Computational modelling of self-reported dietary carbohydrate intake on glucose concentrations in patients undergoing Roux-en-Y gastric bypass versus one-anastomosis gastric bypass

The prevalence of obesity is increasing at alarming rates and, according to estimates, by year 2030 over 500 million adults worldwide will struggle with obesity [1]. Obesity is a global public health priority [2,3], as it can lead to various physical and metabolic comorbidities and thereby increase the risk of mortality [4,5].

Leading to sustained weight loss and improved survival, bariatric surgery is the most effective treatment of morbid obesity [6,7]. The first paper on gastric bypass was published in the 1960s [8]. Thereafter, several modifications to the bypass techniques have been applied. The laparoscopic Roux-en-Y gastric bypass (RYGB) technique was introduced in 1994 [9]. Since then, the RYGB has been considered a gold standard in bariatric surgery. In 2001, the first series of one anastomosis gastric bypass (OAGB), were published [10]. A number of studies have reported superior weight-loss and diabetes remission related to the OAGB [11–14].

In addition to weight loss, reduced plasma glucose concentrations are observed after gastric bypass procedures [15]. Indeed, higher peak glucose concentrations and more rapid decline to or below basal blood glucose concentration, have been well described [16]. Various phenomena, including improved hepatic and peripheral insulin sensitivity, improved beta-cell glucose sensitivity, increased insulin secretion, reduced basal glucose production and reduced carbohydrate intakes contribute to the improved glycaemia following the operation [15]. Of clinical importance, in individuals with obesity and type 2 diabetes, major long-term improvements in glycaemic control and resolution of diabetes are commonly reported [17,18]. To the best of our knowledge, the acute metabolic effects and the short-term glycaemic responses to the dietary carbohydrate intake, comparing RYGB and OAGB, have not been investigated.

Combining the often noisy dietary data with high-frequency time-series from continuous glucose monitoring devices forms a major statistical challenge. To address this, we adapted a previously published machine learning technique [19], with a time-series model incorporating noisy covariates. An important feature of the model is its ability to correct for the uncertain user-reported meal times, without which the relation between nutrient intake and glucose time-series would be obscured. This allows us to precisely quantify the size of the glucose response in terms of the height and width of the response curve, for a given quantity of carbohydrate intake.

In this paper, our aim was to investigate the use of machine learning for modelling the real-life postprandial glucose concentrations in morbidly obese patients undergoing RYGB or OAGB.

This study is part of an open-label prospective randomized controlled RYSA trial (registered in www.clinicaltrials.gov↗NCT02882685↗ on 30 August 2016), where patients were randomized to undergo either RYGB or OAGB surgery in one of the two sites in Finland (Helsinki University Hospital and Oulu University Hospital) [20]. The randomization procedures have been previously explained in detail [20]. For investigating the applicability of artificial intelligence for modelling glycaemia, we included data from patients without diabetes operated between November 2016 and April 2019 at Helsinki University Hospital. To be included in the analyses, patients needed to have flash glucose measurements, food diaries, and weight measurements at baseline and 2 weeks postoperatively (Figure 1). The patients operated at Oulu University Hospital were excluded as they did not use the flash glucose measurement devices. The study protocol was approved by the Ethics Committee of Helsinki and Uusimaa Hospital District (HUS/1706/2016) and by Helsinki University Hospital research review board (HUS269/2017). All procedures were conducted in accordance with the 1964 Declaration of Helsinki and its later amendments. Informed consent was obtained from all study participants prior to their inclusion in the study.

Data from 17 participants (8 men, mean ± SD age 47 ± 8 years, age range 28–56 years) were included in the study. Ten (6 men) and 7 (2 men) patients were randomized to RYGB and OAGB, respectively (eTable 1↗). The two groups were comparable with respect to a number of baseline variables, including BMI, fasting serum lipid and lipoprotein concentrations, HbA_1c, hsCRP, and fasting glucose and insulin concentrations. Pre- and post-operation weight losses were comparable between the groups. Besides the higher preoperative fat (E%) intake in those randomized to the RYGB, the dietary intakes were also comparable (eTable 2↗).

