Consider the form. The international Medical Certificate of Cause of Death is a single page, and its current version was updated in 2016 — the first revision in more than fifty years. The World Health Organization notes that changes are infrequent because they require updates to national legislation and information systems. That conservatism is not bureaucratic inertia. It is the price of comparability: a form that changes every few years cannot produce a time series that stretches across decades.
But the form is not a neutral recording surface. It is a genre with a forced narrative shape. It asks a physician to convert a tangle of conditions into a linear causal train, and then a classification system selects one link in that train as the cause. What a death “is” statistically is the product of a document and a rule set, not a transparent reading of the body.
The architecture of the form
The certificate has two parts. Part 1 asks for a causal sequence: the immediate cause, any intermediate causes, and the underlying cause. Part 2 asks for co-existing conditions that contributed to death but do not fit into the causal sequence in Part 1. The medical practitioner signing the certificate is responsible for indicating which morbid conditions led directly to death and for stating any antecedent conditions giving rise to the underlying cause.
This is a specific cognitive demand. The physician must decide, at the bedside or shortly after, which of perhaps a dozen interacting conditions initiated the train of events. Was it the pneumonia that killed the patient, or the dementia that made the pneumonia fatal, or the hip fracture that precipitated the decline that led to the pneumonia? The form permits only one answer in the underlying-cause slot.
The 1948 decision that made the statistic possible
The pivotal move is not in the form itself but in the rule for reading it. At the Sixth Decennial International Revision Conference in 1948, it was agreed that the cause of death for primary tabulation should be designated the underlying cause of death. The WHO defines that as “the disease or injury which initiated the train of morbid events leading directly to death, or the circumstances of the accident or violence which produced the fatal injury.”
This definition is doing enormous work. It selects one condition from a sequence and elevates it to the status of the cause. The immediate cause — the condition that was actually present at the moment of death — is not the one that gets tabulated. The intermediate causes are not tabulated either. Only the initiating condition counts. A death certificate that reads “immediate cause: respiratory failure; due to: pneumonia; due to: immobility; due to: Parkinson’s disease” produces a statistic of Parkinson’s disease, not pneumonia, not respiratory failure.
This is a defensible epidemiological choice. If you want to know what is killing populations, you want the condition that started the chain, not the final mechanism. But it is a choice, and it has consequences for what the statistic can say.
From certificate to code: the coding step
The certificate records words. The statistic requires codes. The ICD translates written diagnoses of diseases and other health-related problems into alphanumeric codes — diabetes mellitus becomes 5A14 in ICD-11, for example. This translation is not mechanical. A coder must read the physician’s phrasing, match it to the classification’s terminology, and assign the appropriate code.
Then comes the selection step. In mortality coding, a second step checks the medical certificate of cause of death against twelve coding rules — designated SP1 through SP8 and M1 through M4 — to select the underlying cause of death. These rules handle the cases where the certificate’s causal sequence is ambiguous, where the physician has recorded conditions in an order that does not reflect the ICD’s assumptions, or where the recorded sequence requires modification to produce a single underlying cause.
The rules are not secret. They are published in the ICD reference guide and taught in coding courses. But they are technical, and they encode editorial positions about which condition “initiated” the train. A rule that says, for example, that certain conditions are always to be considered the underlying cause when they appear anywhere on the certificate is a codified judgment about causal priority. It can be revised, tested, and contested — but it is not a neutral reading of the physician’s intent.
Automation and the relocation of judgment
The WHO has developed DORIS — the Digital Open Rule Integrated Cause of Death Selection tool — to apply ICD mortality rules automatically. The tool, according to the WHO, applies the rules “transparently and reproducibly,” with detailed steps exposed so users can explain how the selection was made. It is available online and offline, and it is designed to facilitate coding even in low-resource settings.
