Human sensing infrastructures and global public health security in India’s Million Death Study
Global health is increasingly approached as a security issue, especially in the context of unpredictable global epidemics that can be anticipated and monitored with novel, digital sensing techniques. Advocates of a more human-centred notion of health security, who claim that everyday health threats should be of greater concern, have challenged this ‘securitization’ of health. Along these lines of argumentation, this chapter considers the Million Death Study in India, an effort to gather better insight into the most common causes of death by means of an interview-based method called verbal autopsy. I argue that this method’s reliance on ‘human sensors’ requires a sociotechnical sensing infrastructure that steers humans towards their task of symptomatic data collection and diagnosing discrete causes of death. This infrastructure includes various kinds of devices and sets of instructions that shape how human sensors work. This infrastructure produces statistics on mortality in India that prioritize a decontextualized, medical-statistical perspective on mortality in India. Through its intersections with state-based and global health infrastructures, this prioritization has particular political effects in the context of global public health security. While the study provides valuable evidence of the most common, often underrepresented causes of death in India, it is less suited for addressing the structural dimensions of health security in the global South.
Introduction: Sensing the common in global public health security
Global health is increasingly approached as a security issue, to the extent that perceptions of health in terms of threats and required responses can be considered one of the dominant frames of global health governance (McInnes et al. 2012). This ‘securitization’ of global public health foregrounds concerns about the global spread of newly emergent infectious diseases such as SARS and avian flu, the global HIV/AIDS epidemic and bioterrorism (Collier and Lakoff 2008; Hanrieder and Kreuder-Sonnen 2014; Jin and Krackattu 2011). This security-based approach to global public health is both product and instigator of the application of novel approaches to disease detection that supposedly imply a shift in disease monitoring ‘from actuary to sentinel’ (Lakoff 2015). With these characterisations Lakoff points to a shift he identifies in global public health, from calculating averages and common health problems towards early detection and preparedness for extremely rare, but potentially catastrophic health risks. Such a focus on anticipating the exceptional is in turn said to be enabled by new forms of data gathering and processing. These include the use of algorithms that can distil data with potential relevance for health from large sets of data that is not primarily diagnostic (Bengtsson et al. 2019; Roberts and Elbe 2017). From this point of view, global public health threats are increasingly perceived as exceptional events that can be anticipated with novel digital sensing methods.
This paradigm of preparedness is contested for its focus on highly visible, but exceedingly rare health risks. Critical security scholars have questioned the close association between this framing of global public health and traditional state-centred notions of security (Nunes 2018). While they see value in considering health in terms of security, they simultaneously argue for a human-centred approach to global public health security. These scholars point out that ‘[a]t the simplest level, premature and unnecessary loss of life is perhaps the greatest insecurity of human life’ (Chen and Narasimhan 2003: 183). Common causes contributing to this loss of life should therefore receive more attention from security scholars, who should also pay more attention to the insecurities caused by structural threats, and inequalities in access to curative and preventive health services (MacLean 2008). They thereby advance the argument that the things that people are most likely to die from, form a significant security threat in their own right.
The prevention of premature death requires its own strategies for detection of (common) health threats. The Million Death Study (MDS) in India is an example of such a strategy. Researchers argue this study is necessary due to the supposed lack of representative and reliable insight into the causes of premature death in existing public health monitoring in India and other low and middle income countries (Gomes et al. 2017; Jha 2012, 2014). Even though this initiative is situated in the Indian context, researchers thus seek to establish it as a model for public health monitoring and interventions in a broader global health context. The MDS aims to attend to this knowledge gap through an effort to produce representative and cause-specific national mortality statistics by way of an interview-based method called verbal autopsy (VA). In stark contrast to the use of electronic data-processing capabilities entangled with a growing emphasis on anticipating unpredictable epidemics in some approaches to global public health security, the MDS pursues forms of innovation in the detection of common health threats by employing what may be considered human sensors.
The Million Death Study provides important insights into the challenges of developing sensors for signalling structural, common health threats. While study initiators emphasize the idea of ‘a simple, respectful conversation’ as the core of verbal autopsy, we will see how the study employs a broader sensing infrastructure to make human sensors work. I develop this notion of how infrastructures sensitize human sensors in this chapter to understand the joint emergence of a particular mode of knowledge production and specific formations of public health threats in India. In the next section, I lay the groundwork for this notion, which I locate at the intersection of global health security, critiques of quantification in global health and the ‘mediation’ between human and non-human elements of sensing. This discussion forms the basis of an exploration of the Million Death Study, in which I address how data is generated and processed; how human sensors are sensitized towards that process; and how this produces particular mortality statistics and views of major public health issues in India. In conclusion, I consider how understanding the study in terms of sensing provides insight into the MDS’ modes of knowledge production, which create a novel infrastructure for addressing statistical deficiencies, while simultaneously perpetuating deficiencies in considering the structural aspects of global public health security.
Conceptualizing global public health security and its sensing infrastructures
The increasingly prominent approach to security and global health focused on preparedness for novel, as-yet unknown risks builds on long-standing concerns about the global circulation of health threats from the periphery to global centres of power (King 2002), but takes on a distinct contemporary form. Current global health threats are considered to be quintessentially mobile across borders, requiring tailored methods of early detection and intervention. The most prominent of these threats, such as new infectious pathogens, the spread of HIV/AIDS and bioterrorism, are considered to be difficult to address with traditional epidemiological and public health approaches. Global health practitioners therefore argue that new institutional responses, styles of thinking and techniques for measurement and intervention are required. These approaches emphasize preparedness for unknown and unexpected events (Collier and Lakoff 2008). These approaches to global health thereby constitute a deviation from the forms of biopolitical governing Foucault describes as ‘the calculated management of life’ (Foucault 2013: 44). This approach to government, which Foucault himself has described in terms of security (Foucault 2007), involves the measurement and calculation of collective probabilities and averages within a population and is focused on normal distributions of, for example, disease within a population.
