Leveraging Artificial Intelligence and Participatory Modeling to Support Paradigm Shifts in Public Health: An Application to Obesity and Evidence-Based Policymaking

Abstract

The Provincial Health Services Authority (PHSA) of British Columbia suggested that a paradigm shift from weight to well-being could address the unintended consequences of focusing on obesity and improve the outcomes of efforts to address the challenges facing both individuals and our healthcare system. In this paper, we jointly used artificial intelligence (AI) and participatory modeling to examine the possible consequences of this paradigm shift. Specifically, we created a conceptual map with 19 experts to understand how obesity and physical and mental well-being connect to each other and other factors. Three analyses were performed. First, we analyzed the factors that directly connect to obesity and well-being, both in terms of causes and consequences. Second, we created a reduced version of the map and examined the connections between categories of factors (e.g., food production, and physiology). Third, we explored the themes in the interviews when discussing either well-being or obesity. Our results show that obesity was viewed from a medical perspective as a problem, whereas well-being led to broad and diverse solution-oriented themes. In particular, we found that taking a well-being perspective can be more comprehensive without losing the relevance of the physiological aspects that an obesity-centric perspective focuses on.

1. Introduction

The majority of Canadians aged 18 and older are either obese (26.8%) or overweight (36.3%), based on the Canadian Community Health Survey [1]. The high level and increasing prevalence (over time) of obesity and overweight is a global phenomenon. Other studies have shown that obesity doubled in 73 countries within a few decades, and the growth was even more pronounced for countries with a low-middle degree of development [2]. The Obesity Society published a position paper, defining obesity as a disease in 2008; its updated position views obesity as a “multi-causal chronic disease” that may be accompanied by ‘structural abnormalities’ (e.g., musculoskeletal disorders), functional abnormalities (e.g., insulin resistance), various symptoms (e.g., obstructive sleep apnea), and several comorbidity risks (e.g., cardiovascular diseases and type 2 diabetes) [3]. Similar statements have recently been made by other organizations, such as the Canadian Medical Association or the American Heart Association [4]. Such statements are rooted in a medical perspective on obesity, which contrasts with the positions of other organizations “who believe that defining obesity as a disease is unhelpful”, such as the British Psychological Society [5]. The latter has emphasized social factors (e.g., weight stigma) and environmental factors, grouped under the wider notion of an ‘obesogenic environment’ (e.g., food deserts, neighborhoods that do not support walking) [5].

This tension echoes a deeper schism between a medical perspective emphasizing physical health and one that also considers mental well-being. Indeed, some groups view weight loss as an ‘obsession’, noting that many individuals who attempt weight loss become heavier and their self-esteem shattered, while the general population may have a higher bias and stigma for those living with obesity [6]. Put simply, one side suggests that “success is reframed as improvements in health, function and quality of life” [6], whereas another side emphasizes weight loss as the core outcome of obesity treatment [7]. Opponents to a medical view on obesity include the Health at Every Size movement, which is weight neutral, thus rejecting weight as an indicator of health and opposing weight loss as an end goal [8]. These views resulted in today’s body positivity movement [9] and proposals to characterize ‘overweight’ and ‘obese’ as stigmatizing and normative labels [10]. Collectives of feminists, fat activists, and health professionals similarly participate in ‘anti-diet movements’ [11]. Caught between different camps, policymakers are urged to reframe health policies [12], particularly as obesity policies have occasionally been designated as stigmatizing [13], and the ensuing weight bias has a demonstrated negative impact on mental health [12,14]. Thus, policymakers need to exert great care to satisfy different stakeholders when setting targets for population health [15], carefully consider the language within health policies (e.g., use of person-first versus disease-first terminology), and handle a complex problem driven by a massive number of factors.

In this paper, we use techniques from artificial intelligence (AI) and participatory modeling to support policymakers in balancing different perspectives on obesity and compare their implications for public policies. Our study is rooted in the growing “recognition that obesity management should be about improved health and well-being, and not just weight loss” [16]; hence, we integrate both perspectives in a comprehensive representation of knowledge on obesity. In particular, our study is motivated by a discussion paper from the Provincial Health Services Authority (PHSA) of British Columbia, in Canada. The paper, titled “From weight to well-being: time for a shift in paradigms?” acknowledged the long-established and well-documented links between obesity, the broader environmental context, and medical conditions [17]. At the same time, it suggested that shifting from weight-focused to well-being-focused approaches in practice and policy had the potential to improve population health and address the problems associated with obesity in ways that protect and promote mental well-being as well as physical well-being. Our study uses AI and participatory modeling to investigate this proposed paradigm shift and its implications for obesity policies.

AI has a rich history in obesity research [18]. Although it often evokes the use of machine learning to predict or diagnose obesity [19,20,21,22], other facets of AI underpin the foundations of numerous systems [23], such as the use of large language models (e.g., GPT) for personalized assessment [24] and treatment [25,26]. For instance, AI applications in smartphones support efficient ecological momentary assessments to collect precise data on human behavior, thus examining how individual and environmental characteristics shape eating and physical activity behaviors [27]. Serious games have also been used in obesity research [28], for instance, by incorporating elements of AI (e.g., recommender systems) to tailor interventions [29]. Simulations have also played an essential role in policymaking on obesity, for example through agent-based models [30,31], which are increasingly used in digital government research [32]. In an agent-based model, the AI components aim to replicate core aspects of human decision-making within virtual agents. In this paper, we focus on a traditional aspect of AI, which is to represent and analyze knowledge regarding a system. The Foresight Obesity Report [33] represented the system of obesity as a conceptual map, in which weight-related factors were captured as nodes and their interrelationship as edges. This representation has been adopted by many studies [34,35,36,37], often with a participatory modeling approach, whereby the knowledge embodied in the graph was obtained through human participants. Our work contributes to this literature in two ways:

• We create a concept map of obesity with 19 experts to combine different views on obesity, physical well-being, and mental well-being. With 98 nodes and 174 edges, it is currently the most comprehensive expert-driven model of obesity that conciliates medical perspectives with those emphasizing well-being.
• We analyze knowledge on obesity using tools from network science and natural language processing and examine the implications for public policymaking.

