System to classify cause of deaths in HIV-positive persons

Although the risk of death in populations with access to antiretroviral therapy (ART) has dramatically declined, AIDS is still a leading cause [1], as it remains in populations unable to benefit from this life-saving intervention [2]. A large proportion of deaths due to AIDS are as a result of late diagnosis and/or late entry into care, [3,4] with individuals often presenting with life threatening opportunistic diseases. Conversely, if ART is started earlier in the course of chronic HIV infection, or if ART leads to recovery of the immune function in immunodeficient persons, the profile of underlying causes of deaths diversifies. It is critical to classify the components of this diversification for at least three principle reasons. First, ART has now been used for 15 years, but the longer term outcome for using these drugs remains unclear. Emerging trends in various types of deaths may reveal unexpected potential problems associated with longer term ART use. As such, careful studies of underlying causes of death can be used as a safety signal generator. Second, the causes of death can give guidance to the prioritization of public health interventions. Whereas only one intervention prevents the development of all infections that constitute AIDS (namely ART), this cannot be said for what is now collectively named ‘serious non-AIDS diseases’ (i.e. cardiovascular disease, viral hepatitis, chronic kidney disease or non-AIDS cancers). For these diseases, disease-specific preventive interventions [5] and use of ART [6] are required. Prioritization among the many possible interventions would be better investigated by studying the risk of nonfatal manifestations as well, but experience suggests that it is considerably more difficult to reliably ascertain nonfatal events in large populations except if done in a prospective and carefully monitored cohort study setting [1]. Finally, serious non-AIDS diseases develop slowly over time, and risk is hence linked to aging. To date, very little is known regarding outcomes of aging persons with chronic but well treated HIV infection. In all of the above, a correctly determined underlying cause of death remains the most sensitive indicator of intrinsic morbidity [7]. To ascribe the cause of death appropriately, a reliable algorithm is required. The algorithm needs to be able to differentiate deaths that can be attributed to AIDS from non-AIDS deaths, and further be able to differentiate between the different non-AIDS causes. The algorithm must also be able to identify contribution from unsafe lifestyle factors seen in some HIV-positive persons (e.g. illicit drug intake), as well as identify deaths that are directly due to adverse drug reactions. The two main algorithms currently used are the International Classification of Diseases (ICD-10) classification, and the ‘Coding Causes of Death in HIV’ (CoDe) classification [7]. The former is used to classify causes of death in the general population, whereas the latter was specifically developed for, and is exclusively used in, HIV-positive populations. Hernando et al.[7] report in this issue of AIDS on a comparison of these two classification systems; the ICD-10 coding was extracted from the national death registry and the information used for the CoDe classification was derived from data collected in a cohort study. Compared to the CoDe system (and consistent with the reasons that prompted its development in the first place), the ICD-10 coding overestimated the proportion of deaths with AIDS as the underlying cause. This was and remains an expected outcome, as a physician completing the death certificate will invariably have access to less information regarding a person's clinical history than a physician ascribing a cause of death using data collected by the cohort in which the person was followed prior to death. Therefore, these physicians are more likely to attribute an underlying cause of AIDS when a HIV-positive person dies. As can be seen, the CoDe classification is thus preferential to use, and this is the case in both completed [5,8] and ongoing randomized controlled trials (RCT) (e.g. the START study: http://insight.ccbr.umn.edu/start/) and large cohort studies [e.g. the Data Collection on Adverse events of Anti-HIV Drugs (D : A : D) study] [1]. Of note, the CoDe classification has been applied in different ways by various studies, as depicted in the literature. The original application calls for a standardized collection of a comprehensive and extensive set of relevant data from the clinic or alternatively from the cohort database, where design permits. Documents, procedures and software to extract the relevant information can be retrieved free of charge (http://www.cphiv.dk/CoDe/tabid/55/Default.aspx). The choice of central classification of cause of death is then based on this information, determined by two external and independent reviewers. A process of arbitration is sometimes required in the case of divergent views between reviewers. [9,10]. This approach is customarily used in the clinical endpoint RCT setting by clinical endpoint review committees as part of best practice, as the external review process optimizes the chances of an unbiased classification [11]. However, the system is labour intensive due to the large amount of data extraction required and the need for two external reviewers. Therefore, a variation of this approach is to merely extract the available information from the cohort database without a further review of the clinical notes, and to then assign a cause of death through internal (rather than external) review. This alternate approach was used by Hernando et al. in their article (entitled ‘revised CoDe’: see www.art-cohort-collaboration.org and [12]). This ‘revised CoDe’ system is less labour intensive because it neither requires additional data collection by the clinical site, nor the process of external review. A direct comparison of the extent of differences in classification of underlying cause of death using the two approaches has not been performed yet, but it is unlikely to be extensive, provided that the data retrievable from cohort databases are comprehensive [13,14]. Cohorts studying mortality in HIV-positive populations are encouraged to use the CoDe classification to ensure harmonization and standardization across studies in the HIV research field. Acknowledgements J.D.L. wrote the editorial together with J.D.K. and received critical input from C.S. Conflicts of interest There are no conflicts of interest.