The focus on the exceptional in emerging forms of global health security thus stands in marked contrast to Foucault’s writing, causing some authors to conceptualize it as a biopolitical transformation. They do no longer consider probabilistic national statistics to be core instruments for the exercise of power. Rather, the emergence of new measuring techniques and new kinds of social collectives supposedly relocate the management of life to sites other than national populations (Rabinow and Rose 2006). New conceptualisations of populations further relate to shifts from mortality statistics to measuring health, quality of life and the burden of disease on the one hand (Wahlberg 2007; Wahlberg and Rose 2015), and from averages and probabilities to preparedness for unique, catastrophic events (Lakoff 2015) on the other.
Scholars focusing on preparedness in global health argue that these transformed understandings of populations and health threats emerge in conjunction with new techniques for identifying and calculating risks. These are often seen as a manifestation of broader transformations in data generating and processing capabilities, captured under the sweeping label of ‘big data’. Big data allegedly ‘enables an entirely new epistemological approach for making sense of the world’ by attempting to ‘gain insights “born from the data”’ (Kitchin 2014: 2). Data technologies ‘generate different populations as objects of concern and intervention’ (Ruppert 2011: 219) and thereby contribute to the particular shape of global health security focused on the anticipation of exceptional events. Data collection devices and their attendant infrastructures of information collection and processing thereby produce a particular ‘gaze’ on their social environment. They have their own social and political consequences (Bates et al. 2016; Ruppert et al. 2017), rooted in the broader cultural authority of numerical evidence and quantification (Rieder and Simon 2016). In the context of global health security, this authority often does not come from the data directly, but from what Amoore calls ‘data derivatives’ (Amoore 2011); the identification of associations within the data that enable the identification of unknown and absent objects that need to be acted upon.
These characteristics of big data analytics are often positioned in contrast to more ‘traditional’ approaches to statistics and diagnostics in the context of global public health. Nevertheless, critics of this perspective argue that averages and regularities in public health continue to be vital for a global public health security agenda focused on threats to the individual (Chen and Narasimhan 2003; MacLean 2008). Statistical notions of the normal distribution of health and disease are a product of the intricate historical relations between the state and medical science in developing diagnostic classifications. Medical classifications such as the International Classification of Diseases (ICD) that is used in the Million Death Study, were established and further solidified and transformed out of states’ interests in gathering information on the health status of their citizens (Bowker and Star 1999). As Sætnan and colleagues characterize statistics: ‘the act of counting its citizens, territories, resources, problems and so on is one of the acts by which the State participates in creating both itself, its citizens and the policies, rights, expectations, services and so on that bind them together’ (Sætnan et al. 2011: 2, emphasis original). This includes the health status of the population, in the form of distinct categories of disease. Disease classification relies on what Rosenberg calls the ‘structuring act’ of diagnosis, which facilitates the consideration of disease in the aggregate, detached from individual sufferers (Rosenberg 2002). Diagnosis hence serves ‘to establish similarities via classifications in order to promote population health, allocate resources, focus research, and so on’ (Jutel 2011: 201). At the same time, quantification, including population health statistics, helps to bring the reality it describes into being, making a particular set of social problems present that can be expressed through numbers (Espeland and Stevens 2008). Lorway and Khan, for example describe how HIV/AIDS monitoring in India produces new social configurations, individual identities and ideas of vulnerability and (in-)security (Lorway and Khan 2014).
The distinction between big data approaches to epidemics and statistical perspectives on population health is far from absolute. Both ways of identifying health threats attribute a central role to the notion of population, which often implies an abstract, decontextualized and reductionist representation of human life (Krieger 2012; Murphy 2017). Populations in global health emerge from the particular methods of data collection and processing that produce particular, quantitative collectives and their characteristics. These result from what Reubi has called ‘epidemiological reason’, which is characterized by the ambition to save the largest possible number of lives on the basis of rigorous data collection and processing and a global notion of population (Reubi 2018). Yet the predominance of quantitative understandings of population health in a global context is contested for the ways it decontextualizes and generalizes health problems and medical interventions, which obscures the complex, situated social and political dimensions of health and disease (Adams 2013; Birn 2009). Since distinct approaches to the generation and processing of data for monitoring health produce different populations-at-risk, it thus becomes important to ask how research projects ‘see’ (Biruk 2012). In other words, how are populations and the health threats that affect them constructed through particular methods for sensing health problems?
This question is particularly pertinent in light of the Million Death Study’s explicit claim to address the shortcomings of existing population health statistics in India, using the verbal autopsy method. The Million Death Study presents a particular form of sensing in the context of global public health. Sensing in the MDS partly resembles a traditional, nationally oriented emphasis on population health statistics, while it is simultaneously presented as an innovative approach to data generation and processing that aims to circumvent the limitations of medical-statist knowledge production infrastructures in low and middle income countries. The study hence carries an inherent paradox between informing health policy and the simultaneous critique of the state’s inability to produce reliable population health data that is implied in its methodological approach. Despite the emphasis on the human ability to determine cause of death that undergirds verbal autopsy, this approach includes various devices aimed at sensitizing human sensors. These devices include different kinds of instructions for human data collectors and analysts on how to gather symptomatic information and determine causes of death. It is therefore fruitful to think of ‘human’ sensors in the MDS as part of a broader, sociotechnical infrastructure that mediates (Latour 1999) between human and technical contributions to the making of mortality statistics. Drawing on infrastructure studies, we may see how the MDS facilitates the circulation of symptomatic data and diagnoses from households, to the Indian government and research institutes. At the same time, it is worth noting how infrastructures are always established in relation to pre-existing infrastructures, with particular political effects (Anand et al. 2018; Slota and Bowker 2017). We will see how the sensing infrastructure of the MDS steers human sensors towards the study’s particular purpose, how it intersects with existing infrastructures and how it thereby obscures other potential insights with particular implications for the politics of global public health security.