The remainder of this paper is structured as follows. To make this manuscript self-contained, we start in Section 2 with a brief introduction to the principles of participatory modeling; readers familiar with this field may skip the section. Based on these principles, Section 3 explains our methods to elicit knowledge in semi-structured interviews with experts and structure it into a conceptual model. Since the knowledge representation is only as strong as the experts who participate in this exercise, we provide the full list of participants for transparency in Appendix A. Our analysis of participants’ knowledge is explained in Section 4, covering both an examination of the map using network science (Section 4.2) and the use of natural language processing over the transcribed interviews (Section 4.3). Since the map is a massive model spanning multiple domains (e.g., built environment, psychology, physiology), space limitations preclude its inclusion in the main body of this paper; it is instead detailed in Appendix B. Finally, we discuss the implications for policymaking in obesity and potential extensions so that improved AI solutions can further support government services and policies.

2. Background: A Primer on Participatory Modeling for Causal Mapping

Participatory Modeling (PM) is an approach that actively identifies and engages participants in knowledge elicitation and integration. In contrast to other fields of participatory research (e.g., citizen science), PM tends to focus on motivating changes and is considered a researcher-driven approach to taking action [38]. Core application areas of PM include policy and planning [39]. Although there is a wide variety of tools and methods for PM [40], many of them represent a conceptual system in schematic form. The schema may be loosely structured (e.g., rich pictures) or follow a prescribed structure (for a comparison on the level and nature of the constraints, we refer the reader to Table 5 in [41]) such as a map (e.g., cognitive map, causal loop diagram) [42,43,44,45,46], in which case, the approach is specifically known as ‘participatory systems mapping’ [47] and serves to assist with public policy decisions in contrast with group-based methods that may seek to empower communities (e.g., community-based systems modeling) [48]. Note that participatory systems mapping is distinct from ‘participatory mapping’ [49], which is a term used for collecting volunteered geographic information; they share the notion of contributions from participants and an application to participatory planning and decision-making, but participatory mapping focuses on spatially referenced data and involves geographic information systems.

In this paper, we represent knowledge in the form of a causal map, where weight-related factors are represented by nodes and their relationships are captured by edges. Any factor can be represented, but it must be clearly able to increase or decrease; for example, ‘weather’ or ‘built environment’ would be poor choices for factors, but ‘rain level’ or ‘availability of sidewalks’ could be used. A relationship denotes causality; hence, it is directed and typed (either positive or negative) to indicate that an increase in a node either increases (‘+’) or decreases (‘−’) another node. For instance, cardiovascular diseases $\stackrel{+}{\to }$ fear of engaging in physical activity indicates that individuals with cardiovascular diseases have a higher risk of fearing physical activity. Conversely, fear of engaging in physical activity $\stackrel{-}{\to }$ physical activity states that the higher the fear of engagement, the lower the level of physical activity. Hasty conclusions should be avoided on the basis of a single relationship: for example, we cannot conclude that all individuals with a fear of engagement will automatically have a low level of physical activity. Rather, a map represents knowledge regarding an entire system; hence, we need to consider the joint effects of all protective and risk factors [50] (e.g., the fear of engagement may be counterbalanced by a protective factor for physical activity). Note that there are variations of this process [47] in which types include ‘unclear’ (we know that there is a relationship but its type is unknown) or ‘complex’ (when it is non-linear or depends on other considerations). For example, weight-based discrimination may not exist at certain levels of weight but rises rapidly from a certain level onward, which shows a non-linear relationship that would only be typed as ‘+’ under certain conditions [51]. Such variations are not used in the present work, where edges are strictly categorized as increasing or decreasing.

There are several core steps to arrive at a causal map through a PM process. The specific number of core steps can differ across papers depending on the granularity chosen by the authors. For example, some may split the elicitation stage into a development phase (e.g., pilot testing) and its application with participants [52]. Despite differences, practitioners generally agree on two aspects of the steps. The sequence starts with ➀ a definition of the objective and scope, followed by the ➁ stakeholder selection, ➂ knowledge generation, and ➃ qualitative aggregation into a map [53]. These steps account for four facets, known as the four Ps [54]: purpose (why this model and a PM approach?), partnership (who participates?), process (how are they involved?), product (what is the result?).

To begin, ➀ setting a focal point and a problem boundary contributes to operationalizing the purpose of the model. The focal point can consist of a single variable (e.g., obesity [51], smart city [55]) of a set of variables (e.g., suicide ideation, attempt, and death [50]). Before the creation of a representation of a system, the boundaries provide high-level guidance to exclude or include certain concepts. For example, a model for adult obesity should not include childhood obesity. After the system is represented as a map, its boundaries become more clearly delineated by its actual content. That is, “the boundary of such a conceptual system is made of the concepts contained/described within it.” [56]. Although a study usually starts by setting high-level boundaries, there are exceptions in which boundaries are only set after interviews have been transcribed [57].