Automation does not remove judgment. It relocates and formalizes it. The rules themselves are the judgment, codified. What DORIS does is make the reasoning inspectable: you can see which rule fired, which condition was selected, and why. That is a genuine improvement over a coder’s tacit knowledge. But it also means that the editorial position embedded in the rule becomes harder to see as a position. It looks like a procedure.
The WHO’s own implementation guide notes that initial reviews of the first release version of DORIS identified shortcomings, and that the organization will continue to improve the system. This is not a scandal. It is what happens when you try to formalize a rule set that was developed for human coders and apply it to the messy variety of actual certificates. The rules are general; the certificates are specific. The gap between them is where discretion lives.
What the statistic cannot say
The ICD has been the basis for comparable statistics on causes of mortality and morbidity between places and over time for more than a century. That comparability is a real achievement. It is also purchased by forcing a causal narrative into a single selected code.
What gets lost? Part 2 of the certificate — the co-existing conditions — is recorded but not selected. A death from pneumonia in a patient with advanced dementia and a recent hip fracture produces a statistic of pneumonia, and the dementia and the fracture disappear from the primary tabulation. They may be available in multiple-cause data, but the headline statistic does not carry them.
Ill-defined causes are another gap. The WHO’s quality-analysis tool, ANACoD, flags ill-defined causes and unusual patterns. These are the deaths where the certificate says something like “cardiac arrest” or “old age” — conditions that are not really causes in the ICD’s sense. They are recorded, but they do not tell you what initiated the train. The quality tools exist precisely because the statistic needs to know how much of its own count is noise.
And then there are the deaths that never enter the system at all. The WHO notes that implementation of the ICD for mortality requires civil registration, recording, reporting, information flows, quality assurance and feedback, and training. Where those systems are weak, deaths go uncounted. The statistic is silent about its own absences.
Reading a cause-of-death table
The practical implication is that a cause-of-death table is not a transparent window onto what killed people. It is a record of what certifying physicians wrote, what coders selected, and what rules determined. Reading it well means reading the form’s grammar, the selection rules, and the absences alongside the counts.
This does not make the statistic useless. It makes it legible. A rise in deaths coded to a particular condition may reflect a real change in mortality, a change in coding practice, a change in the rules, or a change in how physicians fill out the form. The ICD’s revision history — the shift from one version to another, the movement of categories between chapters, the addition of new codes — is part of what the statistic means. Comparability across time is not automatic; it is maintained by mapping tables and dual-coding exercises that try to bridge the gap between revisions.
The death certificate is a small document. The ICD is a large classification. Between them sits a chain of decisions — the physician’s causal sequence, the coder’s translation, the rule’s selection — that determines what a death counts as. The statistic is the output of that chain. To know what it says, you have to know how the chain was built.
FAQ
Who decides what goes on the death certificate?
The medical practitioner signing the certificate is responsible for indicating which morbid conditions led directly to death and for stating any antecedent conditions giving rise to the underlying cause. The form provides the structure; the physician provides the content.
What is the difference between the immediate cause and the underlying cause?
The immediate cause is the condition present at the moment of death — respiratory failure, for example. The underlying cause is the disease or injury that initiated the train of events leading to death. The ICD tabulates the underlying cause, not the immediate cause.
Can a death have more than one underlying cause?
Not in the primary tabulation. The ICD selects a single underlying cause from the certificate’s causal sequence. Co-existing conditions are recorded in Part 2 but are not selected as the underlying cause. Multiple-cause data may preserve more of the certificate’s content, but the headline statistic uses one code.
What happens when the cause of death is unclear?
The certificate may record an ill-defined cause, such as “cardiac arrest” or “old age.” Quality-analysis tools like ANACoD flag these cases. They are counted, but they do not tell you what initiated the train of events. The proportion of ill-defined causes is one measure of the statistic’s quality.
Does automation change what counts as a cause of death?
Automation applies the same rules that human coders use, but it makes the rule application inspectable and reproducible. The rules themselves are the editorial position; automation does not change them, but it does make them harder to see as a position rather than a procedure.