DOI: 10.31038/EDMJ.2023734

Abstract

101 older female respondents evaluated different sets of messages (vignettes) dealing with diabetes, obesity, and how to deal with them. Regression analysis related the ratings to the presence/absence of the messages. Most messages were perceived as appropriate to the respondent, suggesting a positive blandness and ineffectiveness of messaging. Clustering the respondents by the pattern of coefficients revealed three different mind-sets of approximately equal size, respectively; Focus on medical indicators, Focus on lifestyle, and Focus on drinking water. A mind-set assigner comprising six questions was developed to identify new people, who can then be messaged more convincingly based upon the mind-set to which they are assigned.

Introduction

It is clear that the rampant increase in the obesity followed by diabetes is becoming the most threatening problem to world health. Our lifestyle, the foods that we eat, the reduction in exercise, and the more sedentary behavior has changed affected our health. Overweight is rampant world-wide, and diabetes is not far behind. Both the popular press and the scientific literature abound with material on the spread of obesity [1-3].

There are drugs on the market which reduce intake [4], and a whole industry of dieticians who, for fee, can personally monitor a person, suggesting what to do in order for the person to be heathier. For those who can afford it, both financially and physically, there are retreats [5], dietetic coaches [6], new ways of reducing weight such as the buddy system for dieting [7], and of course bariatric surgery [8]. It may be possible to reduce the deleterious effects of being overweight through the judicious application of the arsenal of tools.

The issue addressed by this paper is whether it might be possible to enhance the messaging to those who are diabetic using the newly emerging technology of AI, artificial intelligence, coupled with a traditional research approach, conjoint measurement. The objective of the effort is to see what can be done in the space of 24 hours, using available technologies, with the goal of creating a systematized, industrial-sale system which teaches how to communicate health issues with people.

The origins of this paper come from three different sources.

1. The use of personal computers and the internet to make consumer research fast, easy, and inexpensive. The approach used here, Mind Genomics, traces its history over a 30+ year period from the early 1990’s. The goal was to take a well-accepted research techniques in consumer science, conjoint measurement, and port it over to a DIY, do-it-yourself system. Summarizing Mind Genomics, one can think of presenting combinations of messages, elements, to a person, getting the person to react to these combinations, and then determining what specific feature or element drives the rating. The objective here is to simulate nature, which presents people with combinations, scenarios, to which the person reacts. Responding to mixtures is more nature, more ‘ecologically valid’ than responding to single ideas.
2. The desire to create better messaging for the world of medicine. Messaging is the way that people can be encouraged to comply. Messaging is the way that the doctor can tell the patient what the problem is, and what may need to be done, as well as the specific words to say it. One might think that with the advances of medicine the messaging will take care of itself, but the reality is that often messaging is the ‘last mile,’ and fails the patient. The messaging may be relevant but motivating to the respondent. Or, the messaging may be incorrect, or even both. Years of experience can fine tune the messaging, but what happens when the person is a new graduate, and dozen have years.
3. The ongoing effort to create a ‘wiki of the mind’, a systematized database of how people respond to the different aspects of everyday life. The idea is to explore the world of the everyday in a systematized manner, almost in the manner of a cartographer of the mind. The approach, called grounded research theory (REF), stands in contrast to the more popular method of hypothetical deductive reasoning, which posits how the world works, sets up a study to confirm or disconfirm, and then proceeds to implementing the experiment. The systematized exploration of the everyday experience does not lend itself to an explanation of how the world works; but rather just ‘what’s actually going on.’

Beyond Simple Questionnaires to Probing the Mind with Mind Genomics

The conventional way of understanding people’s thinking is through simple questionnaires. We are inundated with questionnaires, the modern era suffused with the desire to gain feedback on every interaction with a person and a service provider. Whereas previously questionnaires were often long documents with the respondent asked to score different facets of life, or an experience, today’s questionnaires are short, limited to a topic of one’s experience. Whether one works with a short questionnaire after an interaction, or with a long A&U (attitude and usage) questionnaire, the objective is the same, namely understand at an intellectual level how a person feels about a topic. These questionnaires parse the experience or topic into different, cleanly and surgically divided sections, instructing the respondent to think of each topic or each question, one topic at a time, and in isolated thought rate aspects of that topic. Along with the effort to administer these questionnaires there is an accompanying cadre of analytics, mainly summarizations of the data, data reduction principal component factor analysis [9], and clustering [10], to paint a picture.

The Evolution and Contribution of Mind Genomics to the Issue of Understanding and Messaging

The emerging science of Mind Genomics can be summarized as an experimenting science which understands how people respond to the ordinary world, such understanding promoted by the evaluation of systematically created vignettes of daily life, and the deconstruction of responses to those vignettes into the ‘driving power’ of the components of the vignettes, the ‘elements’. The description of Mind Genomics in this way hints at Mind Genomics as an experimenting science, rather than an observational science. It is the experimentation with features of the ‘ordinary’ which enables the Mind Genomics researcher to craft a new understanding of how people think when they are exposed to the world of ordinary life. In a Mind Genomics study there is no need to alter reality to establish a principle. Rather, the altered realities are simply combinations of features of the everyday, recombined into simple combinations that are evaluated, the pattern of responses showing how ‘nature is working’.

Mind Genomics plays a role in understanding the messaging about weight and diabetes because the messages are relevant to the topic. As will be shown below, the different combinations of messages end up judged in different ways by a person. The carefully created set of combinations of messages, the vignettes, put together through experimental design, present slightly different ‘realities’ to the respondent. It is the pattern of response to these different realities which allows for an immediate understanding of how people react to the messaging. A key benefit of the Mind Genomics approach is its ability to prevent a respondent from ‘gaming’ the system. Thus, Mind Genomics avoids biases which plague many survey methods, especially those dealing with sensitive topics [11,12].

The original work of Mind Genomics dealt with commercial products [13]. These studies were inspired by the pioneering work of the late Professor Paul Green of the Wharton School, University of Pennsylvania [14]. It was Green and his colleagues, especially Yoram Wind, who took the rather esoteric method of conjoint measurement, and, simplifying the notion of understanding ideas by studies of mixtures, brought conjoint measurement into common use. Mind Genomics made the system even simpler and more robust by having the respondents each evaluate different sets of combinations of ideas, each set created to allow for subsequent statistical analysis at the level of the individual [15]. The approach evolved to a DIY (do it yourself) system, with automated analysis and reporting [16,17]. The approach presented here is an example of that latest evolution. A history of Mind Genomics can be found in a variety of published papers [13,18].

Applying Mind Genomics to the Study of How to Communicate Weight Control

The remainder of this paper is devoted to an explication of the Mind Genomics approach to uncovering what to say to people to encourage weight control. The Mind Genomics approach follows a series of templated steps, so that the discoveries presented in this paper end up being simply empirically ‘fleshed out’ data tables and figures. That is, standardized approach to Mind Genomics studies provides the researcher with a tool that can be applied quickly and usually productively to a problem.

Step 1 – Create a name for the study, or really for the experiment. This first step may seem obvious, but the reality of research is that the novice researcher often fails to crystallize the reason for the study, and the topic. Rather, the novice attempts to put into the title the entire research project, rather than separating out the general topic from the specific method. The study here was simply called Diabetes Weight, a name which allowed the researchers to focus on different ways to think about the topic.

Step 2 – Develop four questions pertaining to the topic. The actual test stimuli for Mind Genomics comprise four sets of four phrases each. In turn, each set of four phrases represent four alterative ‘ideas’, or messages about the topic. The structure of four questions allows the researcher to answer each question. It will be the answers to the questions which constitute the test material that the respondent will evaluate. Thus, the structure of the question/answer is a way to focus the mind of the researcher, with the questions serving as an aid to thinking, and as a bookkeeping device.

It is at the stage of creating four questions that many prospective researcher find the task to be difficult. Our education does not teach people to think critically, focusing as it does on answering questions rather than developing them. It should come as no surprise that at this early stage in the Mind Genomics process many prospective researchers become frustrated and give up. It is a tribute to grade school and high students that they find this challenge to be fun, rather than frustrating, and proceed to come up with questions far more readily than do older people, and even far more readily than professionals.

In order to address the long-standing issue of ‘creating the raw material’, viz., questions but also answers, recent versions of the Mind Genomics platform have incorporated artificial intelligence, AI, as a TACT, Technical Aide to Creative Thought. That term was first used almost 60 years ago by the late Professors Anthony Gervin Oettinger at Harvard University, but it is appropriate here. The researcher need not come up with the actual questions, but instead must come up with a ‘squib’ or short paragraph about the topic. It is that ‘squib’ which will become the prompt for embedded AI, GPT3.5 [19].

Figure 1 shows screenshots which allow the researcher either to put in her or his questions, or to request that the AI embedded in “Idea Coach” suggest questions. Not surprising is the observation that having a ‘coach’ to help with the questions changes the set-up, so that it becomes intriguing and fun, rather than frightening. Panel A shows the set of four placeholders in which the researcher is to type a question. Panel A is stark, forbidding, a tabula rasa, a blank sheet, with little to support the researcher. The researcher who wants to use AI need only press the Idea Coach button, and be quickly guided away to a safer place, the box in Panel B. Of course, the researcher must talk about the project, but as we will see, the Idea Coach is forgiving, letting the researcher iterate.

