Elizabeth Toon

Text Mining the History of Medicine.

Historical text archives constitute a rich and diverse source of information, which is becoming increasingly readily accessible, due to large-scale digitisation efforts. However, it can be difficult for researchers to explore and search such large volumes of data in an efficient manner. Text mining (TM) methods can help, through their ability to recognise various types of semantic information automatically, e.g., instances of concepts (places, medical conditions, drugs, etc.), synonyms/variant forms of concepts, and relationships holding between concepts (which drugs are used to treat which medical conditions, etc.). TM analysis allows search systems to incorporate functionality such as automatic suggestions of synonyms of user-entered query terms, exploration of different concepts mentioned within search results or isolation of documents in which concepts are related in specific ways. However, applying TM methods to historical text can be challenging, according to differences and evolutions in vocabulary, terminology, language structure and style, compared to more modern text. In this article, we present our efforts to overcome the various challenges faced in the semantic analysis of published historical medical text dating back to the mid 19th century. Firstly, we used evidence from diverse historical medical documents from different periods to develop new resources that provide accounts of the multiple, evolving ways in which concepts, their variants and relationships amongst them may be expressed. These resources were employed to support the development of a modular processing pipeline of TM tools for the robust detection of semantic information in historical medical documents with varying characteristics. We applied the pipeline to two large-scale medical document archives covering wide temporal ranges as the basis for the development of a publicly accessible semantically-oriented search system. The novel resources are available for research purposes, while the processing pipeline and its modules may be used and configured within the Argo TM platform.

Text-Mining and the History of Medicine: Big Data, Big Questions?