In the next section, I reconstruct the making of cause-specific mortality statistics for India in the Million Death Study. While the securitization of public health in this study is not explicit (i.e. there is no mention of security as such), it is fundamentally informed by a logic of identifying threats to the health and survival of the Indian population and formulating adequate (health policy) responses. My analytical narrative is based on interviews with thirteen researchers, government officials and other public health actors and an analysis of study documents and publications. The interviews took place in India and Canada in the second half of 2013 and the first half of 2014. I interviewed some of my respondents twice, while some of the interviews took place in a collective setting with multiple respondents. I analysed the interviews with a process coding approach that focuses on respondents’ description of work practices within the study (Saldaña 2011). I applied a narrative analysis (Czarniawska 2004) to the coded interviews in order to reconstruct how researchers and government officials construct knowledge making processes and the relation between human sensors and their sensitizing devices. The documents I analysed include study protocols, guidelines and instruction manuals, as well as publications in scientific journals on the study’s importance, methods and results. I consider these documents active participants in the making of the study (Shankar et al. 2017). I therefore analysed them not only in terms of how they describe the study’s content, but also for how they contribute to defining the issue of lacking mortality statistics in India and to developing a response to that issue (Asdal 2015). After providing an introduction of the study’s broad outline and relevance to global public health, I divide my account of the study and its human sensors into three subsections. First, I introduce how data is collected, processed and used for determining mortality statistics. Second, I turn to the theme of human sensors and the infrastructure that ‘sensitizes’ them towards the task of producing mortality statistics. Third, I discuss how this infrastructure contributes to the production of a particular perspective on major public health threats in India that MDS actors frame as policy relevant – and turn to both the strengths and weaknesses of this infrastructure in the conclusion.
Sensing population mortality in India’s Million Death Study
The Million Death Study in India aims to produce representative cause of death statistics for the Indian population through a method called verbal autopsy (VA). Researchers in the study characterize this method as an interview-based, structured investigation of the circumstances and symptoms occurring around the time of death (Jha 2012). Study initiators claim that this effort and its particular approach is needed because the available cause-specific information on mortality is insufficient. They claim it is either based on hospital statistics, which cover only a small portion of all fatalities in India, or on self-assigned causes of death that are often inaccurate. Since the late 1990s, the Office of the Registrar General of India (RGI), which is responsible for the census, and the Centre for Global Health Research at the University of Toronto in Canada, in collaboration with research institutes in India, have therefore developed, piloted and refined a method for generating cause of death statistics that does not rely on the medical system (Jha et al. 2006). The study is funded by the Indian government, grants from research councils in Canada, India and the United States, as well as organisations such as the Bill and Melinda Gates Foundation. Researchers claim it is a cheap study, supposedly costing less than a dollar per year for each household being monitored. According to one senior researcher, it seeks to correct assumptions about public health in India – thereby potentially contributing to the achievement of the Millennium Development Goals (Bhutta 2006). While the study focuses on India, researchers explicitly position its approach to data collection as a model for collecting reliable mortality statistics in other low- and middle-income countries (Jha 2014). For global health research, the study is therefore considered an example for the advancement of evidence-based global health (Birbeck et al. 2013).
From symptomatic data points to mortality statistics
Million Death Study researchers explain the urgency of the study in terms of the limited availability of reliable cause of death information, and the importance of such information for public health. As one senior researcher explained, the study:
tries to answer a very easy question: how do people die? And there is an extraordinary widespread degree of ignorance about how people die and less surprising but also true is that people don’t really understand how important understanding causes of death is to improving health of the living.
Researchers cite the circumstances under which most people in India die as an important factor contributing to this ignorance. They regularly repeated to me that, since the vast majority of people dies at home, in rural areas and without medical attention, there is no information on what causes their deaths. Moreover, the information that is available predominantly comes from hospitals in urban areas, which cannot reasonably be extrapolated to the population as a whole. Since, as one researcher claimed, ‘in the immediate future there will be no clinically certified deaths in India’, researchers and government officials collectively developed an alternative approach to identifying causes of death. Central to this approach is the use of an existing demographic survey, the so-called Sample Registration System (SRS), which already monitors population dynamics on an annual basis.
The SRS was presented to me in interviews as a ‘unique’ infrastructure that was introduced to keep track of changes in overall population size during the ten-year intervals of India’s full census. The SRS sample includes about 0,8% of India’s population (which amounts to a cohort of 8 million people). The sample is distributed over randomly selected units from all over India that collectively are supposed to provide a representative overview of population change. Sample size is based on the census (for example, the current SRS sample, in use since 2014, is based on the 2011 census) and calculations of the total number of participants required to make what demographers consider statistically robust claims about the least common event of interest (which presently is infant mortality). The current SRS sample consists of 8861 units of less than 2000 inhabitants. The SRS aims to trace changes to the population in each of its units, including migration, births and deaths. This is done via ‘dual enumeration’. So-called part-time enumerators, who are local residents, continuously collect relevant information for their respective unit. The numbers they produce are matched with the results of secondary, retrospective enumerations produced by government officials. These visit all households in the roughly ten units they are responsible for twice a year to collect the same information on the preceding six months. The verbal autopsy method of the Million Death Study was added to this existing procedure. Wherever a death is reported, the secondary government enumerators are now required to also collect further details on how this person died.
The MDS expands on the SRS’ use of human enumerators who develop a relationship with the people they survey over time. This should allow them to engage with these people to elicit information needed for diagnosing causes of death. The MDS hence establishes a novel approach to medical research that is – in part – based on the non-medical nature of its data collection. Government enumerators are not expected to diagnose, but to ‘prompt or probe’ family members of the deceased in order to gather more details on what occurred preceding death (SRS Collaborators of the RGI-CGHR 2011: 2). The most important part of this data collection is what is called the ‘narrative’, in which respondents are expected to give their own account (instigated and directed by the enumerator) on the events preceding death. One researcher describes the scenario as follows:
If you are asking me; “how did your grandfather die?” I start at; “yesterday night he was having chest pain”. Chest pain is one symptom. Then he suddenly vomited – vomiting is another symptom. Basically, these signs and symptoms doctors use to assign cause of death.