Figure 1: The completed set of four questions, after being edited and polished by the researcher

Once the researcher has composed the squib and put it into Figure 1, Panel B, the Idea Coach returns with a proposed set of questions. Each iteration of the Idea Coach generates 15 different questions. The researcher can iterate as many times as desired, changing the squib or keeping the squib the same. The AI will return a number of new questions each time, and occasionally repeat a question. At the end of the iteration the research will have developed four questions and can edit them. Figure 2 shows the completed set of four questions after they have been developed by Idea Coach, and ‘edited’ in preparation for the answers.

Figure 2: The completed set of four questions, after being edited and polished by the researcher

After the squib is constructed and put into Figure 1, Panel B, Idea Coach creates a set of 15 questions from a squib in approximately 10-20 seconds. The researcher can select a question or several questions, dropping them into the study, and editing them. When the researcher does not find a question to drop into the study it is easy to re-run the Idea Coach, either with the same squib or with a revised squib. The process can go on several times, until the researcher has uncovered four questions, using Idea Coach and/or developing the question(s) oneself. Idea Coach truly becomes a coaching tool after the researcher gains experience.

At the end of the development of questions, and later of answers, the Mind Genomics platform stores each of the iterations of 15 questions (and 15 answers) on a separate Excel tabulation. As the study is being closed, the platform subjects each set of 15 questions (or answers) to an extensive summarization, using AI, with specific prompts. Table 1 presents the AI summarization of the first set of 15 questions.

Table 1: First results from the AI-powered Idea Coach, attempting to develop questions from the squib

Once the researcher has completed the set-up of the questions, the next step is to set up the four answers to each question. In this case the researcher must edit the questions in order to ensure that the answers are short and understandable. Such editing occurs as the researcher completes the creation of the four questions. It becomes a simple matter to add a phrase to the question as part of the prompt to the AI-driven Idea Coach, that prompt guiding the style of the answer provided. Table 2 shows the first set of 15 answers to Question #1 after the question has been edited by the researcher.

Table 2: First results from the AI-powered Idea Coach, attempting to develop 15 answers from the first question, that question edited to direct the style and comprehension level of the answers.

As we finish this section of the set-up it is well to keep in mind that the incorporation of the AI into the creation of questions and answers ends up producing an ‘Idea Book.’ This Excel file becomes a rich source of information about the topic, summarized by key issues and questions. The Idea Book moves from the list of questions, itself valuable for the researcher, onto issues of points of view, suggestions of what’s missing, and even suggestions about innovations. As noted above, the entire process of creating each logical page should be no more than a minute, with the creation of a 20-page book on the topic requiring less than 20 minutes.

Table 3 shows the final set of questions and answers. The answers have been edited by the researcher in order to be more readable when they are presented to the respondents in various combinations of answers. Henceforth, the answers will be referred to as ‘elements.’ The questions themselves will not play any role in the actual evaluation by respondents, nor in the analysis of the results. Rather, the questions are simply used to allow the researcher or Idea Coach to create a set of related but different ‘elements’, viz., different answers.

Table 3: The final questions and elements (answers to the question), after editing by the researcher

Setting Up the Mind Genomics Study

Mind Genomics studies comprise several parts, all templated. Table 4 shows these parts. Table 4 is provided as a standard output of every Mind Genomics project:

1. Key words selected by the researcher in order to make searching for the study results easier in the database of Mind Genomics studies.
2. Respondent orientation
3. Self-profiling questions. Gender and age are standard questions in every study, and so are not shown in Table 4.
4. The rating question and the different answers.

Table 4: Study information

The Respondent Experience in the Mind Genomics Study

Once the study set up has been completed it is the task of the researcher to provide respondents. With the advent of the Internet, it has become increasingly easy to obtain respondents, although at a slight cost to the researcher. It is the business of companies in the market research ‘space’ to provide motivated respondents for the many hundreds of thousands, perhaps millions or dozens of millions of studies on the internet. Across the entire world there have emerged companies which, for a fee, offer their members the chance to participate in on-line studies. It is from one of these companies, Luc.id Inc., in Louisiana, that the respondents are obtained, with the respondents satisfying specific criteria: Age 45 or older, Lives in one of three states: New York, New Jersey or Connecticut.

Luc.id Inc. is set up to provide these individuals, doing so with the actual experience taking approximately five minutes for a respondent. The respondents are recompensed by Luc.id, ad are totally anonymized. The respondent opts in to participate in the particular study. In turn, the research guarantees not t accept any specific identifying information during the course of the study, In situations where the researcher wants specific information, that information is requested in the ‘open ended’ questions, with a disclaimer that the information offered is entirely optional and left to the respondent.

Luc.id sent out email invitations to the prospective respondents, doing so in ‘waves’ until the quota of 101 respondents was filled. The quota comprised females, ages 45-70, living in New York, New Jersey or Connecticut, respectively. For easy to find respondents, such as those participating in this Mind Genomics study, the quota of 101 individuals is typically filled within 90 minutes. The Mind Genomics platform orients the respondents, presents the test materials, acquires the ratings, creates a database, performs the relevant analysis, and then creates the report along with AI summarization, generally within 15-30 minutes after the completion of the field work.

The actual Mind Genomics experiment begins with a short ‘hello’, and proceeds to the self-profiling classification shown in Figure 3, Panel A. The self-profiling classificaiton presents a clean screen to the respondent. The alternative answers to each question appear when the respondent reaches that question. The happy consequence of this pull-down screen is that the respondent is made to feel comfortable, rather than being confronted by a wall of words.

Figure 3: The respondent experience in the Mind Genomics experiment. Panel A shows the pull-down menu for the self-profiling classification. Panel B shows an example of a vignette, accompanied by a short introduction and the rating scale, both at the top.

Afterwards, the respondent is exposed to 24 different vignettes, similar to that shown in Figure 3, Panel B. The vignettes are shown one vignette after another. The respondent is told simply to rate the combination. Exit interviews with respondents over the years as well as observations of colleagues participating in the Mind Genomics interview continue to point to the fact that professionals try to ‘outsmart’ the system, whereas ordinary respondents simply fall into a relaxed mode, and end up saying that they ‘guess.’ The results below will show that the respondents are performing quite well, and that despite their statement that the feel they are ‘guessing’, the opposite is true. The typical respondent, the non-professional, ends up realizing that it is impossible to ‘game the system’, and settle for a relaxed experience. This relaxed, almost non-involved evaluation is felt to more validly represent the real judgment of the respondent, the judgment made when no one seems to be looking.

Creating the Database and Transforming the Ratings for Subsequent Statistical Analyses

The Mind Genomics study presents vignettes to the respondent, and acquires both the rating on the two-sided scale, as well as the response time. The response time is defined as the elapsed time between the presentation of the vignette and the respondent’s selection of the rating, Figure 4 shows the three vignettes from respondent (participant) #27. As one might surmise, there are 2424 vignettes in total, each vignette different in composition from every other vignette, accompanied by the rating and the response time. When the response time was longer than 9 seconds, BimiLeap automatically made the response time 9 seconds. The reason for this seemingly arbitrary rule is that the respondents were unsupervised and might have been doing other things. Thus, it is important to truncate the range of response times in order to avoid accepting response times of say 36 seconds. Clearly that response time reflects other behaviors besides the respondent participation in the study.

Figure 4: Example of input information to the database showing the respondent, the text of the vignette, the rating assigned to the vignette, and the number of seconds elapsing between the presentation of the vignette and the assignment of a rating.

Relating the Presence/Absence of the Elements

The essence of Mind Genomics is relating the presence/absence of the elements in the vignettes to the ratings. The vignettes themselves vary from respondent to respondent, and in actuality are only vehicles in which to embed the 16 elements. It is the elements which convey the actual information. In turn, the rating scale is the device by which the respondent can communicate feelings. Both the vignettes and the rating scale, however, need to be translated into a set of variables that can be analyzed by statistics.

Figure 5 shows the part of the database prepared for the subsequent statistical analysis. Figure 5 divides into three distinct panels.

Figure 5: The database for vignettes 1-11 presented to and rated by Panelist (respondent) #1

Panel A shows the row number of the database, and then information about the respondent, as provided by the respondent.

Panel B shows the order of testing (1-24) for the specific vignette, and then the deconstruction of the composition of the vignette into a series of 1’ and 0’s. Recall that the vignettes comprise combinations of elements, done so according to an experimental design. The particular design used here, the permuted experimental design, uses one basic structure, but changes the specific combinations through a permutation scheme [15]. The benefit is that each respondent’s data can be analyzed as either part of a group, or analyzed separately, one respondent at a time. The transformation of the experimental design is done by so-called ‘dummy variable’ coding [20]. The coding creates 16 columns, one for each element. When the element is present tin the vignette the cell has the value ‘1’. When the cell is absent from the vignette the cell has the value ‘0’. The 1’s and 0’s are called dummy variables because nothing is known about these variables other than presence or absence. Any analysis with these dummy variables simply shows the contribution to a criterion variable which occurs when the element is placed into the vignette. The ‘why’ is unknown. Only the ‘what’ is known Panel C shows the response time (time elapsed between the appearance of the vignette on the screen and the response), as well as the five-point rating assigned to the vignette. The remaining columns show the transformations of the ratings into the five single responses, the key being R5x (For me, Easy to do), and then two ‘combined’ responses R54 (For me), R52 (Easy to do).

