Probabilistic Cause-of-death Assignment using Verbal Autopsies

Mortality variability and differentials by age and causes of death in rural South Africa, 1994-2018

INTRODUCTION

Understanding mortality variability by age and cause is critical to identifying intervention and prevention actions to support disadvantaged populations. We assessed mortality changes in two rural South African populations over 25 years covering pre-AIDS and peak AIDS epidemic and subsequent antiretroviral therapy (ART) availability.

METHODS

Using population surveillance data from the Agincourt Health and Socio-Demographic Surveillance System (AHDSS; 1994-2018) and Africa Health Research Institute (AHRI; 2000-2018) for 5-year periods, we calculated life expectancy from birth to age 85, mortality age distributions and variation, and life-years lost (LYL) decomposed into four cause-of-death groups.

RESULTS

The AIDS epidemic shifted the age-at-death distribution to younger ages and increased LYL. For AHDSS, between 1994-1998 and 1999-2003 LYL increased for females from 13.6 years (95% CI 12.7 to 14.4) to 22.1 (95% CI 21.2 to 23.0) and for males from 19.9 (95% CI 18.8 to 20.8) to 27.1 (95% CI 26.2 to 28.0). AHRI LYL in 2000-2003 was extremely high (females=40.7 years (95% CI 39.8 to 41.5), males=44.8 years (95% CI 44.1 to 45.5)). Subsequent widespread ART availability reduced LYL (2014-2018) for women (AHDSS=15.7 (95% CI 15.0 to 16.3); AHRI=22.4 (95% CI 21.7 to 23.1)) and men (AHDSS=21.2 (95% CI 20.5 to 22.0); AHRI=27.4 (95% CI 26.7 to 28.2)), primarily due to reduced HIV/AIDS/TB deaths in mid-life and other communicable disease deaths in children. External causes increased as a proportion of LYL for men (2014-2018: AHRI=25%, AHDSS=17%). The share of AHDSS LYL 2014-2018 due to non-communicable diseases exceeded pre-HIV levels: females=43%; males=40%.

CONCLUSIONS

Our findings highlight shifting burdens in cause-specific LYL and persistent mortality differentials in two populations experiencing complex epidemiological transitions. Results show high contributions of child deaths to LYL at the height of the AIDS epidemic. Reductions in LYL were primarily driven by lowered HIV/AIDS/TB and other communicable disease mortality during the ART periods. LYL differentials persist despite widespread ART availability, highlighting the contributions of other communicable diseases in children, HIV/AIDS/TB and external causes in mid-life and non-communicable diseases in older ages.

Burden of child mortality from malaria in high endemic areas: Results from the CHAMPS network using minimally invasive tissue sampling

BACKGROUND

Malaria is a leading cause of childhood mortality worldwide. However, accurate estimates of malaria prevalence and causality among patients who die at the country level are lacking due to the limited specificity of diagnostic tools used to attribute etiologies. Accurate estimates are crucial for prioritizing interventions and resources aimed at reducing malaria-related mortality.

METHODS

Seven Child Health and Mortality Prevention Surveillance (CHAMPS) Network sites collected comprehensive data on stillbirths and children <5 years, using minimally invasive tissue sampling (MITS). A DeCoDe (Determination of Cause of Death) panel employed standardized protocols for assigning underlying, intermediate, and immediate causes of death, integrating sociodemographic, clinical, laboratory (including extensive microbiology, histopathology, and malaria testing), and verbal autopsy data. Analyses were conducted to ascertain the strength of evidence for cause of death (CoD), describe factors associated with malaria-related deaths, estimate malaria-specific mortality, and assess the proportion of preventable deaths.

FINDINGS

Between December 3, 2016, and December 31, 2022, 2673 deaths underwent MITS and had a CoD attributed from four CHAMPS sites with at least 1 malaria-attributed death. No malaria-attributable deaths were documented among 891 stillbirths or 924 neonatal deaths, therefore this analysis concentrates on the remaining 858 deaths among children aged 1-59 months. Malaria was in the causal chain for 42.9% (126/294) of deaths from Sierra Leone, 31.4% (96/306) in Kenya, 18.2% (36/198) in Mozambique, 6.7% (4/60) in Mali, and 0.3% (1/292) in South Africa. Compared to non-malaria related deaths, malaria-related deaths skewed towards older infants and children (p < 0.001), with 71.0% among ages 12-59 months. Malaria was the sole infecting pathogen in 184 (70.2%) of malaria-attributed deaths, whereas bacterial and viral co-infections were identified in the causal pathway in 24·0% and 12.2% of cases, respectively. Malnutrition was found at a similar level in the causal pathway of both malaria (26.7%) and non-malaria (30.7%, p = 0.256) deaths. Less than two-thirds (164/262; 62.6%) of malaria deaths had received antimalarials prior to death. Nearly all (98·9%) malaria-related deaths were deemed preventable.

INTERPRETATION

Malaria remains a significant cause of childhood mortality in the CHAMPS malaria-endemic sites. The high bacterial co-infection prevalence among malaria deaths underscores the potential benefits of antibiotics for severe malaria patients. Compared to non-malaria deaths, many of malaria-attributed deaths are preventable through accessible malaria control measures.

Tree-informed Bayesian multi-source domain adaptation: cross-population probabilistic cause-of-death assignment using verbal autopsy

Determining causes of deaths (CODs) occurred outside of civil registration and vital statistics systems is challenging. A technique called verbal autopsy (VA) is widely adopted to gather information on deaths in practice. A VA consists of interviewing relatives of a deceased person about symptoms of the deceased in the period leading to the death, often resulting in multivariate binary responses. While statistical methods have been devised for estimating the cause-specific mortality fractions (CSMFs) for a study population, continued expansion of VA to new populations (or "domains") necessitates approaches that recognize between-domain differences while capitalizing on potential similarities. In this article, we propose such a domain-adaptive method that integrates external between-domain similarity information encoded by a prespecified rooted weighted tree. Given a cause, we use latent class models to characterize the conditional distributions of the responses that may vary by domain. We specify a logistic stick-breaking Gaussian diffusion process prior along the tree for class mixing weights with node-specific spike-and-slab priors to pool information between the domains in a data-driven way. The posterior inference is conducted via a scalable variational Bayes algorithm. Simulation studies show that the domain adaptation enabled by the proposed method improves CSMF estimation and individual COD assignment. We also illustrate and evaluate the method using a validation dataset. The article concludes with a discussion of limitations and future directions.

Agreement between cause of death assignment by computer-coded verbal autopsy methods and physician coding of verbal autopsy interviews in South Africa

BACKGROUND

The South African national cause of death validation (NCODV 2017/18) project collected a national sample of verbal autopsies (VA) with cause of death (COD) assignment by physician-coded VA (PCVA) and computer-coded VA (CCVA).

OBJECTIVE

The performance of three CCVA algorithms (InterVA-5, InSilicoVA and Tariff 2.0) in assigning a COD was compared with PCVA (reference standard).

METHODS

Seven performance metrics assessed individual and population level agreement of COD assignment by age, sex and place of death subgroups. Positive predictive value (PPV), sensitivity, overall agreement, kappa, and chance corrected concordance (CCC) assessed individual level agreement. Cause-specific mortality fraction (CSMF) accuracy and Spearman's rank correlation assessed population level agreement.

RESULTS

A total of 5386 VA records were analysed. PCVA and CCVAs all identified HIV/AIDS as the leading COD. CCVA PPV and sensitivity, based on confidence intervals, were comparable except for HIV/AIDS, TB, maternal, diabetes mellitus, other cancers, and some injuries. CCVAs performed well for identifying perinatal deaths, road traffic accidents, suicide and homicide but poorly for pneumonia, other infectious diseases and renal failure. Overall agreement between CCVAs and PCVA for the top single cause (48.2-51.6) indicated comparable weak agreement between methods. Overall agreement, for the top three causes showed moderate agreement for InterVA (70.9) and InSilicoVA (73.8). Agreement based on kappa (-0.05-0.49)and CCC (0.06-0.43) was weak to none for all algorithms and groups. CCVAs had moderate to strong agreement for CSMF accuracy, with InterVA-5 highest for neonates (0.90), Tariff 2.0 highest for adults (0.89) and males (0.84), and InSilicoVA highest for females (0.88), elders (0.83) and out-of-facility deaths (0.85). Rank correlation indicated moderate agreement for adults (0.75-0.79).

CONCLUSIONS

Whilst CCVAs identified HIV/AIDS as the leading COD, consistent with PCVA, there is scope for improving the algorithms for use in South Africa.

