Machine Diagnosis

In early 2020, Damo Academy, the research unit of Chinese multinational technology company Alibaba, announced the release of an AI system for the diagnosis of the COVID-19 (coronavirus).1 The system is said to detect COVID-19 in patient computed tomography (CT) scans with 96 per cent accuracy over ordinary viral pneumonia. The algorithm was trained with sample data from over 5,000 confirmed coronavirus cases, as well as data on treatment guidelines and recently published health studies. The algorithm was first released to medical professionals at Qiboshan Hospital in the Zhengzhou Province of China, with plans for adoption across additional provinces. In addition to great accuracy, the new algorithm is said to complete the virus detection process in 20 seconds or under, compared to the five to 15 minutes it would take a doctor to analyse more than 300 images in a CT scan.2

Alibaba is among many tech companies, from AlphaGo’s Deepmind to surveillance company BlueDot, that are rapidly developing new machine learning and artificial intelligence (AI) algorithms for COVID-19 patient diagnosis. These algorithms are in addition to the creation of so-called ‘smart’ devices, such as the Kinsa ThermoScan 5 Ear smart thermometer and app, which combine user temperature readings with other user-generated data to provide real-time detection; and – as founder Inder Singh says – ‘real-time response’.3 The ThermoScan device for COVID-19 is powered by a repurposed algorithm that was originally developed to predict and detect the spread of the traditional flu virus. The ThermoScan algorithm is not the only one that has been repurposed for use in the COVID-19 pandemic. Damo used a previous machine-learning algorithm to develop a public health service app that could answer inquiries about the SARS-CoV-2 coronavirus and its spread. Deepmind has announced that they are releasing a new version of the AlphaFold system to help detect protein structures that can promote research around COVID-19. Toronto-based health surveillance company BlueDot Surveillance is compiling data from various online sources, including airline flight information, to make predictions on where the coronavirus might spread next. SenseTime’s facial recognition algorithms are being used to scan the faces of people wearing masks to help promote contactless identification of suspected carriers. The SenseTime temperature detection algorithm has already been deployed in underground train stations, schools, and other public spaces in and around Beijing, Shanghai, and Shenzhen.4 

Due to the time-sensitivity of the COVID-19 pandemic, the scientific community has in response built on decades of research, organized around the sharing of open access data, as well computationally predicted data structures. However, traditional peer-review and verification processes are being circumvented in order to contribute to larger global scientific research. ‘Normally we’d wait to publish this work until it has been peer-reviewed for an academic journal’, states the DeepMind blog.5 The protein structure predictions in their latest version of AlphaFold ‘have not been experimentally verified’. Yet, they ‘hope they may contribute to the scientific community’s interrogation of how the virus functions, and serve as a hypothesis generation platform for future experimental work in developing therapeutics’.

At risk in the increased demand for vital statistics are issues of privacy, most significantly how the consolidation of personal data might result in increased centralization of surveillant technologies. Data consolidation is data generated from many disparate sources and in various different formats. These data are collected from any available source and not limited to any specific time, location, or purpose. Data consolidation offers important benefits in the race for COVID-19 detection and prevention, making it easier to access, manipulate, and analyze a greater set of data. In one sense, data consolidation adds value by achieving enormous time savings, decreased costs, and higher algorithmic efficiency. On the other hand, data consolidation is an irreversible process. Once personal data is consolidated into new algorithmic training sets, they lose their original context and become part of new software scripting. Scripts can, as in the examples stated above, be repurposed for surveillant algorithms that extend beyond their original context – in this case, the tracking of the COVID-19 virus. In efforts to solve one contextual machine-learning problem, the champions of any new consolidated data pipeline (for instance, Google DeepMind, Damo Academy, or even government security agencies) might claim proprietary ownership over the personal data well beyond the immediate COVID-19 pandemic. Even if data are released to the public as open source, ownership of these data sets are most often transferred under existing claims of propriety. 

