Inclusion strategies
A total of 71 papers were first identified in which 25 were identified through database search using the three different concept combinations, and 39 were identified through the MITS surveillance alliance site and Google scholar. Seven additional studies were included through citation mining. There were 20 duplicates leaving 51 records eligible. After reading the abstracts, ten were excluded, 41 full text articles were then assessed. One recently published article that met the eligibility criteria was recommended and included. A total of 21 articles were finally included in the review. See Fig. 1.
Flow diagram of publication screening and identification
The main characteristics of the identified publications were extracted (additional file 1) and the list of the 16 variables has been mentioned above in the methods section.
Summary of articles
There were five articles that had full VA validation studies [16,17,18,19,20]. Only one [20] had physician coded verbal autopsy (PCVA), the rest were computer coded verbal autopsies (CCVA). Since MITS process involves a VA component that contributes to its final COD, its studies were included since VA validation studies in relation to MITS are so few. There were eight MITS validation studies [21,22,23,24,25,26,27,28] and seven MITS focused studies [9, 13, 29,30,31,32,33]. One of the VA validation studies [16] also has a MITS validation component and one of the MITS focused studies [13] had a VA validation component. One article focused only on CDA [34]. Ten of the studies dealt purely with hospital facility deaths [16, 17, 21,22,23,24, 26,27,28, 34] whilst the other 11 had combined community and facility deaths [9, 13, 18,19,20, 25, 29,30,31,32,33].
Location of the study
Most of the studies included in this review have been conducted in various CHAMPS sites within seven countries. These are Bangladesh, Ethiopia, Kenya, Mali, Mozambique, Sierra Leone, and South Africa. Six of these included publications are MITS focused studies that came from all seven CHAMPS sites [9, 13, 29, 31,32,33]. There were ten studies that were conducted only in Maputo Central Hospital, Mozambique [16,17,18,19, 22,23,24, 26,27,28], this brings the total number of studies from Mozambique to 16. There were two studies from Brazil [20, 21], one separately from Ethiopia [30] thus bringing Ethiopia’s contribution to seven, one from Astana, Kazakhstan [25], and one from Mumbai, India [34].
Study populations and timing of the procedure
All death categories from still births to adults were represented in the various study populations, however most studied neonates and children since CHAMPS work focuses more on under-five year old children. Five articles included maternal deaths in their study sample and four other articles included non-pregnant women of child-bearing age. The total number of deaths studied across the 21 articles is 19,388. The VA instrument most used in the studies were either 2012 or 2016 WHO VA questionnaire. VA data were mostly collected within four weeks after death. The MITS samples were mostly collected within 24 h of death or up to 36 h if the body was preserved through refrigeration. The CDA procedure was mostly done within 24 h after the MITS procedure or clinical data extraction.
Analysis techniques
The measures mostly used to compare the methods included descriptive statistics, sensitivity, specificity, positive and negative predictive value, and Kappa statistic. Almost all the studies used multiple measures. Sensitivity, specificity, positive and negative predictive values were used in eleven studies [13, 16,17,18,19,20,21,22,23,24, 26], Kappa statistic was used in seven studies [17, 21, 23,24,25,26,27], and chance corrected concordance (CCC) was used in two studies [17, 20]. Cause-specific mortality fraction (CSMF) was used in four studies [13, 17,18,19], CSMF accuracy (CSMFA) and chance-corrected CSMFA (CCCSMFA) were used in two studies [17, 20]. Eleven studies used mean, median, frequencies and percentages [9, 21, 23, 25, 28,29,30,31,32,33,34], χ2 analysis was done in five studies [9, 29, 32,33,34], McNemar’s test was done in two studies [25, 27], Fisher’s exact test and odds ratios (ORs) were used in three studies [22, 32, 33].
Quality of the study
The quality and applicability of the included diagnostic accuracy studies were assessed using the QAREL checklist (additional file 2). The overall quality of all the included studies was good. None of the studies recorded a ‘no’ response which would have caused the study to have a poor quality. For items 5 – 7, the answer ‘not applicable’ was applied for those MITS focused studies since the clinical data and VA information were required to arrive at the COD and there was no reference standard in these studies.