R54x=100 when the rating is either 5 or 4, R54=0 when the rating is 3, 2 or 1

R52x=100 when the rating is either 5 or 2, R52=0 when the rating is 5,3 or 1

Once the database has been constructed, the Mind Genomics platform creates a set of simple linear equations, viz., linear models, first for the total panel, and then for each self-defined subgroup, self-defined age grouping, and finally gender. The analysis is called linear regression [21], with the variables taking on only one of two levels, 0 for absent, 1 for present, so-called dummy variables.

The equation for each binary variable is expressed schematically as: Binary Dependent variable=k₁(A1) + k₂(A2) … k₁₆(D4). The equation just presented shows how each of the 16 elements ‘drives’ the binary rating. Typically, the standard error is about 7-8 for these studies, suggesting that coefficients around 16 or higher should be significant in a statistical sense.

The Mind Genomics platform ends up providing a great deal of data, almost a ‘wall of numbers.’ In order to identify underlying patterns, the convention is to shade the so-called ‘strong performer.’ ‘Strong’ means meaningful, based upon insights gleaned from previous Mind Genomics experiments. We move beyond the conventional criterion of ‘statistically significant’ to the following:

For the binary transformed variable, R54, typically the key ‘evaluative variable, such as ‘for me’ or ‘I will buy’, we shade coefficients of 21 or higher.

For all other binary transformed variables, e.g., R52, R5, etc., we shade coefficients of 16 or higher.

For RT, response time, we shade coefficients of 1.5 or higher, the foregoing criteria are not ‘fixed in stone.’ Rather, they reflect simple heuristics which allow the researcher to identify emergent themes. Indeed, the search for themes is so important in Mind Genomics that an alternative heuristic might be considered, namely deleting all coefficients which are lower than the cut-off criterion. That alternative ends up typically creating a sparse table, since in conventional Mind Genomics the coefficients are usually far lower than what we see here.

Table 5 shows the coefficients for the linear equations relating the presence/absence of the 16 elements to the binary transformed variables of R5x, R54x, and R52x, respectively, along with the coefficients for the RT (response time) equation. Ordinarily, R5x (For me and Easy) would not appear in the table. Rather, R54 (For Me) would become the focus of the analysis. Table 5 shows the surprising finding that for this study on obesity and diabetes virtually all of the elements (15 of 16) are strong performers, with coefficients of 21 or higher. In light of this remarkably strong performance of virtually all elements we move the key evaluative variable to R5x, (For me and Easy), which shows the more typical pattern observed in the results for other topics, not medical ones but rather products and services.

Despite the strong performance of many elements, Table 5 does not give up secrets about the underlying patterns of strong performing elements. For example, three of the four strongest performing elements for R5x deal with water and drinking, but drinking water must be associated with a reason, not just be present by itself. When present by itself, without any outcome, however, drinking water performs poorly.

Table 5: Coefficients for important transformed binary variables, and for response time. The elements are sorted in descending order by the coefficient for R5x.

Fact: Drinking water helps your organs, like kidneys, work properly.

Fact: Drinking water boosts your energy levels.

Losing weight is important for older diabetic women because it helps control blood sugar and reduces health risks.

Fact: Water keeps your skin healthy and hydrated.

In a similar fashion patterns for R52X (Easy to do) and RT (response time) are elusive. No clear pattern emerges. We see differences in two variables when we deal with respondents who declare themselves normal weight versus those who declare themselves as overweight. Table 6 shows the dramatic differences between these groups, attributable to the rating of ‘easy’. Those who declare themselves overweight find most of the elements to refer to themselves (high coefficients for ME) but harder to do (lower coefficients for R52X, corresponding to easy). These results suggest the promising use of ‘Easy” as a key variable to consider.

Table 6: Coefficients for important transformed binary variables, and for response time, shown for two self-defined groups, Normal Weight versus Overweight.

Mind-Sets: Uncovering Deep Differences among People in Order to Drive Effective Messages

A continuing theme in Mind Genomics is the emergence of mind-sets, clusters of people who perceived the world in clearly different ways, in ways which make sense and point to divergence what these clusters to feel to be important. Table 6 gives us a sense that when it comes to ‘Hard vs Easy’, those who declare themselves to be overweight find many statements to be not quite as easy as those who declare themselves to be normal weight. This difference in people makes sense but does not satisfy. The differences are there, but do not strike us as a compelling story.

One way to identify these possibly more profound differences is by statistics alone, by considering the pattern of the 16 coefficients and searching for different patterns. The search is not based on the self- definition of the respondent, nor based on the meaning of the elements. Rather, the search is based purely on the mathematical structure of the 16 coefficients. If the mathematics reveals clearly different, easy to interpret mind-sets, the research will have provided a powerful tool to identify what to communicate, and to whom.

The method used to create these clusters is called k-means clustering [22]. The objective is to put the different objects, our 101 respondents, into groups based upon an objective criterion. The criterion is that the ‘distance’ between pairs of people in a cluster should be ‘small’, but the distances between the centroid (average) of the 16 coefficients should be large. The k-means clustering is not exact but tries to satisfy these conditions.

The study uses the coefficient for R5X (ME and Easy to do) as the variable on which to do the clustering exercise. We can imagine 101 rows of coefficients, one row for each respondent, and in turn 16 columns of coefficients. Table 7 shows the strongest performing elements for each mind-set or cluster, respectively Mind-Set 1 (focus on medical indicators), Mind-Set 2 (Focus on Lifestyle), and Mind-Set 3 (Focus on drinking water). Each table is sorted by the coefficients for R5X (For ME and Easy). Each mind-set is clearly different, and interpretable, remarkable in view of the actual ‘difficulty’ of doing the study on the part of the respondent. Recall that each respondent had evaluated 24 different vignettes, with 2-4 elements, seemingly randomly put together. Despite what one might think, the respondents actually perform quite well.

Table 7: Coefficients for important transformed binary variables, and for response time. The elements are sorted in descending order by the coefficient for R5x, for each of the three emergent mind-sets.

In Their Own Words

Our three mind-sets suggest radically different words to which they are sensitive. Do these mind-sets transcend the simple statistical analyses which brought them to our attention? In other words, beyond statistics which tell a ‘nice story’ can we find anything else. One way is to look at how they describe their experience with diabetes, whether their own, their family experience, or just their knowledge.

Each respondent was instructed to write bout diabetes from their own point of view, doing this exercise AFTER having done the evaluation of the 24 unique vignettes. An AI ‘summarizer’, QuillBot [23], took the open-ended answers, and summarized them. Table 8 shows three rather different summaries, suggesting that these mind-sets do actually think about diabetes in different ways. The objective of showing the summarizations is to give a sense of the different ‘morphologies’ or structures in the way people of mind-sets write about their experience, what they say, how they say it.

Table 8: AI summarization of open-ended answers about experience with diabetes from the three emergent mind-sets

Diabetes is a terrible disease that can lead to limb loss and death. It is not a common issue in the family, but it is a serious concern that should be addressed by children. The author believes that children should be more proactive in preventing and helping others who are aware of the disease. They also mention that their family is concerned about obesity and is aware of sugar complications. The author believes that children can learn from their grandparents’ experiences and be more proactive in preventing diabetes. They also mention their own family members who have had diabetes and have lost limbs. They all try to maintain healthy lifestyles and avoid complications. Despite not having children, the author believes that everyone should prevent diabetes and help others.

Identifying the Mind-sets in the Population

An important outcome of dividing people into mind-sets is the ability to tailor the proper message to each person. By understanding the different messages to which people are sensitive, there is the possibility of increasing the overall health of a population. Proper messaging has been shown to substantially reduce the number of within-30-day hospital readmissions after discharge for patients with congestive heart failure [24], as well as double the number of colorectal cancer screenings in an underserved area in the Philadelphia area [25].

How then do we assign a given person to one of the mind-sets? Once the person is properly assigned to a mind-set it becomes much simpler to identify the set of messages likely to communicate properly and effect the appropriate behavior response. The traditional approach has been to look at easy-to-measure factors, such as gender, age, and superficial cultural cues. The experienced physician may attend to more cues, as would an experienced salesperson. But what about the situation wherein one cannot involve an experienced professional. Is there any way that the staff at the ‘front desk’ might be able to assign a patient to a mind-set in a perhaps a minute or less, and attach that information to the patient chart, along with the relevant material to communicate, as well as well as what to avoid.

During the past decade author Moskowitz and colleagues have worked on an identification system, a mind-type assigner or personal viewpoint identifier [16,17]. The objective is to create a short set of questions such as that shown in Figure 6. With such a tool, it becomes possible to have a patient fill out the form in a minute or two. Panel A shows the background information that the patient fills out. Panel B shows the six questions, taken from the study, and answered with a two-point scale. The pattern of responses to the six questions ends up assigning the patient to the most appropriate mind-set.

Figure 6: The first two portion of the mind-set assigner. Panel A shows the background information. Panel B shows the six questions, which are presented in randomized order for each respondent.

Table 9 shows set-up for the ‘assigner’ tool’, including the name of the mind-set, and the feedback given either to the patient and/or t the medical professional. The mind-set assigner further provides the ability to send the respondent to a website with a video.