Carbohydrate intakes and glucose concentrations reduced significantly in both groups following the operation (eTable 2↗ and eFigure 1↗). The maximum glucose concentration decreased in the RYGB group (from 10.0 ± 1.8 to 8.8 ± 0.9 mmol/l, p = .042) and the minimum glucose concentration decreased in the OAGB group (from 3.5 ± 0.4 to 2.4 ± 0.3 mmol/l, p = .002, eTable 2↗). Variability of the glucose concentrations were comparable prior and after the operation, in both groups.

Following the operation, we observed a marked shift in the glucose profile in both groups (Figure 2). In particular, the relative time spent in hypoglycaemia (<4.0 mmol/l) increased in both groups, and the times spent in ≥5 to <6, ≥6 to < 7, and ≥7 mmol/l correspondingly decreased (Table 1). The surgery-induced reduction of the lowest glucose concentration was significantly greater in the OAGB group compared to the RYGB group (from 3.5 ± 0.4 to 2.4 ± 0.3 mmol/l vs. 2.6 ± 0.6 to 2.5 ± 0.4 mmol/l, p = .003, Table 1). Other than this, the two groups were comparable.

Using artificial intelligence based modelling of the longitudinal data [19], we investigated the glucose responses following the meals. Figure 3 shows the glucose measurements from the FreeStyle Libre (dots), the carbohydrate intakes from each meal (bars), and the fitted response curve (line) that represents the relationship between these two for one example participant from both groups before and after the operations. The response curves are estimated for each patient separately and they allow us to infer the parameter βp for each patient. This parameter tells how much the height of the response curve increases postprandially for each gram of carbohydrate. Figure 4(a,b) show the βp values separately for each patient in both surgery types. As seen in the figure, the value of the β_p parameter increased considerably in almost all individuals regardless of the surgery type. Thus, following the operation, the same amount of carbohydrates resulted in a larger increase in the postprandial glucose response. The average parameter values of the estimated response curves, that is the average height parameter of the glucose peak (β) and the average width (duration of the peak) (α) across all cases are summarised in eTable 3↗. Following the operation, the height of the glucose response increased (increased β) but the duration of glucose response was shortened (decreased α) in both surgery types. Namely, the mean value of β increased from 0.046 to 0.086 mmol/l/g in RYGB, and from 0.034 to 0.088 mmol/l/g in OAGB, following the operation. Furthermore, mean value of α (length-scale of the glucose response) decreased from 22.80 to 19.60 min in RYGB, and from 20.56 to 18.33 min in OAGB, postoperatively. The average glucose response to a unit of carbohydrate intake is depicted in Figure 4(c,d), and shows increased height and decreased duration of the response in both surgery types. While the 95% confidence intervals for the difference between these parameters in the two groups are non-overlapping, their values are nevertheless relatively close to each other, indicating that there are no clinically meaningful differences between the two surgeries in their glycaemic responses to meals.

In a sample of morbidly obese individuals undergoing either RYGB or OAGB, using a novel method for estimating the impact of carbohydrate intake on glucose concentration we observed, in a real-life setting, that following the operation the same amount of dietary carbohydrates resulted in a larger increase in the postprandial glucose response. Moreover, while the height of the glucose response increased postoperatively, the duration of the response was significantly shortened. The changes in glucose responses were similar between the two operations, suggesting that in subjects with morbid obesity but without diabetes, both RYGB and OAGB have comparable short-term effects on glycaemia.

