Data narration is the activity of crafting narratives supported by facts extracted from data analysis, using interactive visualizations. It allows the transmission of findings in the data, by visual means, in order to facilitate their reception by a target audience. Despite its recognized utility in public health, data narratives are typically limited to the transmission of treatment recommendations to educate the general public. This paper describes the crafting of a data narrative about tuberculosis pandemic in Gabon, intended to an audience of health professionals and authorities. Specifically, we describe and illustrate all phases of the crafting process, combining best practices in data narration and epidemic intelligence.
for data narration in public health.
In this paper we describe the crafting of a data
narraData narration, i.e. narrating with data visualisations tive about tuberculosis (TB) pandemic in Gabon. This
[
Despite its recognised usefulness in public health [
ration in the field of epidemiology. • A summary of the main messages learned from
the data, which form the heart of the narrative. • An example of usage of data narration as a decision support tool to enable decision makers to define a better tuberculosis control strategy.
related work on data narration and epidemic intelligence processes. Section 3 describes the main phases of the crafting process and the obtained results. Finally, Section 4 presents lessons learned and Section 5 concludes.
In this sub-section we provide some background on tuberculosis and we describe related works on data narration and EI processes. We also review the use of data narration in public health.
synthesis, in which the analyst assembles and organizes information pieces to be communicated, facilitating the telling of visual analytical findings in a compelling narrative.
The same three phases (but named diferently) are described by Lee et al. [7]: (i) explore data, to retrieve ifndings among data, (ii) make a story, to turn findings into a sequence of narrative pieces and build the plot of the narrative, and (iii) tell a story, to render the plot via visual means.
Recently, a conceptual model of data narrative [
We highlight that despite naming discrepancies, the described proposals agree in the major phases of the data narration process.
terium, Mycobacterium Tuberculosis, which is transmitted by air when an infected person coughs or sneezes.
In the most frequent form of TB (pulmonary TB) the pathogen attacks the lungs, but in more complex forms (extra-pulmonary TB) it can also afect other organs (e.g.lymph nodes, bones, kidneys, etc.). Typical symptoms of TB are fever with night sweats, chronic cough, fatigue, shortness of breath and loss of appetite.
Tuberculosis continues to be a serious public health problem whose magnitude requires global attention. According to the World Health Organization (WHO), this pandemic infected 10 million people around the world in 2019. In the African region, that of Sub-Saharan Africa is the most infected with nearly two million cases reported each year. Among these countries, Gabon is the sec- 2.3. Epidemic intelligence process ond most afected with an annual incidence of 428 cases per 100,000 inhabitants, behind Zimbabwe (562 cases per 100,000 inhabitants). The LBVO-A health region alone accounts for more than 50% of tuberculosis cases notified in Gabon.
The World Health Organization (WHO) proposed an EI process [9], with several protocols depending on whether data is collected through Indicator-based surveillance (IBS) or Event-based surveillance (EBS). While the former deals with data that has been previously validated, the latter focuses on new data, mostly based on rumours 2.2. Data narration process and unverified information, therefore requiring a validaData narration, i.e. the crafting of a data narrative, is a tion process. Given the retrospective character of data complex process at the crossroads of several domains : narration and the importance of data validation, we condata processing, data analysis, data visualisation, commu- sider only IBS protocols, relaying on sources in the health nication, among others. Despite numerous contributions sector. in each of these fields, few works propose global method- The process is organized into five main phases: (i) ologies, describing the whole data narration process. detection of raw data, which consists in selecting data
Firstly, Chen et al. [
Apart from the communication phase, the rest of the phases of the EI process proposed by WHO correspond to the data exploration phase of the data narration process.
Indeed, nothing is advised concerning the way of crafting messages (explaining the learned and validated signals) and structuring them in a coherent discourse. Finally, the communication phase of the EI process corresponds to the presentation phase of the data narration process, but also to efective dissemination of the information.
In particular, the WHO suggests that various supports may be used to share information, but the use of visual artifacts is not detailed. 2.4. Data narration in public health two visual stories, one on vaccine safety and the other on cancer immunotherapy. Both examples, based on data, combine multiple media (photographs, illustrations, choropleth maps, tables, graphs, and diagrams) with text to create powerful visual stories for selected target audiences.