Enumerators receive strict instructions on what kind of information to collect, and how to do so in order to make the conversation as insightful as possible for the study. The household interview is primarily conceived as a tool to collect the relevant symptomatic ‘data points’, which Armstrong describes as the ‘smallest possible piece of information’ (Armstrong 2019: 103). In the context of the MDS this refers to the discrete and specific elements of the various events preceding a person’s death. Since the sequence of and coherence between these events matter in the process of diagnosis, the key issue is to gather all of them, in the right order, and with the right degree of detail. This is how enumerators become sensors; their ability to engage in conversation allows them to generate the data required for diagnosing deaths. Their role in this context is only to record symptoms, not to interpret them, which is why they deliberately receive no medical training in the context of the study – since researchers believe that medical understanding of the symptoms might interfere with enumerators’ ability to carefully listen and record.
Symptomatic data are recorded with four different forms, numbered 10 A through D, which respectively apply to neonates, children, adults and maternal deaths. Each form is adapted to generate the symptomatic data considered relevant to diagnosing deaths in that particular (age) group. The SRS sample produces about fifty thousand records of death a year, all of which are subjected to quality control. Data collection of a further randomly selected 5% of forms is repeated, in order to ascertain that symptomatic data points are accurately recorded. The RGI collects these forms before they are made available to MDS researchers for further diagnosis. All of these measures contribute to creating a feedback loop within the study that is supposed to improve accuracy and to provide insight into the functioning of the study’s human sensors.
A panel of a few hundred physicians using the International Classification of Disease (ICD-10) interprets the symptomatic data to categorize deaths and diagnose their causes. Physicians on this panel are recruited via medical schools, professional networks and medical journals and are paid Rs40 for each ‘record’ they ‘code’. A ‘record’ of symptoms described in a household interview gets randomly assigned to two physicians. They ‘code’ the narrative independently, using symptomatic data points in the record to consider potential causes of death and ultimately arrive at a diagnosis. One researcher describes the ensuing coding process as ‘a three-step check and balance’. The first step is the independent coding by two physicians. If they agree on the diagnosis, the cause of death is registered as such. In the 25% of cases where they disagree, a second step of so-called ‘reconciliation’ follows. In this step, both physicians receive each other’s diagnosis and are asked to reconsider. If disagreement continues, which happens in 10% of cases, a third step of ‘adjudication’ follows. This means that a third, more experienced physician receives both diagnoses and is asked to decide.
Records for which a cause of death has been determined are aggregated by RGI, which first publishes a report on the findings. Such reports include nationwide overviews of the percentage of deaths that may be attributed to particular causes, as well as the distribution of causes by gender, age groups, and regions. Only after publication of this report is the data made available to global health researchers, who publish on the results in often cause-specific papers (e.g. on cardiovascular disease, cancer, malaria, traffic accidents, etc.) written in international working groups. One researcher describes this process as follows:
After a record’s completion, we analyse it, provide the information to government and government publishes a report. If the government publishes the report, then on the basis of this data we publish papers. These papers have a policy impact.
Like many of his colleagues, he emphasizes ‘policy impact’ as a core objective of the study. Not only is the MDS supposed to generate more accurate and representative cause of death statistics for India, this is explicitly done to secure better health for the Indian population – and beyond. Collectively, the report, scientific papers and policy interventions based on the study ‘perform’ a particular version of India’s population (Law 2009; Ruppert 2011). This version does not only show the normal distribution of health threats and insecurities overall, but also categorizes the population in terms of how certain indicators (age, gender, and location as proxy of socio-economic status) are expressed in divergence of health threats. This points towards the ways in which the human sensing approach in the study enacts public health security concerns.
Sensitizing a human sensing infrastructure
The reliance on human sensors in the MDS is itself a response to researchers’ conviction that clinical means of monitoring mortality are insufficient. They are often not available, and therefore produce results that are neither representative nor very reliable. The use of human sensors is an attempt to access data on causes of death closer to the source and to interpret that data more accurately. Yet the study operates through a continuous tension between human capabilities and the need to ascertain consistency and quality in the study’s knowledge production processes. The study handles this tension primarily by sensitizing its human sensors towards producing the particular forms of symptomatic data and diagnostic specificity it aims for. Despite researchers’ insistence on the ‘simple conversation’ as the supposed core of the study, its human sensors are thereby made part of a more elaborate sensing infrastructure that consists of various material devices, standards, and procedural rules and instructions that enable the work of human sensors. One researcher, for example, describes how diagnostic criteria in the context of verbal autopsy were developed:
I worked with several experts on cancer and then we took what could be symptoms visualized at time of death. Based on that we produced a list of 10 to 15 symptoms and circulated that to reviewers who said: yes, this could be the best thing. And that would be possible to record, of course. You can’t say we do a CT scan. So, symptoms that were possible to ask and were the most pronounced symptoms.
Not only diagnostic criteria are made useable for human sensors. The forms used in the study, its software platforms, instruction videos and manuals, as well as training are all involved in sensitizing the human sensor towards how to ‘correctly’ gather and interpret data.
One of the main challenges the MDS has to account for is the inability to rely on the medical system, or established medical research infrastructures to gather the kind of representative data the study is after. One senior researcher explains: ‘it can’t be done like a research methodology. It has to be incorporated in the [SRS] system.’ He argues that only RGI and its extensive workforce of about 800 enumerators can sustain a study of this kind. The study’s focus on deaths, rather than the burden of disease, also enables the central role of the non-medical staff. ‘The goal for us is proper reporting of the symptoms’, one researcher explains, before continuing:
Because the one thing you must realize is that many symptoms are not clear when a disease is progressing. When you are developing lung cancer, symptoms may be very confusing. But when you are dying of a cancer it is very obvious.
The most important part is the training. This could not have been achieved, results would be uncertain everywhere if you don’t have proper training. So, we first developed training for the interviewer, at the government level at the various census offices.
This training is supposed to help enumerators in asking the right questions and noting down the relevant ‘data points’. In a training video a researcher showed me, this is explained in the form of ‘five simple steps’ to follow. These include (1) to carefully listen and take notes, (2) to note which out of a total of twelve ‘cardinal symptoms’ (including things like fever, breathlessness, chest pain, and urinary problems) were present, (3) to probe further details about these symptoms, (4) to confirm that none of the remaining symptoms occurred, (5) to confirm the narrative with the respondent, especially clarifying the duration and sequence of all the symptoms. Following these steps, the video maintains, enables anyone to write a good narrative that presumably includes all the data needed for determining a diagnosis.