Table 9: Set up information for the mind-set assigner

The final feature of the mind-set assigner is the ability to ask a set of 16 different questions, each with 2-4 answers. Table 10 shows the set of 16 questions for this study. The mind-set assigner with its set of 16 questions enables the researcher to work with many new respondents, both assigning each respondent to a mind-set, and obtaining more data about the respondent. In this way the research into mind-sets moves beyond simply understanding the person in terms of the particular situation, but actually allows the researcher to deepen knowledge of the person and the covariation of that knowledge with mind-type membership.

Table 10: The 16 questions and their associated answers. The Questions and Answers constitute the third ‘leg’ of the mind-set assigner.

Discussion and Conclusions

The study on diabetes and obesity presented here constitutes the second of a planned series of studies on the communication between the doctor and the patient, using the emerging science of Mind Genomics. Several searches through the published literature and a discussion with young medical professionals revealed again and again that there is a well-established clinical literature on many medical topics, but very little on the words that doctors should say to patients, and the meanings of what patients say to doctors. The oft-given reason, perhaps excuse, is that this sensitivity to the patient can come only with experience, and that the doctor must learn to listen [26].

As the medical world becomes increasingly subject to the financial structures of capitalism, the time that a doctor spends with a patient necessarily decreases. It is the doctor’s time which ‘costs’, and the goal of a business such as the business of health care is to reduce these topline costs. It may be possible to optimize the throughput of diagnostics, and increase the efficiencies in hospital stays, especially surgery, but what about the ability of the doctor or other medical professional to communicate with the patient. If it takes years to educate someone to become an actor as well as years to become a doctor. Where then is the time to become a listener, and an empathic conveyor of news, often bad news. It is the attention to this topic, empathic information exchange, to which this new effort of Mind Genomics is dedicated, and in which spirit this early experiment and paper in that effort have been done and written. The discovery that there are really three mind-sets, not just one general mind-set with the ‘most right messaging’ becomes the important discovery here, one which, if not addressed, can become an ongoing issue in mammography, perhaps getting worse as the medical system becomes increasingly driven to efficiency by driving out the human component.

References

1. Boles A, Kandimalla R, Reddy PH (2017) Dynamics of diabetes and obesity: Epidemiological perspective. Biochimica et Biophysica Acta (BBA)-Molecular Basis of Disease 1863: 1026-1036. [crossref]
2. Mokdad AH, Ford ES, Bowman BA, Dietz WH, Vinicor F, et al. (2003) Prevalence of obesity, diabetes, and obesity-related health risk factors, 2001. JAMA 289: 76-79. [crossref]
3. Mozaffarian D (2016) Dietary and policy priorities for cardiovascular disease, diabetes, and obesity: a comprehensive review. Circulation 133: 187-225. [crossref]
4. Ling H, Lenz, TL, Burns, TL, Hilleman, DE (2013) Reducing the risk of obesity: defining the role of weight loss drugs. Pharmacotherapy: The Journal of Human Pharmacology and Drug Therapy 33: 1308-1321. [crossref]
5. Cohen MM Fiona Elliott, Liza Oates, Adrian Schembri, Nitin Mantri (2017) Do wellness tourists get well? An observational study of multiple dimensions of health and well-being after a week-long retreat. The Journal of Alternative and Complementary Medicine 17: 140-148. [crossref]
6. Nagy A, McMahon A, Tapsell L, Deane F, Arenson D (2018) Therapeutic alliance in dietetic practice for weight loss: Insights from health coaching. Nutrition & Dietetics 75: 250-255. [crossref]
7. Incollingo Rodriguez AC, Rodriguez A, Callahan LC, Saxbe D, Tomiyama, AJ (2019) The buddy system: A randomized controlled experiment of the benefits and costs of dieting in pairs. Journal of Health Psychology 24: 1945-1954. [crossref]
8. Elder KA, Wolfe BM (2007) Bariatric surgery: a review of procedures and outcomes. Gastroenterology 132: 2253-2271. [crossref]
9. Greenacre M, Groenen PJ, Hastie T, d’Enza AI, Markos A, et al. (2022) Principal component analysis. Nature Reviews Methods Primers 2: 100.
10. Omran MG, Engelbrecht AP, Salman A (2007) An overview of clustering methods. Intelligent Data Analysis 11: 583-605.
11. Grimm P (2010) Social desirability bias. Wiley International Encyclopedia of Marketing.
12. Krumpal I (2013) Determinants of social desirability bias in sensitive surveys: a literature review. Quality & Quantity 47: 2025-2047.
13. Moskowitz HR, Gofman A, Beckley J, Ashman H (2006) Founding a new science: Mind Genomics. Journal of Sensory Studies 21: 266-307.
14. Green PE, Krieger AM, Wind Y (2001) Thirty years of conjoint analysis: Reflections and prospects. Interfaces 31: S56-S73.
15. Gofman A, Moskowitz H (2010) Isomorphic permuted experimental designs and their application in conjoint analysis. Journal of Sensory Studies 25: 127-145.
16. Porretta S (2021) The changed paradigm of consumer science: from focus group to mind genomics. In Consumer-based New Product Development for the Food Industry. The Royal Society of Chemistry 21-39.
17. Porretta S, Gere A, Radványi D, Moskowitz H (2019) Mind Genomics (Conjoint Analysis): The new concept research in the analysis of consumer behaviour and choice. Trends in Food Science & Technology 84: 29-33.
18. Moskowitz HR (2012) ‘Mind Genomics’: The experimental, inductive science of the ordinary, and its application to aspects of food and feeding. Physiology & Behavior 107: 606-613. [crossref]
19. Cropley D (2023) Is artificial intelligence more creative than humans?: ChatGPT and the Divergent Association Task. Learning Letters 2: 13-13.
20. Hardy MA (1993) Regression with dummy variables (Vol. 93). SAGE.
21. Burton AL (2021) OLS (Linear) regression. The Encyclopedia of Research Methods in Criminology and Criminal Justice 2: 509-514.
22. Likas A, Vlassis N, Verbeek JJ (2003) The global k-means clustering algorithm. Pattern Recognition 36: 451-461.
23. Fitria TN (2021) QuillBot as an online tool: Students’ alternative in paraphrasing and rewriting of English writing. Englisia: Journal of Language, Education, and Humanities 9: 183-196.
24. Gabay G, Moskowitz HR (2019) “Are we there yet?” Mind-Genomics and data-driven personalized health plans. The Cross-Disciplinary Perspectives of Management: Challenges and Opportunities 7-28.
25. Oyalowo A, Forde KA, Lamann A, Kochman ML (2022) Effect of patient-directed messaging on colorectal cancer screening: A randomized clinical trial. JAMA Network Open 5: e224529-e224529. [crossref]
26. Bombak AE, Riediger ND, Bensley J, Ankomah S, Mudryj A (2020) A systematic search and critical thematic, narrative review of lifestyle interventions for the prevention and management of diabetes. Critical Public Health 30: 103-114.

DOI: 10.31038/EDMJ.2023733

Abstract

Background: Type 2 Diabetes Mellitus (T2DM) and Metabolic Syndrome (MS) continuously rise among South-Asian women. However, few studies explored the association between these chronic conditions and modifiable risk factors among South-Asian women. Therefore, this study evaluated the incidence of T2DM and MS and its association with modifiable risk factors among women with and without a history of Gestational Diabetes Mellitus (GDM).

Methods: We conducted the study using a retrospective and prospective follow-up component (an ambidirectional cohort study). We retrospectively identified women with GDM from 1999 to 2005 from the medical record system of the Aga Khan University (AKU) Hospital Karachi, Pakistan. We prospectively enrolled 226 women with GDM and 423 without GDM (1:2) between 2008 and 2010. The outcomes were the development of T2DM or impaired glucose tolerance (IGT) and MS among women with vs. without GDM and their association with modifiable risk factors, such as body mass index (BMI), body fat%, diet, and physical activity. Using multivariable logistic regression, we created two models to understand the association between modifiable risk factors and the development of T2DM and MS. In the first model, we used BMI as a measure of obesity, while in the second; we replaced it with body fat percentage. Other variables (diet and physical activity) were present in both models.

Results: We observed 10.91 times (CI 4.57-26.04) higher odds of developing T2DM/IGT among women with vs. without GDM. In addition, BMI (kg/m²) (OR 1.09, CI 1.02-1.15) and diet scores (high in fat, sugar, and bakery items) (OR 1.22, CI 1.01-1.49) were found significant. In the second model of T2DM, women with GDM (OR 11.02, CI 4.63-26.22) and body fat% (OR 1.10, CI 1.04-1.17) were found significant. For MS, we observed 2.78 times (CI 1.04-7.41) higher odds of developing MS among women with vs. without GDM. In addition, BMI (kg/m2) (OR 1.32, CI 1.22-1.42) and body fat% (OR 1.35, CI 1.23-1.48) were found significant in the first and second models, respectively.

Conclusions: BMI or body fat%, and, possibly, diet are potential modifiable risk factors for T2DM/IGT and MS among women with GDM.

Keywords

Gestational diabetes mellitus, Modifiable risk factors, Type 2 diabetes mellitus, Metabolic syndrome, South-Asian women, Low-and middle-income countries

Background

Diabetes is a major growing public health concern affecting about half a billion people worldwide [1]. The recent global estimates on diabetes prevalence report that 537 million adults aged 20 to 79 (one out of ten people) are affected by diabetes [2]. One in every four adults in Pakistan has diabetes mellitus (26.7%), reaching the highest national prevalence globally [2].

