Adjudication of Cause of Death in Older Adults: learnings for death certification from the ASPirin in Reducing Events in the Elderly study

Abstract

Background:

Death certificates are a legal requirement for a body to be buried or cremated. Many randomized clinical trials utilise disease-specific causes of death as key outcomes informed by these documents. Determination of cause of death is commonly assigned to an adjudication committee, a time and resource intensive process which becomes more difficult in trials of older individuals. We sought to assess if a simple transcription from a death certificate would be adequate to meet trial requirements and if the certification process itself can be improved to meet public health and research requirements.

Methods:

Random audit of 100 death certificates from ASPREE, a randomised controlled trial conducted in Australian general practice and US community research centres. Participants were Australians (aged 70+ years) and Americans (65+) living in the community and without life-limiting medical conditions at baseline, recruited between 1 March 2010 – 31 December 2014 for the ASPREE trial. Outcome of interest was misclassification of death rate.

Results:

There were 2,757 deaths in the 19,114 study population. In a random sample of 100 of these deaths, misclassification of cause of death was identified through the trial adjudication process in 9% of death certificates.

Conclusions:

In trials without the resources of ASPREE a coding method can be accepted. Based on our experience, we recommend changes to the structure and process of death certification to better serve both clinical research and public health needs, particularly in older, multimorbid populations.

Background

It is well established that the attribution of a specific ‘cause of death’ established at autopsy may differ from that recorded by treating clinicians due to a combination of antemortem misdiagnosis, missed diagnoses and competing causes.^{1, 2} The accuracy of a death certificate, particularly the principal cause of death, is an important index guiding public health interventions. There are several reasons why the accuracy and utility of death certificates will face future challenges as a higher proportion of individuals die in advanced age. Foremost amongst these is the difficulty in selecting the principal contributor when an older person dies with multimorbidity. This is compounded by the natural reluctance of clinicians to pursue a diagnostic workup for older patients at the end of life and the absence of a systematic approach to attributing death to conditions such as diabetes and frailty. Death certifiers, often junior doctors in public hospitals or family physicians in nursing homes are required to specify “Disease or condition directly leading to death”, “Antecedent causes”, and “Other significant conditions” with limited guidance appropriate for older patients.

Experience gained from death adjudications in a large randomised controlled study of aspirin for healthy ageing (ASpirin in Reducing Endpoints in the Elderly - ASPREE)^3–6 has encountered these problems and provided insights into how they might be resolved. The study had a composite primary endpoint (persistent disability-free survival) and eight secondary endpoints [the three individual components of the primary endpoint, i.e. all-cause mortality, dementia, and persistent physical disability, plus fatal and nonfatal cardiovascular disease (including stroke), fatal and nonfatal cancer, mild cognitive impairment, depression, and major haemorrhage (including clinically significant bleeding and haemorrhagic stroke)]. While death as a part of the composite primary endpoint was simply determined by proof of death (two sources such as a death certificate, physician report, etc), a cause of death required adjudication to establish the principal cause of death, an approach more rigorous than typical pragmatic trials which may rely solely on coded registry data.

Although the ASPREE cohort was initially relatively healthy due to study selection criteria, over time their disease burden increased with many developing multiple concurrent illnesses. Each secondary endpoints had an independent adjudication committee, and therefore the possibility arose of each attributing conflicting causes of death. To address this problem a ‘Death Adjudication Committee’ was formed. This committee was informed of the various adjudication committee deliberations on secondary endpoints, but it was the final arbiter of the cause of death. Current members are the authors. The process of adjudication was participant deaths first being identified by ASPREE staff through active surveillance or notification by family member or physician. Death was then confirmed by a second source and relevant data collected from public sources and medical records by an ‘Endpoint captain’. Data was then presented to both adjudicators online for independent adjudication. Concordant adjudications were accepted and discordant adjudications resolved by consensus.

The Death Adjudication Committee took note of information from death certificates, but supplemented this with clinical records collected from hospitals, clinics, nursing homes, coroners, next of kin and information provided by participants during the ASPREE study. Several issues were encountered including: attributing a cause of death in participants with multiple potential causes, e.g., a cardiovascular event in an individual with end-stage renal disease and cancer; death certificates where the “disease or condition directly leading to death” was recorded as a terminal event such as ‘cardiorespiratory arrest’ irrespective of the preceding medical history; probabilistic attributions in settings such as when individuals were ‘found dead’ or died suddenly without a relevant medical history; deaths in individuals in residential care where limited or no investigation had been undertaken, for example with a known abdominal mass where cancer suspected; and variable choice of broader disease categories such as dementia versus more specific entities such as Parkinson’s disease.