Community perception of causes of death using verbal autopsy for diabetes mellitus in Saudi Arabia

Diabetes mellitus is a serious global health issue which significantly impacts public health and socioeconomic development. Exploring how the community perceives the causes of death and their associated risk factors is crucial for public health. This study combines verbal autopsy (VA) with the Type 2 Diabetes Mellitus (T2DM) register to explore community perceptions of causes of death and associated influential factors in Makkah province, Saudi Arabia. 302 VA interviews were conducted with relatives or caregivers of deceased who died between 2018 and 2021 based on T2DM medical register from Alnoor Specialist Hospital in Makkah City, Saudi Arabia. Cause-specific mortality fractions (CSMFs) obtained from the VA using the InterVA-5 model were utilized to assess community perception. We used a multivariable logistic regression model to determine factors influencing community perceptions of causes of death. Lin's CCC with 95% CI was used to analyze the concordance for the CSMFs from verbal autopsy causes of death (VACoD) as a presumed reference standard and family-reported causes of death (FRCoD). The outcomes of this study demonstrate a generally broad spectrum of community perceived mortalities, with some critical misconceptions based on the type of death and other vital events like marital status, with an overall CCC of 0.60 (95% CI: 0.20-1.00; p = 003). The study findings demonstrate that community perception is weak if the deceased was male compared to female (aOR: 0.52; 95% CI: 0.26-1.03) and if the deceased was > = 80 years compared to 34-59 years (aOR: 0.48; 95% CI: 0.16-1.38), but it significantly improves among married compared to single (aOR: 2.13; 95% CI: 1.02-4.42). Exploring community perception of causes of death is crucial as it provides valuable insights into the community's understanding, beliefs, and concerns regarding mortality. Higher or lower community perception is attributed to how people may perceive risk factors associated with the causes of death, which can guide public health planning and interventional programs. The study findings further emphasize the need to employ robust and standardized VA methods within the routine medical services for a systemized assessment of families' reported causes of death.

Cause of Death estimation from Verbal Autopsies: Is the Open Response redundant or synergistic?

Civil registration and vital statistics systems capture birth and death events to compile vital statistics and to provide legal rights to citizens. Vital statistics are a key factor in promoting public health policies and the health of the population. Medical certification of cause of death is the preferred source of cause of death information. However, two thirds of all deaths worldwide are not captured in routine mortality information systems and their cause of death is unknown. Verbal autopsy is an interim solution for estimating the cause of death distribution at the population level in the absence of medical certification. A Verbal Autopsy (VA) consists of an interview with the relative or the caregiver of the deceased. The VA includes both Closed Questions (CQs) with structured answer options, and an Open Response (OR) consisting of a free narrative of the events expressed in natural language and without any pre-determined structure. There are a number of automated systems to analyze the CQs to obtain cause specific mortality fractions with limited performance. We hypothesize that the incorporation of the text provided by the OR might convey relevant information to discern the CoD. The experimental layout compares existing Computer Coding Verbal Autopsy methods such as Tariff 2.0 with other approaches well suited to the processing of structured inputs as is the case of the CQs. Next, alternative approaches based on language models are employed to analyze the OR. Finally, we propose a new method with a bi-modal input that combines the CQs and the OR. Empirical results corroborated that the CoD prediction capability of the Tariff 2.0 algorithm is outperformed by our method taking into account the valuable information conveyed by the OR. As an added value, with this work we made available the software to enable the reproducibility of the results attained with a version implemented in R to make the comparison with Tariff 2.0 evident.

Evaluation of methods for assigning causes of death from verbal autopsies in India

Background

Physician-coded verbal autopsy (PCVA) is the most widely used method to determine causes of death (COD) in countries where medical certification of death is low. Computer-coded verbal autopsy (CCVA), an alternative method to PCVA for assigning the COD is considered to be efficient and cost-effective. However, the performance of CCVA as compared to PCVA is yet to be established in the Indian context.

Methods

We evaluated the performance of PCVA and three CCVA methods i.e., InterVA 5, InSilico, and Tariff 2.0 on verbal autopsies done using the WHO 2016 VA tool on 2,120 reference standard cases developed from five tertiary care hospitals of Delhi. PCVA methodology involved dual independent review with adjudication, where required. Metrics to assess performance were Cause Specific Mortality Fraction (CSMF), sensitivity, positive predictive value (PPV), CSMF Accuracy, and Kappa statistic.

Results

In terms of the measures of the overall performance of COD assignment methods, for CSMF Accuracy, the PCVA method achieved the highest score of 0.79, followed by 0.67 for Tariff_2.0, 0.66 for Inter-VA and 0.62 for InSilicoVA. The PCVA method also achieved the highest agreement (57%) and Kappa scores (0.54). The PCVA method showed the highest sensitivity for 15 out of 20 causes of death.

Conclusion

Our study found that the PCVA method had the best performance out of all the four COD assignment methods that were tested in our study sample. In order to improve the performance of CCVA methods, multicentric studies with larger sample sizes need to be conducted using the WHO VA tool.

Population cause of death estimation using verbal autopsy methods in large-scale field trials of maternal and child health: lessons learned from a 20-year research collaboration in Central Ghana

Low and middle-income countries continue to use Verbal autopsies (VAs) as a World Health Organisation-recommended method to ascertain causes of death in settings where coverage of vital registration systems is not yet comprehensive. Whilst the adoption of VA has resulted in major improvements in estimating cause-specific mortality in many settings, well documented limitations have been identified relating to the standardisation of the processes involved. The WHO has invested significant resources into addressing concerns in some of these areas; there however remains enduring challenges particularly in operationalising VA surveys for deaths amongst women and children, challenges which have measurable impacts on the quality of data collected and on the accuracy of determining the final cause of death. In this paper we describe some of our key experiences and recommendations in conducting VAs from over two decades of evaluating seminal trials of maternal and child health interventions in rural Ghana. We focus on challenges along the entire VA pathway that can impact on the success rates of ascertaining the final cause of death, and lessons we have learned to optimise the procedures. We highlight our experiences of the value of the open history narratives in VAs and the training and skills required to optimise the quality of the information collected. We describe key issues in methods for ascertaining cause of death and argue that both automated and physician-based methods can be valid depending on the setting. We further summarise how increasingly popular information technology methods may be used to facilitate the processes described. Verbal autopsy is a vital means of increasing the coverage of accurate mortality statistics in low- and middle-income settings, however operationalisation remains problematic. The lessons we share here in conducting VAs within a long-term surveillance system in Ghana will be applicable to researchers and policymakers in many similar settings.

Systematic estimates of the global, regional and national under-5 mortality burden attributable to birth defects in 2000-2019: a summary of findings from the 2020 WHO estimates

OBJECTIVES

To examine the potential for bias in the estimate of under-5 mortality due to birth defects recently produced by the WHO and the Maternal and Child Epidemiology Estimation research group.

DESIGN

Systematic analysis.

METHODS

We examined the estimated number of under-5 deaths due to birth defects, the birth defect specific under-5 mortality rate, and the per cent of under-5 mortality due to birth defects, by geographic region, national income and under-5 mortality rate for three age groups from 2000 to 2019.

RESULTS

The under-5 deaths per 1000 live births from birth defects fell from 3.4 (95% uncertainty interval (UI) 3.1-3.8) in 2000 to 2.9 (UI 2.6-3.3) in 2019. The per cent of all under-5 mortality attributable to birth defects increased from 4.6% (UI 4.1%-5.1%) in 2000 to 7.6% (UI 6.9%-8.6%) in 2019. There is significant variability in mortality due to birth defects by national income level. In 2019, the under-5 mortality rate due to birth defects was less in high-income countries than in low-income and middle-income countries, 1.3 (UI 1.2-1.3) and 3.0 (UI 2.8-3.4) per 1000 live births, respectively. These mortality rates correspond to 27.7% (UI 26.6%-28.8%) of all under-5 mortality in high-income countries being due to birth defects, and 7.4% (UI 6.7%-8.2%) in low-income and middle-income countries.

CONCLUSIONS

While the under-5 mortality due to birth defects is declining, the per cent of under-5 mortality attributable to birth defects has increased, with significant variability across regions globally. The estimates in low-income and middle-income countries are likely underestimated due to the nature of the WHO estimates, which are based in part on verbal autopsy studies and should be taken as a minimum estimate. Given these limitations, comprehensive and systematic estimates of the mortality burden due to birth defects are needed to estimate the actual burden.

Effect of Neonatal Azithromycin on All-Cause and Cause-Specific Infant Mortality: A Randomized Controlled Trial

Mass azithromycin distribution reduces all-cause childhood mortality in some high-mortality settings in sub-Saharan Africa. Although the greatest benefits have been shown in children 1 to 5 months old living in areas with high mortality rates, no evidence of a benefit was found of neonatal azithromycin in a low-mortality setting on mortality at 6 months. We conducted a 1:1 randomized, placebo-controlled trial evaluating the effect of a single oral 20-mg/kg dose of azithromycin or matching placebo administered during the neonatal period on all-cause and cause-specific infant mortality at 12 months of age in five regions of Burkina Faso. Neonates were eligible if they were between the ages of 8 and 27 days and weighed at least 2,500 g at enrollment. Cause of death was determined via the WHO 2016 verbal autopsy tool. We compared all-cause and cause-specific mortality using binomial regression. Of 21,832 infants enrolled in the study, 116 died by 12 months of age. There was no significant difference in all-cause mortality between the azithromycin and placebo groups (azithromycin: 52 deaths, 0.5%; placebo, 64 deaths, 0.7%; hazard ratio, 0.81; 95% CI, 0.56-1.17; P = 0.30). There was no evidence of a difference in the distribution of causes of death (P = 0.40) and no significant difference in any specific cause of death between groups. Mortality rates were low at 12 months of age, and there was no evidence of an effect of neonatal azithromycin on all-cause or cause-specific mortality.