A COVID-19 monitoring dashboard that provides data exploration and visualization capabilities was designed in [18]. One of its applications is to support data storytelling.
According to the state of the art, data storytelling techniques have been used many times in the field of public health, both applied to population awareness and education, and recently also to EI. The proposed data narratives were crafted in an ad-hoc way, not following neither EI nor data narration processes, and only the presentation phase (data storytelling) was reported. To the best of our knowledge, this article is the first to describe the complete process of crafting a data narrative about public health phenomena.
in public health. They mainly concern the communication phase, dealing with storytelling issues.
In the field of public health, there are many commu- 3. Crafting of a data narrative nication strategies, but the most efective is storytelling about tuberculosis [10].
There is a growing trend to use storytelling as a re- In this section we describe the crafting of a data narrative
search and intervention tool on public health issues, espe- to tell tuberculosis pandemic situation to health
authoricially those with a strong disease prevention component ties in Gabon. The crafting process is inspired by classical
[11]. data narration ones [
Indeed, storytelling is used to educate populations and enriched with particular tasks from the WHO EI proon health protection practices, to advocate for improved cess [9]. The following sub-sections describe each of the clinical care and to encourage eforts to combat infectious phases and the obtained results. diseases [12].
For example, in [13], authors tested a new commu- 3.1. Goal setting nication modality to promote prevention messages and colorectal cancer screening among Latinos. In [14], au- In this phase, the goal of the data narrative is defined. thors conducted a survey to determine the efectiveness From this objective, one or more analytical questions can of a rational emotional digital storytelling therapy on be derived.
HIV/AIDS knowledge and risk perception among school children in Enugu State, Nigeria. Objective definition The objective of this data narra
In [15], authors developed a digital storytelling inter- tive is to describe the epidemiological situation of tubervention on diabetes to raise awareness among immigrant culosis in the LBVO-A health region of Gabon between and refugee populations with limited English proficiency. 2016 and 2018. Specifically, it aims to provide answers
In another study conducted in rural Alaskan commu- to decision-makers on the epidemiological profile of the nities [16], authors investigated whether digital stories disease, in order to enable them to better orient their could influence participants’ feelings about cancer, and decisions for an efective response. whether viewing the digital stories led to a change (or intention to change) in health behavior. Analytical questions formulation By refining this
These studies showed that storytelling can increase objective and conducting interviews with various ofipatients’ positive responses and increase their level of cials, we obtained the following initial list of analytical knowledge. questions.
To address the need for compelling and successful information visualizations in biomedical sciences, authors Q1: What are the epidemiological characteristics of of [17] propose a theoretical framework for visual sto- tuberculosis? We aim to describe the profile (frerytelling and illustrate its potential application through quency, variations) of tuberculosis according to the characteristics of the tuberculosis patients; Q2: What is the spatial and temporal distribution of patients? Identifying the most afected areas is essential to target response actions.
Data exploration consists in exploring an epidemiological dataset for trends, patterns and correlations. Considering tasks proposed in EI and data narration processes, we proceed in 7 steps: (i) data collection, (ii) data processing (to solve data quality problems), (iii) data analysis (descriptive and analytical epidemiology), (iv) data interpretation (qualitative assessment of findings),
(v) verification (cross-checking and comparison with the state of the art), (vi) epidemic risk assessment, and (vii) message formulation (key messages, based on findings, to the target audience). This steps are no supposed to be executed sequentially, and many back and forth transitions may be necessary.