Similarly, for physicians to serve as diagnostic sensors, they also receive a set of instructions on how to approach coding. When doctors apply to become part of the coding panel, they first have to go through three phases of training and tests. These include exercises in coding a set of narratives, including some that were subject to reconciliation – which are presumably more complex. Since the aim of the study is to identify a single underlying cause for every death, the study manual instructs physicians how to distinguish this from what is called the ‘terminal event’ as well as risk factors and ‘contributory causes’ that lie outside of the main chain of events leading to death. Physicians are further instructed to rely on the symptoms described in the record and their own expert judgement: ‘you are expected to provide an opinion on the cause of death to the best of your knowledge and belief, based on the information available to you’ (SRS Collaborators of the RGI-CGHR 2011: 21). Physicians are expected to find the middle ground between the most specific cause they consider possible, while keeping it at a level general enough to be defensible. This means that physicians are expected to think in terms of public health, looking for common rather than exotic causes. With the aim of attributing a single cause to each death, all of these instructions are meant to increase the chance that the right one gets chosen.
One researcher I interviewed took out his laptop to show how he might go about coding a narrative. He showed me how the respondent in this particular case thought that a kidney problem had been the cause of death and highlighted some symptomatic data (‘swelling of the legs’) in the narrative on screen. This he used to search in the study’s coding software for kidney diseases expressed in the form of swollen legs. However, he also indicated that the software could provide alternative explanations to consider (in this case heart failure) on the basis of the keywords he provided. Finally, he indicated how physicians can rate each record for quality and their own diagnosis for degree of certainty – adding another feedback opportunity that may be used to improve the study’s sensing infrastructure.
This example shows not only how a physician may code, but also how the so-called Central Medical Evaluation software platform plays a key role in the diagnostic process. It allows physicians to search within more than 2000 ICD-codes, provides suggestions for differential diagnoses to consider, and may occasionally correct a physician’s tendencies. This platform forms only half of the software infrastructure incorporated into the Million Death Study, with the other half responsible for performing quality checks on the data supplied on the verbal autopsy forms. This includes controls for whether the data is complete and consistent. Once forms are approved, the Central Medical Evaluation platform takes care of the distribution of records. It allocates no more than ten records at a time to any physician, taking into account which of the eighteen languages used for the forms they master. The software platform is supposed to assist the coding physician in assigning an accurate cause of death, and to do so as quickly as possible. It does so in various ways. It allows physicians to search ICD-codes and blocks certain ICD-codes that are impossible or extremely unlikely – such as prostate cancer for female deaths or many chronic conditions for young children. Yet the software intervenes most actively by suggesting diagnoses it considers more or less likely for specific cases:
For instance, if certain symptoms are in place and that is noted by the system and the physician tries to assign an ICD code that is not intuitive to the system, the system will try to correct the physician and say: this doesn’t make sense. The person had a fever, the person was vomiting and you are saying they died of diabetes.
The software does this by using data from previously diagnosed deaths, which my interview respondent characterized as ‘metadata’, to suggest causes of death it considers likely in similar cases. The aim is to develop ‘machine learning’ for the study, described as follows:
The machine itself is becoming smarter as it codes. So it is sort of like a human being, you learn from your experiences and you make mistakes, but you correct these mistakes as you go forward.
The hope is that this will result in a web-based coding system that increases speed, without compromising quality.
Despite the desire to further automate coding and critiques that the verbal autopsy method is subject to human bias (Butler 2010), studies of available coding algorithms by MDS researchers suggest that human coders are still the most reliable (Desai et al. 2014). As one researcher explains:
As far as VAs are concerned, the physician standard is much better than the machine standard at this point. The machine is not smart enough at this point to replace the physician. And it would have to be smart enough and be able to learn on its own. To keep up with the human being.
Another researcher attributes the difference to the supposedly uniquely human ability to consider the history and relations between symptoms in a way that algorithms cannot. Nevertheless, for the human to be able to gather relevant data and provide accurate analysis, the study prescribes a set of almost mechanistic steps for the collection and treatment of data to its human workers. These steps serve to build a sensing infrastructure in which human sensors are made ‘sensitive’ towards the study’s aim of producing accurate and representative mortality statistics. By doing so, the infrastructure prioritizes particular forms of data and of interpretation over others, with particular consequences for how its insights into, and contributions to, public health in India are framed.
Making mortality profiles, framing public health policy
The aim of the data collected and analysed in the MDS is ‘to yield cause-specific mortality profile at the national level’ (Office of the Registrar General of India 2009: vi) and thereby gain a better understanding of the particular health threats facing the Indian population. An important lesson that researchers believe to have learned over the years is that the VA method works (see also Aleksandrowicz et al. 2014). As one researcher explains:
we made use of the representative strength of SRS. So, there is the possibility of being representative of the Indian population. That was a real strength we had. In spite of the specificity that may be lost [sometimes].
Implicit in this argument is that mortality statistics based on a representative sample are infinitely more reliable and accurate than those produced in hospitals. This holds true to MDS researchers despite the ‘lost specificity’ for rare causes of death in smaller geographical areas, for which the statistics produced within the SRS are not precise enough. The employment of human sensors in the MDS is therefore considered a reliable solution for the failure of national medical infrastructures that cannot produce equally solid mortality statistics.
Researchers additionally believe that the VA method and its sensing infrastructure may be expanded, considering it an innovative form of knowledge production with various other potential contexts of application. Some examples mentioned in interviews include studies of ‘point of death diagnostics’ that compare VA with regular autopsy, of the role of nutrition in disease, and of the impact of providing mortality data on the delivery of health services. More broadly speaking, findings from the MDS are treated as hypotheses for further research on what ‘causes the causes’. This idea is above all supported by observations about regional differences in mortality, which researchers translate into a belief in the contributions the study can make to preventing causes of death that are common in some areas, but rare in others.