The openVA Toolkit for Verbal Autopsies

Verbal autopsy (VA) is a survey-based tool widely used to infer cause of death (COD) in regions without complete-coverage civil registration and vital statistics systems. In such settings, many deaths happen outside of medical facilities and are not officially documented by a medical professional. VA surveys, consisting of signs and symptoms reported by a person close to the decedent, are used to infer the COD for an individual, and to estimate and monitor the COD distribution in the population. Several classification algorithms have been developed and widely used to assign causes of death using VA data. However, the incompatibility between different idiosyncratic model implementations and required data structure makes it difficult to systematically apply and compare different methods. The openVA package provides the first standardized framework for analyzing VA data that is compatible with all openly available methods and data structure. It provides an open-source, R implementation of several most widely used VA methods. It supports different data input and output formats, and customizable information about the associations between causes and symptoms. The paper discusses the relevant algorithms, their implementations in R packages under the openVA suite, and demonstrates the pipeline of model fitting, summary, comparison, and visualization in the R environment.

Performance evaluation of machine learning and Computer Coded Verbal Autopsy (CCVA) algorithms for cause of death determination: A comparative analysis of data from rural South Africa

Computer Coded Verbal Autopsy (CCVA) algorithms are commonly used to determine the cause of death (CoD) from questionnaire responses extracted from verbal autopsies (VAs). However, they can only operate on structured data and cannot effectively harness information from unstructured VA narratives. Machine Learning (ML) algorithms have also been applied successfully in determining the CoD from VA narratives, allowing the use of auxiliary information that CCVA algorithms cannot directly utilize. However, most ML-based studies only use responses from the structured questionnaire, and the results lack generalisability and comparability across studies. We present a comparative performance evaluation of ML methods and CCVA algorithms on South African VA narratives data, using data from Agincourt Health and Demographic Surveillance Site (HDSS) with physicians' classifications as the gold standard. The data were collected from 1993 to 2015 and have 16,338 cases. The random forest and extreme gradient boosting classifiers outperformed the other classifiers on the combined dataset, attaining accuracy of 96% respectively, with significant statistical differences in algorithmic performance (p < 0.0001). All our models attained Area Under Receiver Operating Characteristics (AUROC) of greater than 0.884. The InterVA CCVA attained 83% Cause Specific Mortality Fraction accuracy and an Overall Chance-Corrected Concordance of 0.36. We demonstrate that ML models could accurately determine the cause of death from VA narratives. Additionally, through mortality trends and pattern analysis, we discovered that in the first decade of the civil registration system in South Africa, the average life expectancy was approximately 50 years. However, in the second decade, life expectancy significantly dropped, and the population was dying at a much younger average age of 40 years, mostly from the leading HIV related causes. Interestingly, in the third decade, we see a gradual improvement in life expectancy, possibly attributed to effective health intervention programmes. Through a structure and semantic analysis of narratives where experts disagree, we also demonstrate the most frequent terms of traditional healer consultations and visits. The comparative approach also makes this study a baseline that can be used for future research enforcing generalization and comparability. Future study will entail exploring deep learning models for CoD classification.

Causes of neonatal mortality using verbal autopsies in rural Southern Nepal, 2010-2017

The burden of neonatal mortality remains high worldwide, particularly in South Asia. Verbal Autopsy is a method used to identify cause of death (COD) where vital registration capabilities are lacking. This study examines the causes of neonatal mortality in a large study population in rural Southern Nepal. The data used is from a larger cluster-randomized community-based trial. The study includes 984 neonatal deaths with complete verbal autopsy information which occurred between 2010 and 2017. The InterVA-5 software was used to identify COD. COD included severe infection (sepsis, pneumonia, meningitis/encephalitis), intrapartum related events (identified as birth asphyxia), congenital malformations, and other. The neonatal mortality rate was 31.2 neonatal deaths per 1000 live births. The causes of neonatal mortality were identified as prematurity (40%), intrapartum related events (35%), severe infection (19%), congenital abnormalities (4%), and other (2%). A high proportion, 42.5% of neonatal deaths occurred in the first 24 hours after birth. Over half (56.4%) of deaths occurred at home. This large prospective study identifies population level neonatal causes of death in rural Southern Nepal, which can contribute to national and regional COD estimates. Interventions to decrease neonatal mortality should focus on preventative measures and ensuring the delivery of high risk infants at a healthcare facility in the presence of a skilled birth attendant.

Assessing the Diagnostic Accuracy of Physicians for Home Death Certification in Shanghai: Application of SmartVA

Approximately 30% of deaths in Shanghai either occur at home or are not medically attended. The recorded cause of death (COD) in these cases may not be reliable. We applied the Smart Verbal Autopsy (VA) tool to assign the COD for a representative sample of home deaths certified by 16 community health centers (CHCs) from three districts in Shanghai, from December 2017 to June 2018. The results were compared with diagnoses from routine practice to ascertain the added value of using SmartVA. Overall, cause-specific mortality fraction (CSMF) accuracy improved from 0.93 (93%) to 0.96 after the application of SmartVA. A comparison with a “gold standard (GS)” diagnoses obtained from a parallel medical record review investigation found that 86.3% of the initial diagnoses made by the CHCs were assigned the correct COD, increasing to 90.5% after the application of SmartVA. We conclude that routine application of SmartVA is not indicated for general use in CHCs, although the tool did improve diagnostic accuracy for residual causes, such as other or ill-defined cancers and non-communicable diseases.

Comparison of the Causes of Death Identified Using Automated Verbal Autopsy and Complete Autopsy among Brought-in-Dead Cases at a Tertiary Hospital in Sub-Sahara Africa

BACKGROUND

Over one-third of deaths recorded at health facilities in Zambia are brought in dead (BID) and the causes of death (CODs) are not fully analyzed. The use of automated verbal autopsy (VA) has reportedly determined the CODs of more BID cases than the death notification form issued by the hospital. However, the validity of automated VA is yet to be fully investigated.

OBJECTIVES

To compare the CODs identified by automated VA with those by complete autopsy to examine the validity of a VA tool.

METHODS

The study site was the tertiary hospital in the capital city of Zambia. From September 2019 to January 2020, all BID cases aged 13 years and older brought to the hospital during the daytime on weekdays were enrolled in this study. External COD cases were excluded. The deceased's relatives were interviewed using the 2016 World Health Organization VA questionnaire. The data were analyzed using InterVA, an automated VA tool, to determine the CODs, which were compared with the results of complete autopsies.

RESULTS

A total of 63 cases were included. The CODs of 50 BID cases were determined by both InterVA and complete autopsies. The positive predictive value of InterVA was 22%. InterVA determined the CODs correctly in 100% cases of maternal CODs, 27.5% cases of noncommunicable disease CODs, and 5.3% cases of communicable disease CODs. Using the three broader disease groups, 56.0% cases were classified in the same groups by both methods.

CONCLUSION

While the positive predictive value was low, more than half of the cases were categorized into the same broader categories. However, there are several limitations in this study, including small sample size. More research is required to investigate the factors leading to discrepancies between the CODs determined by both methods to optimize the use of automated VA in Zambia.

Care-seeking and health insurance among pregnancy-related deaths: A population-based study in Jember District, East Java Province, Indonesia

Background

Despite the increased access to facility-based delivery in Indonesia, the country’s maternal mortality remains unacceptably high. Reducing maternal mortality requires a good understanding of the care-seeking pathways for maternal complications, especially with the government moving toward universal health coverage. This study examined care-seeking practices and health insurance in instances of pregnancy-related deaths in Jember District, East Java, Indonesia.

Methods

This was a community-based cross-sectional study to identify all pregnancy-related deaths in the district from January 2017 to December 2018. Follow-up verbal and social autopsy interviews were conducted to collect information on care-seeking behavior, health insurance, causes of death, and other factors.

Findings

Among 103 pregnancy-related deaths, 40% occurred after 24 hours postpartum, 36% during delivery or within the first 24 hours postpartum, and 24% occurred while pregnant. The leading causes of deaths were hemorrhage (38.8%), pregnancy-induced hypertension (20.4%), and sepsis (16.5%). Most deaths occurred in health facilities (81.6%), primarily hospitals (74.8%). Nearly all the deceased sought care from a formal health provider during their fatal illness (93.2%). Seeking any care from an informal provider during the fatal illness was more likely among women who died after 24 hours postpartum (41.0%, OR 7.4, 95% CI 1.9, 28.5, p = 0.049) or during pregnancy (29.2%, OR 4.4, 95% CI 1.0, 19.2, p = 0.003) than among those who died during delivery or within 24 hours postpartum (8.6%). There was no difference in care-seeking patterns between insured and uninsured groups.

Conclusions

The fact that women sought care and reached health facilities regardless of their insurance status provides opportunities to prevent deaths by ensuring that every woman receives timely and quality care. Accordingly, the increasing demand should be met with balanced readiness of both primary care and hospitals to provide quality care, supported by an effective referral system.

Neonatal mortality in two districts in Indonesia: Findings from Neonatal Verbal and Social Autopsy (VASA)

Background

The Government of Indonesia is determined to follow global commitments to reduce the neonatal mortality rate. Yet, there is a paucity of information on contributing factors and causes of neonatal deaths, particularly at the sub-national level. This study describes care-seeking during neonates’ fatal illnesses and their causes of death.