Most of us have heard about ‘Big Data’, often as part of discussions about the information collected about us as consumers and citizens, and the increasingly sophisticated tools that analyse such information. But can we as historians of medicine benefit from thinking about our historical sources as ‘Big Data’, and ‘mine’ this data by adapting the tools used by commerce, computing science and intelligence? What possibilities for historical scholarship would such tools open up; what challenges do they present? These questions motivated our involvement in a collaborative project using text mining tools with medical history sources. In early 2014 we joined University of Manchester colleagues from the National Centre for Text Mining (NaCTeM) to work on a project funded by the UK Arts and Humanities Research Council, under their Digital Transformations theme.1 As their name suggests, NaCTeM2 develops text mining tools, mostly for academic use.3 Our team set out to create a semantic search engine, one that would go beyond finding a specific keyword or string of text in a document. Semantic searches consider the context of use in order to locate terms and their variants representing particular concepts. We wanted to explore how such a search could provide new ways of working with series of medical texts covering a long period of major change in medical knowledge, practice and language. We chose two sources to form our corpus, as large-scale collections of structured text are known in digital humanities: the digitised run of the British Medical Journal from 1840 onwards, and the more recently digitised London-area Medical Officer of Health Reports that form the Wellcome Library’s London’s Pulse collection. Text mining (TM) uses digital tools to detect the structure of textual information, then find and recognise patterns and relationships in the structured data.4 For instance, one TM task is to find and compare the number of instances of particular terms over time in a defined corpus, as Google’s N-Gram viewer does using the ‘millions and millions of books’ that Google has digitised or has access to in digital form.5 Another common TM application is finding the relative frequency of the words in a text and then visualising these in a way that makes the different frequencies apparent, for example, by size and position in word clouds. Other TM tools track and compare the relative locations of terms and their variants in texts, or, in the case of topic modelling, identify groups of terms that tend to be representative of a given topic. As Tom Ewing’s use of these approaches demonstrates, they can provide insights that are not readily apparent in traditional reading, however intensive and analytical.6 Collaborating with NaCTeM allowed us to take advantage of even more complex approaches to TM, where systems can be ‘taught’ to recognise textual data as representing entities of different types, such as place names, medical conditions, etc., as well as specific types of relationships between these entities, for example, which symptoms are presented in the text as being caused by a condition. When combined with other tools and approaches used by digital humanities scholars, such as visualisation tools and GIS mapping, TM allows sources to be interrogated in ways that build upon and complement our traditional reading and analysis. Its proponents claim that automated technologies can do this not only faster and more thoroughly with very large data sets, but in ways that reveal new and interesting historical findings. Is this the case with big medical history data? Before we applied TM tools, we needed to make sure that our digitised corpus was sufficiently correct to be effectively mined, and this was no small task. As Tim Hitchcock has pointed out, many historians do not recognise the extent of OCR errors in the digitised texts that our existing search systems query.7 The recently created London’s Pulse is relatively error free, but in the BMJ files, which were digitised and OCRed several years ago, up to thirty per cent of the words have errors. Our NaCTeM colleagues devised a customised approach to correcting OCR errors in medical historical texts,8 which means our system provides a significant improvement on full-text BMJ searches. We then worked with NaCTeM colleagues to analyse sample text, identifying entities and relationships so we could teach our system how to carry out that identification on its own. We began by considering the kinds of entities and relationships historians might want to search for in this corpus. After experimenting with a very large, complex scheme with many subcategories, we decided on a streamlined scheme with seven entity categories (Anatomical; Biological Entity; Condition; Environmental; Sign or Symptom; Subject; and Therapeutic or Investigational) and two relationship categories (Affect and Cause). A team marked up a large sample of text, highlighting where these entities and relationships occurred. We then submitted this sample to a system equipped to ‘learn’ how to recognise annotations of different types, based on language patterns in the text. The ‘trained’ system was able to use these learned patterns to recognise entities and relationships in the un-annotated remainder of the corpus – more than 150 years’ worth of weekly issues of the BMJ, and more than 5000 reports by London-area Medical Officers of Health. Teaching the system to discriminate between the entities historians consider important in historical medical texts proved much more difficult than teaching it to identify simpler entities like named locations. First, terms such as disease names that have been used to describe similar phenomena have changed over time, but using TM techniques the system was able to learn, for instance, that ‘infantile paralysis’ and ‘poliomyelitis’ were different terms used in overlapping time periods for a reasonably similar phenomenon. However, some terms have multiple and changing meanings and uses, depending not only on temporal but also textual context, reflecting the very changes in medical thinking we want to examine. One example is the term ‘inflammation’: as an entity, is it a Condition? a Sign or Symptom? Or is it a characteristic of a body part and thus Anatomical? Any categorisation decision we made, and the rules we devised for making sense of the context, would have to be clear enough for the search system to ‘learn’ and apply. Yet that decision would still need to reflect the term’s changing and indeterminate use in a way that would satisfy historian users of the search system. These tricky and problematic decisions have been built into our system, and we are of course anxious that our users be aware that, thanks to such decisions and to the complexity of the overall task, they need to be critical as they approach the results our system gives. In fact, the system facilitates critical engagement by the ease and speed of making alternative and cross-checking interrogations. As we trial our ‘beta-version’ with our Advisory Group, we are excited by the possibilities that this system, and that TM and indeed digital humanities tools as a whole, open up. First, our system speeds up searches dramatically, and allows more focused searches than would be possible even with fairly sophisticated Boolean searching. By searching for Condition: ‘tuberculosis’, for example, the user gets results where the system has recognised the term as referring to tuberculosis as a condition, rather than finding every instance of the word ‘tuberculosis’ in the text (in phrases like ‘National Tuberculosis Association’, or ‘tuberculosis nurse’). But semantic searching is about much more than convenience. The user can find all instances of a particular entity category: one can, for example, locate all articles published in 1892 where a Biological Entity (including non-human animals and microorganisms) is mentioned, and find the frequency with which each Biological Entity is mentioned. Combining entity searches and relationship searches enables the user to find instances where one entity is said to cause another: by asking what Condition entities are said to cause the entity Sign or Symptom: ‘swelling’ in the entity Anatomical: ‘feet’, the user can find case reports and reviews that discuss which ailments were understood to cause the feet to swell. (By contrast, consider the overwhelming flood of results the searcher would get by searching for the terms ‘feet’ and ‘swelling’.) This capacity is particularly useful for those who want to investigate relatively common, everyday phenomena that would stymie the best intentions of researchers because they are difficult to find in text, too numerous to manage easily, or easily overlooked by the all-too human researchers. We thus expect this tool not only to speed up searching and make it more precise, but also to help us see things that would otherwise be too difficult to see or too easy to miss, or that we might not even have known we were looking for. It will never provide easy and obvious answers to big questions, and it requires that the user know something about how it works. Nevertheless, we hope that as a tool that can facilitate exploration and new ways of encountering existing resources, it will be valuable both as a resource in its own right, and as a means of introducing our colleagues to TM tools and some of the possibilities of digital humanities.