Type 2 Diabetes Mellitus (T2DM) is a progressive disease resulting from insufficient insulin secretion or resistance [3]. Several environmental risk factors, such as unhealthy diet, physical inactivity, and obesity, are further implicated in the pathogenesis of T2DM [4,5]. One of the metabolic risk factors for the development of T2DM is Gestational Diabetes Mellitus (GDM) [4]. It has been estimated that about 30% to 70% of women with GDM develop T2DM within approximately 15 years of the index pregnancy [6,7]. Among South Asian women, about 17% to 33% develop T2DM within 5 to 10 years after the index pregnancy [8].

GDM is defined as high blood glucose levels first recognized during pregnancy or subsequent pregnancies [9]. GDM is a highly prevalent metabolic issue during pregnancy, affecting about 1.8% to 31.5% of all pregnancies worldwide, depending on the screening methods, diagnostic criteria, and population characteristics [10]. The prevalence of GDM among low-and middle-income countries (LMICs) varies between 9.2% to 12.7% [11]. There has been a rise in the prevalence of GDM in Pakistan over the last few decades, with an increase from 6.3% in 2003 to 19% in 2018 [12].

GDM, on the one hand, increases the risk of developing many short-term complications, including preeclampsia, caesarean section, preterm births, macrosomia, and prenatal and perinatal mortality [13]. On the other hand, GDM tends to increase the risk of developing adverse health in the long term, such as T2DM, Cardiovascular Diseases (CVDs), and Metabolic Syndrome (MS) [7].  MS is defined as the presence of biological indicators such as abnormal waist circumference, high systolic and diastolic blood pressure, increased triglyceride levels, low high-density lipoprotein (HDL) levels, and impaired blood glucose levels [14].

Since these long-term issues can be preventable, lifestyle factors, i.e., diet, physical activity, and weight reduction, play a pivotal role in preventing and delaying the onset of T2DM and MS [15]. However, few studies have explored the association of lifestyle factors with T2DM and MS in women with GDM [16,17]. In addition, the existing evidence mainly based on the West, like Yang et al., demonstrated a lower risk of developing T2DM among women with a history of GDM who effectively managed modifiable risk factors [18]. There is a dearth of evidence, particularly from LMICs, on the role of modifiable risk factors in the development of T2DM and MS among women with a history of GDM. Therefore, this study evaluated the incidence of T2DM and MS and its association with modifiable risk factors among women with and without a history of GDM attending a tertiary healthcare facility in Karachi, Pakistan.

Methods

This study involved retrospective and prospective follow-up components (an ambidirectional cohort study). We retrospectively identified women from the medical record system who identified as having GDM based on International Classification of Diseases (ICD) 10 code and received prenatal care during 1999-2005 at the Aga Khan University Hospital (AKUH). All women identified through the Medical Record (MR) system with singleton birth in the index pregnancy, spoke Urdu, and residents of Karachi were included in the study. Those living outside Pakistan, with incomplete medical records, who used drugs influencing blood glucose concentration, such as glucocorticoids, antipsychotic drugs, or metformin, were excluded. We also excluded women diagnosed with T2DM before the initiation of the study as that was the key outcome, and they might have modified their lifestyle after a diagnosis of T2DM which may bias the association between the risk factors and development of T2DM.

All potential GDM women were contacted through mail as well as via phone calls as part of the recruitment process. Both mailing addresses and phone numbers were obtained from the MR system. All GDM women were first issued an invitation letter outlining the aims and methods of the study, a consent form, and a pre-paid, self-addressed mail-back envelope. The enclosed consent form was to be filled out by the women, and then mailed back. All GDM women were contacted by phone in addition to being mailed out. The phone calls were made a week after the mailings were completed. The letters were sent out, and the phone calls were made by the trained research staff. The detailed recruitment plan and participant response rate were published elsewhere [19]. During the phone calls, initial consent was obtained from the women to review their MRs for further basic medical information.

Women without GDM (matched for age at pregnancy and gestational age) were also identified from the MR system and contacted via phone to recruit them for the study. Women with GDM were classified as exposed, and those who did not develop GDM were considered non-exposed. We enrolled women as exposed and non-exposed in the 1:2 ratio. Those who consented and were eligible to participate were invited to AKUH to provide written informed consent and for further study assessments during 2008-2010. Structured questionnaire was used to assess sociodemographic, dietary intake, and physical activity through a face-to-face interview. In addition, anthropometric assessments and blood tests for evaluating T2DM or Impaired Glucose Tolerance (IGT) and MS status were also done.

Retrospective Data Collection

Data collected from MRs included maternal age, parity, known hypertension, mode of delivery, gestational age, use of insulin/diet for treating GDM, and GDM during a subsequent pregnancy.

Prospective Data Collection

Blood Work

All participants underwent an oral glucose tolerance test (OGTT) which involved a fasting blood sample followed by a 2-hour post-glucose sample (75-gram glucose liquid). At the same time, a 12-hour fasting lipid profile test was also performed to assess the status of dyslipidemia and MS. The details of the blood work and their cut-off values are described in additional file (see Additional file 1).

Anthropometry

Stature

Height and weight were measured at the time of the interview by the research officer. Height was measured using a wall-mounted measuring scale to the nearest cm, while weight was measured using a Tanita Body composition analyzer (Tanita Corp. CA. USA).

Waist Circumference and Waist-to-Hip Ratio

Waist and hip circumferences were measured using non-stretchable tape. These measurements were used to estimate waist circumference and waist-to-hip ratio among participants. The research officer was trained in measuring the circumferences at the correct point, i.e., waist circumference was measured between the uppermost part of the hip bone and the lowest rib margin (tenth rib), and the hip circumference was measured at the widest point over the buttocks.

Body Composition

Body composition was measured by a body fat analyzer. Total body fat was measured by a non-invasive Tanita body composition analyzer BC 310 (Tanita Corp. CA. USA), which measured fat mass, body fat percentage, abdominal fat mass, and fat-free mass.

Questionnaires and Interview

Two research officers were involved in data collection using an Urdu-translated interviewer-administered questionnaire for the following components:

Sociodemographic Questionnaire

A sociodemographic questionnaire collected information on participants’ educational level, occupation, household income, and socioeconomic status.

Food frequency Questionnaire

Diet was assessed using a Food Frequency Questionnaire (FFQ) that was developed and validated among Pakistani women [20]. The questionnaire has food items with their frequencies and portion sizes. The frequency of all food items consumed was multiplied by the portion sizes and then converted into the daily intake.

Using principal component analysis (PCA), we reduced dietary intake data to generate one variable to assess the diet-disease association. The new diet variable loadings were high in fat, sugar, and other bakery items (see Additional file 2).

Physical activity Questionnaire

To assess participants’ physical activity, a Monica Optional Study of Physical Activity (MOSPA) questionnaire was used. The questionnaire has been adapted from the World Health Organizations’ Monitoring Trends and Determinants of Cardiovascular Disease study [21] and has been validated among Pakistani women [22]. The MOSPA questionnaire assessed physical activity in four broader categories, including leisure, occupational, transportation, and household chores. Detailed information on calculating Energy Expenditure (EE) was provided in additional file (see Additional file 3).

Outcomes and Their Assessments

The outcomes were defined as the development of T2DM or IGT and MS among women with and without a history of GDM and their association with modifiable risk factors such as Body Mass Index (BMI), body fat%, diet, and physical activity. Women with fasting blood sugar ≥126 mg/dL and 2-hour post glucose ≥200 mg/dL were considered T2DM, whereas those with one value abnormal, either fasting blood sugar ≥126 mg/dl or 2-hour post glucose ≥200 mg/dl considered having IGT [3]. Metabolic Syndrome (MS) was defined as women with waist circumference > 80 cm with any one of the following conditions (HDL <50 mg/dl and Triglyceride >150 mg/dl) OR (HDL <50 mg/dl and FBS >100 mg/dl) OR (Triglyceride >150 mg/dl and FBS >100 mg/dl) [14].

Modifiable Risk Factors and Their Association with T2DM/IGT and MS

We evaluated the independent effect of BMI, body fat%, diet, and total reported physical activity (energy expenditure (kcal)/day) as modifiable risk factors on the development of T2DM/IGT and MS while adjusted for education, wealth index, and family history of diabetes.

Sample Size

Based on the findings of Feig et al. [23] and Cianni et al. [24] for the risk of developing T2DM and MS among women with GDM vs. without GDM, respectively, a sample size of 62 in exposed and 124 in non-exposed (1:2) for T2DM, whereas a sample size of 140 in exposed and 280 in non-exposed (1:2) for MS was required assuming 80% power and 5% level of significance.

However, we prospectively enrolled 226 women with a history of GDM who were eligible and consented to participate in the study as exposed and their comparators 423 women without a history of GDM as non-exposed (1:2) matched on age at the time of pregnancy and gestational age.

Statistical Analysis

Sociodemographic characteristics of women with and without a history of GDM were presented as mean ± SD for age at the time of pregnancy and age at the time of follow-up and median and range for the number of children. The frequencies and percentages were reported for education, occupation, wealth index, parity, mode of delivery, family history of diabetes, hypertension, treatment of GDM during pregnancy, and GDM during a subsequent pregnancy. The anthropometric and body composition distribution among the two groups were presented as mean ± SD or frequencies and percentages as appropriate. A chi-square test for all categorical variables, whereas an independent t-test for all continuous variables was computed to compare any differences between the two groups.

