Diabetes & its Complications

Author(s): Howard R. Moskowitz¹, Stephen D. Rappaport², Sharon Wingert³, Taylor Mulvey⁴, Petraq Papajorgji⁵, Kenneth Tomaro¹ and Martin Mulvey¹

¹Cognitive Behavioral Insights, LLC, NY, USA.

²Stephen D. Rappaport Consulting LLC, CT, USA.

³Tactical Data Group, VA, USA.

⁴St. Thomas More School, CT, USA.

⁵Profinit Consulting, Tirana, Albania.

^*Correspondence:

Howard R. Moskowitz, Cognitive Behavioral Insights, LLC, Albany, NY, USA.

Received: 02 Mar 2024 Accepted: 11 Apr 2024

Citation: Moskowitz HR, Rappaport SD, Wingert S, et al. Mind-Sets Regarding Diabetes: How AI + Mind Genomics Thinking Provides an Empowering New Approach. Diabetes Complications. 2024; 8(2); 1-9.

Introduction

The prevalence of technology in the world of food and the increasing abundance of food in many areas leads to a variety of nutritional and other health issues. One of the biggest and perhaps the most looming of all is diabetes. Today's medical world is replete with issues about “diabesity”, the made-up word to describe the “double whammy” combination of diabetes and obesity. The literature is filled with articles on diabetes. Just in the past year alone, in 2023, the number of articles on diabetes exceeded 159,000 according to Google Scholar. The total set of scientific publications referred to in Google Scholar for diabetes is 4,180,000. When we turn to Google itself, the number is almost seven billion, presumably due to the recognition by the public that diabetes is a health problem and at the same time, presents a looming problem in economic welfare. The concerns of people toward the life changes incurred by diabetes are matched by the implications of what it will cost to take care of these people as diabetes becomes a worldwide epidemic.

This study was undertaken as a demonstration of how artificial intelligence — the new large language models such as ChatGPT and others — can provide us with a new way to “understand” the mind of a person who is reading about all the diabetes issues. The underlying idea was that “creating” a mind-set using AI enables the user to understand the topic more deeply. We chose to use the large language models because in the last several years, these models have become increasingly sophisticated. With such sophistication, it appeared that one might be able to harness the power of AI to synthesize what does not yet necessarily even exist, and even present the synthesized information in the form of recognizable questions and answers [1-3].

Mind Genomics and the Organizing Idea of Mind-Sets

Mind Genomics is an emerging science that deals with how we make decisions about our everyday experiences. The organizing principle emerging from Mind Genomics is the recognition that for virtually any topic of daily life, people divide into mind-sets. A mind-set is operationally defined as an interpretable pattern of what is important to the individual. Furthermore, this division of people into mind-sets occurs from the “bottom-up,” viz., at the granular level, rather than representing universal ways of thinking about the world across almost any topic. That is, rather than dividing the entire world of people into a limited number of “general” groups, as some methods do such as Claritas [4], Mind Genomics “creates” these groups for each topic, doing so in a simple, repeatable manner, using well-accepted methods such as regression and clustering [5,6].

The actual process is simple, founded on pure empirical work. The actual Mind Genomics process presents “respondents,” viz., survey takers, with vignettes, viz., combinations of “elements,” i.e., messages about the topic, created by experiment design [7], obtain their rating of the vignettes, and deconstruct the ratings into the part-worth contribution of each element. To the typical respondent these messages look like a blooming, buzzing confusion, but the reality is quite different. What appears to happen, let's say in the case of diabetes, is that the respondent casually inspects a vignette, this combination of elements, and assigns a rating to the combination, often doing so in a state of relaxation, seeming disinterested. That disinterest is generated by the aforementioned frustration at not being able to discover the pattern, and thus not being able to assign the “correct rating.” The final step applies cluster analysis to the set of individual coefficients, one set per each respondent the system generates. From the aforementioned regression analysis, one automatically deconstructs the rating into the part-worth contribution of looks at these vignettes, these combinations, assigns a rating, from that rating emerges scores for every one of the elements, the messages, and from that analysis, the mind genomics statistics ends up clustering these people, these survey takers, into different clusters. The system is rigorous, efficient, and impossible to game [8,9].