To ➁ select stakeholders, the overarching principle is that “researchers purposefully select participant stakeholders based on expertise or vested interests in the research” [39]. Most studies systematically identify the knowledge areas needed to solve the problem [58]. The corresponding identification of participants is rarely systematic in PM studies and involves a variety of methods, such as sampling (which may include referrals for snowball sampling), nomination by a committee, or the application of selection criteria (e.g., who qualifies as an ‘expert’?) [58]. Although selecting stakeholders is primarily a matter of partnership, it also starts to venture into the process. For instance, when participants are invited, an email can explain the process to them. This helps to set the tone and ‘manage expectations’, which is one of several characteristics of successful PM projects [59].

The ➂ knowledge generation step is often the most detailed in studies and also the most time-consuming since it is process-intensive. Participants could be engaged individually via semi-structured interviews (e.g., in person or remotely, through a facilitator or a device [60]) or as a group through workshop activities such as writing factors on post-it notes, consolidating them and linking them collaboratively [47]. Each approach needs a specific skill set from the facilitator; for instance, workshops require paying particular attention to social and group dynamics such that there is equity among participants. The nature of the engagement is important because the co-production of knowledge at this step can be a strong benefit of a PM approach. Indeed, knowledge co-production contributes to greater trust in the model [61], and endorsements of a model by community champions and/or leading experts can contribute to an efficient translation from knowledge to action. Finally, we arrive at an ➃ aggregate map, which invokes both a product and a process (how are participants involved in validating and using the map?). The map may have been directly constructed from a group in a workshop or will result from combining the maps of individuals. Neither approach is easier than the other; rather, they depend on the needs of the project and the skills of the team. For instance, a workshop activity may require an expert facilitator to promote equity and avoid conflicts. In contrast, individual interviews may be simpler, but combining individual maps requires either accounting for the heterogeneity of the terminology used by participants [62], or limiting it by only allowing for certain pre-selected terms [63]. The aggregate map may also be augmented based on the literature and data [64,65].

3. Knowledge Representation by Participatory Systems Mapping

3.1. Key Steps of Our Methodology

The four stages of the methodology (➀–➃) refer to the steps explained in the previous section. Our methods are summarized in Figure 1 and detailed below.

Three elements constitute our focal point ➀: physical well-being, mental well-being, and obesity. We focus on adults within Canada; hence, childhood obesity is excluded from our map. The relationships between mental well-being and obesity are detailed in a report by the Provincial Health Services Authority (PHSA) of British Columbia, so we use their report as a point of departure. Specifically, we transform their report into a map and analyze it to unveil areas that are already well-covered and identify those needing expert input. This is similar to the approach of Moon and Browne, who disaggregated their complex system “into smaller subsystem modules that each represented a functioning unit, about which an individual is likely to have more comprehensive knowledge” [52]. Note that the notion of ‘subsystem modules’ or ‘sub-models’ is common in participatory systems mapping but these components are sometimes found after individual maps are combined [66]. Identifying the components before the interviews helps us distribute the knowledge elicitation across participants and ensure that we have sufficient participants for each facet of the problem.

Given the desired areas of expertise, we select stakeholders ➁ who can cover these areas based on either significant research expertise or relevant experience in policymaking. We do not require that participants live in Canada, particularly when it comes to medical expertise (e.g., how inflammation or cardiovascular diseases are caused). We do require knowledge of the Canadian setting for constructs that are place-specific, such as the built environment (e.g., the existence of food deserts). Stakeholders were primarily identified by a leading researcher (see acknowledgments) with additional referrals; this is the same approach as in past studies employing purposive sampling facilitated by the positioning of a researcher, supplemented by snowball sampling [67]. There is no ‘gold number’ on how many individuals should be invited or ultimately interviewed. For example, recent PM studies have interviewed 7 researchers [68], 11 [69] or 14 stakeholders [66], 15 subject-matter experts [70], or as many as 30 individuals [71]. With 19 participants, our project is on the higher end of the distribution for PM projects performing individual interviews.

All experts contacted agreed to participate and provide their names for full transparency. As shown in Appendix A, participants provided a breadth of expertise and significant experience on topics that directly contribute to our core topics. For example, physical well-being was addressed by our multiple experts on physical activity, as well as experts in human physiology. Mental well-being was covered through expertise in domains such as social determinants or healthy living. Per the terms of the IRB approval granted by Simon Fraser University (approved by the Office of Research Ethics under the application titled ‘From weight to well-being: assessing drivers and mechanisms’), we cannot associate the identity of a participant with any specific element (e.g., cannot attribute a quote or construct in the map). Every participant received the same IRB-approved email invitation for a Skype interview, including study details and the consent form. Participant interviews were recorded on Skype, including the provision of oral consent.