Binary logistic regression analysis was performed using the development of T2DM/IGT and MS as the dependent variable and women with a history of GDM vs. Non-GDM as an exposure variable, whereas four modifiable risk factors as an independent variable (BMI, body fat%, diet, and physical activity). A multivariable logistic regression analysis was performed to assess the association between modifiable risk factors and the risk of developing T2DM/IGT and MS among women with and without a history of gestational diabetes. Univariate analysis was performed to compute crude regression coefficients with 95% CIs.  A stepwise approach was used during multivariable analysis while adjusting for confounders such as education, wealth quintiles, and family history of diabetes. A p-value of <0.05 was considered significant. Data were analyzed using Stata (V.17, Statacorp).

Results

Sociodemographic Characteristics of the Study Population

The sociodemographic characteristics of women with vs. without a history of GDM were compared in Table 1. Both groups were comparable in terms of almost all variables, except for number of children (p=0.020), mode of delivery (p=0.001), family history of diabetes (p=0.001), and presence of hypertension (p=0.009).

Table 1: Sociodemographic characteristics of the study population

	GDM n=226 Mean ± SD or n (%)	Non-GDM n=423Mean ± SD or n (%)	P-value
Age at the time of pregnancy (years)	31.18 ± 4.87	31.06 ± 4.75	0.77
Age at the time of follow-up (years)	37.15 ± 5.16	37.38 ± 5.05	0.58
Education			0.58
Primary to secondary (Class 6-10)	23 (10.2)	42 (9.9)
Intermediate (Class 12)	44 (19.5)	69 (16.3)
Graduation (Class 14 and above)	159 (70.4)	312 (73.8)
Occupation			0.14
Housewife	196 (86.7)	369 (87.6)
Employed	30 (13.3)	46 (10.9)
Others	0	6 (1.4)
Wealth Index (quintiles)			0.89
Lowest	91 (40.3)	171 (40.4)
Middle	58 (25.7)	101 (23.9)
Highest	77 (34.0)	151 (35.7)
Parity			0.12
Primiparous	48 (21.6)	69 (16.6)
Multiparous	174 (78.4)	346 (83.4)
Numbers of children; Median (range)	1 (0-7)	1 (0-6)	0.020*
Mode of delivery			0.001*
Spontaneous vaginal delivery	114 (50.7)	267 (65.3)
Assisted (Forceps/vacuum)	110 (48.9)	140 (34.2)
Caesarean section	1 (0.4)	2 (0.5)
Family history of diabetes	175 (77.8)	273 (65.0)	<0.001*
Hypertension			0.009*
Essential hypertension	19 (8.4)	20 (4.7)
Pregnancy-induced hypertension	24 (10.6)	24 (5.7)
Treatment of GDM during pregnancy
Insulin	34 (15.5)	–
Diet	186 (84.5)	–
GDM during subsequent pregnancy	41 (18.7)	–

GDM: Gestational diabetes
^*P-Value <0.05

Incidence of T2DM and MS among the Study Population

With a median follow-up of 6 years, we found 34 (15%) incident cases of T2DM/IGT among women with a history of GDM, whereas 7 cases (1.7%) among women without a history of GDM (p <0.001). We also evaluated the status of MS in our study population and found that one out of every 15 women was diagnosed with MS in the GDM group. In contrast, among the non-GDM group, the diagnosis was less (one out of every 36 women) (6.7% vs. 3.1%, p <0.033). Moreover, women in the GDM group were found with a higher level of fasting blood glucose (41.3% vs. 14.7%), total cholesterol (25.2% vs. 16.4%), triglycerides (30.2% vs. 14%), and LDL (72.1% vs. 61%) and a lower level of HDL (70.3% vs. 64.6%), all with a p-value of <0.05 except for the HDL (Figure 1).

Figure 1: Incidence of T2DM/IGT, MS, and other conditions among study population.
T2DM: Type 2 Diabetes Mellitus, IGT: Impaired Glucose Tolerance, MS: Metabolic Syndrome, GDM: Gestational Diabetes, LDL: Low-density lipoprotein, HDL: High-density lipoprotein
^#also included women with one value abnormal, i.e., IGT
^MS defined as women with waist circumference > 80 cm with any one of the following conditions (HDL <50 mg/dl & Triglyceride >150 mg/dl) OR (HDL <50 mg/dl & FBS >100 mg/dl) OR (Triglyceride >150 mg/dl & FBS >100 mg/dl)
*P-value <0.05
**P-value <0.001.

Modifiable Risk Factors among the Study Population

While comparing the anthropometric and body composition of the study population, there were significant differences between the two groups in terms of BMI (kg/m²) (28.4 vs. 27.4, p 0.017), waist circumference (cm) (66.7 vs. 64.9, p 0.036), body fat (%) (36.1 vs. 34.5, p 0.007), and visceral fat (%) (6.6 vs. 5.9, p 0.025) in women with vs. without GDM, respectively.

There were differences in physical activity between groups in occupation and household chores-related activity; however, none of these differences were found to be statistically significant (Table 2).

Table 2: Description of modifiable risk factors among the study population

		GDM n=226 Mean ± SD or n (%)	Non-GDM N=423Mean ± SD or n (%)		P-value
Anthropometric
Weight (kg)		68.31 ± 12.63	66.60 ± 13.16		0.11
BMI (kg/m2)		28.42 ± 5.29	27.37 ± 5.33		0.017*
Normal weight (<23)		28 (12.4)	77 (18.3)		0.014*
Overweight (23-26.9)		67 (29.8)	150 (35.6)
Obese (≥ 27)		130 (57.8)	194 (46.1)
Waist circumference (cm)		66.69 ± 9.80	64.85 ± 11.06		0.036*
High (≥ 80)		19 (8.4)	32 (7.6)		0.70
Waist-to-hip ratio		0.82 ± 0.08	0.81 ± 0.10		0.059
High (≥ 0.8)		133 (58.8)	218 (51.5)		0.075
Body Composition
Body fat (%)		36.12 ± 6.18	34.51 ± 7.63		0.007*
Visceral fat (%)		6.56 ± 2.73	5.95 ± 2.70		0.025*
Muscle mass (kg)		40.47 ± 35.16	42.96 ± 34.24		0.38
Bone mass (kg)		2.20 ± 0.23	2.39 ± 1.97		0.24
Total body water (kg)		31.57 (2.90)	31.18 (3.22)		0.15
Physical Activity (energy expenditure; kcal/day)
Total reported physical activity		655.47 ± 26.48	688.67 ± 21.29		0.35
Occupation		620.96 ± 33.64	521.14 ± 33.92		0.06
Transportation		135.25 ± 22.29	172.69 ± 19.11		0.24
Household chores		463.17 ± 19.18	514.24 ± 16.81		0.06
Leisure time		154.89 ± 18.87	158.97 ± 15.53		0.87

GDM: Gestational diabetes, BMI: Body mass index
^*P-Value <0.05

Modifiable Risk Factors and the Risk of Developing T2DM/IGT

We created two models to understand the association between modifiable risk factors and the development of T2DM and MS in women with vs. without GDM. In the first model, we used BMI as a measure of obesity, while in the second, we replaced it with body fat percentage. Other variables, such as diet and physical activity were present in both the models. We observed that women with a history of GDM had 10.91 times (CI 4.57-26.04) higher odds of developing T2DM/IGT than women without a history of GDM. Furthermore, among women with a history of GDM, with every one unit increase in the BMI (kg/m²) and diet scores (high in fat, sugar, and other bakery items), the odds of developing T2DM/IGT increased by 1.09 (CI 1.02-1.15) and 1.22 (CI 1.01-1.49), respectively. In the second model, the odds of developing T2DM/IGT among women with GDM (OR 11.02, CI 4.63-26.22) and body fat% (OR 1.10, CI 1.04-1.17) were significant. Both models were adjusted for education, wealth index, and family history of diabetes (Table 3).

Table 3: Association of T2DM/IGT and modifiable risk factors among women with vs. without a history of GDM

Variables		T2DM/IGT
		Model 1	Model 2a
	Crude OR (95% CI)	ORadj (95% CI)	ORadj (95% CI)
Group
Non-GDM	Ref	Ref	Ref
GDM	10.52 (4.58-24.17)*	10.91 (4.57-26.04)*	11.02 (4.63-26.22)*
BMI	1.09 (1.03-1.15)*	1.09 (1.02-1.15)*	NA
Body fat%	1.10 (1.05-1.16)*	NA	1.10 (1.04-1.17)*
Diet	1.08 (0.91-1.28)	1.22 (1.01-1.49)*	1.21 (0.99-1.48)^
PA expenditure (kcal/day)	1.00 (0.99-1.01)	1.00 (0.99-1.01)	1.00 (0.99-1.01)

T2DM: Type 2 diabetes mellitus, IGT: Impaired glucose tolerance, GDM: Gestational diabetes, BMI: Body mass index, CI: Confidence interval, NA: Not applicable, OR: Odds ratio, OR_adj: Adjusted odds ratio, PA: Physical activity,
PCA included diet high in fat, sugar, and other bakery items
Women were matched on age and time of delivery
Models 1 and 2 also adjusted for education, wealth index, and family history of diabetes
^aIn model 2, BMI was replaced by fat%
OR for BMI and PCA are for one unit increase, and for fat% and physical activity are for a 10% and 10 minutes increase, respectively
^*P-value <0.05
^{^}P-value=0.05

Modifiable Risk Factors and the Risk of Developing MS

Similarly, in the first model, we observed that women with a history of GDM had 2.78 times (CI 1.04-7.41) higher odds of developing MS than women without a history of GDM. Furthermore, among women with a history of GDM, with every one-unit increase in the BMI (kg/m²), the odds of developing MS increased by 1.32 (CI 1.22-1.42). When we replaced BMI with body fat% in the second model, we found that the odds of developing MS among women with a history of GDM (OR 2.80, CI 1.12-7.03) and body fat percentage (OR 1.35, CI 1.23-1.48) were significant (Table 4).