Data collection Sociodemographic and clinical data order to solve several data quality problems, including were collected from 7968 medical records of patients un- disambiguation of empty fields, standardization of data dergoing TB treatment at the Nkembo Specialized Hospi- types, correction of geographic boundaries (e.g. for overtal in Libreville. They were completed with geographical lapping districts). Age groups were computed and prodata about the administrative boundaries (neighbour- fessional status were aggregated. hoods and districts) of the LBVO-A region. Clean data were stored in a PostgreSQL spatial data
Patients and their treatments are described in terms of warehouse. the following dimensions:
1According to the WHO, a "lost to follow-up" is a patient whose treatment has been interrupted for at least two consecutive months and for whom no treatment outcome has been assigned (including patients transferred to another treatment unit and those whose treatment outcome is not known) • Time: The year of the treatment (2016 to 2018); Data analysis Data were analyzed using the interac• Type of TB: Clinical form (pulmonary, extra- tive BI tool Tableau Desktop2, devising many queries and pulmonary, multidrug-resistant or unknown); noting findings. • Age: The age of the patient, organised in two Below, we present four examples of the analyzes carlevels: age and age group; ried out. Since tuberculosis mostly afects vulnerable people, we began by studying univariate distributions of • Gender: The gender of the patient (male or fe- the dimensions most concerned by this aspect (profesmale); sion, age, HIV status and geography). Findings set of • Profession: The profession status of the patient; new queries. • Geography: The patient’s place of residence, or- The distribution of cases by professional status, illusganised in two levels: neighbourhood and district; trated in Figure 1, shows a predominance of students • HIV status: The patient’s HIV status, whether (21.94%) and unemployed patients (17.81%). The proportested (negative, positive) or unknown; tion of patients with unknown profession is also high • Treatment outcome (cured, completed, failure, de- (22.39%).
ceased, lost to follow-up1, discontinued and trans- The distribution of cases by age group (see top part ferred). of Figure 3) shows more patients among young adults (20-34 years) and mature adults (35-64 years).
The spatial and spatio-temporal distributions of cases by district and by neighbourhood (figures omitted because of their size) did not show any spatial correlation,
which is very common in other spatial epidemiology population is under 22 years old. studies. In order to investigate further, we studied the These two results allow the calculation of prevalence distribution by neighbourhood typology (see Figure 2), (number of patients in the study period per 1000 inhabwhere the trends are more marked (p-value = 0.01). Ad- itants) by age group (bottom part of Figure 3). It can ditional data sources were selected for this analysis, and be seen that children are little afected by tuberculosis, additional statistical tools were used. but young adults are only slightly more afected than
The distribution of cases by treatment outcome re- older adults. Finally, it should be noted that the small vealed that 74.60% of patients are lost to follow-up. This number of senior patients does not allow the formulation alarming finding lead to a new analytical question: of statistically-valid insights.
Q3: What are the epidemiological characteristics of patients lost to follow-up?
half of the patients (57.71%) come from poor neighbourData interpretation Hereafter, we present a detailed hoods (see Figure 2). This result is in line with those of example of data interpretation. [20].
The high proportion of patients being students could Other results (not described in data analysis parabe linked to the high consumption of drugs in schools. graphs above) were also verified. For example, men are Indeed, a study carried out at the Paul Idjendjé Gondjout more afected by tuberculosis (61.60% of patients), which School in Libreville [19] showed that among the surveyed is confirmed by other studies [ 21, 22]; pulmonary forms pupils, aged between 14 and 20 years, 40% consumed of TB were observed in a large majority of patients (90%) alcohol, 20% smoked tobacco and 15% smoked cannabis. and drug-resistant forms were marginal (0.28%), which
To put these findings into perspective, given relatively is in line with professional knowledge. young population of Gabon, we studied the age pyramid Other results deviate from the state of the art. For exin the study area (middle part of Figure 3), taken from the ample, the proportion of patients lost to follow-up (75%) is Gabon’s 2013 General Population and Housing Census too high compared to other countries in the Sub-Saharan (RGPL 2013), which evidences that more than half of the African sub-region (e.g. in Mali [23]). This discrepancy must be also informed. In addition, the proportion of unemployed patients (17.81%), even being substantial, is very low compared to those found in other studies [24, 25]. This proportion must be completed by the high number of patients with unknown profession (see Figure 1).
tuberculosis among the student population.
Patients lost to follow-up may lead to a generalized TB epidemic in the country. Geographically disadvantaged areas are more risky.
M2: The proportion of unemployed patients (17.81%) is lower than in other African countries.
proportion of cases (27.64%).
M14: In all the districts, the general trend is a slight drop in prevalence. However, two districts (1st of Akanda and 2nd of Owendo) show particular trends and greater variations, partialy explained by their small populations.