The approach adopted in the ASPREE trial was to document both a ‘trajectory to death’ and a ‘mode of death’ where the trajectory attempted to specify the most likely underlying cause, and the mode referred to the terminal event. The trajectory equated to the underlying cause and the mode of death has to this point not been used in any ASPREE research output. Both were accompanied by decision rules to avoid inconsistency (Table 1). We also adopted a pragmatic epidemiological approach to sudden deaths (observed or unobserved) as being due to vascular events unless there was a more likely alternative diagnosis. Discordant adjudications were resolved through consensus.

Table 1.

Decision Rules for Death Adjudications

1.	If death is a result of a fall/head trauma with intra-cranial but extra-cerebral bleeding (and including subarachnoid hemorrhage), the underlying cause of death will be classified as a ‘Clinically Significant Bleed’.
2.	If death is due to immediate cause of bleeding, then regardless of the reason for the bleeding the underlying death should be classified as ‘Clinically Significant Bleed’. The exception to this rule is #3.
3.	If death is a result of a hemorrhagic stroke, underlying cause of death should be classified as ‘Stroke’ (not CSB).
4.	Rules for unwitnessed unexpected death, subject found dead. If there is a death certificate with:  i. No autopsy then the underlying cause of death (COD) should be classified as according to proffered COD providing there is ante mortem evidence to support this otherwise classify as Other/unknown;  ii. With autopsy then the underlying COD should be classified as according to proffered COD providing there is postmortem evidence to support this otherwise classify as Other/unknown.
5.	If death is result of myeloproliferative disease, underlying cause of death is ‘Cancer’.
6.	If there is no supporting information available directly related to death event, recent illness or treatment (no medical record, no death certificate), underlying cause of death should be Other/unknown.

This adjudication approach had a high burden and cost to the study, principally in retrieving and collating medical records. For future studies involving older participants, a relevant question is whether a simple transcription of a death certificate could have met study requirements. To investigate this, we conducted a post-hoc audit of death adjudications during the study and considered whether our observations might have lessons for improvements in death certification.

Methods

We collected data on all adjudicated deaths from trial commencement in 2010 into the post-trial period to the beginning of 2025. A random selection of 100 cases from these deaths were identified using the following method. Random selection of the first case was done by an online random number generator.⁷ If an identified case did not have a death certificate, sequential cases were searched until the next case with a death certificate was identified. Our study focussed on discordance between the principal cause of death listed on the death certificate and the adjudicated trajectory to death established by the adjudication committee.

Results

The ASPREE study had 19,114 enrolled participants, 16,703 (87.4%) from Australia and 2,411 (12.6%) from the US. There were 3,507 known deaths, 2,757 of whom have been adjudicated to date [2,583 (93.7%) in Australia and 174 (6.3%) in the USA]. A copy (60%) or extract (40%) of a death certificate was available for 85.8% of adjudicated cases. Regardless of whether a death certificate was available, adjudications were informed by coroner’s reports, medical documents, and/or US and Australian National Death Index linkage. Adjudicated deaths was higher in Australia (15.5%) compared with the US (7.2%) over the observed period. This may be due to age differences driven by US minority eligibility criteria. There was no suggestion of a significant difference in detection rates between the countries as deaths identified only from NDI matching are similar in Australia (1.4%) and the US (1.7%) and the acquisition of source documents in 750 unadjudicated deaths, 689 (91.9%) in Australia and 61 (8.1%) in the US were similar to adjudicated case numbers.

Of the one hundred selected death certificates, ninety were from Australia and ten from the US. All were physician certified. Trajectories to (causes of) death were cancer 29, cardiovascular disease (not stroke) 18, dementia 14, infection (including COVID-19) 13, stroke 7, neurological disease (not stroke) 6, non-specific diseases of ageing 4, renal failure 3, gastrointestinal disease 2, haemorrhage 2, and respiratory disease and trauma one each. Eighty-three of the adjudications were essentially unchanged. Changes were identified in 17 cases (Table 2). Of these only nine cases (2–7, 12, 15 and 17) would the underlying cause of death have changed. These involved adjudicated cases of dementia, heart failure, myocardial infarction, cancer and stroke missing from the death certificate. In most of these cases the discrepancy resulted from the certificate listing an immediate cause of death rather than the underlying trajectory.

Table 2.

