodemographic data of the person who died, and go through the history of injuries and accidents. Then come the list of symptoms with the duration of each, the health services used before death, the certificate and medical documents kept at home, if any, and finally the free narrative of the events leading up to the death, with the cause in the informant's opinion.

This free narrative is considered essential when the analysis is done by physicians (Nichols et al., 2018). It allows the order of events to be seen, something a list of symptoms alone does not show.

The WHO instrument

The 2016 version of the WHO instrument was designed to serve as an international standard and to be read directly by automated programs such as InterVA, InSilicoVA and Tariff 2.0. It comes with manuals for interviewers and supervisors and is available in a template for electronic collection on the Open Data Kit platform (Nichols et al., 2018). Collection can be done on paper or on a tablet (de Savigny et al., 2017).

Shortening without losing information. A research group systematically developed a reduced version of the questionnaire used in the validation study of the PHMRC consortium. The ideal size was chosen by analyzing how performance fell as the instrument shrank, for adults, children and newborns (Serina and colleagues, 2015). This reduced version, together with Tariff 2.0, gave rise to SmartVA, prepared for use at scale on tablets and phones (Nichols and colleagues, 2018).

From the report to the cause: doctors and algorithms

After the interview, someone has to assign a cause to the answers. One way is review of the forms by doctors. Another is the use of algorithms. A large study compared five automatic methods, InterVA-4, Random Forest, Simplified Symptom Pattern, Tariff and King-Lu, with review by doctors, using 12,535 deaths from diverse populations with a reliably established cause (Murray and colleagues, 2014).

InterVA relies on Bayes' rule (Nichols and colleagues, 2018). The InterVA-5 version was developed to align with the WHO 2016 standard (Byass and colleagues, 2019).

The three most cited families of algorithms are InterVA, based on Bayes' rule, InSilicoVA, one of the programs that read the WHO 2016 instrument, and Tariff 2.0 with SmartVA, an algorithm packaged with the reduced PHMRC questionnaire (Nichols and colleagues, 2018).

All three accept data from the WHO 2016 instrument, which allows comparing results between countries and between methods (Nichols and colleagues, 2018).

The reasoning behind the algorithm

Bayes' rule is the same principle as pretest probability: you start from the chance of each hypothesis and adjust that chance as each new piece of information arrives. The difference is that the clinician does this in their head, and the program does it on its own. The text on base rate in diagnosis shows why the initial prevalence weighs so heavily in this calculation.

How often the estimated cause is right

The most important question about any diagnostic method is how often it is right. To answer it, a reference standard is needed. The PHMRC consortium created rigorous diagnostic criteria, with laboratory, pathology and imaging findings, to identify deaths with a confirmed cause in health services, and developed a verbal autopsy questionnaire based on WHO standards (Murray and colleagues, 2011). The goal was to gather well-defined reference deaths to validate the methods of cause assignment.

With this database, the researchers compared the methods. In adults, accuracy in the fraction of causes for the automatic methods ranged from 0.764 to 0.770, against 0.680 for review by doctors and 0.625 for InterVA (Murray and colleagues, 2014). In children, Tariff reached 0.783, review by doctors 0.678 and InterVA 0.520 (Murray and colleagues, 2014). In newborns, the values were 0.817, 0.719 and 0.629 (Murray and colleagues, 2014).

Comparing with complete autopsy

Study	Reference used	Main finding
Murray 2011	Deaths in health services with clinical, laboratory and imaging criteria	Created the PHMRC validation database
Murray 2014	PHMRC database	Automatic methods surpassed medical review in the fraction of causes
Menéndez 2020	Complete diagnostic autopsy in Maputo	Low agreement for tuberculosis, malaria and pneumonia

The hardest test came from Maputo, in Mozambique. Researchers compared InterVA with complete diagnostic autopsy in 316 deaths. None of the deaths from tuberculosis were identified by verbal autopsy, and sensitivity was 33 for malaria and 36 for pneumonia, per 100 deaths from these causes. The algorithm assigned 26 deaths to eclampsia, but only 4 had this cause at autopsy (Menéndez and colleagues, 2020). Sensitivity was also zero for diarrheal diseases and disseminated infections, and 32 per 100 for HIV. In 16 deaths, the model did not reach a conclusive result.

The authors conclude that these results do not help build confidence in current estimates of causes of death (Menéndez et al., 2020). They also note that malaria transmission in Maputo is low, which weighs on the interpretation of the numbers for this cause.

Population and individual

These results seem contradictory, but they measure different things. Accuracy in the fraction of causes measures whether the method gets the distribution of causes right in a large group. Case-by-case agreement measures whether it gets the cause of each death right. In the PHMRC study, the chance-corrected agreement of automated methods in adults was close to half, below the accuracy in the fraction of causes (Murray et al., 2014). A method can estimate well the proportion of deaths from a cause in a region and still get wrong which deaths those were.