M15: There are two hot districts for TB, Les-PK and Nzeng-Ayong, with respectively 13.% and 11.58% of patients.
M16: More than half of the patients (57.71%) come from poor neighbourhoods.
M17: In the distribution of patients lost to follow-up by type of neighbourhoods, we find the same trend as for all patients.
M18: TB afects all social classes and genders, with most patients being adults and living in poor neighbourhoods.
M19: The profile of the patients lost ot follow-up, in terms of gender, age, occupation and location, is very similar to that of other patients.
M3: A high proportion of patients (45.81%) are be- In this phase, messages are assembled and ordered in tween 20 and 34 years old. a coherent and logical plot in order to facilitate their M4: The highest prevalence values (from 16.73 to 17.42 understanding and attract the audience. We started by per 1000 inhabitants, in 3 years) also happens for setting the audience and selecting the messages to convey patients between 20 and 34 years old, but preva- to such audience. Then, we choose the narrative structure lence slightly drops for older patients. and organized messages.
M5: Children are little afected by TB (prevalence is
lower than 2.22 per 1000 inhabitants, in 3 years). Audience choice The target audience for this TB data M6: The proportion of patients lost to follow-up narrative is composed of health authorities in Gabon, (74.60%) is alarming. It is higher than in other including the head of the National Tuberculosis Control African countries. Program (PNLT), the head of the Institute of EpidemiM7: There is no spatial nor spatio-temporal correla- ology and Endemic Control (IELE), policy makers, as tion. well as epidemiologists and doctors involved in the fight M8: The spatial and temporal evolution is correlated against TB.
with the typology of the neighbourhoods (p-value = 0.01), with precarious and mixed neighbour- Message selection All messages formulated during hoods being significantly more afected, but show- data exploration are selected to be communicated to the ing a slight downward trend. target audience.
M9: A very large majority of patients (90%) sufer from the pulmonary form, which is consistent with professional knowledge. Multidrug-resistant TB has been recorded in 0.28% of patients.
M10: The distribution by clinical form of those lost to
follow-up is very similar to that of all patients.
M11: Among the patients, there is a male predominance
(61.60%).
M12: The proportion of patients with unknown HIV status is very high (65%). This does not allow this criterion to be considered in the patient profile.
The plot is organized in 8 acts, listed in the first columns of Table 1.
The first act introduces the study intention. The second act presents the salient messages and the next 6 acts focus on a dimension describing the patients, respectively, occupation, gender, age, HIV status and geography. The ifnal act presents conclusions and recommendations.
We have opted for this structuring, as it facilitates the understanding of a patient profile, according to the various dimensions that make it up. It is intended that the audience should first look at the introduction and general presentation, but then they should be able to navigate between the following acts according to their needs. The order of navigation has no impact on the conclusions.
This structure is well known in modern storytelling as Martini-Glass [26]. It combines several types of interactivity in a balanced way: The author lays down his narrative path (Acts I and II) and the reader interacts and explores the available paths to better understand the data (Acts III to VIII). This makes navigation through the story lfexible, according to the needs of the audience.
tion (e.g. interactive dashboards, infographics, slideshow, video) and the setting of visual artifacts (graphics, colors, text, etc.) for telling acts and episodes, attracting the audience attention.
Visual representation choice We designed and implemented two versions of the data narrative with different visual rendering: (i) an interactive narrative, composed of interconnected interactive dashboards, and (ii) a video3, capturing a particular navigation through the interactive narrative, with audio explanations. We remark that both visual versions are implemented in French; the ifgures and data samples presented in this paper were translated for ensuring readability.
We used Tableau Desktop to render the interactive narrative and OBS Studio for recording the video.
Dashboard implementation We tested several visualisation possibilities, set up the various charts, maps and tables, and supplemented them with explanatory text, visual efects and audio explanations.
Each act is represented by an interactive dashboard, with the exception of Act VII (geography), which needing to display several maps, is broken down into several dashboards. Each episode (concerning a message) is represented by one or more visual artefacts (e.g. charts, maps, tables, text, images, audio).