Methods

We conducted a cross-sectional community-based study to identify all neonatal deaths in Serang and Jember Districts, Indonesia. Follow-up interviews were conducted with the families of deceased neonates using an adapted verbal and social autopsy instrument. Cause of death was determined using the InSilicoVA algorithm.

Results

The main causes of death of 259 neonates were prematurity (44%) and intrapartum-related events (IPRE)-mainly birth asphyxia (39%). About 83% and 74% of the 259 neonates were born and died at a health facility, respectively; 79% died within the first week after birth. Of 70 neonates whose fatal illness began at home, 59 (84%) sought care during the fatal illness. Forty-eight of those 59 neonates went to a formal care provider; 36 of those 48 neonates (75%) were moderately or severely ill when the family decided to seek care. One hundred fifteen of 189 neonates (61%) whose fatal illnesses began at health facilities were born at a hospital. Among those 115, only 24 (21%) left the hospital alive–of whom 16 (67%) were referred by the hospital.

Conclusions

The high proportion of deaths due to prematurity and IPRE suggests the need for improved management of small and asphyxiated newborns. The moderate to severe condition of neonates at the time when care was sought from home highlights the importance of early illness recognition and appropriate management for sick neonates. Among deceased neonates whose fatal illness began at their delivery hospital, the high proportion of referrals may indicate issues with hospital capability, capacity, and/or cost.

A FLEXIBLE BAYESIAN FRAMEWORK TO ESTIMATE AGE- AND CAUSE-SPECIFIC CHILD MORTALITY OVER TIME FROM SAMPLE REGISTRATION DATA

In order to implement disease-specific interventions in young age groups, policy makers in low- and middle-income countries require timely and accurate estimates of age- and cause-specific child mortality. High-quality data is not available in settings where these interventions are most needed, but there is a push to create sample registration systems that collect detailed mortality information. current methods that estimate mortality from this data employ multistage frameworks without rigorous statistical justification that separately estimate all-cause and cause-specific mortality and are not sufficiently adaptable to capture important features of the data. We propose a flexible Bayesian modeling framework to estimate age- and cause-specific child mortality from sample registration data. We provide a theoretical justification for the framework, explore its properties via simulation, and use it to estimate mortality trends using data from the Maternal and Child Health Surveillance System in China.

Adapting and validating the log quadratic model to derive under-five age- and cause-specific mortality (U5ACSM): a preliminary analysis

BACKGROUND

The mortality pattern from birth to age five is known to vary by underlying cause of mortality, which has been documented in multiple instances. Many countries without high functioning vital registration systems could benefit from estimates of age- and cause-specific mortality to inform health programming, however, to date the causes of under-five death have only been described for broad age categories such as for neonates (0-27 days), infants (0-11 months), and children age 12-59 months.

METHODS

We adapt the log quadratic model to mortality patterns for children under five to all-cause child mortality and then to age- and cause-specific mortality (U5ACSM). We apply these methods to empirical sample registration system mortality data in China from 1996 to 2015. Based on these empirical data, we simulate probabilities of mortality in the case when the true relationships between age and mortality by cause are known.

RESULTS

We estimate U5ACSM within 0.1-0.7 deaths per 1000 livebirths in hold out strata for life tables constructed from the China sample registration system, representing considerable improvement compared to an error of 1.2 per 1000 livebirths using a standard approach. This improved prediction error for U5ACSM is consistently demonstrated for all-cause as well as pneumonia- and injury-specific mortality. We also consistently identified cause-specific mortality patterns in simulated mortality scenarios.

CONCLUSION

The log quadratic model is a significant improvement over the standard approach for deriving U5ACSM based on both simulation and empirical results.

Validation of SmartVA using conventional autopsy: A study of adult deaths in Brazil

BACKGROUND

Accurate cause of death data are essential to guide health policy. However, mortality surveillance is limited in many low-income countries. In such settings, verbal autopsy (VA) is increasingly used to provide population-level cause of death data. VAs are now widely interpreted using the automated algorithms SmartVA and InterVA. Here we use conventional autopsy as the gold standard to validate SmartVA methodology.

METHODS

This study included adult deaths from natural causes in São Paulo and Recife for which conventional autopsy was indicated. VA was conducted with a relative of the deceased using an amended version of the SmartVA instrument to suit the local context. Causes of death from VA were produced using the SmartVA-Analyze program. Physician coded verbal autopsy (PCVA), conducted on the same questionnaires, and Global Burden of Disease Study data were used as additional comparators. Cause of death data were grouped into 10 broad causes for the validation due to the real-world utility of VA lying in identifying broad population cause of death patterns.

FINDINGS

The study included 2,060 deaths in São Paulo and 1,079 in Recife. The cause specific mortality fractions (CSMFs) estimated using SmartVA were broadly similar to conventional autopsy for: cardiovascular diseases (46.8% vs 54.0%, respectively), cancers (10.6% vs 11.4%), infections (7.0% vs 10.4%) and chronic respiratory disease (4.1% vs 3.7%), causes accounting for 76.1% of the autopsy dataset. The SmartVA CSMF estimates were lower than autopsy for "Other NCDs" (7.8% vs 14.6%) and higher for diabetes (13.0% vs 6.6%). CSMF accuracy of SmartVA compared to autopsy was 84.5%. CSMF accuracy for PCVA was 93.0%.

INTERPRETATION

The results suggest that SmartVA can, with reasonable accuracy, predict the broad cause of death groups important to assess a population's epidemiological transition. VA remains a useful tool for understanding causes of death where medical certification is not possible.

Social and demographic correlates of cardiovascular mortality in the Kassena-Nankana districts of Ghana: a verbal post-mortem analysis

BACKGROUND

The burden of cardiovascular disease (CVD) in Ghana is rising, but details on its epidemiology are scarce. We sought to quantify mortality due to CVD in two districts in rural Ghana using verbal post-mortem (VPM) data.

METHODS

We conducted a proportional sub-hazards analysis of 10 232 deaths in the Kassena-Nankana East and West districts from 2005 to 2012, to determine adult mortality attributed to CVD over time. We stratified results by age, gender and socio-economic status (SES), and compared CVD mortality among SES and gender strata over time. A competing risk model estimated the cumulative effect of eliminating CVD from the area.

RESULTS

From 2005 to 2012, CVD mortality more than doubled overall, from 0.51 deaths for every 1000 person-years in 2005 to 1.08 per 1000 person-years in 2012. Mortality peaked in 2008 at 1.23 deaths per 1000 person-years. Increases were comparable in men (2.0) and women (2.3), but greater among the poorest residents (3.3) than the richest (1.3), and among persons aged 55-69 years (2.1) than those aged ≥70 years (1.8). By 2012, male and female CVD mortality was highest in middle-SES persons. We project that eliminating CVD would increase the number of individuals reaching age 73 years from 35% to 40%, adding 1.6 years of life expectancy.

CONCLUSIONS

The burden of CVD on overall mortality in the Upper East Region is substantial and markedly increasing. CVD mortality has especially increased in lower-income persons and persons in middle age. Further initiatives for the surveillance and control of CVD in these vulnerable populations are needed.

Mortality transition over a quarter century in rural South Africa: findings from population surveillance in Agincourt 1993-2018

Background

Mortality burden in South Africa since the mid-1990s has been characterized by a quadruple disease burden: HIV/AIDS and tuberculosis (TB); other communicable diseases (excluding HIV/AIDS and TB), maternal causes, perinatal conditions and nutritional deficiencies; non-communicable diseases (NCDs); and injuries. Causes from these broad groupings have persistently constituted the top 10 causes of death. However, proportions and rankings have varied over time, alongside overall mortality levels.

Objective

To provide evidence on the contributions of age and cause-of-death to changes in mortality levels in a rural South African population over a quarter century (1993–2018).

Methods

Using mortality and cause-of-death data from the Agincourt Health and Socio-Demographic Surveillance System (HDSS), we derive estimates of the distribution of deaths by cause, and hazards of death by age, sex, and time period, 1993–2018. We derive estimates of life expectancies at birth and years of life expectancy gained at age 15 if most common causes of death were deleted. We compare mortality indicators and cause-of-death trends from the Agincourt HDSS with South African national indicators generated from publicly available datasets.

Results

Mortality and cause-of-death transition reveals that overall mortality levels have returned to pre-HIV epidemic levels. In recent years, the concentration of mortality has shifted towards older ages, and the mortality burden from cardiovascular diseases and other chronic NCDs are more prominent as people living with HIV/AIDS access ART and live longer. Changes in life expectancy at birth, distribution of deaths by age, and major cause-of-death categories in the Agincourt population follow a similar pattern to the South African population.

Conclusion

The Agincourt HDSS provides critical information about general mortality, cause-of-death, and age patterns in rural South Africa. Realigning and strengthening the South African public health and healthcare systems is needed to concurrently cater for the prevention, control, and treatment of multiple disease conditions.