Where the method goes wrong

The accuracy of verbal autopsy depends heavily on the quality of diagnostic criteria, the age of the person who died and the type of disease involved. Deaths with nonspecific signs and symptoms are especially difficult (Fligner et al., 2011). The authors cite the difficulty of separating deaths from malaria from those caused by other febrile diseases, such as septicemia, meningitis, encephalitis and pneumonia. Maternal deaths, perinatal deaths and stillbirths also have poor diagnostic specificity.

The case of eclampsia in Maputo illustrates the problem in practice: the algorithm assigned far more deaths to this cause than the autopsy confirmed (Menéndez et al., 2020). In the opening scene, hypertensive disease of pregnancy is the most eye-catching hypothesis, but the method alone would not confirm it. It is the same shortcut discussed in the text on differential diagnosis with clues: a striking clue can end the search too early.

Sources of error described in the studies

• Nonspecific signs and symptoms (Fligner et al., 2011).
• Fevers from different causes that look alike in the account (Fligner et al., 2011).
• Maternal and perinatal deaths, with poor specificity (Fligner et al., 2011).
• Lack of any medical source of information in many questionnaires (Fligner et al., 2011).
• Zero sensitivity for tuberculosis compared with complete autopsy (Menéndez et al., 2020).
• Deaths with no conclusive result from the model (Menéndez et al., 2020).

When the methods complement each other

Anatomopathological autopsy has long been considered the reference for cause of death (Fligner et al., 2011). The main obstacle is the need for pathologists, assistants and a histology laboratory, which are lacking in many low- and middle-income countries. Another concern is cultural and religious attitudes toward death and the handling of the body, which vary by country and must be addressed case by case.

The authors propose continuously evaluating a portion of deaths with anatomopathological autopsy together with verbal autopsy, to improve the quality of mortality data and identify new or emerging diseases (Fligner et al., 2011). They also note that the same assumptions that lead to error in up to 40 of every 100 deaths certified by physicians are present in verbal autopsy data. The text on morbidity and mortality sessions shows the same logic of learning from error on a smaller scale.

Beyond the cause: the circumstances of death

The biological cause is only part of the story. Researchers in Agincourt, South Africa, created a short set of questions about the path to care and the circumstances and events surrounding the death and integrated it into the WHO 2012 short verbal autopsy form (D'Ambruoso et al., 2016). Among the deaths studied, 39 in every 100 families did not call for help, 36 in every 100 found care too expensive and 33 in every 100 did not go to a health service.

The authors conclude that adding these questions brings information complementary to the distribution of causes and relevant to health planning (D'Ambruoso et al., 2016). The goal was to create indicators on care processes, circumstances and events linked to the health system and the social context.

Questions about the path to care

The themes in the Agincourt set ask the interview to investigate whether the family called for help, whether the person reached a health service, whether the cost weighed on the decision to seek care, what care pathway the person followed and what events happened near the time of death (D'Ambruoso et al., 2016).

Use in public health and in clinical practice

Verbal autopsy data feed burden of disease estimates, program planning and intervention evaluation. The review of 26 papers found examples of the method being used to guide policy in several countries (Thomas et al., 2018). SmartVA was presented as a way to quickly improve knowledge about causes of death in rural populations with few resources, with experiences in Myanmar, Papua New Guinea, Bangladesh and the Philippines (Hazard et al., 2020).

For the clinician, verbal autopsy carries a direct lesson: it shows how much can be inferred from the account alone and how much is lost. The text on signs, symptoms and history discusses how to interpret what arrives through the words of the patient and of those accompanying them.

What verbal autopsy teaches those who provide care

The WHO questionnaire separates the symptom list, with the duration of each one, from the free account and from the cause in the informant's opinion (Nichols et al., 2018). This separation is a good model for the consultation, in which a companion's account also mixes what was seen with what was concluded. The text on diagnostic uncertainty deals with how to weigh clues reported by third parties.

In the consultation, this means recording each symptom with its duration, asking for the free account in the order in which things happened, noting separately what the informant thinks was the cause and asking about medical documents kept at home.

Reasoning about the scene from the start

A teaching sequence for the scene, using the steps described above, could be this:

1. Context. Young woman, seven months pregnant, death at home.
2. Clues. Severe headache, swelling in the legs, seizure.
3. Questionnaire. Record each symptom with its duration and collect the sister's free account.
4. Attribution. Review by a physician or by an algorithm.
5. Limit. Maternal deaths have poor specificity in verbal autopsy, and in Maputo eclampsia was over-attributed.
6. Circumstances. Ask whether the family called for help and whether they reached a service.
7. Record. Choose the ICD-10 code for the estimated cause, knowing about the uncertainty.