The most important promise of the MDS and its verbal autopsy method that researchers emphasize, however, is its potential to contribute to global public health. The first RGI report of study results published in 2009, for example, observes that the findings are ‘not only of national interest but [are] also watched globally’ (Office of the Registrar General of India 2009: 52). This argument follows a similar logic to the one about regional diversity in causes of death. As one researcher explained, India’s is ‘a huge population, and diverse, which reflects the whole developing world’. For example, since the number of deaths from snake-bites in India is underestimated, researchers explicitly argue that global estimates are probably too low as well (see Mohapatra et al. 2011). Additionally, verbal autopsy is presented as a feasible option for other countries that have no comprehensive mortality registration. Innovation in global health to address shortcomings in the collection of mortality statistics beyond India drive MDS leadership to consider, for example, expanding the study to other countries and to formulate the ambition of elevating collection of mortality statistics to a global ambition akin to the United Nations Sustainable Development Goals (see also Jha 2014).
Despite these various arguments in favour of the approach and insights of the MDS, researchers’ opinions differ with regard to the study’s insights they find most valuable. Among the examples of notable findings, they mentioned causes of death as diverse as cancers, HIV/AIDS (which is less common than previously thought), malaria and snakebites (both more common), smoking and cardiovascular diseases. Some researchers thought that the MDS provides the ultimate proof that the epidemiological transition (i.e. the shift from infectious to chronic diseases as predominant causes of death) is well underway in India – a point of view confirmed in the study’s findings (Office of the Registrar General of India 2009). Others noted that there is a significant ‘residual burden’ of diseases like tuberculosis and malaria. One researcher explained that he saw the value of the MDS mostly in the context of those conditions:
Especially in those areas where there is, deaths occur in rural areas, it may be very good. When death occurs in rural areas, without access to medical attention and nobody knows how many deaths occur.
The study results further show that such acute conditions, including snake bites, tuberculosis and malaria, are more likely to kill and remain unattended to in rural regions, among women, and in poorer states.
Besides pointing to diagnostically specific, largely biologically defined causes of death, researchers also indicate that more structural and institutional factors play a significant role. The following scenario may be illustrative. After again pointing out how most rural deaths occur in households, this researcher explains why this is so:
So, this is a particular village and your hospital is located 15 km away. Reaching the hospital is a problem. This is one barrier. And if the hospital is there, the doctor might not be available. If the doctor is available, facilities may not be available. Then this doctor has to refer to another hospital nearby. Then the problem is again distance. And traveling costs. These are the barriers. […] And they are unreported, these deaths.
This scenario is once more confirmed by the differences between states in terms of causes of death and their distribution within the population in the study’s findings, which correlate with differences in socio-economic status (Office of the Registrar General of India 2009). These differences are not only expressed in the burden of infectious diseases. Non-communicable diseases, too, are more likely to affect the poorer population, often with catastrophic financial consequences for those already disadvantaged (Rajan and Prabhakaran 2012). These dimensions of mortality, disease, and its consequences in India reflect widely shared concerns about structural inadequacies in the Indian health system and its contributions to public health (Dreze and Sen 2013; Global Health Watch 2011; Rao et al. 2015). Furthermore, they indicate how the specifically sensitized human sensors in the MDS both hold promise and face limitations in terms of human-centred perspectives on global public health security.
Conclusion: Sensing infrastructures and the contextualization of global public health security
Researchers in the Million Death Study commonly portrayed the study as a simple endeavour, relying on what I termed ‘human sensors’. These are government enumerators who visit households and collect symptomatic information about any fatalities in a ‘simple conversation’ on the one hand, and physicians who deduce a cause of death from the symptoms described for each deceased person in the sample on the other. The objective of the study is to produce reliable and representative numbers on the pattern of mortality in India, which are currently not available. It employs the so-called verbal autopsy method to overcome the problems of incomplete and biased statistics generated within the Indian medical system. Yet while researchers emphasize the simplicity of the VA method, closer consideration of the study reveals that it cannot work without a more extensive sensing infrastructure. Within this infrastructure, human sensors are steered towards contributing to the study’s objectives, which results in data capture and analysis that produces a decontextualized, predominantly biological perspective on mortality in India. The study infrastructure and its particular orientation towards the quantification of death intersects with other global public health infrastructures. I turn to these intersections to further illuminate the political salience of sensing in the context of global public health security.
The political salience of infrastructure emerges in part from the association of infrastructure with social progress, both in response to and in spite of the contentious relation between what is new and what builds on pre-existing infrastructures (Anand et al. 2018). In the Million Death Study, the relations between the study’s sensing infrastructure and statistical infrastructures of the Indian state as well as global health infrastructures are of particular interest. The study’s infrastructure combines human sensors with technical devices in order to address a weakness in national infrastructures for collecting mortality data. Their insufficiency is a key reason why the MDS does not use established medical infrastructures for data collection. Yet the study can only circumvent the medical system by building on a different, existing, demographic government infrastructure. This creates an infrastructural paradox, in which the study’s critique of the state’s deficiency in collecting data that is vital for its public tasks (i.e. securing health) can only be addressed through a state infrastructure that functions. This ambivalent relation between the MDS and the Indian state cannot be seen separate from a second infrastructural intersection between the MDS and global health as a professional and academic field (McGoey et al. 2011). This global health infrastructure is characterized and facilitated by multiple, asymmetrical connections between research institutions, governments and funding agencies in the global North and South (Crane 2010) that involve various kind of non-governmental actors in public health activities in low and middle income countries. Against this background, research and intervention in global health often pursues a decontextualized strategy of knowledge production that allows the circulation of insights at the expense of situated approaches to public health (Biehl and Petryna 2013). Despite the innovative knowledge production infrastructure of the MDS, it reproduces contested aspects of global public health infrastructures that decontextualize health problems and potentially dilute democratic accountability.