We then conducted individual semi-structured interviews ➂. Although a group setting would “enhance mutual learning and problem-solving” [53], mutual learning is not the purpose of our work since participants are selected because they are already experts with aspects of the obesity system. Rather than a social objective, our project was designed to support decision-making, policymaking, and management of the obesity system. In addition, PM research has noted that conflict across shareholders is a potential challenge [59]. Interviewing experts individually is part of how we avoid this conflict, since “one-on-one interviews also allow participants to express views they might be hesitant to express in a group setting” [39]. Each semi-structured interview is performed as detailed in [50] by teasing out causes and consequences. As noted by Sterling and colleagues when extracting a conceptual model from discussions, “modelers must strive to ensure their own views and favored methods do not drive the model development” [59]. Consequently, the interviewer avoids confirmatory questions that may plant a concept in the interviewee’s mind (e.g., ‘do you think that body image increases well-being?’), and instead uses open-ended questions (e.g., ‘what do you think contributes to increasing well-being?’) and summaries (e.g., ‘you stated that body image and physical activity increase well-being; is there anything else?’). The interviewer does not judge the participant and does not distract the participant by showing a map or notes. In line with previous research, “these interviews were then transcribed and coded for key terms. The relationships among the key terms elicited were then visually represented (by the research team) as a concept map depicting the causal linkages mentioned by the interviewee.” [72]. The individual maps resulting from semi-structured interviews are finally ➃ aggregated into a group-level map, which accounts for the variation in terminology. Since the participants provide complementary perspectives on different facets of obesity and well-being, we cannot ‘select’ one participant as a base map and add the others; rather, we focus on integrating them and preserving information. When this group-level map is ready, we document it in a report that includes a rationale for every node and edge and share the report with all participants for feedback and corrections. Soliciting feedback is also part of how we navigate the potential for conflict across stakeholders, as experts are then made aware of the content originating from other participants. Our process of sharing the map along with explanations for every construct is similar to the process by Swierad et al., in which the map was shared for review together with contextual narratives [73]; this goes beyond typical practices for PM studies, in which participants provide feedback by reviewing only a printed version of the map [74].

3.2. A Starting Point: The PHSA Report on Obesity and Well-Being

We manually coded all the relationships mentioned in the PHSA report. Each relationship was directional: a change in one factor triggers a change in another factor. For example, the sentence “bias and stigma have significant negative consequences, including overeating and avoidance of exercise and psychological harm” led to coding “weight stigma impacts overeating”, “weight stigma impacts exercise”, and “weight stigma impacts psychological harm”. Relationships that endorsed associations rather than causations were accounted for as two-directional relationships. For example, the sentence “severe obesity is associated with a lifetime history of binge eating” was coded as “obesity impacts binge eating” and “binge eating impacts obesity”. The set of all relationships constitutes a map, whose characteristics were explored using visual analytics. The maps are provided in Figure 2, Figure 3 and Figure 4, where each circle stands for a concept, and causations are depicted via curves (i.e., the presence of a curve linking two factors means that one causes the other). When the report provided association rather than causation, the map shows two curves denoting causations running both ways. The fact that relationships overwhelmingly depicted causations rather than associations can be seen in the figures, as two factors tend to be connected by a single curve rather than two curves. Note that these maps only aim to depict relationships; they do not state whether the relationships are positive, negative, or part of feedback loops.

In the analysis stage, we first assigned a category to each factor based on the categories used in the Foresight Obesity Report [33]. Coding new maps against the thematic clusters defined in the Foresight Obesity Report has indeed become common practice in PM studies on obesity [75]. In addition to the categories from the Foresight Obesity Report, we created the “Disease” category to account for the fact that numerous comorbidities were mentioned in the PHSA report. There are three main findings from this first analysis (Figure 2). First, it highlights that the PHSA report focused on the social and psychological contributors to obesity and well-being; while numerous comorbidities were cited, they mostly came from one subsection of the report. In comparison, physical activity, and exercise account for relatively few factors, as do food production and consumption. Second, using the categories from the Foresight Obesity Map does not result in coherent groupings of concepts (i.e., clusters); the map in Figure 2 cannot be straightforwardly divided into components based on the color of the nodes. This suggested that more specific categories needed to be developed in order to adequately capture the uniqueness of the PHSA report. Finally, this analysis found that the comorbidities (in brown on the left of Figure 2) were only listed as consequences of obesity and did not influence other factors. This is a major gap in the PHSA report given the wealth of evidence regarding the impact of comorbidities on well-being. For example, the PHSA report accounted for the association between obesity and high blood pressure (i.e., hypertension). However, it did not account for the multiple studies showing the impact of hypertension on people’s well-being, for example, in terms of health-related quality of life [76].

Having noticed that categories from the Foresight Obesity report lead to scattered coverage, the second phase of our analysis searched for categories tailored to the PHSA map. This is similar to previous analyses of systems maps, in which the same system was decomposed into different sets of categories to understand its dynamics in complementary ways [57]. To automatically find categories that fit the PHSA map, we assessed which modules were present in the map. A module is a list of factors that are ‘structurally coherent’; that is, these factors are mathematically more linked to one another than they are to factors in other modules. The results (Figure 3) suggest seven categories that better identify the specificities of the PHSA report. In particular, this points out the strengths of the report regarding its thorough examination of eating disorders and weight stigma. This analysis also uncovers new gaps in the report: the fact that concepts linked to nutrition are relatively isolated indicated the need to bring in more clinical evidence about how nutrition mediates the relationships between factors.

In the final phase, we examined which concepts played an important role in the map. This was conducted by considering two metrics: the degree, which is the number of relationships that a factor is involved in, and betweenness centrality, which is the extent to which a factor ends up on pathways between other factors. Degree and betweenness are the two node centrality measures commonly applied in the analysis of obesity maps [35]. In particular, previous work found a correlation ($R^2=0.51$) between degree centrality and body mass index [77]. The results show that obesity is the most central, followed by themes pertaining primarily to mental well-being (e.g., weight bias, weight bullying, self-esteem, body image); the prominence of such constructs and the lack of precise physiological or environmental notions reinforce the findings from the previous two analyses.

To summarize, our assessment of the PHSA report has helped clarify the strengths on which our own map can build, as well as the areas that need to be expanded. Both aspects are listed in

Table 1. Strengths of the previous report and areas emphasized/peripheral in our work.