Table 4: Association of MS and modifiable risk factors among women with vs. without a history of GDM

Variables		MS
		Model 1	Model 2a
	Crude OR (95% CI)	ORadj (95% CI)	ORadj (95% CI)
Group
Non-GDM	Ref	Ref	Ref
GDM	2.25 (1.05-4.81)*	2.78 (1.04-7.41)*	2.80 (1.12-7.03)*
BMI	1.30 (1.21-1.40)*	1.32 (1.22-1.42)*	NA
Body fat (%)	1.33 (1.22-1.45)*	NA	1.35 (1.23-1.48)*
Diet	1.03 (0.83-1.27)	1.20 (0.92-1.55)	1.17 (0.91-1.51)
PA expenditure (kcal/day)	1.00 (0.99-1.01)	1.00 (0.99-1.01)	1.00 (0.99-1.01)

MS: Metabolic Syndrome, GDM: Gestational diabetes, CI: Confidence interval, NA: Not applicable, OR: Odds ratio, OR_adj: Adjusted odds ratio, PA: Physical activity
PCA included diet high in fat, sugar, and other bakery items
Women were matched on age and time of delivery
Models 1 and 2 also adjusted for education, wealth index, and family history of diabetes
^aIn model 2, BMI was replaced by fat%
OR for BMI and PCA are for one unit increase, and for fat% and physical activity are for a 10% and 10 minutes increase, respectively
^*P-value <0.05

Discussion

This study evaluated the incidence of T2DM and MS and its association with modifiable risk factors among women with and without a history of GDM attending a tertiary healthcare facility in Karachi, Pakistan. We found a higher incidence of T2DM (15%) and MS (6.7%) in women with GDM compared to those without GDM, 1.7%, and 3.1%, respectively. In addition, the two modifiable risk factors, such as BMI and diet, predicted T2DM. Replacing BMI with body fat% in a similar model led to comparable estimates for the two indicators of obesity. However, in our second model, the history of GDM and BMI or body fat percentage were associated with MS.

Gestational diabetes is an independent risk factor for developing T2DM and other metabolic issues [25,26]. Our findings are similar to a meta-analysis that reported almost ten times higher risk (pooled relative risk) of developing T2DM among women with a history of GDM as compared to their counterparts [27]. Similarly, for MS, a study from Italy reported about 9% of women with a previous history of GDM identified with MS (15 out of 166 women) compared to controls (1%, 1 out of 90) [24]. We also observed similar results where women with a history of GDM had higher odds of developing T2DM and MS than women without a history of GDM.

Previous studies have identified the predictive relationship between the modifiable risk factors and the risk of diabetes and MS [28-30]. However, most of these studies have been conducted among women without GDM. Few studies from the West have explored the association of modifiable risk factors with T2DM in women with GDM [18,31]. For example, a longitudinal cohort study among women with a history of GDM found a 90% relative reduction in the risk of T2DM among those with adequate control of five modifiable risk factors, such as BMI, diet, physical activity, smoking, and alcohol use [18]. We also assessed the association of four modifiable risk factors, such as BMI, body fat%, diet, and physical activity, on the development of T2DM and MS among women with vs. without a history of GDM and found an apparent effect of BMI and body fat% on the risk of T2DM and MS.

In general, Asians are more susceptible to fat deposition and metabolic derangement even at a lower BMI and are at a greater risk of developing T2DM and MS [32]. Additionally, BMI is a crude indicator of adiposity compared to other, more direct measures of body composition, i.e., body fat percentage [33], which can be assessed by DEXA scans and the Bio-Electrical Impedance Analysis (BIA) method [34]. A direct assessment of body fat percentage can be more accurate in predicting the association between adiposity and chronic disease outcomes such as T2DM and MS in women with GDM [33]. Therefore, we also developed a regression model with body fat percentage by replacing BMI for T2DM and MS. Our study provides evidence that the two parameters (BMI and body fat%) had similar strength of association in predicting T2DM and MS among South Asian women in our study.

Other modifiable risk factors, such as unhealthy diet and physical inactivity, play a central role in the development of T2DM and other metabolic issues [35]. For example, there is strong evidence of the cause-and-effect association between unhealthy dietary patterns and the risk of T2DM and MS [36,37]. Similarly, the dose and response relationship between physical activity and the risk of T2DM has been established [38]. However, the association of diet and physical activity with the development of T2DM and MS has not been fully explored among women with a history of GDM in South Asia. Hence, we evaluated the effect of diet and physical activity on the risk of T2DM and MS in our study population and found that increased dietary scores (high in fat, sugar, and other bakery products) in the presence of BMI predicted T2DM among women with previous GDM. However, the study findings were insignificant when we replaced BMI with body fat% in the second model. Unfortunately, we did not find any association between physical activity and the risk of T2DM and MS. The reason may be that the physical activity scores were very low in our study population. In addition, as our study sample was based on AKU hospital, which mainly serves the affluent people of the city; therefore, we might not have been able to accurately estimate the impact of some lifestyle factors, such as diet and physical activity, on the outcome due to lack of variability in the data. However, one cannot underestimate the value of physical activity in preventing the development of T2DM and MS.

Our study has several strengths. To the best of our knowledge, this is the first study that assessed the association between modifiable risk factors and the risk of developing T2DM and MS among South Asian women with and without a history of GDM. The prospective nature of the study allowed us to collect information on the modifiable risk factors prior to the occurrence of the outcomes, minimizing recall and interviewer bias during the data collection. We identified women with a history of GDM from the medical record system of AKUH using the International Classification of Diseases (ICD) code 10 for GDM. We objectively assessed the diagnosis of T2DM and MS using the World Health Organization (WHO) [3] and International Diabetes Federation (IDF) [14] criteria, respectively. We included the comparator group (women without GDM) matched on age at the time of pregnancy and gestational age. In addition, we also adjusted our regression models for other known confounders, such as education, wealth quintiles, and family history of diabetes, and hence provide more accurate estimates of the risk of T2DM and MS and their association with modifiable risk factors among women with and without GDM. Our study does have some limitations. We evaluated the modifiable risk factors at a single time point and have not followed them up to gather information on long-term lifestyle habits, which would have enabled us to estimate a more accurate effect of these risk factors on the outcomes. Additionally, we enrolled our study sample from AKU hospital, which mainly serves the affluent population of the city; therefore, we might not have been able to accurately estimate the impact of some lifestyle factors, such as diet and physical activity, on the outcome due to lack of variability in the data. Furthermore, the hospital-based sample might limit the generalizability of our study findings. However, given the nature of the study, there were no community-based registries of pregnant women present at the time this study was being carried out to answer the research question.

Conclusion

Our study findings reveal that potentially modifiable risk factors, such as BMI or body fat% and possibly diet are associated with the development of T2DM/IGT and MS in South Asian women with a history of GDM. These findings hold significant implications, particularly for countries like Pakistan, where T2DM/IGT and MS are rising. Therefore, establishing prevention programs targeted towards women with GDM during the postpartum period is crucial. These programs should prioritize the promotion of healthy lifestyle behaviors, such as healthy eating habits and increasing physical activity, to reduce fat deposition in individuals with high body fat percentages.

Declarations

Ethical Approval and Consent to Participate

The study was approved by the Ethical Review Committee of Aga Khan University (Ref: 1094-CHS/ERC-08). Informed consent was obtained from study participants and before participation.

Funding

This research was funded by The Aga Khan University Research Council (URC); however, the funding is not available to support publication of the research.

Author Contributions

RI conceptualized the study, wrote the study protocol, and secured funding. RI supervised the study. SN and GB were involved in the execution of the study and data acquisition. RQ facilitated data collection from the medical records as well as approaching women. SN, GB, and UK were involved in the analysis, interpretation, and writing the manuscript. RI and RQ reviewed and provided scientific revisions to the manuscript. All authors agreed prior to submission to take responsibility and be accountable for the contents of the manuscript. All authors read and approved the final manuscript.

Acknowledgments

We would like to thank our study participants who took part in this study. We are also thankful to the research team who contributed to this study.

