Topic Timelines for Enabling Close and Distant Reading of Discursive Shifts. A Pilot Case Using Periodicals of European Diabetes Organizations

(1) Context and motivation

We are currently running an interdisciplinary project that develops mixed-methods approaches to historical analysis and applies them to study the history of patient organizations in Europe in the 20th century. Rather than focusing on particular diseases or organizations, the project analyses the emergence, development, and impact of patient organizations as historical phenomena. Hence, we are studying long and complex processes, and, since previous research is scarce, we have little information about when and where important events and developments took place. Therefore, an important part of our work is corpus-driven exploration using unsupervised text mining methods to extract information without relying on previous knowledge of the corpus content.

Simultaneously, we are aware of the risks involved in relying on quantitative analysis of texts. Extensive close reading of parts of the sources is indispensable to understanding patient organizations and their impact on medical reasoning. Our approach is therefore directed towards finding ways in which qualitative and quantitative methods can work productively in tandem.

(2) Dataset description

The four patient periodical corpora of our study were all constructed by applying OCR on printed material as described in Aangenendt et al. (2024). Basic corpora statistics are shown in Table 1. While the texts themselves cannot be shared as open data due to copyright restrictions, we share the output of the text mining experiments conducted here, as well as aggregated data from the texts. The type of data shared is summarized in Table 2. The data listed has reuse potential for historians interested in diabetes in the 20^th century and in popular medical discourse more broadly. However, the main reuse potential we present is the Topic Timeline tool, which can be applied to any corpus that extends over time.

(3) Method

Our aim was to develop and test a visualization technique suited for research in the history of science and ideas involving large text corpora without prior knowledge of their contents and to analyse how these contents develop over time. For this, we needed to combine qualitative and quantitative elements and a temporal dimension. Most methods for visualizing longitudinal variations in topic prevalence solely present the variation in an aggregated form or only let the user access texts representing prototypical examples of the extracted topics. Examples of visualizations typically used for visualizing aggregated topic prevalence or frequency include line charts (Blei et al., 2003), stacked bars (Sheehan et al., 2021), heat maps (Meaney et al., 2022), as well as the Theme River visualization (Havre et al., 2000), which is specifically developed to visualize longitudinal topic variation in an aggregated form. Like many visualizations used within the digital humanities, these approaches are adapted from methods for visualising purely quantitative information (Drucker, 2015). While being useful for showing topic variation, these types of aggregated visualizations do not direct the user to specific places in the corpus. The Topic Timelines tool, therefore, has three main aims: first, it presents an overview of topic variation; second, it shows how the aggregated topic information stems from individual texts by being prominent—that is, important and salient—and/or prevalent in its frequency; third, it directs the user to specific pages in the corpus for further analysis. In this way, the tool combines close and distant reading. It makes it possible for the user to select potentially interesting content from the distant reading view of aggregated timelines and then carry out a close reading of the selected texts. Since we have previously described the design ideas (Skeppstedt et al., 2024), the focus here is instead to show how the tool can be used in practice. The main visual components of the topic timelines are summarized in Figure 1.

Figure 1

The topic modelling functionality on which the visualization is built automatically extracts recurring topics from a text corpus. Each topic is represented by a list of associated words, a list of associated texts, and a measure of the topics’ prominence in these texts. This output, as well as a timestamp associated with each text (e.g., publication date), are used to create the topic timelines visualization. The visualization represents the topic by a semi-transparent circle at the intersection between a vertical line representing the text and a horizontal line representing the topic: the more prominent the topic, the larger the circle. In addition, the more prevalent the topic (i.e., the more texts in which it occurs), the more circles will be included in the graph. This creates a pattern of overlapping circles of different sizes that—in contrast to previous visualization approaches—provides a zoomable visualization of variation in both topic prominence and topic prevalence. The circles also form clickable links to access and read the texts.