Reclassifications and discordance

Case #	Death certificate			Adjudication
	Disease/condition directly leading to death	Antecedent causes	Other	Mode	Trajectory
1	Cardiac arrythmia	Blank	Advanced dementia	Unknown	Dementia
2	Cardiac arrest, ventricular tachycardia	Acute cardiac event, non-ischaemic cardiomyopathy		Clinically significant bleeding	Clinically significant bleeding
3	Klebsiella variicola bacteraemia	Blank	Frailty, Congestive Cardiac Failure, Chronic Obstructive Pulmonary Disorder, Advanced Rheumatoid arthritis	Infection	Trauma
4	Acute on chronic kidney disease	Blank	Non-ischaemic arrythmias induced cardiomyopathy	Heart failure	Renal disease
5	Delirium	Blank		Prostate cancer	Prostate cancer
6	Paroxysmal atrial fibrillation	Mitral valve regurgitation, hypertension, hyperthyroidism		Myocardial infarction	Cardiovascular death
7	Dementia	Hypertension, hyperlipidaemia, intermittent syncope		Stroke	Dementia
8	Bronchopneumonia	Blank	Renal failure, dementia	Infection	Dementia
9	Metabolic encephalopathy	Uterine cancer		Other [need access detail]	Colorectal cancer
10	Voluntary assisted dying	Advanced ovarian cancer		Metastases to vital organ	Cancer Ovary or Endometrium
11	Old age	Blank		Infection	Multiorgan failure/non-specific/Unascertained
12	Broncho pneumonia	Blank		Infection	Trauma
13	Parkinsons disease	Blank		Dementia	Neurological disease (not stroke)
14	Right lower lobe pneumonia, STEMI, rapid atrial fibrillation, acute kidney failure	Dementia, diabetes		Myocardial infarction	Coronary heart disease
15	Parkinsons disease	Blank		Infection	Dementia
16	Voluntary assisted dying metastatic rectal cancer	Metastatic rectal cancer		Metastases to vital organ	Colon/Rectum cancer
17	Pneumonia	Cerebrovascular accident, atrial fibrillation, hypertension		Infection	Dementia

Discussion

In the near 15 years of follow up in the ASPREE study, 18.3% of the cohort had died. Most (86%) had death certification available for adjudication. Our audit of adjudicated deaths found that 83 of the 100 deaths could have been simply transcribed without risk of misclassification. Of the discordant seventeen cases, eight did not have endpoint implications. Three (cases 1, 8 and 11) would be amenable to a simple decision rule to avoid discordance, two (10 and 16) involved voluntary assisted dying, two (13 and 14) were unrelated to cause of death, and one (9) was simply the type of cancer. This leaves nine cases with important study outcomes identified.

Historically, certification of death developed from parish records. It has since become a vital centralised public health asset. In the era of communicable diseases with high infant and adolescent mortality and truncated life expectancy, this allowed public health authorities to monitor and address the contemporary disease burden to allow improvements in duration and quality of life. Dying young in the past drew a confident ‘diphtheria’ or ‘consumption’ attribution or in the more recent past ‘motor vehicle accident’ or ‘myocardial infarction’ in the middle-aged. A more contemporary scenario may be a frail aged care facility resident being found dead in bed by nursing staff on a morning round, which often leaves the responsible doctor putting a probabilistic attribution such as myocardial infarction or stroke, given they are a common cause of death, but with little confidence in the diagnosis.

It is an onerous and costly task in a clinical trial to collect premortem data and go through a rigorous adjudication process. In such a large-scale clinical trial as ASPREE (N = 19,114) was it worth it? The results presented here indicate that it would have resulted in a 9% misclassification of those with an endpoint. This does not address the 14.2% without a copy or extract of the death certificate.

How does this compare with other death classification findings? Flagg and Anderson reported that in 2018 34.7% of US death records had an unsuitable and 2% had an unknown or ill-defined underlying cause of death.⁸ They also reported that 12.7% had an immediate or intermediate cause as the underlying cause of death, and 19.8% had a nonspecific underlying cause of death where some form of re-categorisation would likely be necessary. Izci et al investigated breast cancer patients in a Belgian University hospital and found a 15.3% discordance in death certification compared with patient files with 10.7% false positive and 4.5% false negative rates.⁹

Dementia was an important outcome in a trial in the aged and part of our composite primary and a secondary endpoint. However, we noted substantial variation in how often it was recorded as a contributing cause of death of the underlying cause of death. Ahmad et al reported that in England in 2015 only 53.8 % of people dying with dementia had dementia recorded on their death certificates despite a prior diagnosis of dementia from other sources rising markedly over the same period.¹⁰ Shi found a similar figure of 51.2% in Scotland.¹¹ Gao and colleagues reported their experience in the Medical Research Council Cognitive Function and Ageing Study (CFAS) and CFAS II.¹² Participants in these studies were aged 65 and from England and Wales. The reporting of dementia on death certificates was compared with the study diagnosis of dementia. They found that dementia recording on death certificates was poor with 21.0% sensitivity in CFAS but increasing to 45.2% in CFAS II.