To the best of our knowledge, no prior studies have been conducted comparing the acute glucose effects of RYGB and OAGB. In their study, Lazar et al. reported that, in long-term, symptomatic and asymptomatic hypoglycaemias are common observations following both RYGB and OAGB [23]. However, the authors provided no information on the operations. Therefore, it is unknown whether the lengths of the gastric bypasses were different in that study, and whether the operations are thus comparable. Consistent with those observations, however, the time spent in hypoglycaemic levels also increased in the current study. In a retrospective study of obese participants with type 2 diabetes, Almalki et al. reported that compared to RYGB, OAGB resulted in greater remission rate [24]. However, as continuous glucose monitoring was not conducted, no data were available on more acute fluctuations of the blood glucose levels. In the current study, we observed no major differences in acute effects on glucose metabolism between the two operations. Importantly, in both our study and in that of Almalki et al., bypassed limbs were of identical length (80 + 130 cm in RYGB and 210 cm in OAGB). Moreover, in the current study, the gastrojejunostomies were equally wide in both groups. While not different in short-term, the long-term glycaemic effects of the two operation types, using continuous glucose measurements, will need to be assessed in the future.

A number of mechanisms may explain the postoperative changes seen in the glucose responses, including reduction in dietary intake and entering of the less-digested chyme in larger-than-normal quantities into the small intestine and resulting sudden absorption of nutrients [25]. In addition to the observed immediate postprandial increases in the glucose concentrations, it was also evident that the duration of the postprandial glucose peak was significantly shortened, suggesting improved glucose clearance from the circulation. This phenomenon may partly be explained by the functional changes taken place in the remaining small intestine, such as hyperstimulation of the L-cells and subsequent increase in the secretion of glucagon-like peptide-1 (GLP-1), with known glucose-lowering effects due to increased insulin secretion [26], and improved hepatic insulin sensitivity [27].

Considering that the current participants did not have diabetes, the clinical implications related to the changes seen in the glycaemic responses are not known. While it is likely that fast clearance of glucose from the circulation is beneficial, high initial glucose peak has been associated with higher intima-media thickness [28,29], and higher pulse pressure [30], which are known cardiovascular risk factors. On the other hand, the observed reduction in overall glycaemia is likely beneficial. Reduction in glycaemia was accompanied by an increase in time spent in hypoglycaemia after both operations. This increase in hypoglycaemia, related to gastric bypass surgeries, has also previously been reported in other samples of nondiabetic individuals [31]. Importantly, in healthy subjects, hypoglycaemias increase the levels of pro-inflammatory cytokines, markers of lipid peroxidation, reactive oxygen species, and leucocytosis [32]. In individuals with type 2 diabetes with high risk of cardiovascular complications, instead, hypoglycaemic events increased the risk of cardiac arrhythmias, including bradycardia, atrial ectopic activity, and ventricular premature beats [33]. Despite the observed increases in the above risk factors, bariatric surgery has been shown to reduce the risk of cardiovascular events by approximately 50% [34]. It is, therefore, likely that the surgery-associated beneficial changes related to HbA_1c, weight loss, blood pressure, and concentrations of triglycerides, LDL cholesterol, and HDL cholesterol, overweigh that of daily glycaemic fluctuations [35,36].

The results of this study were made possible by utilizing a machine learning method that explicitly corrected for the noisy user-reported meal times when modelling the time-series from continuous glucose monitoring. In a number of previous studies, factors such as measurement errors related to dietary intake have been addressed [37–39], and in order to improve the quality of the collected dietary data, various statistical approaches have been designed, including the regression calibration [40,41]. On the other hand, the timing uncertainty, accounted for in the current model, has rarely been addressed within this context, while its impact on the accurate estimation of the glucose response may be more dominant [19]. Hence, besides the clinical observations, this study also demonstrated the capability of such a model to yield additional insight into complex data not available by conventional statistical models. Furthermore, the model is likely helpful in applications where the goal is to model time-series data that depend on noisy covariates, for example studies combining patient-reported data with any time-series measurements.

