Edoardo Faggionato (1); Michele Schiavon (1); Laya Ekhlaspour (2); Bruce A. Buckingham (3); Chiara Dalla Man (1)
(1) Department of Information Engineering, University of Padova, Padova, Italy; (2) Pediatrics School of Medicine, University of California, San Francisco, CA, USA; (3) Department of Pediatric Endocrinology, Stanford University, Palo Alto, CA, USA
Objectives:
Individuals with type 1 diabetes (T1D) have to constantly monitor and control their blood glucose (BG) levels through the administration of exogenous insulin since the physiological insulin feedback loop is irreversibly impaired. However, optimal dosing of this hormone is notoriously a challenging task, due to the presence of several disturbing factors affecting BG, such as physical exercise, psychological stress, and food intake. Current insulin dosing strategies require the individual to estimate the carbohydrate (C) amount of the meal to calculate the prandial bolus, without accounting for other macronutrients, in particular for fats (F) and proteins (P) that are known to affect postprandial gastric retention (GR), glucose rate of appearance (Ra) in blood, and insulin sensitivity (SI) [1]. Such variables can be estimated, in clinical settings, from plasma glucose and insulin concentrations, using the Oral glucose Minimal Model (OMM) [2] or, in real-life conditions, from minimally-invasive devices like continuous glucose monitor (CGM) and insulin pump (IP), using the recently-developed Minimally-Invasive OMM (MI-OMM) [3].
Quantifying the effect of meal composition on postprandial glucose excursion would be key in optimizing insulin therapy for T1D management. Therefore, this work aimed to apply the MI-OMM to quantify the effect of macronutrients on clinical variables, such as GR, Ra, and SI, in real-life scenarios.
Methods:
One-hundred-and-twenty participants with T1D (age=15.5±11.5 years, body weight, BW=51.3±28.0 kg) were recruited in three different clinical centers. Each participant underwent one admission ranging between 3 and 5 days in a supervised rental home setting under free-living conditions, wearing a CGM, to collect BG data, and an IP, for insulin administration. Meal timing and content (amount in grams of C, F, and P) were recorded with the assistance of the medical personnel.
From the total of 1211 recorded meals, 261 prandial CGM traces were extracted from 1 h before to 4 h after the time of the meal. These were selected with the following criteria to avoid confounding factors in the analysis: i) more than 85% of the CGM measurements available in the time window; ii) available information on the content of C, F, and P; iii) no other meal recorded within 5 h before and 4 h after the current meal; iv) no exercise session recorded within 2 h before and 4 h after the current meal; v) no unphysiological pattern of CGM profile (e.g., BG rise before the meal); vi) no evident unreported calibrations. Meals in this dataset were classified into low vs. high F content (LF vs. HF) and low vs. high P content (LP vs. HP), trying to keep each H- and L-class sufficiently matched in terms of numerosity, BW, and age of the population. Three different types of classification were attempted: absolute amount, amount per kg of BW, and percentage of F and P in the meal, trying to find a standardized classification for meals among the wide and heterogeneous population of enrolled volunteers.
The MI-OMM was identified on the extracted CGM traces using a Bayesian Maximum A Posteriori estimator implemented in Matlab [4]. Since CGM data are autocorrelated, the covariance matrix of the residual error was constructed using Yule-Walker’s equation [5] for an autoregressive model of order 2. Model assessment was done by checking the distribution of weighted residuals, the precision and the physiological plausibility of parameter estimates.
Thereafter, model-estimated GR, Ra, and SI in HF vs LF and HP vs LP were compared. In particular, SI, the half-life of the GR curve (TGR50%), and the area under the Ra profile in the first 2 h after the meal (AUR2h) were calculated and their distributions were quantitatively compared using the Wilcoxon’s rank-sum statistical test.
Results:
The model was able to well describe the CGM profiles and generally provide physiologically plausible parameter estimates with good precision. Weighted residuals were reasonably uncorrelated passing the runs test for randomness in 186 out of 261 sessions (71%).
None of the 3 meal classifications was completely free of drawbacks; however, the classification based on the absolute amount of F and P in the meal proved to be sufficiently robust, generating classes that were well separated in terms of macronutrient content and a quite homogenous distribution of the individuals among classes. Hence, a meal with F content <15 g was labeled as LF meal (n=108), while a meal with F content >25 g was labeled as HF meal (n=101). Similarly, a meal with P content <15 g was labeled as LP meal (n=91), while a meal with P content >30 g was labeled as HP meal (n=95). Meals that fell neither into the H- nor the L-class were considered medium content and were not used in the comparison.
The comparison of the results obtained for different classes of meals was consistent with the literature [1]. In particular, results showed that a high presence of F or P in the meal significantly slows both GR and Ra and reduces SI. SI dropped from 9.6[5.8-12]·10−4 min−1/(μU/mL) (median[IQR]) for a LF meal to 6.8[4.1-11]·10−4 min−1/(μU/mL) for a HF meal (p=0.01), and from 9.4[6.3-12]·10−4 min−1/(μU/mL) for a LP meal to 6.3[3.7-11]·10−4 min−1/(μU/mL) for a HP meal (p=0.001); the TGR50% increased from 90[65-120] min for a LF meal to 108[81-134] min for a HF meal (p=0.005), and from 89[67-121] min for a LP meal to 105[82-128] min for a HP meal (p=0.023); and the AUR2h reduced from 0.48[0.38-0.62] for a LF meal to 0.41[0.32-0.51] for a HF meal (p=0.008), and from 0.48[0.38-0.63] for a LP meal to 0.41[0.32-0.49] for a HP meal (p=0.005). Of note, F seems to have a slightly stronger impact on GR compared to P, whereas P showed a stronger influence on SI.
Conclusions:
Limitations of this work come mainly from the real-life scenario where the data were collected, hampering the identification of the model (e.g., need to remove CGM autocorrelation) and difficulties in separating F and P effects due to the concomitant presence of both macronutrients in the meal (Spearman’s correlation of F and P amount in the meals was 0.73).
To the best of our knowledge, this is the first analysis where the effect of meal compositions on postprandial glucose excursion was precisely quantified with the use of a model-based methodology in real-life conditions. These results can be used to re-design current insulin therapies making them account also for the presence of macronutrients in meals other than carbohydrates, or in more advanced frameworks, the model itself could be incorporated in algorithms for BG control, or used to inform machine learning techniques for BG prediction. The next steps include the analysis of the MI-OMM with nonlinear mixed effects modeling to incorporate the information on meal composition to describe the variability of model parameters.
References:
[1] C. E. M. Smart, B. R. King, and P. E. Lopez, “Insulin dosing for fat and protein: Is it time?,” Diabetes Care, 2019
[2] C. Dalla Man, A. Caumo, and C. Cobelli, “The oral glucose minimal model: Estimation of insulin sensitivity from a meal test,” IEEE Trans Biomed Eng, 2002
[3] E. Faggionato, M. Schiavon, L. Ekhlaspour, B. A. Buckingham, and C. Dalla Man, “The minimally-invasive oral glucose minimal model: Estimation of gastric retention, glucose rate of appearance, and insulin sensitivity from type 1 diabetes data collected in real-life conditions,” IEEE Trans Biomed Eng, 2024
[4] Matlab, version R2020a (The MathWorks Inc., Natick, Massachusetts, USA)
[5] G. T. Walker, “On periodicity in series of related terms,” Proc. R. soc. Lond. Ser. A-Contain. Pap. Math. Phys., 1931
Reference: PAGE 32 (2024) Abstr 11112 [www.page-meeting.org/?abstract=11112]
Poster: Drug/Disease Modelling - Other Topics