Strengths of the PHSA Report	Areas of Emphasis in Our Work	Areas Peripheral to Our Work
Psycho-social pathways (e.g., consequences of weight stigma)	Clinical pathways (e.g., consequences of comorbidities, impact of nutrition)	Food production
Mental well-being	Physical well-being	Food consumption
Resources impacted by obesity (e.g., job opportunities)	Resources enabling a high level of physical well-being (e.g., the built environment)	Genetics

. Given our primary objective of bringing in solid clinical evidence, our analysis shows that we need to build on the PHSA report in three areas: clinical pathways (e.g., to capture the impact of comorbidities on obesity), physical well-being (e.g., to complement the previous report’s emphasis on mental well-being), and resources that enable a high level of physical well-being (e.g., the environment in which individuals live). While genetics are clearly involved in understanding the clinical components of obesity, they were not proposed as one of the key themes due to the context of supporting efficient policies for well-being. While food production and consumption are heavily influenced by policies, we continue to underemphasize these two components given our focus on physical well-being as it relates to overweight and obesity.

3.3. Extending the Map via Semi-Structured Interviews

As detailed in Section 3.1, our identification of weaknesses and strengths served to set the agenda for identifying participants. We initially identified a total of 54 experts based on recommendations from a leading expert for domains that need to be covered. In addition, 6 experts were identified through referrals. We did not invite all experts at once; rather, we gradually invited experts in each domain until sufficient coverage was obtained. Out of the 60 experts identified, we invited 26 based on our specific needs throughout the process. A total of 22 individuals agreed to participate (i.e., 4 did not return our email), out of whom, 19 set up an appointment time and were interviewed (i.e., 3 agreed to participate but did not return emails to set up a time). All interviewees are listed in

Table A1. List and fields of expertise for the key informants.

Name	Position	Fields of Expertise
Geoff Ball, Ph.D.	Professor and associate chair of research, Department of Pediatrics, Faculty of Medicine & Dentistry, University of Alberta, Canada	Optimize obesity management and prevention for children and families, including via clinical trials or qualitative research
Katherine Cianflone, Ph.D.	Professor Emeritus, former Canada Research Chair on Adipose Tissue (Tier 1), Universite Laval, Canada	Adipose tissue metabolism, factors controlling fat, molecular basis of obesity
Jean-Pierre Chanoine, Ph.D., MD	Clinical Professor, Pediatric Endocrinologist, Department of Pediatrics, BC Children’s Hospital, Canada. Secretary General of Global Pediatric Endocrinology and Diabetes (GPED)	Pediatric endocrinology, capacity building, access to medicine
Jean-Philippe Chaput, Ph.D.	Professor, Department of Pediatrics, University of Ottawa. Research Scientist, Healthy Active Living and Obesity Research Group CHEO Research Institute	Prevention and Treatment of Obesity in Children, Sleep health, Screen time, Physical activity
Jean-Pierre Després, Ph.D.	Professor, Department of Kinesiology, Faculty of Medicine, Universite Laval, Canada Scientific. Director, International Chair on Cardiometabolic Risk, Universite Laval. Innovation and Science Director, Alliance Santé Québec	Adipose tissue distribution, visceral obesity, type 2 diabetes, lipids, lipoproteins, cardiovascular disease, and their prevention through physical activity and healthy living
Jim Frankish, Ph.D.	Clinical Psychologist & Endowed Professor, School of Population and Public Health, UBC (passed away)	Nutrition education, health literacy, community capacity, healthy communities, and health promotion in primary care
Danijela Gasevic, MD, Ph.D.	Associate Professor and Head of Professional Education, School of Public Health and Preventive Medicine, Monash University, Australia	Chronic disease prevention, particularly regarding the effect of physical inactivity and sedentary behavior on health
Carolyn Gotay, Ph.D.	Professor Emeritus, founding Canadian Cancer Society Chair in Cancer Primary, School of Population and Public Health, UBC, Canada	Interventions to reduce modifiable cancer risk factors, quality of life in cancer patients and survivors
Michael Hayes, Ph.D.	Professor Emeritus and former Director, School of Public Health and Social Policy, University of Victoria, Canada	Health inequities, disability, public policy, obesity, health literacy, population health promotion
Terry Huang, Ph.D.	Distinguished Professor and Chair, Department of Health Policy and Management, City University of New York, USA	Chronic disease prevention, design and health, built environment, public-private partnerships, cross-cultural health
David Lau, Ph.D., MD	Professor Emeritus, Department of Biochemistry & Molecular Biology, University of Calgary, Canada. Former Chair of Diabetes & Endocrine Research Group and Director of the Julia McFarlane Diabetes Research Centre	Fat cell biology in health and obesity, development of insulin resistance in obesity and diabetes, and cellular mechanisms of diabetic vascular complications
Scott Lear, Ph.D.	Professor, Pfizer/Heart & Stroke Foundation Chair in Cardiovascular Prevention Research, Faculty of Health Sciences, Simon Fraser University, Canada	Cardiovascular disease prevention, population health, ethnic disparities
Gary Lewis, Ph.D., MD	Professor, Department of Medicine and Department of Physiology, University of Toronto. Director, Division of Endocrinology and Metabolism, University of Toronto	Whole body, integrative, physiological studies in humans
Pablo Monsivais, Ph.D.	Associate Professor, Department of Nutrition and Exercise Physiology, Elson S. Floyd College of Medicine, Washington State University, USA	Public Health, Epidemiology, Social Inequalities, Food and Nutrition
Kim Raine, Ph.D.	Distinguished Professor, School of Public Health, University of Alberta, Canada	How social conditions and people’s behaviors (particularly food and eating behaviors) interact to transmit obesity and chronic diseases through social means
Arya Sharma, Ph.D., MD	Professor of medicine, chair in obesity research and management, University of Alberta. Founder and Scientific Director, Canadian Obesity Network	Evidence-based prevention and management of obesity and related cardiovascular disorders
John Spence, Ph.D.	Professor, Faculty of Kinesiology, Sport, & Recreation, University of Alberta, Canada	Benefits and determinants of physical activity and how physical inactivity and sedentary behavior are related to health
Tom Warshawski, MD	Associate Clinical Professor of Pediatrics, UBC. Chair of the Childhood Obesity Foundation. Former head of pediatrics, Kelowna General Hospital, Canada	Promoting Healthy Active Living in children and youth
James Woodcock, Ph.D.	Professor of Transport and Health Modelling, University of Cambridge, UK	Health impacts of changes to how we travel and how such changes might occur