Malaria mortality VA validation compared with MITS
Three studies were fully focused on malaria [13, 16, 34]. Malaria was mentioned as part of the diagnoses and thus part of the validation in eight studies [8, 17,18,19, 22,23,24, 29].
Malaria was screened for in almost all the studies. The various laboratory tests used for the screening and detection of the malaria parasite included rapid diagnostic test, blood smear microscopy and using real-time quantitative polymerase chain reaction (qPCR) assay. CHAMPS sites additionally do molecular analysis conducted through TaqMan Array cards (this is a customizable diagnostic platform that allows the performance of dozens of real-time PCR reactions simultaneously (TAC); ThermoFisher Scientific, Waltham, MA, USA) [13]. In two studies [16, 17], further search for P. falciparum was conducted through histologic examinations, immunohistochemical stains and also thorough search for haemozoin in macrophages was conducted in all tissues as evidence of the past malaria infection. Malaria was assigned as a COD based on: (a) presence of cerebral malaria or (b) presence of abundant haemozoin deposition in tissues in the absence of other CoD [16, 17]. In the CDA focused study done in India [34], malaria was diagnosed if the ring, schizonts and/or gametocytes of the malaria parasite were identified either before death or in the post-mortem peripheral blood smears or by the presence of malaria pigment in the splenic imprint or in the tissues collected for histopathological examination. It is worthy to note that very high parasitaemia adds credibility to the importance of malaria in the causal pathway to death while very low parasitaemia are less indicative of causality.
In a CHAMPS study [9], malaria accounted for 16% (39 out of 241) of child (age 1 to 5) deaths when only the underlying cause was considered and 5% (13 out of 278) of infant deaths. However, when the full causal chain was considered, malaria accounted for 27% (66 out of 238) of child deaths indicating a 1.7-fold increase.
In another CHAMPS study [29], malaria was one of the six most common underlying COD at 11% (71 out of 632 decedents) in children 1 to 59 months. When the full chain of events was considered, malaria accounted for 20% (123 decedents) of child deaths. It is important to note that not all the seven CHAMPS sites had malaria deaths, it was absent from South Africa, Ethiopia, and Bangladesh sites.
Malaria was responsible for 21% of the febrile deaths in a CDA focused study done in Mumbai, India [34]. Malaria deaths mostly occurred during the monsoon period or rainy season. Plasmodium falciparum was mostly responsible for these deaths and the most common complication or mode of death was cerebral malaria in this study [34].
In a recent article by Ogbuanu et al. [13], malaria accounted for 30.5% (262/858) of the deaths in children aged 1 – 59 months in four countries (Sierra Leone, Kenya, Mozambique and Mali). Sierra Leone had the highest malaria-associated deaths with 42.9% (126/294), followed by Kenya with 31.4% (96/306), Mozambique with 18.2% (36/198) and Mali with 6.7% (4/60) malaria-associated deaths in the full causal chain [13]. No malaria-attributable death was documented among stillbirths and neonates. All these malaria-associated deaths were attributed to P. falciparum however, there was also a high bacterial co-infection (24%). The most likely comorbid condition seen in these malaria cases was anaemia.
See Fig. 2 for the malaria proportion of deaths in various country locations. In one of the studies conducted in Mozambique [16], out of 264 deaths, only 6 (2%) were due to malaria. Using the CDA procedure as the gold standard, the sensitivity and specificity of the VA were 33% and 96% and for the MITS, it was 100% and 100%, respectively. Another study [17] also found a low sensitivity of 33% and a specificity of 97% in identifying malaria using VA (InterVA model) in comparison to the gold standard CDA.
Malaria proportion of deaths in various country locations
The recent CHAMPS study [13] with a correlation component between computer-coded VA and MITS results for malaria diagnosis also found low levels of sensitivity and moderate to high levels of specificity. In the detection of malaria as a CoD based on MITS, VA analysis using the InSilico method had a sensitivity and specificity of 29% and 86.6%, respectively, whereas the InterVA method had a sensitivity and specificity of 18.4% and 95.1%, respectively [13]. The sensitivity and specificity of the VA validation studies above ranged from 18.4% to 33% and from 86.6% to 97%, respectively.