Estimating causes of death where there is no medical certification: evolution and state of the art of verbal autopsy

Over the past 70 years, significant advances have been made in determining the causes of death in populations not served by official medical certification of cause at the time of death using a technique known as Verbal Autopsy (VA). VA involves an interview of the family or caregivers of the deceased after a suitable bereavement interval about the circumstances, signs and symptoms of the deceased in the period leading to death. The VA interview data are then interpreted by physicians or, more recently, computer algorithms, to assign a probable cause of death. VA was originally developed and applied in field research settings. This paper traces the evolution of VA methods with special emphasis on the World Health Organization's (WHO)'s efforts to standardize VA instruments and methods for expanded use in routine health information and vital statistics systems in low- and middle-income countries (LMICs). These advances in VA methods are culminating this year with the release of the 2022 WHO Standard Verbal Autopsy (VA) Toolkit. This paper highlights the many contributions the late Professor Peter Byass made to the current VA standards and methods, most notably, the development of InterVA, the most commonly used automated computer algorithm for interpreting data collected in the WHO standard instruments, and the capacity building in low- and middle-income countries (LMICs) that he promoted. This paper also provides an overview of the methods used to improve the current WHO VA standards, a catalogue of the changes and improvements in the instruments, and a mapping of current applications of the WHO VA standard approach in LMICs. It also provides access to tools and guidance needed for VA implementation in Civil Registration and Vital Statistics Systems at scale.

Comparison of machine learning algorithms applied to symptoms to determine infectious causes of death in children: national survey of 18,000 verbal autopsies in the Million Death Study in India

BACKGROUND

Machine learning (ML) algorithms have been successfully employed for prediction of outcomes in clinical research. In this study, we have explored the application of ML-based algorithms to predict cause of death (CoD) from verbal autopsy records available through the Million Death Study (MDS).

METHODS

From MDS, 18826 unique childhood deaths at ages 1-59 months during the time period 2004-13 were selected for generating the prediction models of which over 70% of deaths were caused by six infectious diseases (pneumonia, diarrhoeal diseases, malaria, fever of unknown origin, meningitis/encephalitis, and measles). Six popular ML-based algorithms such as support vector machine, gradient boosting modeling, C5.0, artificial neural network, k-nearest neighbor, classification and regression tree were used for building the CoD prediction models.

RESULTS

SVM algorithm was the best performer with a prediction accuracy of over 0.8. The highest accuracy was found for diarrhoeal diseases (accuracy = 0.97) and the lowest was for meningitis/encephalitis (accuracy = 0.80). The top signs/symptoms for classification of these CoDs were also extracted for each of the diseases. A combination of signs/symptoms presented by the deceased individual can effectively lead to the CoD diagnosis.

CONCLUSIONS

Overall, this study affirms that verbal autopsy tools are efficient in CoD diagnosis and that automated classification parameters captured through ML could be added to verbal autopsies to improve classification of causes of death.

Verbal autopsy models in determining causes of death

OBJECTIVES

To systematically review current practices, strengths and limitations of existing VA approaches to increase understanding of health system stakeholders and researchers.

METHODS

The review was conducted and reported based on the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) guidelines, in which articles were systematically obtained from the PubMed and SCOPUS online databases. The search was limited to English language journal articles published between 2010 and 2020. The review identified 5602 articles and after thorough scrutiny, 25 articles related to VA approaches were included.

RESULTS

(1) InterVA and Tariff are widely used VA models; (2) Bayes rule is the most common and successful algorithm; (3) the lack of standardised datasets and metrics to evaluate models creates bias in determining VA model performance; (4) performance of the models trained using in-hospital data cannot be replicated in community death; (5) the performance of models among physicians and computer-coded algorithms differs with variation in settings.

CONCLUSION

The physician-certified verbal autopsy (PCVA) approaches are more effective in determining community CoD while computerised coding of verbal autopsy (CCVA) models perform well when the underlying CoD are reliably established using hospital data where data are trained in a similar environment to the target population. Our study recommends the use of hybrid models that combine strengths from various models and using an open standards dataset that includes death from different settings.

A Semantic-Based Framework for Verbal Autopsy to Identify the Cause of Maternal Death

OBJECTIVE

Verbal autopsy is a technique used to collect information about a decedent from his/her family members using questionnaires, conducting interviews, making observations, and sampling. In substantial parts of the world, particularly in Africa and Asia, many deaths are unrecorded. In 2017, globally pregnant women were dying daily around 810 and 295,000 in a year because of pregnancy-related problems, pointed out by World Health Organization. Identifying the cause of a death is a complex process which requires in-depth medical knowledge and practical experience. Generally, medical practitioners possess different knowledge levels, set of abilities, and problem-solving skills. Additionally, the medical negligence plays a significant part in further worsening the situation. Accurate identification of the cause of death can help a government to take strategic measures to focus on, particularly increasing the death rate in a specific region.

METHODS

This research provides a solution by introducing a semantic-based verbal autopsy framework for maternal death (SVAF-MD) to identify the cause of death. The proposed framework consists of four main components as follows: (1) clinical practice guidelines, (2) knowledge collection, (3) knowledge modeling, and (4) knowledge codification. Maternal ontology for the framework is developed using Protégé knowledge editor. Resource description framework application programming interface (API) for PHP (RAP) is used as a Semantic Web toolkit along with Simple Protocol and RDF Query Language (SPARQL) is used for querying with ontology to retrieve data.

RESULTS

The results show that 92% of maternal causes of deaths assigned using SVAF-MD correctly matched manual reports already prepared by gynecologists.

CONCLUSION

SVAF-MD, a semantic-based framework for the verbal autopsy of maternal deaths, assigns the cause of death with minimum involvement of medical practitioners. This research helps the government to ease down the verbal autopsy process, overcome the delays in reporting, and facilitate in terms of accurate results to devise the policies to reduce the maternal mortality.

Bayesian hierarchical factor regression models to infer cause of death from verbal autopsy data

In low-resource settings where vital registration of death is not routine it is often of critical interest to determine and study the cause of death (COD) for individuals and the cause-specific mortality fraction (CSMF) for populations. Post-mortem autopsies, considered the gold standard for COD assignment, are often difficult or impossible to implement due to deaths occurring outside the hospital, expense, and/or cultural norms. For this reason, Verbal Autopsies (VAs) are commonly conducted, consisting of a questionnaire administered to next of kin recording demographic information, known medical conditions, symptoms, and other factors for the decedent. This article proposes a novel class of hierarchical factor regression models that avoid restrictive assumptions of standard methods, allow both the mean and covariance to vary with COD category, and can include covariate information on the decedent, region, or events surrounding death. Taking a Bayesian approach to inference, this work develops an MCMC algorithm and validates the FActor Regression for Verbal Autopsy (FARVA) model in simulation experiments. An application of FARVA to real VA data shows improved goodness-of-fit and better predictive performance in inferring COD and CSMF over competing methods. Code and a user manual are made available at https://github.com/kelrenmor/farva.

Extracting Cause of Death From Verbal Autopsy With Deep Learning Interpretable Methods

The international standard to ascertain the cause of death is medical certification. However, in many low and middle-income countries, the majority of deaths occur outside of health facilities. In these cases, Verbal Autopsy (VA), the narrative provided by a family member or friend together with a questionnaire is designed by the World Health Organization as the main information source. Until now technology allowed us to automatically analyze the responses of the VA questionnaire with the narrative captured by the interviewer excluded. Our work addresses this gap by developing a set of models for automatic Cause of Death (CoD) ascertainment in VAs with a focus on the textual information. Empirical results show that the open response conveys valuable information towards the ascertainment of the Cause of Death, and the combination of the closed-ended questions and the open response lead to the best results. Model interpretation capabilities position the Deep Learning models as the most encouraging choice.

Estimating the mortality burden of large scale mining projects-Evidence from a prospective mortality surveillance study in Tanzania

We set up a mortality surveillance system around two of the largest gold mines in Tanzania between February 2019 and February 2020 to estimate the mortality impact of gold mines. Death circumstances were collected using a standardized verbal autopsy tool, and causes of death were assigned using the InSilicoVA algorithm. We compared cause-specific mortality fractions in mining communities with other subnational data as well as national estimates. Within mining communities, we estimated mortality risks of mining workers relative to other not working at mines. At the population level, mining communities had higher road-traffic injuries (RTI) (risk difference (RD): 3.1%, Confidence Interval (CI): 0.4%, 5.9%) and non-HIV infectious disease mortality (RD: 5.6%, CI: 0.8%, 10.3%), but lower burden of HIV mortality (RD: -5.9%, CI: -10.2%, -1.6%). Relative to non-miners living in the same communities, mining workers had over twice the mortality risk (relative risk (RR): 2.09, CI: 1.57, 2.79), with particularly large increases for death due to RTIs (RR: 14.26, CI: 4.95, 41.10) and other injuries (RR:10.10, CI: 3.40, 30.02). Our results shows that gold mines continue to be associated with a large mortality burden despite major efforts to ensure the safety in mining communities. Given that most of the additional mortality risk appears to be related to injuries programs targeting these specific risks seem most desirable.