Privacy advocates Latanya Sweeney and Ji Su Yoo argue that although personal medical data is considered highly sensitive information in many countries, these data are vulnerable to breaches in privacy and security.8 Despite best efforts to protect sensitive data through anonymization and encryption, many medical records still rely on a unique identifier for patients. These identifiers are often the same identifiers that are used across other official government and administrative records, such as national insurance numbers, employment documents, credit card and banking applications, taxation, and other administrative processes. Sweeney and Su Yoo have shown that although encrypted, these identifiers are presumed to be anonymous when the patient’s name and address are removed, and can be easily de-anonymized. The researchers argue that personal medical data is still at risk for de-anonymization, especially if the data has been transferred in ownership across several companies and agencies – each with their own security protocols and quality control processes. For example, Sweeney and Su Yoo conducted two de-anonymization experiments on 23,163 encrypted Resident Registration Numbers (RNN) from prescription data of South Koreans. The researchers demonstrated that embedded within the prescription data was demographic information that could be traced back to other publicly known numerical patterns such as credit card numbers, phone numbers, email addresses, passwords, and related RNNs. Sweeney and Su Yoo cite several cases of major data breaches at private South Korean companies. For instance, a July 2011 security breach at SK Communications, which owns a popular social networking platform, left many South Koreans at risk of theft and fraud. The researchers demonstrated how the exposed personal data could be used to further de-anonymise user medical records and other patient data. Sweeney and Su Yoo argue that these concerns extend beyond national borders. For instance, in February 2013, a lawsuit was filed by 1,200 physicians and 900 private individuals against IMS Health, a large multinational corporation headquartered in the United States, which collects medical data and RNNs of millions of South Koreans. According to the researchers, IMS Health claimed that they did not violate patient privacy because the data was reasonably encrypted. Sweeney’s and Su Yoo’s ultimate aim was to provide a better understanding of risks to privacy, at a time when many countries become more data-driven, and either restructure or consider adopting national identification systems. 

While the consolidation of vital data, public health analysis, and surveillance might appear to be a twenty-first century problem, contemporary data-driven approaches to societal understanding emerge from a longer history of mathematical application to social investigation. By social investigation, I refer to the process by which human and non-human phenomena are abstracted into fragments of data by means of mathematical theory and analysis. Traces of the sudden application of mathematics to societal phenomena can be seen as early as the seventeenth century, when the aggregation of detailed data on populations became common practice among amateur statisticians in the United Kingdom as well as continental Europe. Although data aggregation and statistical exercises did not originate in Europe, or in the seventeenth century, the century marked an important shift in the collective use of data throughout European countries. Prior to the seventeenth century, there were very few mathematical methods to explain the world in which we live. However, as new statistical methods began to emerge, the world – as well as the people within it – came into view as an exercise in pattern recognition. 

As data aggregation increased in popularity, statistics became the preferred method for empirically driven knowledge production. Life tables expanded into a myriad of subclassifications, such as the rate of urban over rural death, the proportion of tradesmen afflicted by a certain chronic disease, or the excess of male over female births.11 Data and statistics became powerful tools for human classification and the serious study of human-impacted phenomena such the growth of domestic commerce and international trade, the reproduction of labour forces, and the impact of epidemics such as the plague on the political economy. For instance, in his 1662 publication Natural and Political Observations on the Bills of Mortality, English demographer John Graunt collated data from sources such as the Bills of Mortality, an official record of weekly causes of death in London. Graunt’s aim was to create a statistical model that could predict the spread of the bubonic plague, including data on public burials that were found in the public domain and more proprietary government surveys on death records from local villages throughout the United Kingdom. 

