Utilizing the Glucose and Insulin Response Shape of an Oral Glucose Tolerance Test to Predict Dysglycemia in Children with Overweight and Obesity, Ages 8–18 Years

Round 1

Reviewer 1 Report

Aim:

In their article, Renier et al explore how the information in 5 OGTT measurements of glucose and insulin, including the shape of individual glucose and insulin curves, and HbA1c, is associated with future dysglycemia in children with overweight or obesity.

Praise:

Understanding and analysing the complexity in metabolic regulation calls for exploration of data by different statistical models and methods, often beyond methods in standard, applied statistics text books. Methods from functional data analysis are examples of such methods, and it is commendable that such methods are explored in this article, although originally developed for more densely sampled data than 5 measurements per participant.

The topic: exploring dysglycemia in children 8-18, with Overweight and Obesity is of high importance, both to society and the readers of the journal.

The introduction of the article was very well written and nice to read.

The first part of the aim was intriguing. The second part of the aim was somewhat word-rich and not so clear. I assumed that the aim would be clearer as I read on.

Concerns:

The authors have used several statistical methods and analyses and done quite well, but need to do a thorough review of both the text describing the statistical methods and results, and the corresponding tables and figures. Had they not done such a thorough job in the first place, this would have been an even larger job.

My main objection is that there is general lack of clarity and reflection in both language and results from the statistical analyses presented. But based on the thorough and diverse work presented, the authors are certainly able to fix both.

Major issues:

Mixing analyses of the OGTT measurements directly, and the FDA results. These analyses and results were so intertwined that it was a bit hard to pinpoint my objections, but I’ll do my best:

The authors present the OGTT measurements as they are (FPG, 2hrPG; 30, 60 and 90 minutes glucose, and 0,30,60,90,120 min insulin values), with descriptive statistics and figures (e.g. Fig 1).

The curves in Fig 1 are observed values with interpolation lines, and this should be stated in the figure legend:

Figure 1. Observed oral glucose tolerance test biphasic, monophasic, and monotonically increasing glucose shape classifications. The plot shows interpolated curves for individual measurements (thin lines) and means (bold lines) within each classification group: (A) Biphasic  (B) Monophasic (C) Monotonically increasing.

The authors use the original OGTT measurements to do the crude classification into “biphasic”, “monophasic” and “incessant increase” (as a non-native English user, I had to check this term. Why not use the mathematically well-known term “monotonic”, as suggested above?). By definition, the monophasic group must have the peak measurement at 30, 60 or 90 minutes, which make the result interpretation at line 242, p 7 an obvious observation. Similarly, the incessant group must have their peak at 120 minutes, and the biphasic curves, by definition, need to have a dip at 60 or 90 minutes, also making the statements at l 243 and 244 obvious and equally superfluous observations. However, as the classification was done on basis of the glucose measurements only (was it? The unreflected interpretations just mentioned, of the correlations in l 241-244, cause doubt, so this would be good to clarify), the corresponding insulin profile observations at l 245 and 246 are worth mentioning.

As you may have noticed, in my report I have now suggested a new phrase, “insulin profile”. The reason is that since you also do a functional data analysis, the term “curve” should maybe be reserved for the continuous curves, to avoid confusion.

Then to the description of the FDA, p 4, l 148-168. (Minor comment: The name of the first author of ref 20 is misspelled in one way in the text, and in another way in the reference list. This should be fixed.) Although the mathematical details and notation are important for those doing the analysis, most readers of Diabetology may be somewhat unfamiliar with e.g. matrix notation, and this part of the text should preferably be shortened and written in plain text, with the details moved to the appendix. There also seems to be a slight mix-up in concepts, which makes the text almost correct, but not entirely.

For instance:

L 148-149: “To summarize each participant’s glucose and insulin curves as single functional objects, cubic B-splines [30] were fit to the glucose and insulin observations with the fda package in R [31], using K = 7 basis functions fitted around J = 5 knots (measurements).” Well, not exactly. The glucose curves are not observed, they must be estimated from the observations, and they are not really *summarized* by functional objects, more like *represented* as functional objects. However, an evaluation of which words to use can  also be considered a matter of taste. However, this is characteristic for the rest of the M&M and the entire Results part of the article: It contains a lot ow well done work, but the text is imprecise (from a statistical point of view), sometimes a bit wrong (e.g. R is not a parameter), and this diminishes the entire article.

I therefore strongly advice the authors to team up with a fellow statistician with some background in FDA, to revise and rephrase p 4-9 to a more stringent statistical language. Also, some of the analyses and output should be revised by a statistician. I will point out some examples. Note: When I state questions in my report, I do not expect the authors to reply and argue for their choices, but to make changes in the article, so that these questions need not to be asked.

The way the term “B-spline” is used, cause confusion, e.g. P4, l 151, rewrite to “The estimated curve for each participant is a linear combination of the B-spline basis functions…”

Consequently, please replace “B-splines” with “curves”, “fitted curves” or “estimated curves” at p4, l 160 and P1, l 23. At p4, l 162, 163, the B-spline term is confusing, as it is not clear whether it refers to the original basis functions or the estimated glucose (or insulin) curve.