Thorough investigation of the study participants, their randomization into the surgical operations, use of bypass limbs of identical length, modelling the effect of dietary carbohydrate on the glucose curves in response to bariatric surgery, and conducting the study in a real-life setting are important strengths of the study. Although the study sample was small, it is sufficient for investigating the applicability of the novel statistical method used in the current study, as also evidenced by the uniform observations among the study subjects. The reported total energy intake, especially prior to the operations, was low and, indeed, it is well established that individuals with obesity tend to under-report their dietary intake [42]. However, due to a small sample size, we could not afford to exclude any participants based on their reported energy intake. Regardless of the potential underreporting, distinct differences in glycaemic responses before and after the operations were observed, suggesting that the model worked well using the collected data. Moreover, while it is acknowledged that also other dietary constituents, such as fibres, glycaemic index, fats, and proteins impact the postprandial glycaemic fate, only carbohydrate intake was taken into account, in the model. In principle, including additional covariates is straightforward. In practice, however, this leads to some challenges, e.g. unsatisfactory convergence of the Markov chain Monte Carlo algorithm [43] used for model fitting, nonidentifiability of the model with correlated covariates, and consequent lack of interpretability of the results. Given that carbohydrate intake is a well-known causative dietary factor for the glucose response, simplifying the model accordingly ensures the robustness of the findings. Another simplifying assumption is that the model uses a constant response baseline, i.e. glucose measurements between meal times. While this is not physiologically fully accurate, it nevertheless gives a reasonable fit to the data (Figure 3). In addition, our aim was to estimate the causal effect of carbohydrate intake on the preoperative and postoperative glucose responses. Using a varying baseline (e.g. a slow varying Gaussian process regression [19], would make it difficult to separate the variation in the baseline from the variation caused by the actual input (food). Thus, employing a simple constant baseline enables the glucose response to the carbohydrate intake to be estimated. Further improvements of these aspects are within our planned future research. Finally, due to the exclusion of individuals with diabetes, the results are not confounded by diabetes medication or glucose toxicity. Subsequently, our observations may not be directly applied to those with diabetes.

In conclusion, in this study we demonstrated the utility of a novel technique in a clinical study that combined patient-reported and time-series data. In these analyses, marked increase in postprandial glucose concentrations, rapid clearance of glucose from the circulation, and increase in time spent in hypoglycaemia were observed after both RYGB and OAGB, with no clinically significant differences between groups. Whether nondiabetic individuals undergoing gastric bypass surgeries would benefit from monitoring the post-surgery hypoglycaemias and the potential to prevent them by dietary means should be investigated.

The study was funded by the Academy of Finland [grant numbers 335443, 314383, 266286] and the Academy of Finland, Centre of Excellence in Research on Mitochondria, Metabolism and Disease (FinMIT), [grant number 272376]; Finnish Medical Foundation; Gyllenberg Foundation; Novo Nordisk Foundation, [grant numbers NNF20OC0060547, NNF17OC0027232, NNF10OC1013354]; Finnish Diabetes Research Foundation; Orion Research Foundation; Finnish Foundation for Cardiovascular Research; University of Helsinki and Helsinki University Hospital; Government Research Funds. TS was funded by Martti I Turunen Foundation; Finnish Medical Foundation; and Mary and Georg C. Ehrnrooth Foundation. AJ was funded by the Academy of Finland [grant numbers 335447 and 341157].

AJ, TS, and KHP designed the RYSA trial and collected the data. AJ and TS recruited the patients. All authors substantially contributed to the conception and design of the current work. PM, RAA, AJA, and KHP designed the analyses. RAA and AJA conducted the analyses and drafted the first version of the manuscript. All authors contributed to the interpretation of data, and revision of the work critically for important intellectual content. All authors have approved the final version of the manuscript to be published and agree to be accountable for all aspects of the work.

No potential conflict of interest was reported by the author(s).

According to the General Data Protection Regulation of the European Union (679/2016), the principles of data protection should apply to any information concerning an identified or identifiable natural person and that personal data which have undergone pseudonymization, which could be attributed to a natural person by the use of additional information should be considered to be information on an identifiable natural person. Thus, according to the GDPR, all pseudonymized data are considered personal data and cannot be published openly. Therefore, we are bound to the law and to the strict hospital policies, and are unable to share the data. However, the institutional (Helsinki and Uusimaa Hospital District) contact details for potential future data requests are as follows: https://huspalvelu.microsoftcrmportals.com/fi-FI/↗.