To summarize the process, the well-accepted method of clustering ends up dividing people into a limited number of clusters, i.e., groups, based upon the pattern of the coefficients from Ordinary Least Squares (OLS) regression, estimated for each respondent for a specific topic. Thus, clusters comprise individuals whose patterns of responses to these elements are quite similar. The clusters are called “mind-sets,” and the process itself is called “mind-set segmentation.” These groups are easy to interpret. The strong-performing elements, the messages which were part of the vignettes, tell a coherent story, even though the statistical analysis never took into account the meaning of the elements, but just considered the similarity of the patterns of coefficients emerging from the regression and the clustering.

The opportunity facing us was whether we could instruct AI based upon a LLM to tell us about mind-sets for diabetes. AI had been previously incorporated into the Mind Genomics platform, BimiLeap.com, to help users develop questions and answers. The incorporation was through the abovementioned Idea Coach, and is called “Socrates as a Service,” trademarked and abbreviated as SCAS.

Rather than doing the Mind Genomics experiments described above, the approach would be to assume that there exist a specific but unnamed, undefined mind-sets. We would tell the SCAS only that there are a specified number of mind-sets. We would then instruct the SCAS to answer a set of questions for each mind- set. What would the SCAS do in terms of providing meaningful results? Would the “stories” around the SCAS make sense?

Method

The strategy in this project was simply to tell the SCAS that there exist a number of mind-sets relevant to diabetes. The Mind Genomics platform (BimiLeap.com) was used. BimiLeap has a place where it allows the user to type in a request to the linked SCAS, ChatGPT [10]. The “place” to put in one’s request to the SCAS is called the “Idea Coach.” Idea Coach is really a receptacle in which we can put in questions or requests for artificial intelligence.

Table 1 shows the full briefing for discovering the information about mind-sets. If one reads it clearly and easily, one sees that there's no key information about what the mind-set should be, just that there are a certain number of them. Note that the SCAS is given a slight bit of personal information about the user (viz., an 80-year-old). The questions shown in Table 1 are those that a person might ask. There is no request for clinical information. Rather, the type of information requested might be the information for which a person would search on the Internet, or seek in a readable, consumer-oriented publication featuring an article on diabetes.

The important thing to keep in mind when looking at Table 2 is that this is just one of the different mind-sets synthesized by the SCAS. Recall that the SCAS was given no information about these mind- sets and had to synthesize both the name of the mind-set as well as answer the 12 questions. An inspection of Table 2 suggests a high quality of results. The results seem to be meaningful, the results seem to provide new information, and the results make a lot of sense in terms of the topical answers. For example, for the Western Diet, the key messages are that processed foods and added sugars are major contributors, a “factoid” commonly accepted. There is nothing here of a clinical nature, but what's quite exciting is the ability of the SCAS to synthesize all of this information within a period of 15-20 seconds or less, and to do this on a repeated basis should the user want. As we will see in the next sections, there is both a deeper analysis and several more suggested mind-sets.

Once the iterations have been completed and the user provided with the material shown in Table 2 for each mind-set, the study is closed with perhaps 5, 10, 15 or even 20 iterations. For each iteration, SCAS generates nine additional inquiries for higher level post-iteration analyses. These appear in Table 3 for the Western

Diet mind-set. The sections are labeled. The sections range from additional questions to ask, to questions about what's missing, various points of view, and finally, ideas for innovation.