As an example, Figure 4 presents the rendering of Act II (general presentation). It takes the form of a welcome screen, which gives access to the dashboards correspond- Unlike previous works, which target the general public, ing to the following acts (in the top menu). Several types this narration targets experts in epidemiology and public of visualizations are used to highlight the messages. The health, who are decision-makers. interface ofers the possibility of filtering by neighbour- To our knowledge, this article is the first to describe hood, district and year, for in-depth spatio-temporal fo- the process of crafting a data narrative considering the cus, and to zoom in a map region. peculiarities of the epidemiological field and highlighting the main methodological challenges. Concretely, (i) data analysis is completed by a comparison with other state-or4. Lessons learned the-art studies, (ii) messages are supported by statistical tests, (iii) the geographic component is very important, (iv) the restitution must be both guided and interactive, and (v) back and forth transitions among process phases and steps are be necessary.
In addition to Gabonese health authorities, who are fully satisfied by the experience, the data narrative was presented to professionals of other African countries [27]. We hope that this initiative will serve to inspire other teams to reproduce the experience in other health ifelds, in particular to facilitate understanding of the epidemiological situation of other infectious diseases (covid, cholera, dysentery, yellow fever, ebola, etc.).
As perspectives, we plan to work on producing a data narrative based on the simulation results. Concretely, we are studying the spread of the tuberculosis epidemic with a multi-agent system whose parameters will be taken from the underlying spacial data warehouse, and which will be used to test several scenarios and health policies.
and processes [
First, statistical data analysis is not suficient for public health decision making. A systematic comparison with the state of the art, by comparing the figures obtained, is imperative in order to discern global phenomena from regional or seasonal peculiarities. Thus, decision-makers can judge which dimensions of the patient profile are in agreement with the situation in other countries, for which joint actions can be put in place, and which relate to the Gabonese population. Similarly, the results obtained must undergo extensive testing in order to prove its statistical value. As the target audience is predominantly scientific, these results can be communicated in the narrative.
Second, unlike seasonal narratives (frequent in data journalism), in scientific narratives analytical questions are not all known in advance. On the contrary, new questions may arise during data analysis. We illustrated this during the study of the lost to follow-up, in Subsection 3.2. Iterations between goal setting and data exploration phases are often necessary. New findings can also impact previous messages and require updating.
Third, the geographic component is very important in assessing the spatial and spatiotemporal extent of health problems. The restitution in the form of maps is to be privileged, but also, the spatial correlations.
Finally, the data narrative should allow interactive navigation between dashboards. There are diferent proifles among the decision-makers. On the one hand, we ifnd various needs in terms of dimensions and indicators studied, for which a thematic organization (such as that implemented in acts III to VII) is perfectly suited. On the other hand, health authorities need a more comprehensive and guided reading of the narrative. The challenge is to find a good balance for rendering, both guided and interactive.
[7] B. Lee, N. H. Riche, P. Isenberg, S. Carpendale, More on Big Data, 2020, pp. 1380–1387. than telling a story: Transforming data into visu- [19] Y. Mboumba Sambo, Lutte contre la consommation ally shared stories, IEEE Computer Graphics and de la drogue en milieu scolaire au gabon: Cas du Applications 35 (2015). Lycée Paul IDJENDJE GONDJOUT de la commune [8] F. E. Outa, M. Francia, P. Marcel, V. Peralta, P. Vassil- de Libreville, Dissertation, INJS Gabon, 2018. iadis, Supporting the generation of data narratives, [20] E. Engohan Alloghe, M. Toung Mve, S. Ramaroin: ER Forum, Demo and Posters 2020, Vienna, Aus- joana, J. J. Iba Ba, D. Nkoghe, Epidemiologie de tria, 2020. tuberculose infantile au centre antituberculeux de [9] World Health Organization, Early detection, assess- libreville de 1997–2001, Med trop 66 (2006) 469–471. ment and response to acute public health events: [21] B. Melki, S. Saad, H. Daghfous, M. Khelifa, F. Tritar, implementation of early warning and response with Forme grave de la tuberculose : le pyopneumothoa focus on event-based surveillance, Technical doc- rax tuberculeux, Revue des Maladies Respiratoires ument, 2014. 32 (2015). [10] N. Patel, N. Patel, Modern technology and its use [22] B. Larbani, M. Terniche, S. Taright, M. Makhloufi, as storytelling communication strategy in public La prise en charge de la tuberculose pulmonaire health, MOJ Public Health 6 (2017) 338–341. dans une unité de contrôle de la tuberculose d’alger, [11] B. McCall, L. Shallcross, M. Wilson, C. Fuller, Revue des Maladies Respiratoires 34 (2017).