The Topic Timeline tool is not specific to a particular topic modelling algorithm. We employed a previously used workflow based on the topic modelling tool Topics2Themes (Skeppstedt et al., 2018).⁴ We have used this tool in several studies to extract interesting topics (e.g., Skeppstedt et al. 2020, Skeppstedt et al. 2021) and compared its output to manually annotated topics in a temporal corpus (Stede et al., 2023). In the workflow, Topics2Themes (i) uses NMF topic modelling (Non-negative Matrix Factorization), (ii) re-runs the algorithm several times to extract stably occurring topics, and (iii) a user manually adds stop words after inspecting the output, to ensure the topics being based on relevant content words.

We aimed for a timeline visualization that would be as detailed as possible, but also fit on a single page. When configuring the Topics2Themes tool to extract 90 topics, we retrieved around 70–80 stably occurring topics (depending on corpus) that fit on one page. All four corpora were lemmatized—for Swedish using Efselab (Östling, 2018), and for the other three languages using SpaCy (Honnibal & Montani 2017) — before we applied topic modelling. For all four corpora, we used the page as the text unit for the topic modelling algorithm. In a first iteration, we employed a standard stop word list from NLTK (Bird et al 2009). We then manually examined words associated with the topics extracted, and added words to the stop word list which had generated meaningless topics—for example, grammatical words, addresses, titles, and abbreviations. We then ran a second iteration of the topic modelling algorithm with these expanded stop word lists, resulting in the final topic modelling output used for the visualization.

In the next step, we assessed the words included in each topic together with a sampling of texts from each topic to categorize the topics in overarching thematic fields. We assigned the themes separately based on our findings in each periodical. For this reason, the thematic fields we assigned differ between the timeline visualizations. However, all corpora had medicine, food, and administrative matters as prominent thematic fields. We then reordered the timeline visualization so that the topics were grouped according to thematic field. This step made it easier to read the visualization and compare the distributions of related topics over time. For example, it becomes possible to see whether food topics overall become more or less prominent at a particular time, or if urine test topics become less prominent when blood test topics become more prominent. In the visualizations, we have shaded the thematic fields in different colors to make them more visible. The downside of this reordering is that it becomes more difficult to find a particular topic based on its number. Therefore, we have highlighted the specific topics that we discuss in the text.

Da (2019) argues that previously unknown or unexpected topics extracted by topic modelling are only worth taking into account when the algorithm also extracts other known or verifiable topics. Using archival sources, we were able to match the timelines to known events: In Figure 2, topic 32 appears frequently between 1963 and 1982, containing the words “diabetes resort, nordanede, farm, place, crown, person, full board, signed, hereby, relative, registration, of which, name, time, day, open, address, starting, christmas, stina”. This topic concerns the diabetes resort in Nordanede near Sundsvall in northern Sweden, which the SDA operated from 1963 to 1983 and whose matron was Stina Frisk (SE/ARAB/1611/4/1/4; SE/RA/730662/A/A 4/6). Similarly, the timelines align with known shifts in technologies of diabetes management. In Figure 2, topic 71 about insulin pumps appears in the early 1980s, when these devices became available to consumers (Bradwel 2023, pp. 80–82).

Figure 2

In Figure 3 topics 30 and 52 are advertisements for blood sugar monitoring devices, a technology that became widespread among home consumers in the 1980s (Pfaff, 2018). Meanwhile, Figure 4 makes visible the shift from topic 38 (urine testing kits) before 1965, towards the first self-monitoring blood testing technologies (especially digital glucometers) in topics 15, 41, and 60 after 1980. The French materials also show the shift towards self-monitoring blood testing technologies in the 1980s in topics 18, 38, and 60 in Figure 5.

Figure 3

Figure 4

Figure 5

(4) Results and discussion

By showing the distribution of topics in the journals over time, the Topic Timeline tool enabled us to identify trends and turning points in a voluminous source material over a long time period. In this section, we discuss our findings from using the Topic Timeline tool on the four corpora of diabetes periodicals.

Some topics reveal shifts in medical reasoning; by comparing topics to each other, we can see when and how prominently a particular medical issue was discussed and the distribution of similar and related topics. For example, in Figure 3, which represents the German material, topics 40 (cardiovascular illness), 61 (problems with the feet and other extremities) and 63 (retinopathy) all regard common complications of diabetes. However, while topics 61 and 63 are relatively evenly distributed throughout the publication span of the Diabetiker-Journal, the cardiovascular topic becomes more prominent over time. This can be compared to topic 60 in the diet field, which concerns fat and cholesterol. From the visualization, we can see that this topic becomes much more prominent in the 1970s and 1980s during a time of increased attention to cardiovascular problems. The mapping thus shows which diabetes complications and dietary regimes received attention during specific phases of the organization’s history.