Methods for correcting inference based on outcomes predicted by machine learning

Many modern problems in medicine and public health leverage machine-learning methods to predict outcomes based on observable covariates. In a wide array of settings, predicted outcomes are used in subsequent statistical analysis, often without accounting for the distinction between observed and predicted outcomes. We call inference with predicted outcomes postprediction inference. In this paper, we develop methods for correcting statistical inference using outcomes predicted with arbitrarily complicated machine-learning models including random forests and deep neural nets. Rather than trying to derive the correction from first principles for each machine-learning algorithm, we observe that there is typically a low-dimensional and easily modeled representation of the relationship between the observed and predicted outcomes. We build an approach for postprediction inference that naturally fits into the standard machine-learning framework where the data are divided into training, testing, and validation sets. We train the prediction model in the training set, estimate the relationship between the observed and predicted outcomes in the testing set, and use that relationship to correct subsequent inference in the validation set. We show our postprediction inference (postpi) approach can correct bias and improve variance estimation and subsequent statistical inference with predicted outcomes. To show the broad range of applicability of our approach, we show postpi can improve inference in two distinct fields: modeling predicted phenotypes in repurposed gene expression data and modeling predicted causes of death in verbal autopsy data. Our method is available through an open-source R package: https://github.com/leekgroup/postpi.

Effect of Mass Treatment with Azithromycin on Causes of Death in Children in Malawi: Secondary Analysis from the MORDOR Trial

Recent evidence indicates mass drug administration with azithromycin may reduce child mortality. This study uses verbal autopsy (VA) to investigate the causes of individual deaths during the Macrolides Oraux pour Réduire les Décès avec un Oeil sur la Résistance (MORDOR) trial in Malawi. Cluster randomization was performed as part of MORDOR. Biannual household visits were conducted to distribute azithromycin or placebo to children aged 1–59 months and update the census to identify deaths for VA. MORDOR was not powered to investigate mortality effects at individual sites, but the available evidence is presented here for hypothesis generation regarding the mechanism through which azithromycin may reduce child mortality. Automated VA analysis was performed to infer the likely cause of death using two major analysis programs, InterVA and SmartVA. A total of 334 communities were randomized to azithromycin or placebo, with more than 130,000 person-years of follow-up. During the study, there were 1,184 deaths, of which 1,131 were followed up with VA. Mortality was 9% lower in azithromycin-treated communities than in placebo communities (rate ratio 0.91 [95% CI: 0.79–1.05]; P = 0.20). The intention-to-treat analysis by cause using InterVA suggested fewer HIV/AIDS deaths in azithromycin-treated communities (rate ratio 0.70 [95% CI: 0.50–0.97]; P = 0.03) and fewer pneumonia deaths (rate ratio 0.82 [95% CI: 0.60–1.12]; P = 0.22). The use of the SmartVA algorithm suggested fewer diarrhea deaths (rate ratio 0.71 [95% CI: 0.51–1.00]; P = 0.05) and fewer pneumonia deaths (rate ratio 0.58 [95% CI: 0.33–1.00]; P = 0.05). Although this study is not able to provide strong evidence, the data suggest that the mortality reduction during MORDOR in Malawi may have been due to effects on pneumonia and diarrhea or HIV/AIDS mortality.

Probabilistic cause-of-disease assignment using case-control diagnostic tests: A latent variable regression approach

Optimal prevention and treatment strategies for a disease of multiple causes, such as pneumonia, must be informed by the population distribution of causes among cases, or cause-specific case fractions (CSCFs). CSCFs may further depend on additional explanatory variables. Existing methodological literature in disease etiology research does not fully address the regression problem, particularly under a case-control design. Based on multivariate binary non-gold-standard diagnostic data and additional covariate information, this article proposes a novel and unified regression modeling framework for estimating covariate-dependent CSCF functions in case-control disease etiology studies. The model leverages critical control data for valid probabilistic cause assignment for cases. We derive an efficient Markov chain Monte Carlo algorithm for flexible posterior inference. We illustrate the inference of CSCF functions using extensive simulations and show that the proposed model produces less biased estimates and more valid inference of the overall CSCFs than analyses that omit covariates. A regression analysis of pediatric pneumonia data reveals the dependence of CSCFs upon season, age, human immunodeficiency virus status and disease severity. The article concludes with a brief discussion on model extensions that may further enhance the utility of the regression model in broader disease etiology research.

Using Bayesian Latent Gaussian Graphical Models to Infer Symptom Associations in Verbal Autopsies

Learning dependence relationships among variables of mixed types provides insights in a variety of scientific settings and is a well-studied problem in statistics. Existing methods, however, typically rely on copious, high quality data to accurately learn associations. In this paper, we develop a method for scientific settings where learning dependence structure is essential, but data are sparse and have a high fraction of missing values. Specifically, our work is motivated by survey-based cause of death assessments known as verbal autopsies (VAs). We propose a Bayesian approach to characterize dependence relationships using a latent Gaussian graphical model that incorporates informative priors on the marginal distributions of the variables. We demonstrate such information can improve estimation of the dependence structure, especially in settings with little training data. We show that our method can be integrated into existing probabilistic cause-of-death assignment algorithms and improves model performance while recovering dependence patterns between symptoms that can inform efficient questionnaire design in future data collection.

Direct maternal deaths attributable to HIV in the era of antiretroviral therapy: evidence from three population-based HIV cohorts with verbal autopsy

OBJECTIVE

To assess whether HIV is associated with an increased risk of mortality from direct maternal complications.

DESIGN

Population-based cohort study using data from three demographic surveillance sites in Eastern and Southern Africa.

METHODS

We use verbal autopsy data, with cause of death assigned using the InSilicoVA algorithm, to describe the association between HIV and direct maternal deaths amongst women aged 20-49 years. We report direct maternal mortality rates by HIV status, and crude and adjusted rate ratios comparing HIV-infected and uninfected women, by study site and by ART availability. We pool the study-specific rate ratios using random-effects meta-analysis.

RESULTS

There was strong evidence that HIV increased the rate of direct maternal mortality across all the study sites in the period ART was widely available, with the rate ratios varying from 4.5 in Karonga, Malawi [95% confidence interval (CI) 1.6-12.6] to 5.2 in Kisesa, Tanzania (95% CI 1.7-16.1) and 5.9 in uMkhanyakude, South Africa (95% CI 2.3-15.2) after adjusting for sociodemographic confounders. Combining these adjusted results across the study sites, we estimated that HIV-infected women have 5.2 times the rate of direct maternal mortality compared with HIV-uninfected women (95% CI 2.9-9.5).

CONCLUSION

HIV-infected women face higher rates of mortality from direct maternal causes, which suggests that we need to improve access to quality maternity care for these women. These findings also have implications for the surveillance of HIV/AIDS-related mortality, as not all excess mortality attributable to HIV will be explicitly attributed to HIV/AIDS on the basis of a verbal autopsy interview.

Using health facility deaths to estimate population causes of neonatal and child mortality in four African countries

BACKGROUND

Verbal autopsy is the main method used in countries with weak civil registration systems for estimating community causes of neonatal and 1-59-month-old deaths. However, validation studies of verbal autopsy methods are limited and assessment has been dependent on hospital-based studies, with uncertain implications for its validity in community settings. If the distribution of community deaths by cause was similar to that of facility deaths, or could be adjusted according to related demographic factors, then the causes of facility deaths could be used to estimate population causes.

METHODS

Causes of neonatal and 1-59-month-old deaths from verbal/social autopsy (VASA) surveys in four African countries were estimated using expert algorithms (EAVA) and physician coding (PCVA). Differences between facility and community deaths in individual causes and cause distributions were examined using chi-square and cause-specific mortality fractions (CSMF) accuracy, respectively. Multinomial logistic regression and random forest models including factors from the VASA studies that are commonly available in Demographic and Health Surveys were built to predict population causes from facility deaths.

RESULTS

Levels of facility and community deaths in the four countries differed for one to four of 10 EAVA or PCVA neonatal causes and zero to three of 12 child causes. CSMF accuracy for facility compared to community deaths in the four countries ranged from 0.74 to 0.87 for neonates and 0.85 to 0.95 for 1-59-month-olds. Crude CSMF accuracy in the prediction models averaged 0.86 to 0.88 for neonates and 0.93 for 1-59-month-olds. Adjusted random forest prediction models increased average CSMF accuracy for neonates to, at most, 0.90, based on small increases in all countries.

CONCLUSIONS

There were few differences in facility and community causes of neonatal and 1-59-month-old deaths in the four countries, and it was possible to project the population CSMF from facility deaths with accuracy greater than the validity of verbal autopsy diagnoses. Confirmation of these findings in additional settings would warrant research into how medical causes of deaths in a representative sample of health facilities can be utilized to estimate the population causes of child death.

Under 10 mortality patterns, risk factors, and mechanisms in low resource settings of Eastern Uganda: An analysis of event history demographic and verbal social autopsy data

Background

Globally, the under-10 years of age mortality has not been comprehensively studied. We applied the life-course perspective in the analysis and interpretation of the event history demographic and verbal autopsy data to examine when and why children die before their 10^th birthday.

Methods

We analysed a decade (2005–2015) of event histories data on 22385 and 1815 verbal autopsies data collected by Iganga-Mayuge HDSS in eastern Uganda. We used the lifetable for mortality estimates and patterns, and Royston-Parmar survival analysis approach for mortality risk factors’ assessment.