In two multi-cause calibration studies [18, 19], both InSilicoVA and Expert Algorithm VA (EAVA) exhibited high frequency of misclassification for malaria in comparison to MITS even though malaria was identified more often in InSilico VA than EAVA. For InSilico VA, malaria was correctly identified in 48% of the multi-cause MITS malaria deaths. The MITS—coded malaria deaths were misclassified by VA as other infections and pneumonia in 24% and 21% of the deaths respectively [18]. For the EAVA, malaria was correctly identified in only 12% of the multi-cause MITS malaria deaths. The MITS—coded malaria deaths were highly misclassified by VA as other infections and pneumonia in 35% and 34% of the deaths, respectively [18]. These results are similar to the study also conducted by Fiksel et al. [19]. When the data were calibrated, the malaria CSMF estimates for children (1 – 59 months) for the three VA methods including ensemble model increased [18, 19].
MITS validation compared with CDA
MITS was validated against CDA in nine studies across all age groups. The MITS procedure has been shown to have a high sampling success rate ranging from 67% for the kidney to 100% for blood, CSF, lung, liver, and brain [28]. These organs have the most diagnostic yield. In a MITS validation conducted by Castillo et al. in adults [22], there was 76% (85 / 112) overall concordance between the MIA diagnosis and CDA diagnosis. Infectious diseases (79%) and malignant tumours (81%) had higher concordances than other diseases (56%). In the MITS samples, there was also a higher identification rate (84%) of specific microorganisms causing deaths due to infectious diseases. In another MITS validation conducted by Bassat et al. [24] in children, a substantial concordance was seen with the CDA (Kappa = 0.70, 95% CI 0.49–0.92) and the agreement was 75% (36 / 48) of the cases. As with other studies, the MIA or MITS showed a high sensitivity and specificity for infectious diseases (93% and 75%) and malignant tumours (100% and 100%, respectively.
A study in Manaus, Brazil [21], yielded a degree of coincidence or overall concordance of 85% (47 / 55 cases) between the MIA and the CDA diagnoses (Kappa = 0.78, 95% CI 0.68–0.95). This perfect level of coincidence was similarly higher in infectious diseases (90%) and malignant tumours (87%) than for other diseases (67%). There was also a high concordance of 83% when MITS diagnosis in neonates and infants were compared with CDA diagnosis based on the full causal chain in Nur-Sultan, Kazakhstan [25].
In summarizing the studies contributing to this theme, overall concordance rates between MITS diagnosis and CDA diagnosis ranged from 68% in maternal deaths [23] and in neonates [26] to 85% in predominantly adults [21] and 90% concordance in disease categorization. MITS was shown to be relatively accurate and reliable, thus a robust substitute for the CDA.
Importance of clinical information
In all the MITS validation studies listed above, the results were analyzed blindly to clinical data to know how accurate the technique was. In order to know how valuable the clinical data to the diagnostic yield of the MITS is, Fernandes et al. [27] analysed MITS blinded to clinical data (MITSb), MITS enhanced with clinical data (MITSc), CDA blinded to clinical data (CDAb) in comparison to the gold standard CDA with clinical data (CDAc). They found out that clinical data increased diagnostic coincidence between MITS blind to clinical data and the gold standard by 11% (30 / 264 cases) and modified the CDAb diagnosis in 20 (8%) of 264 cases. There was a significant increase in concordance between MITSb and MITSc with the gold standard in neonates, adults, and maternal deaths. In children, even though the increase in k value was not significant, it was still evident. The agreement between MITSb and CDA for children was 89% (48/54 cases; κ value = 0.704, substantial agreement) [27]. When the clinical data were added (MITSc) in comparison with CDA, the degree of coincidence increased to 93% (50/54 cases; κ value = 0.802, almost perfect agreement). Overall, the addition of clinical data increased the diagnostic accuracy of MITS in relation to CDA.
Add Comment