List of Abbreviations:

T2DM: Type 2 Diabetes
GDM: Gestational Diabetes
LMICs: Low-and Middle-Income Countries
CVDs: Cardiovascular Diseases
MS: Metabolic Syndrome
ICD: International Classification of Diseases
AKU: Aga Khan University
IGT: Impaired Glucose Tolerance
OGTT: Oral Glucose Tolerance Test
FFQ: Food Frequency Questionnaire
PCA: Principal Component Analysis
MOSPA: Monica Optional Study of Physical Activity
LDL: Low-Density Lipoprotein
HDL: High-Density Lipoprotein
BMI: Body Mass Index
WHO: World Health Organization
IDF: International Diabetes Federation

References

1. Cho NH, Shaw JE, Karuranga S, Huang Y, da Rocha Fernandes JD, et al. (2018) IDF Diabetes Atlas: Global estimates of diabetes prevalence for 2017 and projections for 2045. Diabetes Res Clin Pract 138: 271-281. [crossref]
2. International Diabetes Federation Brussels, Belgium2021 [10th: [Available from: https://diabetesatlas.org/.
3. Association AD (2018) Classification and Diagnosis of Diabetes: Standards of Medical Care in Diabetes-2019. Diabetes Care 42: S13-S28.
4. Fletcher B, Gulanick M, Lamendola C (2002) Risk Factors for Type 2 Diabetes Mellitus. Journal of Cardiovascular Nursing 16: 17-23.
5. Zheng Y, Ley SH, Hu FB (2018) Global aetiology and epidemiology of type 2 diabetes mellitus and its complications. Nat Rev Endocrinol 14: 88-98. [crossref]
6. Lee AJ, Hiscock RJ, Wein P, Walker SP, Permezel M (2007) Gestational diabetes mellitus: clinical predictors and long-term risk of developing type 2 diabetes: a retrospective cohort study using survival analysis. Diabetes Care 30: 878-883. [crossref]
7. Lauenborg J, Hansen T, Jensen DM, Vestergaard H, Mølsted-Pedersen L, et al. (2004) Increasing incidence of diabetes after gestational diabetes: a long-term follow-up in a Danish population. Diabetes Care 27: 1194-1199. [crossref]
8. Gadve SS, Chavanda S, Mukherjee AD, Aziz S, Joshi A, et al. (2021) Risk of Developing Type 2 Diabetes Mellitus in South Asian Women with History of Gestational Diabetes Mellitus: A Systematic Review and Meta-Analysis. Indian J Endocrinol Metab 25: 176-181. [crossref]
9. Classification and Diagnosis of Diabetes: Standards of Medical Care in Diabetes-2020. Diabetes Care 43: S14-s31.
10. Zhu Y, Zhang C (2016) Prevalence of Gestational Diabetes and Risk of Progression to Type 2 Diabetes: a Global Perspective. Curr Diab Rep 16: 7. [crossref]
11. Wang H, Li N, Chivese T, Werfalli M, Sun H, et al. (2022) IDF Diabetes Atlas: Estimation of Global and Regional Gestational Diabetes Mellitus Prevalence for 2021 by International Association of Diabetes in Pregnancy Study Group’s Criteria. Diabetes Research and Clinical Practice 183: 109050. [crossref]
12. Sheikh A, Sheikh L (2020) Changing prevalence of Gestational Diabetes Mellitus during pregnancy over more than a decade. J pak Med Assoc 70: 1477-1478. [crossref]
13. Group HSCR (2009) Hyperglycemia and Adverse Pregnancy Outcome (HAPO) Study: associations with neonatal anthropometrics. Diabetes 58: 453-459. [crossref]
14. Zimmet P, KG MMA, Serrano Ríos M (2005) [A new international diabetes federation worldwide definition of the metabolic syndrome: the rationale and the results]. Rev Esp Cardiol 58: 1371-1376. [crossref]
15. Goveia P, Cañon-Montañez W, Santos DP, Lopes GW, Ma RCW, et al. (2018) Lifestyle Intervention for the Prevention of Diabetes in Women With Previous Gestational Diabetes Mellitus: A Systematic Review and Meta-Analysis. Front Endocrinol (Lausanne) 9: 583. [crossref]
16. Ratner RE, Christophi CA, Metzger BE, Dabelea D, Bennett PH, et al. (2008) Prevention of diabetes in women with a history of gestational diabetes: effects of metformin and lifestyle interventions. J Clin Endocrinol Metab 93: 4774-4779. [crossref]
17. Pérez-Ferre N, Del Valle L, Torrejón MJ, Barca I, Calvo MI, et al. (2015) Diabetes mellitus and abnormal glucose tolerance development after gestational diabetes: A three-year, prospective, randomized, clinical-based, Mediterranean lifestyle interventional study with parallel groups. Clin Nutr 34: 579-585. [crossref]
18. Yang J, Qian F, Chavarro JE, Ley SH, Tobias DK, et al. (2022) Modifiable risk factors and long term risk of type 2 diabetes among individuals with a history of gestational diabetes mellitus: prospective cohort study. Bmj 378: e070312. [crossref]
19. Iqbal R, Haroon A, Jabbar A, Babar N, Qureshi R (2012) What method of contact works best for recruiting participants in a study: lessons for health care researchers? J Pak Med Assoc 62: 1293-1297. [crossref]
20. Iqbal R, Rafique G, Badruddin S, Qureshi R, Cue R, et al. (2007) Increased body fat percentage and physical inactivity are independent predictors of gestational diabetes mellitus in South Asian women. Eur J Clin Nutr 61: 736-742. [crossref]
21. Pereira MA, FitzerGerald SJ, Gregg EW, Joswiak ML, Ryan WJ, et al. (1997) A collection of Physical Activity Questionnaires for health-related research. Med Sci Sports Exerc 29: 205. [crossref]
22. Iqbal R, Rafique G, Badruddin S, Qureshi R, Gray-Donald K (2006) Validating MOSPA questionnaire for measuring physical activity in Pakistani women. Nutr J 5: 18. [crossref]
23. Feig DS, Zinman B, Wang X, Hux JE (2008) Risk of development of diabetes mellitus after diagnosis of gestational diabetes. Cmaj 179: 229-234. [crossref]
24. Di Cianni G, Lencioni C, Volpe L, Ghio A, Cuccuru I, et al. (2007) C-reactive protein and metabolic syndrome in women with previous gestational diabetes. Diabetes Metab Res Rev 23: 135-140. [crossref]
25. Kaiser K, Nielsen MF, Kallfa E, Dubietyte G, Lauszus FF (2021) Metabolic syndrome in women with previous gestational diabetes. Scientific Reports 11: 11558. [crossref]
26. Diaz-Santana MV, O’Brien KM, Park Y-MM, Sandler DP, Weinberg CR (2022) Persistence of Risk for Type 2 Diabetes After Gestational Diabetes Mellitus. Diabetes Care 45: 864-870. [crossref]
27. Vounzoulaki E, Khunti K, Abner SC, Tan BK, Davies MJ, et al. (2020) Progression to type 2 diabetes in women with a known history of gestational diabetes: systematic review and meta-analysis. Bmj 369: m1361. [crossref]
28. Aravinda J (2019) Risk factors in patients with type 2 diabetes in Bengaluru: A retrospective study. World J Diabetes 10: 241-248. [crossref]
29. Ismail L, Materwala H, Al Kaabi J (2021) Association of risk factors with type 2 diabetes: A systematic review. Computational and Structural Biotechnology Journal 19: 1759-1785. [crossref]
30. Al Shehri HA, Al Asmari AK, Khan HA, Al Omani S, Kadasah SG, et al. (2022) Association between preventable risk factors and metabolic syndrome. Open Med (Wars) 17: 341-352. [crossref]
31. Mishra S, Shetty A, Rao CR, Nayak S, Kamath A (2020) Risk factors for gestational diabetes mellitus: A prospective case-control study from coastal Karnataka. Clinical Epidemiology and Global Health 8: 1082-1088.
32. Deurenberg P, Deurenberg-Yap M, Guricci S (2002) Asians are different from Caucasians and from each other in their body mass index/body fat per cent relationship. Obes Rev 3: 141-146. [crossref]
33. Bhowmik B, Munir SB, Diep LM, Siddiquee T, Habib SH, Samad MA, et al. (2013) Anthropometric indicators of obesity for identifying cardiometabolic risk factors in a rural Bangladeshi population. J Diabetes Investig 4: 361-368. [crossref]
34. Yi Y, Baek JY, Lee E, Jung H-W, Jang I-Y (2022) A Comparative Study of High-Frequency Bioelectrical Impedance Analysis and Dual-Energy X-ray Absorptiometry for Estimating Body Composition. Life 12: 994. [crossref]
35. Raman PG (2016) Environmental Factors in Causation of Diabetes Mellitus. In: Marcelo LL, Sonia S, editors. Environmental Health Risk. Rijeka: IntechOpen 9.
36. Czekajło A, Różańska D, Zatońska K, Szuba A, Regulska-Ilow B (2018) Association between dietary patterns and metabolic syndrome in the selected population of Polish adults-results of the PURE Poland Study. European Journal of Public Health 29: 335-340. [crossref]
37. Pestoni G, Riedl A, Breuninger TA, Wawro N, Krieger J-P, et al. (2021) Association between dietary patterns and prediabetes, undetected diabetes or clinically diagnosed diabetes: results from the KORA FF4 study. European Journal of Nutrition 60: 2331-2341. [crossref]
38. Lee IM, Shiroma EJ, Lobelo F, Puska P, Blair SN, et al. (2012) Effect of physical inactivity on major non-communicable diseases worldwide: an analysis of burden of disease and life expectancy. Lancet 380: 219-229. [crossref]