(Appendix A) with their permission.

A student facilitator was hired for this project. After the first round of selecting applicants based on grades and recommendation letters, applicants were evaluated based on the same set of project-specific questions, which included domain knowledge (e.g., ‘Can you cite three social determinants of obesity?’) and situations (e.g., ‘Imagine that you have to interview one of the most renowned experts of obesity world-wide. How would you prepare?’). Once hired, the facilitator was trained on running semi-structured interviews and practiced on three individuals, who were not among participants but had a sufficient level of awareness on the application field as Ph.D. holders and published scholars in obesity research. After training, semi-structured interviews with participants began. An appointment at a time of their convenience was set for an hour. In line with previous guidance on semi-structured interviews for causal mapping [57], an interview “continues until a sufficient level of depth is reached or until the insights of the interviewee are exhausted, at which point the chain of causality may be built up further during interviews with other participants”. On average, each recorded interview lasted for 40 min, with most taking 27 to 37 min and a few lasting an hour. These durations are similar to those encountered in other works on mapping complex systems, such as suicide [70].

We employed a professional transcription service to turn each recording into a text document. On average, one transcript contained $4592\pm 1452$ words, leading to a total of almost 90,000 words. We manually extracted a map from each transcribed interview. Although there is emerging technology to automatically derive maps from text [78], PM studies primarily use software to help with the manual process of annotating (e.g., nVivo) rather than replacing it entirely with algorithms. We used a dedicated setup to manually combine the maps into one (Figure 5). In this case, emerging technology for combining maps does not offer the same level of accuracy as manual processing; hence, it remains the gold standard [62]. The contribution of participants to the combined map reflects their areas of expertise. For example, an expert on health literacy and health inequities would not contribute to identifying and linking physiological constructs such as adipose tissue metabolism and lipoproteins. As expected, there are substantial differences regarding the parts of the map that are shaped by individual experts.

4. Analyzing Knowledge on Obesity and Well-Being

4.1. Overview of the Map

Our final map is available on the third-party repository of the Open Science Framework [79] at https://osf.io/7ztwu/, under ‘sample maps’. Our map is named Giabbanelli & Macewan; other maps on obesity are provided for comparison and include the Foresight Obesity Map as well as the work of Verigin et al. [80]. A criticism of participatory modeling approaches in health is that “model assumptions and structures tend not to be opened up to scrutiny” [81]. Given this specific concern, the assumptions and structure of the many components of the map are detailed in Appendix B; this appendix was also shared with participants as part of gathering feedback. The map consists of 98 concept nodes and 174 directed–typed relationships. The size of our map is about average for a systems map related to obesity, between smaller maps obtained for childhood obesity in South Africa (47 factors) [74] or Manhattan’s Chinatown (39 factors) [73], and larger maps such as the Foresight map (108 factors) [33] of the work of McGlashan et al. (114 factors) [75].

This section focuses on the analysis of the map. Although network analysis (as shown for the PHSA report in Section 3.2) is the common approach employed in systems maps [82], our work goes one step further by also employing natural language processing to examine the transcribed audio recordings of stakeholder interviews. A manual content analysis of the interviews is fairly common in PM studies [41]. In studies, such as the study by Radonic or LaMere et al., recordings served as a form of indirect elicitation to verify that the map had the right content and level of details [83,84]. In other works, such as Kropf et al. [85], an aggregated map was constructed from a workshop, and then semi-structured interviews were performed, with content analysis informing refinements of the aggregated map. Our use of NLP has similar objectives but promotes a computational approach that scales up with the data volume of almost 90,000 words.

4.2. Network Analysis of the Conceptual Map

We used the conceptual map to compare the perspectives around well-being with those around obesity (Figure 6). We found that different sets of factors were involved in these two perspectives (

Table 2. Examples of factors that connect to either obesity or well-being.

	Obesity	Well-Being
Causes	Medications	Perceived environmental safety
	Overeating	Presence of a vibrant community
	Physical activity	Resilience
	Diabetes	Ability to engage in physical activity
Consequences	Short sleep duration	Antipsychotics
	Cancer	Medications
	Dysfunctional adipose tissue
	Weight bias

). Two specific differences were noted between these perspectives. First, obesity evokes a medical focus on weight as a problem, whereas well-being leads to a more solution-focused approach that is broad and diverse.