The material returns from post-project “analysis” approximately 15 minutes after the study has been completed. The material for this secondary analysis is presented in the form of an Excel workbook, called the Idea Book. Each iteration generates its own worksheet in this Idea Book. Thus, for a user with the patience to go through 30 different iterations, some more successful than others (taking about 15-20 minutes in total), the Idea Book will comprise 30 worksheets, one worksheet for each iteration. The exercise itself creates a very rich database and encourages critical thinking. Once again, the effort is not to look at clinical materials, but rather to look at the way ordinary people think about a topic, material that would be in the world to which the LLM has access for training.

Discussion and Conclusions

The richness of the data in Tables 2 and 3 are tribute to the increased power of AI, especially the AI generated from large language models. One might say that there is nothing new here, that all the information was already known and that there are no new discoveries to be made from these tables. That would be perfectly acceptable. The objective of this approach is to gather in minutes the types of questions and answers and information that could be not gathered even in months. The objective is to structure critical thinking in such a way that the user can begin to explore the topic with far greater information than was hitherto possible. The tables in the appendix provide questions and answers to an additional eight mind-sets. Not shown are the deeper analyses for each mind-set, material shown in Table 3 for the Western Diet mind-set. Space does not allow for the inclusion of eight additional tables the size of Table 3, which presents the detailed post-iteration analysis by AI of the material generated for one iteration for a specific mind-set.

One can only imagine the types of advances to be made if before each project with people, the investigator were to go through this type of exercise. Perhaps the exercise would not be limited to 10 iterations, but rather with 100 iterations, the effort lasting as much as 4-5 hours. During those iterations the user would end up learning new things, changing the questions to follow a newly emerging line of thinking, and following one’s intuition by changing the query presented in the Idea Coach. Within reason, the effort would itself provide the user with an education, that education turbocharged by the reality of seeing the results of the iteration within 15-20 seconds. The subsequent Idea Book, generated in the 30-60 minutes after the effort is concluded, would be witness to the evolution of one’s thinking. The evolution in the change of questions would be the witness to how the user is deepening knowledge about the topic. The Idea Book would contain those questions in the section dealing with the query (see Table 1), then the results shown immediately (see Table 2), and finally the deeper analyses by AI (see Table 3).

One can only imagine what might happen if this type of thinking were available all around the world for a variety of health and other issues. And that it would simply be a matter of using artificial intelligence, this SCAS, Socrates as a Service, based on large language models, to explore a topic. Take a half a day, or even as much as a day, to run through a hundred or two hundred iterations and see what one gets.

Acknowledgement

The authors gratefully acknowledge the help of Vanessa Marie B. Arcenas for preparing this manuscript for publication.

1. Kasneci E, Seßler K, Küchemann S, et al. ChatGPT for good On opportunities and challenges of large language models for Learning and individual differences. 2023; 103: 102274.
2. Thirunavukarasu AJ, Ting DSJ, Elangovan K, et al. Large language models in medicine. Nature medicine2023; 29: 1930-1940.
3. Wei J, Tay Y, Bommasani R, et Emergent abilities of large language models. arXiv preprint. 2022.
4. Scheufele EL, Hodor B, Popa G, et Population Segmentation Using a Novel Socio Demographic Dataset. Online journal of public health informatics. 2022; 14.
5. Burton AL. OLS regression. The encyclopedia of research methods in criminology and Criminal 2021; 2: 509-514.
6. Milligan GW, Cooper MC. Methodology review Clustering Applied psychological measurement. 2021; 11: 329-354.
7. Cash P, StankoviÄ? T, Štorga M, et al. Experimental design Springer International Publishing.
8. Moskowitz HR. Mind genomics the experimental inductive science of the ordinary and its application to aspects of food and Physiology behavior. 2012; 107: 606-613.
9. Gere A, Bellissimo N, Harizi A, et Non Meat Analogs A Mind Genomics Cartography of their perceived health benefits. Woodhead Publishing. 2023; 569-588.
10. Kalyan A survey of GPT-3 family large language models including ChatGPT and GPT-4. Natural Language Processing Journal. 2023; 100048.