Other findings helped identify ways in which particular shifts in medical reasoning were enacted in material concerning lifestyles and habits rather than in explicit discussions of diagnosis or etiology. In Figure 2 (Swedish material), a conspicuously dense cluster appears in the diet field in the mid-1950s (topic 69). Whereas topics 30 and 53, which are about diabetes diets, are relatively evenly distributed over the entire publication span, and the recipe topic (9) has two prolonged phases in the 1970s and 80s in which it is prominent, topic 69 caught our attention due to its concentration in 1956–1957. Clicking on the texts, we found weekly meal plans. These are not recipes or discussions of diet but rather a kind of guide for food intake overall. In order to understand its brief and intense peak in the late 1950s, we turned to archival sources. In the early days of the SDA, the organization rallied around a leading diabetes physician at the time named Jakob Möllerström. He was a proponent of the so-called free diet, which enjoyed popularity in Sweden and elsewhere in the 1940s. Its advocates argued that the strict diets that had previously been used to control diabetes were, due to insulin, no longer necessary. Instead, if the diabetic used insulin to prevent ketoacidosis, there was no need to be overly concerned about elevated blood sugar levels and try to achieve sugar-free urine through special diets (Bradwel, 2023, pp. 66–71). According to Möllerström, “[s]ugar as such is not a substance foreign or harmful to the organism, but on the contrary a vital substance which must always be present in the blood in sufficient quantities” (1943 års sockersjukutrednings betänkande, 1948).

In the short term this approach enhanced the quality of life for many diabetics, until it became obvious that elevated blood sugar levels increased the risk of long-term complications. To members of the early SDA, who had in many cases themselves experienced and struggled with the strict starvation diets of the pre-insulin era, the free diet was a monumental relief. However, it was also a political issue. The free diet was, according to an appeal by the SDA from 1951 to allow diabetics to be government employees, “a precondition for full ability to work”, and since it was now the standard, there were no medical grounds to discriminate against diabetics (SE/RA/730662/F/F7:14). By this time, however, Möllerström’s reputation had started to dwindle in medical circles, but he was still supported by the SDA, which was led by some of his patients. Then, in 1956, the organization elected its first non-diabetic president, the Social Democratic member of parliament Nancy Eriksson, who immediately began a reform of the organization that tied it more closely to scientific advice and abandoned Möllerström’s ideas. At this point, the professionalization and scientization of diabetic diets became a core aim, in particular through establishing the dietitian profession (Söderfeldt, forthcoming). This cluster, then, appears as an important bridge into a new way of conceiving of diabetes within the organization.

Other topics help us identify shifts in organizational priorities and profile. In Figure 3, topic 15 is about local, in-person meet-ups, which appear to have had a heyday in the mid-1980s. In Figure 2, topic 24 regards personal experiences of living with diabetes. It appears frequently in the 1950s and then becomes less prominent before becoming more prominent again in the 1980s. These results are useful pointers towards sections of the corpora to be studied more closely in an analysis of the organizations and their publications.

Another example of how the timelines help in identifying trends beyond the “dense” discussion found in the more explicitly medically-oriented content is topic 12 in Figure 4 (representing the British material). It first appears to be a random collection of verbs and names (“say, melissa, would, go, think, get, feel, henry, day, know”) clustering in the 1970s and 1980s. Closer investigation reveals that the first decade of this topic represents content from the children’s pages of the magazine. “Melissa” and “Henry” were characters in a recurring story feature between 1973 and 1982 revolving around a young diabetic girl. Melissa’s adventures offered young diabetics didactic lifestyle advice and medical pedagogy: through her experiences, readers learned the necessity of carrying spare syringes, proper injection procedure, the correct dose of glucose after their insulin, and the importance of regulating behaviour to mitigate the risk of hypoglycaemia. Some of the verbs that cluster in topic 12 are thus merely the lemmatized forms of common narrative verbs (“said Melissa”). But they also strongly suggest how children were addressed both as readers and diabetics. The verbs “think” and “feel” imply that paying attention to emotions and physical experiences were integral (in the eyes of adult writers) to successfully managing diabetes in children, and that the child diabetic’s relation to the condition was best shaped by articulating and drawing on these experiences.