We performed an analysis to check these findings by creating a reduced version of our conceptual map (Figure 7). The idea of reducing a map of obesity based on its communities was pioneered by Finegood [86] and then used in studies by other scholars [75]. Examining the map at a higher level of abstraction allows us to observe patterns between themes, which may not be apparent when working at the level of specific concepts. Indeed, “there is growing evidence of the effectiveness of applying multilevel network methods in diverse fields to identify meaningful patterns of interdependencies across multiple levels of a system, relative to what can be understood via simplified, single layer network analyses.” [87]. To obtain a reduced version, factors were gathered by categories, and a relationship existed between two categories if and only if there was a relationship between factors in these categories. For example, a relationship existed in our conceptual map between socioeconomic status and mental well-being; hence, there is a relationship in the reduced map between the categories of determinants and well-being. Note that there is no gold standard on the number of categories that should be used. From a small number of PM studies using reduced maps, it appears that the number of categories may scale with the number of factors. A map of the obesogenic environment for children in Pennsylvania had 3 clusters for 32 features [77], one on obesity in Manhattan’s Chinatown had 4 themes for 39 factors [73], and another in South Africa had 5 domains for 47 factors [74]. In a different domain, a map for HIV in sub-Saharan Africa had 4 or 5 domains for 71 nodes. Since our map is larger, we decomposed it into more domains accordingly.

This reduced map shows that well-being is directly connected to all other categories. From an obesity-centric perspective, food production, and social factors are more indirect and distal. That is, obesity is certainly impacted by food production but this was viewed through the lens of food consumption. Consequently, this reduced map confirms the focused nature of an obesity-centric perspective compared to the more diverse themes produced by a well-being-centric perspective.

4.3. Natural Language Processing for the Interviews

We analyzed the interviews to compare the language used when discussing obesity compared to well-being. The unit of analysis was at the level of a respondent’s answer. That is, each interview was divided into a set of answers. These answers were then categorized into two sets: answers that included obesity and answers that included well-being. Note that these sets are not exclusive since an answer may speak of both. In these two sets, standard text analysis procedures were used to transform the words. Specifically, words were stemmed (e.g., “problems” and “problem” are seen as the same) and common English words (e.g., “and”, “or”, “that”) were removed. Then, we counted the frequency of each word in each set. Comparing the frequencies (Figure 8 and Figure 9) reiterates the finding that obesity emphasizes the problem from a medical perspective, whereas well-being triggers a more diverse solution-oriented language.

5. Discussion and Future Work

5.1. Structuring the Relations between Physical and Well-Being in the Context of Obesity

The representation and analysis of knowledge are fundamental themes within artificial intelligence (AI). In this paper, we used AI through the lens of participatory modeling (PM), network analysis, and natural language processing (NLP) to examine the tension between different views on obesity and well-being. We started with a previous report as a point of departure, where mental well-being was emphasized. Our work disentangles the different components of well-being by accounting for both physical and mental well-being. Interviewees overwhelmingly reported that these two facets are closely interrelated:

“You don’t have two separate drawers, one is mental health, one is physical health. They’re really part of the same thing.”

However, most of the interviewees hinted at a temporal separation between these two concepts. Factors can first influence physical health, which in turn impacts mental health, or vice versa.

Table 3. Factors that affect mental and physical well-being (listed by interviewees).

Mental Well-Being	Physical Well-Being	Both
Bias, bullying, culture of eating that promotes healthy choices instead of weight loss, deep relationships, discrimination, feeling able to contribute, feeling comfortable, feeling valued, happiness, physical activity, presence of a vibrant community, psychological stability, self-confidence, stigma	Ability to be physically active, appetite, broken bones, built environment, cardiovascular health, diabetes risk, eating behavior, energy balance, food quality, hormonal systems, metabolic health, nutrition, overweight/obesity, physical activity, sleep	Availability of healthy food options, built environment (pollution level, aesthetics, infrastructure), community ties, the culture of eating healthy food, eating behavior, economy, exercise, family (where kids go, how they arrive there, what they eat, how they spend their time), income, perceptions of neighborhood safety, political climate, public health messaging, school environment, work environment

summarizes the factors that interviewees thought would influence physical or mental well-being first, and the factors that would influence both at the same time. When examining how mental and physical well-being affect each other, interviewees were more likely to provide examples of mental well-being influencing physical well-being than the other way around. This may be due to a cognitive bias: if there are several intermediate factors mediating the impact of physical well-being on mental well-being, then this impact may feel indirect and interviewees can be less aware of its existence. It is also possible that framing the discussion around the components of well-being may have influenced interviewees’ responses toward mental well-being since physical well-being is commonly referred to as physical health. Furthermore, prompting interviewees to think about the differences between physical and mental well-being could have biased their answers toward an assumption that there is a difference. However, this does not seem to be the case. For example, an interviewee began by saying that “the difference is huge” but, like all other interviewees, eventually concluded that the two are interrelated: “I believe that they can’t go, they shouldn’t be looked in isolation”.

Two interviewees raised the importance of including other dimensions of well-being. In particular, they raised the need to take into account social well-being. However, they defined this concept differently. One defined social well-being as having the resources to play a role in society:

“That would entail having access to the resources that are required for everyday living. Access to shelter, to clothing, to reasonable foodstuffs, to entertainment, meet entertainment needs so that people could participate meaningfully in the society or the culture in which they find themselves”.

The other interviewee viewed social well-being more in terms of social capital, in the context of close family ties or community ties. Overall, this suggests that future studies could expand on well-being by going beyond the physical and mental aspects that were considered here.