This does not mean that the MDS gets death ‘wrong’. Its insights are vital and its method suggests an innovative way of breaking through the vicious circle of deficient infrastructures for both medical care and knowledge production in low and middle income countries. Nevertheless, there are aspects of global public health security that deserve more sustained attention. The case of the Million Death Study points to the political consequences of the mutual configuration of security devices (or infrastructures) and the threats they identify (Amicelle et al. 2015) in the context of global public health security. In terms of a human-centred approach to global public health security, it advances insight into the common health threats affecting people in the global South, and how these are often misrepresented. Yet global health security scholars also point to the structural dimensions of health threats (Chen and Narasimhan 2003; MacLean 2008), and although MDS researchers recognize their importance, the study’s infrastructure is less suitable for making these visible. Issues such as the structural violence that make the socially marginalized more vulnerable to harm (Farmer et al. 2006), the absence of basic medical services and the complex determinants of health (including nutrition, housing, and many others) (Pfeiffer 2013), are all of vital importance to health security. Million Death Study researchers indicate that issues of this kind are widely presented in the narratives enumerators produce; the challenge is to integrate such qualitative, experiential evidence into the study’s efforts to sense the major threats to global public health security.
Adams, V. (2013). Evidence-Based Global Public Health. Subjects, Profits, Erasures. In J. Biehl and A. Petryna (Eds), When People Come First. Critical Studies in Global Health. Princeton: Princeton University Press, pp. 54–90.
Aleksandrowicz, L., Malhotra, V., Dikshit, R., et al. (2014). Performance Criteria for Verbal Autopsy-Based Systems to Estimate National Causes of Death: Development and Application to the Indian Million Death Study. BMC Medicine 12(21): 1–14.
Amicelle, A., Aradau, C., and Jeandesboz, J. (2015). Questioning Security Devices: Performativity, Resistance, Politics. Security Dialogue 46(4): 293–306.
Amoore, L. (2011). Data Derivatives. On the Emergence of a Security Risk Calculus for Our Times. Theory, Culture & Society 28(6): 24–43.
Anand, N., Gupta, A., and Appel, H. (Eds) (2018). The Promise of Infrastructure. Durham: Duke University Press.
Armstrong, D. (2019). The Social Life of Data Points: Antecedents of Digital Technologies. Social Studies of Science 49(1): 102–117.
Asdal, K. (2015). What is the Issue? The Transformative Capacity of Documents. Distinktion: Scandinavian Journal of Social Theory 16(1): 74–90.
Bates, J., Lin, Y., and Goodale, P. (2016). Data Journeys: Capturing the Socio-Material Constitution of Data Objects and Flows. Big Data & Society (July–December): 1–12.
Bengtsson, L., Borg, S., and Rhinard, M. (2019). Assembling European Health Security. Epidemic Intelligence and the Hunt for Cross-Border Health Threats. Security Dialogue: 1–16.
Bhutta, Z. (2006). The Million Death Study in India: Can It Help in Monitoring the Millennium Development Goals? PLoS Medicine 3(2): e103.
Biehl, J., and Petryna, A. (2013). Critical Global Health. In J. Biehl and A. Petryna (Eds), When People Come First. Critical Studies in Global Health. Princeton: Princeton University Press, pp. 1–20.
Birbeck, G., Wiysonge, C., Mills, E., Frenk, J., Zhou, X., and Jha, P. (2013). Global Health: The Importance of Evidence-based Medicine. BMC Medicine 11: 223.
Birn, A. (2009). The Stages of International (Global) Health: Histories of Success or Successes of History? Global Public Health: An International Journal for Research, Policy and Practice 4(1): 50–68.
Biruk, C. (2012). Seeing Like A Research Project: Producing ‘High Quality Data’ in AIDS Research in Malawi. Medical Anthropology: Cross-Cultural Studies in Health and Illness 31(4): 347–366.
Bowker, G., and Star, S. L. (1999). Sorting Things Out. Classification and its Consequences. Cambridge: The MIT Press.
Butler, D. (2010). Verbal Autopsy Methods Questioned. Nature 467 (28 October): 1015.
Chen, L., and Narasimhan, V. (2003). Human Security and Global Health. Journal of Human Development 4(2): 181–190.
Collier, S., and Lakoff, A. (2008). The Problem of Securing Health. In A. Lakoff and S. Collier (Eds), Biosecurity Interventions. Global Health and Security in Question. New York: Columbia University Press, pp. 7–32.
Crane, J. (2010). Unequal ‘Partners’. AIDS, Academia and the Rise of Global Health. Behemoth: A Journal of Civilization 3(78–97).
Czarniawska, B. (2004). Narratives in Social Science Research. Thousand Oaks: Sage Publications.
Desai, N., Aleksandrowicz, L., Miasnikof, P., et al. (2014). Performance of Four Computer-Coded Verbal Autopsy Methods for Cause of Death Assignment Compared with Physician Coding on 24,000 Deaths in Low- and Middle-Income Countries. BMC Medicine 12(20): 1–8.
Dreze, J., and Sen, A. (2013). An Uncertain Glory. India and Its Contradictions. Princeton: Princeton University Press.
Espeland, W., and Stevens, M. (2008). A Sociology of Quantification. European Journal of Sociology 49(3): 401–436.
Farmer, P., Nizeye, B., Stulac, S., and Keshavjee, S. (2006). Structural Violence and Clinical Medicine. PLoS Medicine 3(10): 1686–1691.
Foucault, M. (2007). Security, Territory, Population. Lectures at the College de France 1977–1978. New York: Picador.
—— (2013). Right of Death and Power over Life. In T. Campbell and A. Sitze (Eds), Biopolitics: A Reader. Durham: Duke University Press, pp. 41–60.
Global Health Watch (2011). Global Health Watch 3: An Alternative World Health Report. London: Zed Books.
Gomes, M., Begum, R., Sati, P., et al. (2017). Nationwide Mortality Studies to Quantify Causes of Death: Relevant Lessons from India’s Million Death Study. Health Affairs 36(11): 1887–1895.
Hanrieder, T., and Kreuder-Sonnen, C. (2014). WHO Decides on the Exception? Securitization and Emergency Governance in Global Health. Security Dialogue 45(4): 331–348.
Jha, P. (2012). Counting the Dead is One of the World’s Best Investments to Reduce Premature Mortality. Hypothesis 10(1).
—— (2014). Reliable Direct Measurement of Causes of Death in Low- and Middle-Income Countries. BMC Medicine 12(19): 1–10.