In doing this, the BDA was not unusual as far as child and adolescent health promotion was concerned, in particular during the 1980s (Elizabeth, 2021; Kelly, 2023). Representations of childhood were in flux and highly politicized during this decade, and were especially visible in health promotion campaigns around and targeting children (Mold and Elizabeth, 2019). The BDA had been seeking to engage child readers for a much longer period of time, with the goal that the diabetic child would grow up into the “well-balanced” diabetic adult (Moore, 2020). Earlier issues of the journal, such as those from the 1950s, often targeted children through letters written by adults about their own child’s diabetes and through celebrations of adult diabetics who had been able to become parents. Later issues featured an increasing number of letters and articles from children themselves and fictional content (like the Melissa stories). Viewing the distribution of topic 12, we see that it becomes most prominent in the 1980s and continues well after the Melissa and Henry stories had ceased publication. The topic cluster in the 1970s and 1980s (and particularly after the Melissa stories stopped in 1982) is not only drawing from the emotive language of this serial but from other material on the same youth pages of Balance. Such texts enacted a shared experience of being a child diabetic in schools, youth clubs, summer holidays, and camps. At the same time, the clickable functionality of the timeline makes clear to the researcher that the increasing volume of personal narrative content collected in topic 12 does not come from the children’s or youth pages at all, but instead from other parts of the magazine. This might suggest one way in which the discursive trends of the children’s pages preceded and anticipated the later emergence of personal narrative content across the periodical, such as emotional testimony around clinical encounters or the loss of eyesight (e.g., Cooper 1989/1990).

In all of the timeline visualizations, a large number of topics are derived from advertisements. These constitute 29 of the 72 topics in the British corpus, 34 out of 82 topics in the Swedish one, 24 out of 72 in the German, and 21 out of 72 in the French. Ads also appear as parts of many other topics. From a linguistic and semantic viewpoint, this is to be expected. Advertisements are relatively stable texts, with brand names, slogans, and advertising copy remaining unchanged across several years of publication; topic modelling has a high probability of successfully recognizing these collocations. By grouping the advertisements into thematic fields, we were able to see changes over time in the rise and decline of particular brands or products and in promotional design and language alongside other thematic fields for medical and lifestyle topics. In this way, the timeline visualizations called attention to new enactments of diabetes in the relationships between medical, promotional, and colloquial discourses. In the following, we will show how the language in advertisements reflects broader historical developments in diabetes enactment.

We grouped together similar advertisements across the British and French material (Figures 4 and 5). In the British timeline (Figure 4), topics 7 and 14 group advertisements for insulin. Below this is a cluster of topics that capture a variety of different technologies that came into being at specific moments during the lifetime of the magazine: 38 and 15 contain urine and blood glucose testing kits, 41 and 60 capture the rise of the digital glucometer, while 54, 35, 59 and 61 offer a window onto syringes and associated devices (such as sterilization equipment and carry cases). The remainder of the advertisement topics indicate ads for artificial sweeteners, foodstuffs, and miscellaneous services like insurance companies. Topics 38 and 15 largely read as continuations of each other, and our reordering of the Topic Timeline in thematic fields, as discussed in section 3, shows a shift in testing regimens as new and digital measurement technologies for blood glucose (topic 15) displaced the older and analogue ones for urine (topic 38). Similarly in the French material (Figure 5), topics 4, 57, and 64 group advertisements for older insulin and glucose testing technologies (syringes and urine testing kits). Topics 18, 27, 38, 58, 60, and 72 group advertisements for self-monitoring blood glucose technologies. Topics 42 and 67 are especially useful, since they show how the technologies and promotional material for specific products—Iniematic syringes and Glucagon emergency kits—changed with the introduction of self-monitoring blood glucose testing.

