Verbal autopsy (VA) is used in areas with limited medical certification of death to obtain information on causes of death (CODs) [1–3]. VA tools typically consist of a structured survey administered to a close relative or associate of the deceased by a trained field worker, to record the signs and symptoms that occurred before death. This information is used to assign the most probable COD, most often via physician-certified verbal autopsy coding (PCVA).

PCVA has limitations in inter- and intra-observer differences in coding [3], but remains widely useful, particularly in establishing population-based estimates of the major CODs [1]. There has been interest in the use of computer-coded VA (CCVA) methods to automate COD assignment [3]. CCVA methods are, in theory, cheaper, faster and more consistent over time - but their performance against PCVA and against each other has not yet been assessed reliably.

This is the largest comparison study yet done of CCVA and PCVA methods. We found that, at an individual level, ORF, OTM and InterVA-4 replicated the coding of physicians comparably, but that the average agreement level for the leading COD was about 50%. Agreement with the physician-assigned code rose substantially for all CCVA methods if the three most probable CODs were considered, and generally improved with larger subsets within a given study. On a population level, the KL method performed best in terms of CSMF accuracy and error, and replicated the CSMF distribution of PCVA in the original datasets fairly well. ORF did not outperform KL, even on the IHME dataset, and did not perform better than InterVA-4, despite claims to the contrary [18]. At the individual level, InterVA-4, which does require training on a dataset, produced broadly comparable results to the methods that do require training.

Comparison to physician coding as a reference standard poses several methodological challenges. Importantly, our study focused on CCVA replication of physician codes (and errors), and not whether the underlying COD assigned by the physician was correct. Validation of PCVA is limited by the lack of a true reference standard in countries where verbal autopsy is performed [1–3, 9]. Nonetheless, PCVA with good quality control can yield useful results on COD distributions in countries where medical certification remains uncommon [1]. The studies we included in the comparisons had physician coding done with reasonably good levels of quality control [4–7, 9] as shown by a low proportion of ill-defined deaths. Physician coding that contains large amounts of random errors would reduce agreement on a COD in the dual-physician coding system, and would tend to increase the number of unspecified CODs in the International Classification of Diseases (ICD-10) [19], such as ‘senility’ (ICD-10 code R54) or ‘cause not determined’ (ICD-10 code R99). This would in turn make it harder for CCVA methods to identify specific causes. Moreover, the size of the errors or biases in CCVA methods depends on the inherent errors and biases of PCVA results. This is particularly relevant for machine learning, as its accuracy requires learning on “true” class labels. High misclassification rates in the training set will also affect performance in the testing set across datasets, as noted recently on the full IHME dataset [20]

The performance of each CCVA method at individual assignment improved when trained and tested on a larger number of cases, most likely due to a greater number of cases from which to learn the distinct relationships between specific symptom profiles and CODs. The differences in the field and coding methods across studies would tend to reduce the observed differences in PCVA and CCVA between the various comparisons (and more likely so at the individual level than at the population level). This might have contributed to the observed comparability of the results for the four CCVA methods. With larger studies and more standardized field and physician coding methods, any real, underlying differences between various CCVA methods may become apparent. Finally, we note that InterVA-4 has a threshold of probability for designating the most probable cause (that is, the most probable cause must also have a probability above 50%, otherwise the death is classified as indeterminate), whereas ORF and OTM select the most probable causes without applying any thresholds.

On an individual level, no single method was able to replicate physician assignment of COD more than about half the time. At the population level, the King-Lu method was the best method to estimate CSMFs, though it does not assign individual CODs. However, good population-level agreement accuracy does not guarantee good individual agreement [21, 22]. A key methodological feature is the need to avoid the false gold standard of hospital-based deaths [1, 3]. Reliance on hospital or urban-based deaths for training of automated methods may lead to learning of symptom patterns and other features that are not representative of populations without medical attention. Indeed, the CSMFs between home and hospital deaths are dissimilar, as demonstrated in India [23].

Our study is the largest cross-country comparison of current CCVA methods versus PCVA, covering about twice as many deaths as an earlier multi-country study [7], and including a mix of various ages, and community and hospital deaths. Nonetheless, we faced certain limitations. First, we could not compare the original IHME random forest and tariff algorithms, though the original methods were re-created to the best of our abilities from the published descriptions [12, 13], yielding broadly similar results (Figures 1 and 2). Second, access during the analyses phase to the full IHME hospital-based dataset of 12,000 records would have allowed more robust comparisons. Similarly, the China dataset was also somewhat limited by the small sample size.

Different CCVA methods have various strengths and weaknesses depending on the study scenario and study objective. An ideal solution could involve a combination of automated methods to obtain robust individual- and population-level estimates. In the medium term, it appears unwise and certainly premature to recommend that automated systems replace physicians in coding VAs. CCVA methods could be used in parallel with physician coding to increase speed, efficiency and quality of coding. Future work may focus on the performance of a combination of various automated methods, and must extend to larger datasets and explore specifics for important age groups (children, maternal, adult), by gender, and across various settings of home- and hospital-based deaths. Future studies need to also place specific emphasis on testing computer based methods on a random sample of deaths in countries, as this would be much more useful in determining the true underlying CSMF at the population level [24].