Results

The under-10 and 5–9 years of age mortality probabilities were 129 (95% Confidence Interval [CI] = 123–370) per 1000 live births and 11 (95% CI = 7–26) per 1000 children aged 5–9 years, respectively. The top four causes of new-born mortality and stillbirth were antepartum maternal complications (31%), intrapartum-related causes including birth injury, asphyxia and obstructed labour (25%), Low Birth Weight (LBW) and prematurity (20%), and other unidentified perinatal mortality causes (18%). Malaria, protein deficiency including anaemia, diarrhoea or gastrointestinal, and acute respiratory infections were the major causes of mortality among those aged 0–9 years–contributing 88%, 88% and 46% of all causes of mortality for the post-neonatal, child and 5–9 years of age respectively. 33% of all causes of mortality among those aged 5–9 years was a share of Injuries (22%) and gastrointestinal (11%). Regarding the deterministic pattern, nearly 30% of the new-borns and sick children died without access to formal care. Access to the treatment for the top five morbidities was after 4 days of symptoms’ recognition. The childhood mortality risk factors were LBW, multiple births, having no partner, adolescence age, rural residence, low education level and belonging to a poor household, but their association was stronger among infants.

Conclusions

We have identified the vulnerable groups at risk of mortality as LBW children, multiple births, rural dwellers, those whose mother are of low socio-economic position, adolescents and unmarried. The differences in causes of mortalities between children aged 0–5 and 5–9 years were noted. These findings suggest for a strong life-course approach in the design and implementation of child health interventions that target pregnant women and children of all ages.

Non-confirming replication of "Performance of InSilicoVA for assigning causes of death to verbal autopsies: multisite validation study using clinical diagnostic gold standards," by Flaxman et al

BACKGROUND

A verbal autopsy (VA) is an interview conducted with the caregivers of someone who has recently died to describe the circumstances of the death. In recent years, several algorithmic methods have been developed to classify cause of death using VA data. The performance of one method-InSilicoVA-was evaluated in a study by Flaxman et al., published in BMC Medicine in 2018. The results of that study are different from those previously published by our group.

METHODS

Based on the description of methods in the Flaxman et al. study, we attempt to replicate the analysis to understand why the published results differ from those of our previous work.

RESULTS

We failed to reproduce the results published in Flaxman et al. Most of the discrepancies we find likely result from undocumented differences in data pre-processing, and/or values assigned to key parameters governing the behavior of the algorithm.

CONCLUSION

This finding highlights the importance of making replication code available along with published results. All code necessary to replicate the work described here is freely available on GitHub.

Automated verbal autopsy: from research to routine use in civil registration and vital statistics systems

BACKGROUND

The majority of low- and middle-income countries (LMICs) do not have adequate civil registration and vital statistics (CRVS) systems to properly support health policy formulation. Verbal autopsy (VA), long used in research, can provide useful information on the cause of death (COD) in populations where physicians are not available to complete medical certificates of COD. Here, we report on the application of the SmartVA tool for the collection and analysis of data in several countries as part of routine CRVS activities.

METHODS

Data from VA interviews conducted in 4 of 12 countries supported by the Bloomberg Philanthropies Data for Health (D4H) Initiative, and at different stages of health statistical development, were analysed and assessed for plausibility: Myanmar, Papua New Guinea (PNG), Bangladesh and the Philippines. Analyses by age- and cause-specific mortality fractions were compared to the Global Burden of Disease (GBD) study data by country. VA interviews were analysed using SmartVA-Analyze-automated software that was designed for use in CRVS systems. The method in the Philippines differed from the other sites in that the VA output was used as a decision support tool for health officers.

RESULTS

Country strategies for VA implementation are described in detail. Comparisons between VA data and country GBD estimates by age and cause revealed generally similar patterns and distributions. The main discrepancy was higher infectious disease mortality and lower non-communicable disease mortality at the PNG VA sites, compared to the GBD country models, which critical appraisal suggests may highlight real differences rather than implausible VA results.

CONCLUSION

Automated VA is the only feasible method for generating COD data for many populations. The results of implementation in four countries, reported here under the D4H Initiative, confirm that these methods are acceptable for wide-scale implementation and can produce reliable COD information on community deaths for which little was previously known.

BAYESIAN FACTOR MODELS FOR PROBABILISTIC CAUSE OF DEATH ASSESSMENT WITH VERBAL AUTOPSIES

The distribution of deaths by cause provides crucial information for public health planning, response and evaluation. About 60% of deaths globally are not registered or given a cause, limiting our ability to understand disease epidemiology. Verbal autopsy (VA) surveys are increasingly used in such settings to collect information on the signs, symptoms and medical history of people who have recently died. This article develops a novel Bayesian method for estimation of population distributions of deaths by cause using verbal autopsy data. The proposed approach is based on a multivariate probit model where associations among items in questionnaires are flexibly induced by latent factors. Using the Population Health Metrics Research Consortium labeled data that include both VA and medically certified causes of death, we assess performance of the proposed method. Further, we estimate important questionnaire items that are highly associated with causes of death. This framework provides insights that will simplify future data.

Regularized Bayesian transfer learning for population-level etiological distributions

Computer-coded verbal autopsy (CCVA) algorithms predict cause of death from high-dimensional family questionnaire data (verbal autopsy) of a deceased individual, which are then aggregated to generate national and regional estimates of cause-specific mortality fractions. These estimates may be inaccurate if CCVA is trained on non-local training data different from the local population of interest. This problem is a special case of transfer learning, i.e., improving classification within a target domain (e.g., a particular population) with the classifier trained in a source-domain. Most transfer learning approaches concern individual-level (e.g., a person’s) classification. Social and health scientists such as epidemiologists are often more interested with understanding etiological distributions at the population-level. The sample sizes of their data sets are typically orders of magnitude smaller than those used for common transfer learning applications like image classification, document identification, etc. We present a parsimonious hierarchical Bayesian transfer learning framework to directly estimate population-level class probabilities in a target domain, using any baseline classifier trained on source-domain, and a small labeled target-domain dataset. To address small sample sizes, we introduce a novel shrinkage prior for the transfer error rates guaranteeing that, in absence of any labeled target-domain data or when the baseline classifier is perfectly accurate, our transfer learning agrees with direct aggregation of predictions from the baseline classifier, thereby subsuming the default practice as a special case. We then extend our approach to use an ensemble of baseline classifiers producing an unified estimate. Theoretical and empirical results demonstrate how the ensemble model favors the most accurate baseline classifier. We present data analyses demonstrating the utility of our approach.

Robustness of the Tariff method for diagnosing verbal autopsies: impact of additional site data on the relationship between symptom and cause

Background

Verbal autopsy (VA) is increasingly being considered as a cost-effective method to improve cause of death information in countries with low quality vital registration. VA algorithms that use empirical data have an advantage over expert derived algorithms in that they use responses to the VA instrument as a reference instead of physician opinion. It is unclear how stable these data driven algorithms, such as the Tariff 2.0 method, are to cultural and epidemiological variations in populations where they might be employed.

Methods

VAs were conducted in three sites as part of the Improving Methods to Measure Comparable Mortality by Cause (IMMCMC) study: Bohol, Philippines; Chandpur and Comila Districts, Bangladesh; and Central and Eastern Highlands Provinces, Papua New Guinea. Similar diagnostic criteria and cause lists as the Population Health Metrics Research Consortium (PHMRC) study were used to identify gold standard (GS) deaths. We assessed changes in Tariffs by examining the proportion of Tariffs that changed significantly after the addition of the IMMCMC dataset to the PHMRC dataset.

Results

The IMMCMC study added 3512 deaths to the GS VA database (2491 adults, 320 children, and 701 neonates). Chance-corrected cause specific mortality fractions for Tariff improved with the addition of the IMMCMC dataset for adults (+ 5.0%), children (+ 5.8%), and neonates (+ 1.5%). 97.2% of Tariffs did not change significantly after the addition of the IMMCMC dataset.

Conclusions

Tariffs generally remained consistent after adding the IMMCMC dataset. Population level performance of the Tariff method for diagnosing VAs improved marginally for all age groups in the combined dataset. These findings suggest that cause-symptom relationships of Tariff 2.0 might well be robust across different population settings in developing countries. Increasing the total number of GS deaths improves the validity of Tariff and provides a foundation for the validation of other empirical algorithms.

Automatically determining cause of death from verbal autopsy narratives

Background

A verbal autopsy (VA) is a post-hoc written interview report of the symptoms preceding a person’s death in cases where no official cause of death (CoD) was determined by a physician. Current leading automated VA coding methods primarily use structured data from VAs to assign a CoD category. We present a method to automatically determine CoD categories from VA free-text narratives alone.

Methods

After preprocessing and spelling correction, our method extracts word frequency counts from the narratives and uses them as input to four different machine learning classifiers: naïve Bayes, random forest, support vector machines, and a neural network.

Results

For individual CoD classification, our best classifier achieves a sensitivity of.770 for adult deaths for 15 CoD categories (as compared to the current best reported sensitivity of.57), and.662 with 48 WHO categories. When predicting the CoD distribution at the population level, our best classifier achieves.962 cause-specific mortality fraction accuracy for 15 categories and.908 for 48 categories, which is on par with leading CoD distribution estimation methods.