Phenylbutazone (Bute, PBZ, EPZ): one drug across two species

AbstractIn this article we explore the different trajectories of this one drug, phenylbutazone, across two species, humans and horses in the period 1950–2000. The essay begins by following the introduction of the drug into human medicine in the early 1950s. It promised to be a less costly alternative to cortisone, one of the “wonder drugs” of the era, in the treatment of rheumatic conditions. Both drugs appeared to offer symptomatic relief rather than a cure, and did so with the risk of side effects, which with phenylbutazone were potentially so severe that it was eventually banned from human use, for all but a few diseases, in the early 1980s. Phenylbutazone had been used with other animals for many years without the same issues, but in the 1980s its uses in veterinary medicine, especially in horses, came under increased scrutiny, but for quite different reasons. The focus was primarily the equity, economics, and ethics of competition in equine sports, with differences in cross-species biology and medicine playing a secondary role. The story of phenylbutazone, a single drug, shows how the different biologies and social roles of its human/animal subjects resulted in very different and changing uses. While the drug had a seemingly common impact on pain and inflammation, there were inter-species differences in the drug’s metabolism, the conditions treated, dosages, and, crucially, in intended clinical outcomes and perceptions of its benefits and risks.

Book Reviews: Contested Medicine: Cancer and the Military.

Contested Medicine brings a fresh perspective to a notorious and important story. Drawing upon his experience as a radiation medicine specialist, the historian Gerald Kutcher examines Eugene Saenger’s 1960s and early 1970s work with total-body irradiation (TBI) at the University of Cincinnati. Saenger and his colleagues traced the metabolic and psychological effects TBI had on patients with advanced cancers; this work was funded by the US Department of Defense, which wanted to know what would happen to the combat performance of American soldiers exposed to radiation. Kutcher uses the TBI story to anchor his consideration of two fundamental and intertwined elements of post-war biomedicine: the contested nature of therapeutic research amidst new systems of knowledge production (the clinical trial), and the development of biomedical ethics as a form of governance and a set of practices. By examining how Saenger’s work was supported, justified, experienced, rationalised, scrutinised, and judged, Kutcher also helps us reconsider how we make sense of historical medical scandals, both in their initial contexts, and as they have been understood and used by later actors. The book begins with three short chapters establishing the context for Saenger’s TBI work and the themes of Kutcher’s analysis. The first outlines how the clinical trial came to dominate post-war medical investigation, while the second reviews medical discussions among mid-century medical authorities about what constituted ethical research conduct and how it could be sustained. Kutcher then reviews the melding of military and medical questions in the 1950s discussions of radiotherapy for sick patients, and of radiation injury to healthy soldiers. The bulk of the book’s analysis, though, comes in its middle section, which considers what the TBI studies meant to multiple constituencies, including the doctors and researchers who conducted the studies, and the peer review committees that recast the studies to pass new governmental research regulations. Chapter 5 is especially insightful and original, using one patient’s experience to show what TBI meant to and for those who served unknowingly as ‘proxy soldiers’. Here, Kutcher’s medical expertise enhances his analysis, as he reconstructs patient experience through fine detail and thoughtful speculation. Finally, the book concludes by tracking how Saenger’s work was recast yet again by those criticising it, first in the exposes of the 1970s and then again in the 1990s by a new set of authorities – the bioethicists of the Advisory Commission on Human Radiation Experiments (ACHRE). Kutcher parses the ACHRE’s deliberations to show that bioethicists also found it nearly impossible to determine whether Saenger’s work was medical or military, whether it was motivated primarily by therapeutic concerns or by research questions, and what ethical criteria could be used to judge past conduct. The fluid identity and ever-changing nature of the TBI studies meant they defied historical and ethical attempts to classify them, and ultimately, to deliver a definitive verdict on their moral status. That fluidity is far from unique in biomedicine – which, as Kutcher concludes, means that the prescriptive rules usually offered by bioethics ‘are limited in what they can accomplish’ (p. 211). In Contested Medicine, Kutcher has produced a book that successfully demonstrates how researchers, institutions, and ethical authorities managed (or failed to manage) the ‘tensions between research imperatives and therapeutic necessities’ (p. 6) characteristic of biomedicine. At times, Kutcher summarises what his sources say when the reader might want to hear more from the source materials themselves, but on the whole, the book is very well written. Contested Medicine will thus be a valuable resource for scholars interested in post-war medicine and science and, though its focus is on an American story, the book’s analytical framework is strong enough to make it of interest to those who work on other national contexts.