Jha, P., Gajalakshmi, V., Gupta, P., et al., RGI-CGHR Prospective Study Collaborators (2006). Prospective Study of 1 Million Deaths in India: Rationale, Design and Validation Results. PLoS Medicine 3(2): e18.
Jin, J., and Krackattu, J. (2011). Infectious Diseases and Securitization: WHO’s Dilemma. Biosecurity and Bioterrorism: Biodefense Strategy, Practice, and Science 9(2): 181–187.
Jutel, A. (2011). Classification, Disease, and Diagnosis. Perspectives in Biology and Medicine 54(2): 189–205.
King, N. (2002). Security, Disease, Commerce: Ideologies of Postcolonial Global Health. Social Studies of Science 32(5/6): 763–789.
Kitchin, R. (2014). Big Data, New Epistemologies and Paradigm Shifts. Big Data & Society (April–June): 1–12.
Krieger, N. (2012). ‘Who and What is a ‘Population’? Historical Debates, Current Controversies, and Implications for Understanding ‘Population Health’ and Rectifying Health Inequalities. The Milbank Quarterly 90(4): 634–681.
Lakoff, A. (2015). Real-Time Biopolitics: The Actuary and the Sentinel in Global Public Health. Economy and Society 44(1): 40–59.
Latour, B. (1999). Pandora’s Hope. Essays on the Reality of Science Studies. Cambridge, MA: Harvard University Press.
Law, J. (2009). Seeing Like a Survey. Cultural Sociology 3(2): 239–256.
Lorway, R., and Khan, S. (2014). Reassembling Epidemiology: Mapping, Monitoring and Making-Up People in the Context of HIV Prevention in India. Social Science & Medicine 112, 51–62.
MacLean, S. (2008). Microbes, Mad Cows and Militaries: Exploring the Links Between Health and Security. Security Dialogue 39(5): 475–494.
McGoey, L., Reiss, J., and Wahlberg, A. (2011). The Global Health Complex. BioSocieties 6(1): 1–9.
McInnes, C., Kamradt-Scott, A., Lee, K., et al. (2012). Framing Global Health: The Governance Challenge. Global Public Health: An International Journal for Research, Policy and Practice 7(Sup2): S83–S94.
Mohapatra, B., Warrell, D., Suraweera, et al. (2011). Snakebite Mortality in India: A Nationally Representative Mortality Survey. PLoS Neglected Tropical Diseases 5(4): e1018.
Murphy, M. (2017). The Economization of Life. Durham: Duke University Press.
Nunes, J. (2018). Critical Security Studies and Global Health. In C. McInnes, K. Lee, and J. Youde (Eds), The Oxford Handbook of Global Health Politics. Oxford: Oxford University Press.
Office of the Registrar General of India. (2009). Report on Causes of Death in India 2001–2003. Retrieved from https://censusindia.gov.in/Vital_Statistics/Summary_Report_Death_01_03.pdf
Pfeiffer, J. (2013). The Struggle for a Public Sector. PEPFAR in Mozambique. In J. Biehl and A. Petryna (Eds), When People Come First. Critical Studies in Global Health. Princeton: Princeton University Press, pp. 166–181.
Rabinow, P., and Rose, N. (2006). Biopower Today. Biosocieties 1: 195–217.
Rajan, V., and Prabhakaran, D. (2012). Non-Communicable Diseases in India: Transitions, Burden of Disease and Risk Factors – A Short Story. India Health Beat. Supporting Evidence-Based Policies and Implementation 6(1).
Rao, M., Godajkar, P., Baru, R., et al. (2015). Draft National Health Policy 2013: A Public Health Analysis. Economic & Political Weekly 50(17): 94–101.
Reubi, D. (2018). A Genealogy of Epidemiological Reason: Saving Lives, Social Surveys, and Global Population. BioSocieties 13(1): 81–102.
Rieder, G., and Simon, J. (2016). Datatrust: or, the Political Quest for Numerical Evidence and the Epistemologies of Big Data. Big Data & Society (January–June): 1–6.
Roberts, S., and Elbe, S. (2017). Catching the Flu: Syndromic Surveillance, Algorithmic Governmentality and Global Health Security. Security Dialogue 48(1): 46–62.
Rosenberg, C. (2002). The Tyranny of Diagnosis: Specific Entities and Individual Experience. The Milbank Quarterly 80(2): 237–260.
Ruppert, E. (2011). Population Objects: Interpassive Subjects. Sociology 45(2): 218–233.
Ruppert, E., Isin, E., and Bigo, D. (2017). Data Politics. Big Data & Society (July–December): 1–7.
Saetnan, A. R., Lomell, H. M., and Hammer, S. (2011). Introduction: By the Very Act of Counting: The Mutual Construction of Statistics and Society. In A. R. Saetnan, H. M. Lomell, and S. Hammer (Eds), The Mutual Construction of Statistics and society. (pp. 1–17). London: Routledge.
Saldaña, J. (2011). Fundamentals of Qualitative Research. London: Routledge.
Shankar, K., Hakken, D., and Østerlund, C. (2017). Rethinking Documents. In U. Felt, R. Fouche, C. Miller, and L. Smith-Doerr (Eds), The Handbook of Science and Technology Studies. Fourth edition. Cambridge, MA: The MIT Press, pp. 59–85.
Slota, S., and Bowker, G. (2017). How Infrastructures Matter. In U. Felt, R. Fouche, C. Miller, and L. Smith-Doerr (Eds), The Handbook of Science and Technology Studies. Fourth Edition. Cambridge, MA: The MIT Press.
SRS Collaborators of the RGI-CGHR. (2011). Prospective Study of Million Deaths in India: Technical Document No VIII: Health Care Professional’s Manual For Assigning Causes of Deaths Based on RHIME Reports. Toronto: RGI-CGHR, University of Toronto.
Wahlberg, A. (2007). Measuring Progress. Distinktion. Scandinavian Journal of Social Theory 8(1): 65–82.
Wahlberg, A., and Rose, N. (2015). The Governmentalization of Living: Calculating Global Health. Economy and Society 44(1): 60–90.