Conclusions

Our narrative-based machine learning classifier performs as well as classifiers based on structured data at the individual level. Moreover, our method demonstrates that VA narratives provide important information that can be used by a machine learning system for automated CoD classification. Unlike the structured questionnaire-based methods, this method can be applied to any verbal autopsy dataset, regardless of the collection process or country of origin.

Electronic supplementary material

The online version of this article (10.1186/s12911-019-0841-9) contains supplementary material, which is available to authorized users.

Automated versus physician assignment of cause of death for verbal autopsies: randomized trial of 9374 deaths in 117 villages in India

BACKGROUND

Verbal autopsies with physician assignment of cause of death (COD) are commonly used in settings where medical certification of deaths is uncommon. It remains unanswered if automated algorithms can replace physician assignment.

METHODS

We randomized verbal autopsy interviews for deaths in 117 villages in rural India to either physician or automated COD assignment. Twenty-four trained lay (non-medical) surveyors applied the allocated method using a laptop-based electronic system. Two of 25 physicians were allocated randomly to independently code the deaths in the physician assignment arm. Six algorithms (Naïve Bayes Classifier (NBC), King-Lu, InSilicoVA, InSilicoVA-NT, InterVA-4, and SmartVA) coded each death in the automated arm. The primary outcome was concordance with the COD distribution in the standard physician-assigned arm. Four thousand six hundred fifty-one (4651) deaths were allocated to physician (standard), and 4723 to automated arms.

RESULTS

The two arms were nearly identical in demographics and key symptom patterns. The average concordances of automated algorithms with the standard were 62%, 56%, and 59% for adult, child, and neonatal deaths, respectively. Automated algorithms showed inconsistent results, even for causes that are relatively easy to identify such as road traffic injuries. Automated algorithms underestimated the number of cancer and suicide deaths in adults and overestimated other injuries in adults and children. Across all ages, average weighted concordance with the standard was 62% (range 79-45%) with the best to worst ranking automated algorithms being InterVA-4, InSilicoVA-NT, InSilicoVA, SmartVA, NBC, and King-Lu. Individual-level sensitivity for causes of adult deaths in the automated arm was low between the algorithms but high between two independent physicians in the physician arm.

CONCLUSIONS

While desirable, automated algorithms require further development and rigorous evaluation. Lay reporting of deaths paired with physician COD assignment of verbal autopsies, despite some limitations, remains a practicable method to document the patterns of mortality reliably for unattended deaths.

TRIAL REGISTRATION

ClinicalTrials.gov , NCT02810366. Submitted on 11 April 2016.

An Expectation Conditional Maximization approach for Gaussian graphical models

Bayesian graphical models are a useful tool for understanding dependence relationships among many variables, particularly in situations with external prior information. In high-dimensional settings, the space of possible graphs becomes enormous, rendering even state-of-the-art Bayesian stochastic search computationally infeasible. We propose a deterministic alternative to estimate Gaussian and Gaussian copula graphical models using an Expectation Conditional Maximization (ECM) algorithm, extending the EM approach from Bayesian variable selection to graphical model estimation. We show that the ECM approach enables fast posterior exploration under a sequence of mixture priors, and can incorporate multiple sources of information.

Bayesian Joint Spike-and-Slab Graphical Lasso

In this article, we propose a new class of priors for Bayesian inference with multiple Gaussian graphical models. We introduce Bayesian treatments of two popular procedures, the group graphical lasso and the fused graphical lasso, and extend them to a continuous spike-and-slab framework to allow self-adaptive shrinkage and model selection simultaneously. We develop an EM algorithm that performs fast and dynamic explorations of posterior modes. Our approach selects sparse models efficiently and automatically with substantially smaller bias than would be induced by alternative regularization procedures. The performance of the proposed methods are demonstrated through simulation and two real data examples.

An integrated approach to processing WHO-2016 verbal autopsy data: the InterVA-5 model

BACKGROUND

Verbal autopsy is an increasingly important methodology for assigning causes to otherwise uncertified deaths, which amount to around 50% of global mortality and cause much uncertainty for health planning. The World Health Organization sets international standards for the structure of verbal autopsy interviews and for cause categories that can reasonably be derived from verbal autopsy data. In addition, computer models are needed to efficiently process large quantities of verbal autopsy interviews to assign causes of death in a standardised manner. Here, we present the InterVA-5 model, developed to align with the WHO-2016 verbal autopsy standard. This is a harmonising model that can process input data from WHO-2016, as well as earlier WHO-2012 and Tariff-2 formats, to generate standardised cause-specific mortality profiles for diverse contexts. The software development involved building on the earlier InterVA-4 model, and the expanded knowledge base required for InterVA-5 was informed by analyses from a training dataset drawn from the Population Health Metrics Research Collaboration verbal autopsy reference dataset, as well as expert input.

RESULTS

The new model was evaluated against a test dataset of 6130 cases from the Population Health Metrics Research Collaboration and 4009 cases from the Afghanistan National Mortality Survey dataset. Both of these sources contained around three quarters of the input items from the WHO-2016, WHO-2012 and Tariff-2 formats. Cause-specific mortality fractions across all applicable WHO cause categories were compared between causes assigned in participating tertiary hospitals and InterVA-5 in the test dataset, with concordance correlation coefficients of 0.92 for children and 0.86 for adults. The InterVA-5 model's capacity to handle different input formats was evaluated in the Afghanistan dataset, with concordance correlation coefficients of 0.97 and 0.96 between the WHO-2016 and the WHO-2012 format for children and adults respectively, and 0.92 and 0.87 between the WHO-2016 and the Tariff-2 format respectively.

CONCLUSIONS

Despite the inherent difficulties of determining "truth" in assigning cause of death, these findings suggest that the InterVA-5 model performs well and succeeds in harmonising across a range of input formats. As more primary data collected under WHO-2016 become available, it is likely that InterVA-5 will undergo minor re-versioning in the light of practical experience. The model is an important resource for measuring and evaluating cause-specific mortality globally.

Automated verbal autopsy classification: using one-against-all ensemble method and Naïve Bayes classifier

Verbal autopsy (VA) deals with post-mortem surveys about deaths, mostly in low and middle income countries, where the majority of deaths occur at home rather than a hospital, for retrospective assignment of causes of death (COD) and subsequently evidence-based health system strengthening. Automated algorithms for VA COD assignment have been developed and their performance has been assessed against physician and clinical diagnoses. Since the performance of automated classification methods remains low, we aimed to enhance the Naïve Bayes Classifier (NBC) algorithm to produce better ranked COD classifications on 26,766 deaths from four globally diverse VA datasets compared to some of the leading VA classification methods, namely Tariff, InterVA-4, InSilicoVA and NBC. We used a different strategy, by training multiple NBC algorithms using the one-against-all approach (OAA-NBC). To compare performance, we computed the cumulative cause-specific mortality fraction (CSMF) accuracies for population-level agreement from rank one to five COD classifications. To assess individual-level COD assignments, cumulative partially-chance corrected concordance (PCCC) and sensitivity was measured for up to five ranked classifications. Overall results show that OAA-NBC consistently assigns CODs that are the most alike physician and clinical COD assignments compared to some of the leading algorithms based on the cumulative CSMF accuracy, PCCC and sensitivity scores. The results demonstrate that our approach improves the performance of classification (sensitivity) by between 6% and 8% compared with other VA algorithms. Population-level agreements for OAA-NBC and NBC were found to be similar or higher than the other algorithms used in the experiments. Although OAA-NBC still requires improvement for individual-level COD assignment, the one-against-all approach improved its ability to assign CODs that more closely resemble physician or clinical COD classifications compared to some of the other leading VA classifiers.

Automated verbal autopsy classification: using one-against-all ensemble method and Naïve Bayes classifier

Verbal autopsy (VA) deals with post-mortem surveys about deaths, mostly in low and middle income countries, where the majority of deaths occur at home rather than a hospital, for retrospective assignment of causes of death (COD) and subsequently evidence-based health system strengthening. Automated algorithms for VA COD assignment have been developed and their performance has been assessed against physician and clinical diagnoses. Since the performance of automated classification methods remains low, we aimed to enhance the Naïve Bayes Classifier (NBC) algorithm to produce better ranked COD classifications on 26,766 deaths from four globally diverse VA datasets compared to some of the leading VA classification methods, namely Tariff, InterVA-4, InSilicoVA and NBC. We used a different strategy, by training multiple NBC algorithms using the one-against-all approach (OAA-NBC). To compare performance, we computed the cumulative cause-specific mortality fraction (CSMF) accuracies for population-level agreement from rank one to five COD classifications. To assess individual-level COD assignments, cumulative partially-chance corrected concordance (PCCC) and sensitivity was measured for up to five ranked classifications. Overall results show that OAA-NBC consistently assigns CODs that are the most alike physician and clinical COD assignments compared to some of the leading algorithms based on the cumulative CSMF accuracy, PCCC and sensitivity scores. The results demonstrate that our approach improves the performance of classification (sensitivity) by between 6% and 8% compared with other VA algorithms. Population-level agreements for OAA-NBC and NBC were found to be similar or higher than the other algorithms used in the experiments. Although OAA-NBC still requires improvement for individual-level COD assignment, the one-against-all approach improved its ability to assign CODs that more closely resemble physician or clinical COD classifications compared to some of the other leading VA classifiers.