5.2. Limitations and Suggestions for Future Studies

Our participants were chosen for their expertise either in obesity research or in obesity policymaking, and participants were gradually added to the study to cover domains that had to be expanded from previous works. Despite these strengths, every study that uses a participatory modeling approach to synthesize knowledge on a subject is necessarily limited to reflecting the perspectives of its participants. Complementary perspectives would be offered by directly interviewing individuals with different lived experiences of overweight and obesity. Previous studies have indeed shown that the mental models of individuals directly concerned with a problem could differ from those who are experts in the management of this problem, thereby resulting in different maps [88]. Our previous studies further showed that individuals have very heterogeneous perspectives on obesity [89]; hence, accounting for lived experiences is unlikely to result in just ‘an expert map’ and a single ‘map of lived experiences’; rather, several maps may be produced by identifying shared features [63] or trajectories among lived experiences.

Our network analysis was guided by the focus of our research question on a proposed paradigm shift from obesity to well-being. The publicly released map that we have constructed could be analyzed in numerous other ways, in response to other specific questions. For instance, network analysis often serves to identify “critical components within an interconnected system” [90], thus providing insight for public health in terms of setting intervention targets. Thus, our map could be analyzed with respect to several node centrality algorithms since hundreds of them are now available and applicable to sparse networks such as systems maps [91]. There have also been manuscripts dedicated to finding feedback loops in an obesity map [67], or identifying data gaps in obesity research [92]. Using the analysis of a map as a proxy for designing interventions on a complex system has been a growing trend, as scholars noted in 2019 that these “whole-of-system interventions represent the next step in the science of obesity prevention” [87]. While our focus is on obesity and well-being, we note that obesity research is an interdisciplinary field and similar mapping efforts have been undertaken in related areas such as physical activity or nutrition. Indeed, there is also growing literature in nutrition that conceptualizes and discusses systems components or use-system-oriented frameworks to assist stakeholders with design and evaluation [93]. Such works illustrate that the practice in PM and obesity has often been to detail the process of knowledge representation and publicly release a map (as we have done here), such that other scholars can benefit from it.

Similarly, our use of natural language processing sought to shine a light on the differences between the language used in obesity and well-being-centric perspectives. A variety of other NLP analyses are available and we have employed them elsewhere [94], such as sentiment analysis, topic modeling, or sarcasm detection. Many of these analyses have been conducted regarding obesity in social media, for instance, to study stigma in public opinion [95] or the relationships between health and place [96]. Since our expert sample would not be suitable for studies of public opinion or geography, these NLP techniques would not be directly applicable here. However, future works may use the content of our map to guide data collection efforts in social media, thus revealing how the general population thinks about various facets of the obesity system [97] and how interventions may be framed for public communication [98].

Through the interviews, experts expressed that relationships differed in strength. For example, it was said that pain is “certainly one of the most important” biomedical barriers to physical activity, or that having harmonious relationships with neighbors “would be really important” in shaping people’s perceptions of their neighborhood. While the map presented here only articulates the relationships between numerous factors, including the strength of these relationships would allow practitioners to utilize the map for “what-if” scenarios. For example, practitioners could ask what would happen if an intervention were to be implemented on weight bias: the weighted map would then show how changes in weight bias would diffuse through the whole system. Consequently, a weighted map could serve as a decision support tool that is particularly valuable when facing the complexity of a phenomenon such as obesity and well-being. Previous endeavors have produced weighted maps for obesity [51,80], but they have always been smaller (i.e., fewer factors); hence, there is still a need for a systems map that precisely characterizes each relationship.

5.3. Implications of the Map for AI Solutions Focused on Well-Being

Numerous solutions seek to leverage advances in AI to improve well-being in the context of obesity [18]. Although the goals of these solutions may vary (e.g., prevention vs. treatment), many of them focus on constructs that are proximal to individuals [19,20,21,22]. This emphasis on individual responsibility is perhaps expected for self-management tools, and it is also found in other complex health problems [50]. However, the social–ecological framework has made it clear that well-being is not only a matter of individual determinants, it is also about how individuals engage with their peers, the communities, and environments in which they live, and the norms conveyed in society. Few AI solutions recognize that individual behaviors are shaped by their environment [27]. In this context, our map can help to avoid an exceedingly reductionist perspective. By serving as a knowledge repository, the map can, thus, guide the design of new tools. For example, apps for self-management can continue with asking questions about the individual (e.g., patterns pertaining to eating and exercising, sleeping patterns, self-esteem), and they can now follow up on such questions either within a level (e.g., to identify forms of exercise that are appropriate for an individual’s capacity and improve their health) or across levels (e.g., to assess whether low self-esteem is the product of a broader issue in interactions with peers or societal norms). The potential for the map to improve knowledge representation and reasoning related to well-being and obesity is also valuable for tools building on emerging technologies, such as large language models [24,25,26]. As shown in recent case studies [99], models such as GPT do not inherently encode causality related to obesity; hence, they need to be fine-tuned. The content of our map can, thus, fine-tune GPT, providing it with the knowledge of domain experts before using it to guide individuals in a holistic manner.

6. Conclusions

Obesity-centric and well-being-centric perspectives offer two different paradigms through which to frame public policy interventions. We created and analyzed a systems map of obesity and well-being, as well as analyzed key-informant interviews with experts in the field. We found several key differences between focusing on obesity versus focusing on well-being. A well-being-centric perspective tended to be more comprehensive and included all factors identified by experts as pertinent to the topic, whereas an obesity-centric perspective was less connected to several factors, notably food production and social factors. Furthermore, discussions of the well-being-centric perspective tended to generate solution-focused language, whereas the obesity-centric perspective generated problem-focused language. In conclusion, this research demonstrated that the paradigm through which weight is viewed can impact which factors are considered relevant to the problem and the type of solutions that may be proposed.