With nature and mathematics in tow, Graunt’s cause to connect issues of birth, mortality, labour, economics, and health became attractive tools for the realization of political governance, and more specifically population control. Although mass data aggregation originated in amateur practice for private use, the practice gained prominence as a wider political strategy. Shichiro Matsukawa notes that Graunt dedicated his work in Natural and Political Observations on the Bills of Mortality to Lord J. Roberts, minister of state under King Charles II who was an avid patron of statistics as the ‘new science’.12 Matsukawa writes: 

It is undeniable that Graunt’s study, motivated by his personal interest, was too crude in construction to inquire into the true nature of the subject observed and too obscure in basic thinking to clarify the socio-economic relationship. This notwithstanding, his discovery of regularities in social phenomena was of such significance at that time, especially because the regularities were numerically demonstrated with indisputable evidence, that Graunt, shop-keeper as he was, was given the membership of the Royal Society with the King himself as recommender.13

Although Graunt’s numerical models were never fully realized, his methods did inspired new belief in the statistical study of governance based on population and societal data.14 A prominent example is political arithmetic, introduced by English politician and physician Sir William Petty around 1671 or 1672. Political arithmetician Charles Davenport describes political arithmetic as ‘the art of reasoning by figures, upon things relating to government’.15 The definition political arithmetic summarizes the extensive use of numerical analysis to all matters related to the state, from the survey of households and family demographics to the calculation of revenues, live stock, taxation, poverty relief, military power, and so forth.

As such, statistics provided a practical utility for the understanding of human and non-human populations, which was thought to provide a more disciplined and efficient means of economic and political administration, as well as a more granular investigations into public life, which could now be surveilled in ‘Number, Weight, or Measure’.17 More so, statistics was thought to provide a much more comprehensive evaluation of institutional structures than humans alone, freeing governance from ‘prejudice, credulity, and superstition’, as John Arbuthnot explains.18

Matsukawa argues that Graunt’s efforts as a merchant and self-taught statistician were primarily motivated by personal interest rather than an intention to solve social problems.19 Graunt’s intention was to develop a universal measure for trade. Matsukawa posits that as commercial arithmetic (later political arithmetic), or the building of predictive data models using of mathematics, had progressed in England into the first decades of the seventeenth century, statistical methods were primarily derived for purposes outside of academic study. Statistical methods and the study of mathematics were instead seen as cultivators of a social economy that could be universalized into a series of causal factors that might explain a potential loss in productivity and manufacture. By quantifying regularities in population phenomena, Graunt attempted to demonstrate that human patterns were inherently economic, and could therefore be numerically regulated by the laws of nature, or more explicitly the laws of God. Thus, any human or economic regularity could be imagined as the Divine Truth of nature that could be, as result, anticipated with a degree of certainty.

Alain Desrosières posts that, while Graunt (a tradesman) and Petty (a physician, mathematician, and member of the Parliament of England) saw potential in statistical as a new kind of God-sanctioned science, they more so positioned themselves as experts in all matters of population phenomena ‘with a precise field of competence who suggests techniques to those in power while trying to convince them that, in order to realize their intentions, they must first go through them’.20 Desrosières suggests that the distinction between research and the promotion of expertise is a matter of a particular course of action, in which any knowledges produced in the former are wilfully consolidated into the latter as a display of individual power. 

Desrosières argues that keeping records on baptisms, marriages, burials, and other characteristics and life phenomena is linked to concerns with determining individual identity as that intended for economic and administrative means.23 Donna Jones posits that life, understood under these terms, ‘has now becomes nothing more interesting than a specific kind of information in an information age’ that fulfils the fantasy of empirical provenance.24 Jones argues that despite early attempts at revealing the mysteries of life (or death) by means of empiricism, the actual condition of being human is incomprehensible. Life is enlisted as the subject of technological experimentations and discovery, yet remains distant from the lived experiences of those observed and quantified. Statistics serves as a medium for the transmission of a simulated reality conveyed as natural law, whereas in actual ‘actions of a people are mediated by the culture that they themselves have created, they exhibit a heightened form of freedom from natural mechanical causality that a purposive organism exhibits in its life activities’.25 Within this regime, life is abstracted into empirical method and recast as that which has value only in as much the fragments of existence can be collected, traced, and measured.

