This research analyzed the CoD of 1366 adult and 205 child BiD cases who were brought into a third-level health facility in the capital city of Zambia by using an automated VA tool, called SmartVA, for the first time. The results indicate that the top CoD among the adult cases were infectious diseases, including AIDS, TB, and malaria, followed by NCDs, such as stroke, CVDs, and DM, and that those among the child cases were mainly infectious diseases, such as pneumonia, diarrhea, and malaria, followed by accidents. These trends are similar to the distribution of the CoD data of the Global Burden of Disease [26].
The comparison of the CoD between the SmartVA and death notification form showed that there are discrepancies in the distributions of the CoD among the adult cases; that is, there were fewer HIV/AIDS cases and more malaria cases in the death notification form. There were also fewer HIV/AIDS cases among the child cases in the death notification form. One possible reason is that the coders misclassified the CoD if a death case had two or more health conditions that lead to death. The WHO’s instruction manual for ICD-10 coding [29] states in page 35, “When more than one condition is entered on the certificate, select the condition entered alone on the lowest used line of underlying causes of death only if it could have given rise to all the conditions entered above it.” However, the coders for the death notification form may have reported the more immediate health conditions leading to death as the CoD. In that case, HIV/AIDS would be less likely to be selected as the final CoD.
In addition, there were more cases with an undetermined CoD on the death notification form compared with the SmartVA. This may be due to the healthcare staff’s lack of knowledge about recording CoD. The symptoms, such as pain or fever, should not be noted as the CoD and unspecified health conditions, such as natural death or sudden death, should also be avoided. Thus, creating a standardized procedure for recording the CoD might be necessary.
Furthermore, the concordance rate between the SmartVA and death notification form was low especially among the adult BiD cases. This trend was the same among the different age categories, although the concordance rate was slightly higher in the middle-aged category (20–59 years) than in people aged 13–19 years and 60 years and above. One of the main reasons behind this discrepancy could be the larger number of undetermined cases and misclassification on the death notification form. Especially, the small number of AIDS cases as the CoD on the death notification form is not consistent with what is already known about this disease burden in Zambia [27]. Since these CoD on the death notification form are reflected in the national vital statistics data, it is a problem that the actual data reported to the Civil Registration Office are skewed. The establishment of a system to improve the accuracy of the CoD among the BiD cases is necessary to capture the accurate vital statistics through the CRVS in a timely manner. Automated VA using the SmartVA is one option for strengthening the vital statistics since it can improve the quality of the determination of the individual CoD and can be used to estimate the population level of the CoD from all samples.
There are several limitations in this research. Firstly, the research period in the UTH was only 4 months because of financial and human resource constraints. However, seasonal fluctuation needs to be considered for some health conditions, such as malaria, diarrhea, and probably some NCDs. Therefore, the results of this research cannot be completely generalized as annual data. However, since most health conditions do not vary by month, according to the existing health information system, these conditions that do not exhibit seasonal differences can be generalized as annual data.
Secondly, there were several BiD cases in the UTH that were not included in this research. For example, some of children BiD cases were not covered because the UTH has another specific ward for children cases and some of these BiD cases were directly carried to the children wards. Such BiD cases in the children wards were unfortunately not well recorded. Furthermore, neonatal BiD cases were not captured because most of them were dealt with as hospital deaths once they were delivered to the neonatal intensive care unit. However, since most of the child BiD cases that were older than 1 month were directly carried to the BiD ward, this research did not just cover neonatal BiD cases. It is desirable to conduct another research study to investigate neonatal BiD cases in the UTH.
Thirdly, there was a substantial number of BiD cases with an undetermined CoD, even when using the SmartVA, although its determination rate was significantly better than that of the death notification form. When the SmartVA was used to analyze the top 10 CoD among the same research samples by calculating the CSMF as the population level, there were discrepancies in the distribution compared to the results of the individual analysis in terms of there being more malaria and more other NCDs (Additional File 2). These differences may derive from the large number of undetermined cases, which account for approximately one fourth of the samples. In order to make the results reliable, the computer algorithm that determines the CoD should be improved to reduce the number of undetermined BiD cases in the future.
Lastly, the validity of the Tariff Method 2.0 for identifying the CoD by using the SmartVA should be considered more. According to Serina et al. [10], the chance-corrected concordance used to assess the extent to which the Tariff Method 2.0 correctly predicts an individual’s CoD and the CMSF’s accuracy for measuring the performance at the population level, when applied to the shortened version of the PHMRC Questionnaire, were 50.0% and 76.6% respectively for the adult population. However, since this validity was investigated by comparing it with the gold standard dataset [19], the applicability to the Zambian setting should be scrutinized by investigating the actual data. In addition, new automated VA tools have been developed, such as the InterVA-5 [30], Naïve Bayes Classifier [31, 32], SilicoVA [33], and so on. We also need to investigate the functionality, feasibility, and validity of these new tools. In the future, a validation study that compares the CoD by using an automated VA tool with those determined by an actual full autopsy is required to estimate the true burden of diseases in Zambia.