A. Hayward, Storytelling as a research tool and [23] A. Sylla, B. Marchou, N. Kassi, N. Ello, T. Aba, intervention around public health perceptions and G. Kouakou, C. Mossou, E. Ehui, S. Eholié, E. Biasbehaviour: a protocol for a systematic narrative sagnéné, Co-infection tuberculose/vih: à propos review, BMJ open 9 (2019). de 717 cas suivis dans un service de maladies in[12] E. K. Tsui, A. Starecheski, Uses of oral history and fectieuses en afrique subsaharienne, Médecine et digital storytelling in public health research and Maladies Infectieuses 47 (2017) S137–S138. practice, Public health 154 (2018) 24–30. [24] G. Tékpa, V. Fikouma, R. M. M. Téngothi, [13] L. K. Larkey, J. Gonzalez, Storytelling for promoting J. de Dieu Longo, A. P. A. Woyengba, B. Kofi, Ascolorectal cancer prevention and early detection pects épidémiologiques et cliniques de la tubercuamong latinos, Patient education and counseling lose en milieu hospitalier à bangui, The Pan African 67 (2007) 272–278. Medical Journal 33 (2019). [14] B. Ezegbe, C. Eseadi, M. O. Ede, J. N. Igbo, A. Aneke, [25] S. NIANG, E. ABDALLAHI, K. THIAM, F. B. R.
D. Mezieobi, G. C. Ugwu, A. U. Ugwoezuonu, E. Eliz- MBAYE, M. CISSE, A. DIENG, N. T. BADIANE, et al., abeth, K. R. Ede, et al., Eficacy of rational emotive Aspects épidémiologiques, diagnostiques et évodigital storytelling intervention on knowledge and lutifs de la tuberculose pulmonaire à microscopie risk perception of hiv/aids among schoolchildren positive au district sanitaire de saint-louis., Revue in nigeria, Medicine 97 (2018). Africaine de Médecine Interne 5 (2018) 65–69. [15] J. W. Njeru, C. A. Patten, M. M. Hanza, T. A. [26] E. Segel, J. Heer, Narrative visualization: Telling Brockman, J. L. Ridgeway, J. A. Weis, M. M. Clark, stories with data, IEEE TVCG 16 (2010). M. Goodson, A. Osman, G. Porraz-Capetillo, et al., [27] R. Ondzigue Mbenga, V. Peralta, T. Devogele, Stories for change: development of a diabetes dig- S. Maghendji, E. B. Ngoungou, Processus de narraital storytelling intervention for refugees and im- tion de données en intelligence épidémique avec apmigrants to minnesota using qualitative methods, plication à la pandémie de tuberculose au gabon, in: BMC public health 15 (2015) 1–11. 8e Journées Camerounaises d’Informatique Médi[16] M. Cueva, R. Kuhnley, L. Revels, N. E. Schoenberg, cale, 2021.
M. Dignan, Digital storytelling: a tool for health promotion and cancer awareness in rural alaskan communities, International journal of circumpolar health 74 (2015). [17] T. Botsis, J. E. Fairman, M. B. Moran, V. Anagnostou,
Visual storytelling enhances knowledge dissemination in biomedical science, Journal of biomedical informatics 107 (2020). [18] A. S. Peddireddy, D. Xie, P. Patil, M. L. Wilson,
D. Machi, S. Venkatramanan, B. Klahn, P. Porebski, P. Bhattacharya, S. Dumbre, et al., From 5vs to 6cs: Operationalizing epidemic data management with covid-19 surveillance, in: 2020 IEEE Int. Conf.