Despite these claims it is imperative that we consider what literal traces of life might behind after COVID-19. It would be foolish to speculate on any certain outcome of what could very well be one of the most comprehensive consolidations of personal data in human history. It remains to be seen if any approach to social investigation during this crisis will materialize into more concentrated modes of governance, or if individual and collective lives as they are might reinforce existing notions of political and economic value beyond the crisis at hand. This is not to undermine the need for more advanced solutions to local and global health crisis, particularly during unprecedented expansion of life-threatening pathogens. We carry a responsibility to actively promote the value of individual and collective life, which may require the amplification and expansion of local and systematic efforts. Within our new, more concentrated assemblage of data, disease, and human abstraction are traces of historical methods of social investigation that attempt to reimagine life and death as phenomena that are motivated by the empirical imaginary. This imaginary has in many instances, as seen in the example of political arithmetic, produced alternative sets of human values organized around the centralization of power and the exercise of social falsification. While throughout the seventeenth century new numerical resources were limited to early statistical theory, we are armed today with an array of powerful machine learning and deep data mining algorithms that can rapidly produce societal insight. 

While it is difficult to establish a clear causal link between political arithmetic and contemporary numerical analysis conducted in the search for COVID-19 solutions, there are traces of political arithmetic thought in government responses to the pandemic. Governmental responses to COVID-19 in the UK, in particular – where there is an unprecedented consolidation of social data – are reminiscent of a numerically led strategy to pandemic recovery. For instance, the UK Department of Health and Social Care (DHSC), which is responsible for government policy on health and adult social care matters in England, and overseer of the English National Health Service (NHS), has issued a five pillar testing strategy to understanding trends and risks to public health so as to control and prevent the spread of COVID-19. The pillars include: 1) the scaling up of NHS swab testing for those in medical need; 2) Mass swab testing for critical key workers; 3) Mass antibody testing to help determine immunity; 4) Surveillance testing to learn more about the disease; and 5) Spearheading a Diagnostic National Effort to build a mass-testing capacity. Concerning pillar 4, specifically, the Department of Health and Social Care states that:

Robust population surveillance programmes are essential to understand the rate of infection, and how the virus is spreading across the country. They help us to assess the impact of measures taken so far to contain the virus, to inform current and future actions, and to develop new tests and treatments.27

Part of the ‘robust’ population surveillance programmes are mandates that general practitioners share confidential coronavirus patient information with the UK Government.28 GPs are asked to adhere to data protection regulations ‘within reason’ to ‘support the secretary of state’s response’ to COVID-19.29 Government-approved IT contractors TPP (which runs a health service app) and EMIS (a clinical support system), are also asked to share consenting patients’ data with the UK Biobank project, a national and international health research repository established in cooperation between the UK government and various health research foundations.30 Although GPs are asked to keep records of all the data processed under the mandates, it is unclear where this data will be stored, how this data will be used post-virus, or under what conditions these data might be subject to the privacy agreements of external organizations. 

Swift movements to consolidate data by agencies such as DHSC risk leaving open future gaps in data privacy, while also centralizing the patient data into government forecasting efforts. Nonetheless, it is premature to speculate on how these new insights might impact the future of social living, economic or political life. Given the brief genealogy of a key moment in numerical analysis and the contemporary response to the COVID-19 pandemic, we must consider, however, what impact data consolidation might have on future, even non health related, surveillance programmes. It must be asked whether the risks involved outweigh the immediacy of crisis; and if so, what traces of political and economic speculation based on our intimate medical data might be left behind? Although reasonable efforts to secure our information seem to remain a priority, history has shown that what might appear to be steps towards the immediacy of crisis mitigation may lead to the propagation of far more widespread than the contemporary biological virus.