The marketing of the newer devices in the 1980s shows shifting relationships to testing and therapeutic technologies and the incorporation of medical terminology into commercial and colloquial vernaculars. Self-monitoring blood glucose devices, especially, are advertised in a way that indicates what historians of technology have called the successful “becoming transparent” of the instrument (Gooding et al., 1989). In this process the instrument becomes “invisible,” allowing the user to perceive and interact with nature through its perfect and unquestioned mediation. As a visual and semantic evocation of this, we often see the absence of the “diabetic” —as a visible person and as an user of technology—in the advertising materials for these types of devices. They are presented directly in front of the reader, as if to indicate an already-existing embodied facility with the objects.

These technologies were also represented through languages of convenience and discretion, indicating enactments of diabetes as a manageable condition allowing full participation in the workforce and a public, busy life. British topic 15 and French topic 56, for example, indicate the importance of speed and ease in blood glucose testing. The Ames digital glucometer offered “simple convenient blood glucose monitoring” with the added advantage of a “privacy switch” that turned off all device noises. This would allow diabetics to make their use of the technology not only conveniently but invisibly and inaudibly when in public (“Super saving spectacular,” 1987). Such advertisements addressed a diabetic user who already trusted the therapeutic technologies and was also a savvy medical consumer with expectations of discretion, comfort, and privacy.

(5) Implications/Applications

The ability of the Topic Timeline tool to illustrate changes in collocations over time makes it especially productive and useful for historical research that combines distant and close reading. Topic timelines allow more efficiency in generating overviews of large corpora and also reveal layers that might go unnoticed in traditional close and distant reading approaches. By clicking on the timeline to access the texts, the historian can both deepen their understanding of the shifts and developments visualized on the timeline through close reading and maintain a critical view on the visualization itself.

The dynamic between the historian’s analogue analysis and the Topic Timeline tool thus provides new perspectives on sources, acting as controls on their own theoretical and historical assumptions. Due to their tendency to repeat sentences and words, for example, advertisements and announcements feature prominently among the topics. Visualizing these topics reveals an important enactment of diabetes in a marketplace involving businesses, patient organizations, and readers. Furthermore, these visualizations show how diabetes was enacted in multiple ways that produced distinct disease objects over time: a pathology of the pancreas, a loss of vision, a lifestyle of regulated food intake, an emotional experience, a condition successfully managed by a shrewd medical consumer, and a community. In this sense, the Topic Timeline tool allows historians to capture multiple enactments and the standardization of disease experiences with the introduction of new treatments, institutions, social relationships, and technologies.

The Topic Timelines tool is, however, suited for research far beyond the history of medicine. In our example, it provided new perspectives on the patient organization periodical as a site of disease enactment and the targeting of different “publics” (Mold, et al., 2019). Material by and for children was published alongside medical articles by researchers and specialist consultants and advertisements for soft drinks, syringes, and blood testing devices. The periodicals show these organizations as heterotopias bringing together the voices of patients, family members, workers, and children with medical and policy debates, lifestyle advice, and the commercial messages of a growing industry. Hence, the Topic Timeline tool is a suitable method for exploring historical corpora that contain a diversity of text types, keywords, and authorial voices—in other words, what is typical for many source materials outside of the scientific and political mainstream.

Data accessibility statement

Data that supports the findings of this study are openly available in Zenodo at https://doi.org/10.5281/zenodo.14244899 and GitHub at https://github.com/CDHUppsala/topic-timelines. As discussed in the dataset description, the full-text data cannot be shared due to copyright restrictions.

Competing Interests

The authors have no competing interests to declare.

Author Contributions

Ylva Söderfeldt: Conceptualization; Funding acquisition; Investigation; Methodology; Project administration; Resources; Supervision; Validation; Writing – original draft; Writing – review & editing.

Andrew Burchell: Investigation; Validation; Writing – original draft; Writing – review & editing.

Julia Reed: Investigation; Validation; Writing – original draft; Writing – review & editing.

Maria Skeppstedt: Data curation; Formal analysis; Investigation; Methodology; Resources; Software; Validation; Visualization; Writing – original draft; Writing – review & editing.