Output from the gtsummary package: Although the package developers claim to provide publication ready tables, this does not mean the tables are flawless, and a critical revision of the content of each table should be done. Examples:

T1: Why report both mean and median, SD and [min, max] for symmetrically distributed variables like age, when such variables can be sufficiently summarized by the mean and SD? Similarly, for skewedly distributed variables, the median and quartiles (Q1 and Q3), potentially supplemented by min and max, are better summaries than mean and SD.

T2: Presenting the p value for a test comparing age at two visits, in a study where all participants per definition are at least 6 months older at follow-up, I must confess made me giggle. Thanks! But this must of course be removed before publication.

T2: Time between…: Are three decimals (for SD and min) really helpful? Maybe this should be presented in months, rather than years?

In general, for the numerous p values (and consequently, the numerous hypothesis tests) presented in the tables: Are they all necessary? Do each p value answer important research questions? A critical revision of which are necessary/useful would be advisable.

If they indeed are all necessary: Why are all the hypothesis tests done parametrically, although descriptive statistics are given as means (SDs)? This is statistically inconsistent.

Baseline comparisons of participants with or without follow-up involve two groups. Yet, all tests are referred to as Kruskall Wallis. Although it is technically correct to use a Kruskal-Wallis test (or the parametric ANOVA) for comparison of two groups, it may ease the reading to refer to the two-sample version (rank sum test, frequently called Mann Whitney test, or the parametric two-sample t-test).

Major issue: I miss clear formulations/statements about what is the response variable in each of the many regression analyses. The way it is written in the article now, make me wonder which variable is the response variable and which is the covariate/predictor/explanatory variable. This must be clarified.

It is also not clear whether the purpose of the regression analyses were to estimate effects, or to predict dysglycemia (misspelled as “glycemia at P4, l193?), and in case of the former purpose (indicated by the word “confounding”, p 8, l 293), whether a DAG (or better, one DAG per regression model) was considered for each multiple analysis.

Fig 3: How were the shape characteristics included in the analyses? Were they included as three dichotomous dummy variables, each coded 1 for the specified shape, and 0 for any other shape? If yes, why were not the univariate regression results in TS2 and TS3 incorporated similarly in the heat map?

Fig 2: The figure legend informs that this visualization is based on the functional data analyses, i. e. analyses of the continuous, fitted/estimated glucose and insulin curves. However, this figure only shows means of the original measurements, with interpolated lines. That is, not the curves, really. Instead, this figure should show (means of) fitted curves.

TS1: The percentages should be given per row, since the numbers across the row are the numbers we are interested in comparing.

A final request for reflection over results presented: The authors choose three principal components for the glucose curves, and three principal components for insulin, without arguing very much for their choice, except this was chosen in ref [20]. Principal component analysis is data driven and may vary between studies and phenomenon studied. While three components might be justified in the case of glucose in this article, I would have chosen otherwise for insulin. There, the two first components account for 99.2% of the variation in the curve trajectories, and it seems strange to keep and discuss a third component.

These are some examples of why I suggest revising the entire text (and analyses) at p4-9 in collaboration with a biostatistician with FDA experience.

I look forward to the next version of the manuscript. 

Introduction very well written. However, the statistical methods and results section should have a makeover in collaboration with a biostatistician with FDA experience, to enhance clarity and disentangle confusion due to not-quite-correct use of statistical terms and concepts. 

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Good study.

It would have been nice had the authors measured plasma insulin also 

ok

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

This study marks a pioneering use of Functional Data Analysis (FDA) in summarizing OGTT glucose and insulin curves. Its primary aim is to assess the correlation between Functional Principal Components (FPCs) and metabolic health markers, as well as their predictive value for future dysglycemia in a diverse adolescent population. The report is not only meaningful but also engaging.

Several questions arise from the presented information. Firstly, the age range of children included in the study spans from 8 to 18 at baseline and 9 to 20 at follow-up, encompassing both childhood and adolescence due to the inclusion of puberty. Secondly, the cohort initially enrolled 671 patients, but only 193 patients were followed, reflecting a considerable loss ratio in follow-up. A loss ratio of 10% is generally deemed acceptable, underscoring the need to explore and address the reasons behind the higher attrition in this study.

Additionally, the reference to studies in adults highlights the potential of random glucose and 1-hour GCT as promising screening tools for prediabetes and diabetes. The inquiry arises whether any data on 1-hour GCT is available in this study, given its utility in detecting dysglycemia in children and adolescents, as indicated by a report published in Diabetes Care (2011 Dec; 34(12): 2597–2602).

Finally, under the Materials and Methods section, there is a mention of Visits 1 and 3, with an apparent oversight regarding Visit 2. Clarification is needed to confirm whether Visit 2 or Visit 3 serves as the designated follow-up visit in the study's protocol. This clarification is essential for a comprehensive understanding of the study's design and timeline.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

See attached file.

Comments for author File: Comments.pdf

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

The manuscript has been modified according to the comments of review.

Author Response

Please see the attachment

Author Response File: Author Response.pdf