Step 7 reminds us that the estimated cause also needs a code. Knowing how to find the right ICD-10 code is part of the work. The text on pandemic preparedness shows another use of surveillance data, and the text on cholera, measles and malnutrition in conflicts recalls how counting deaths guides the response in crises.

How NEXO helps train this reasoning

In NEXO, the case of the day is fictional, with a real clinical presentation and up to six progressive clues, and you propose the diagnosis by searching for the ICD-10 code.

The limits on using the game are described on the NEXO Economy page. If you research the teaching of clinical reasoning, the page for Researchers explains how to get in touch. Professionals who write or review cases find information under Curator and Creators, and institutions can learn about the proposal on the partnerships page.

Frequently asked questions

Is verbal autopsy used where there is no death certificate?

It was proposed as the best alternative where medical certification of death is rare (de Savigny et al., 2017).

Is there a standardized WHO questionnaire?

The 2016 WHO instrument serves as an international standard and can be read by InterVA, InSilicoVA and Tariff 2.0 (Nichols et al., 2018).

Can algorithms replace review by physicians?

In estimating the distribution of causes, studies show an advantage for algorithms. In the PHMRC database, automated methods had higher accuracy than medical review (Murray et al., 2014).

Can verbal autopsy get the cause of a specific death wrong?

In Maputo, comparison with complete autopsy showed low agreement for tuberculosis, malaria and pneumonia (Menéndez et al., 2020).

Can the interview be used to understand failures of the health system?

Questions about the circumstances showed families that did not call for help or found care too expensive (D'Ambruoso and colleagues, 2016).

Take-home messages

• Nearly half of the deaths in the world have no certified cause.
• Verbal autopsy is a structured interview with family members, done after the mourning period.
• The 2016 WHO instrument standardizes the interview and allows automated analysis.
• In the PHMRC database, algorithms surpassed medical review in the fraction of causes.
• Against the complete autopsy in Maputo, case-by-case agreement was low.
• Nonspecific symptoms and maternal deaths are the hardest to classify.
• Questions about the circumstances bring useful information for planning.
• The anatomopathological autopsy in part of the deaths can improve data quality.

Want to practice reasoning with progressive clues? Play the case of the day.

References

1. Byass P, Hussain-Alkhateeb L, D'Ambruoso L, et al. An integrated approach to processing WHO-2016 verbal autopsy data: the InterVA-5 model. BMC Med. 2019. PMID 31146736.
2. D'Ambruoso L, Kahn K, Wagner RG, et al. Moving from medical to health systems classifications of deaths: extending verbal autopsy to collect information on the circumstances of mortality. Glob Health Res Policy. 2016. PMID 29202052.
3. de Savigny D, Riley I, Chandramohan D, et al. Integrating community-based verbal autopsy into civil registration and vital statistics (CRVS): system-level considerations. Glob Health Action. 2017. PMID 28137194.
4. Fligner CL, Murray J, Roberts DJ. Synergism of verbal autopsy and diagnostic pathology autopsy for improved accuracy of mortality data. Popul Health Metr. 2011. PMID 21806831.
5. Hazard RH, Buddhika MPK, Hart JD, et al. Automated verbal autopsy: from research to routine use in civil registration and vital statistics systems. BMC Med. 2020. PMID 32146903.
6. Menéndez C, Quintó L, Castillo P, et al. Limitations to current methods to estimate cause of death: a validation study of a verbal autopsy model. Gates Open Res. 2020. PMID 33145479.
7. Murray CJ, Lopez AD, Black R, et al. Population Health Metrics Research Consortium gold standard verbal autopsy validation study: design, implementation, and development of analysis datasets. Popul Health Metr. 2011. PMID 21816095.
8. Murray CJ, Lozano R, Flaxman AD, et al. Using verbal autopsy to measure causes of death: the comparative performance of existing methods. BMC Med. 2014. PMID 24405531.
9. Nichols EK, Byass P, Chandramohan D, et al. The WHO 2016 verbal autopsy instrument: An international standard suitable for automated analysis by InterVA, InSilicoVA, and Tariff 2.0. PLoS Med. 2018. PMID 29320495.
10. Serina P, Riley I, Stewart A, et al. A shortened verbal autopsy instrument for use in routine mortality surveillance systems. BMC Med. 2015. PMID 26670275.
11. Thomas LM, D'Ambruoso L, Balabanova D. Verbal autopsy in health policy and systems: a literature review. BMJ Glob Health. 2018. PMID 29736271.