Recording was more likely for severe cases or in those living in an institution, but less likely if death occurred in hospital.

Data linkage is another alternate method of capturing outcomes in trials. In the UK Kalsi et al found only 1125 of 1324 women with confirmed ovarian cancer had a relevant cancer registration, i.e. 199/1324 (15%) would be missed.¹³ Our own match with the Victorian Cancer Registry and had a higher degree of concordance.

From our rigorous adjudication process, it became apparent that the death certificate can and should be more accurate. Of interest, the adjudication committee was not necessarily in agreement with the findings documented in coroners’ reports. The primary interest of the coroner was to determine if a death was natural or un-natural. To this end very few received full autopsies although (in Victoria, Australia at least) most undergo full-body CT imaging.

Would our approach be applicable to death certification in the general population? While it may not address all aspects of misclassification, simpler language and illustrative decision rules may lead to more consistent classification and assistance in the era of complex multimorbidity. Use of definitions from the Global Burden of Disease and the ICD-10 codes used for common causes of death would reduce uninformative attributions such as cardiorespiratory arrest. These methods have been built into many electronic health records to enable human entry for search words that then are converted via a SNOMED CT crosswalk ontology into ICD-10 code options for the physician to then narrow down the selection to a known element term. This also happens with CTCAE ontologies for adverse events in clinical trials. So, it is not unheard of to think about some electronic search and selection in death certificates to reduce errors in interpreting free text.

The challenge remains of assigning a single cause of death in older people with complex, competing conditions. Multiple simultaneous contributing causes, a reluctance to institute comprehensive diagnostic workups in older people in an end-of-life scenario leading to limited documentation and difficulties with of attribution granularity, e.g., dying with COVID-19 infection or prostate cancer rather than because of it, and a general resistance to make attributions in a death certificate with personal or family resistance, e.g., dementia, remain. Another identified problem in attributing causation in the aged were falls. Was an intracranial bleed subsequent to, or causative of, the fall?

For clinical trials without appropriate funding, comprehensive adjudication may be unnecessary assuming the misclassification rate affects each randomised group similarly. Trials could use coders and clinician adjudicators only where coding is problematic and when death certificates cannot be obtained. In trials with older people where multiple potential contributors are recognised as potential causes of death, coders will need to capture all the details of a death certificate and not just be concerned for a single cause attribution. They should also be cognisant of premortem serious adverse events.

Limitations

When only the death certificate is available for adjudication, there is no opportunity for reclassification, as it serves as the sole source of information. It can also be argued that the treating physician completing the certificate is the best judge as they have direct knowledge of and interaction with the patient rather than a remote interpretation of a limited dataset. The reality is that certifying doctors (including trial adjudicators) will encounter difficulties in attributing a single cause of death in this age group in the setting of complexity (multimorbidity) and palliative modes of care. For example, a cancer may be suspected, even likely, in a nursing home patient which is not investigated and palliated in line with patient or family wishes. This leaves adjudicators without objective evidence. The chosen trajectory to death also begs the question how far back one goes in the attribution chain of events. For example, someone with type 2 diabetes in middle-age who dies from a myocardial infarction in old age. Where to draw the line? Are risk factors, genetics and environmental exposures on the causal chain or indeed ultimately responsible? We could go all the way back to the intrauterine environment and further into ancestry.

The processes used in clinical trials could lead death certification to online submission with forced classifications and built in decision rules to standardise causation and avoid triviality.

Conclusion

Adjudication is a worthwhile process for study endpoint ascertainment and accuracy but needs to remain informed of changes in certification and is not mandatory for studies with limited resources as the misclassification rate is modest. The adjudication process developed for ASPREE may inform reform of death certificates to more systematically identify the disease process or event leading to death from among the immediate, consequential and contributing descriptors. However, trials in advanced age may need to adopt a more nuanced cause of death outcome listing all major contributors to death, perhaps prioritising one as the principal.