Long-term Relapse of Type 2 Diabetes After Roux-en-Y Gastric Bypass: Prediction and Clinical Relevance

Bariatric surgery (BS) is recommended for patients with a BMI ≥40 kg/m² or ≥35 kg/m² when associated with comorbidities (1), induces major weight loss (2,3), improves glycemic control, and induces partial or complete diabetes remission (DR) (4). DR is defined by fasting plasma glucose (FPG) and HbA_1c normalization (<7 mmol/L and <6% [42 mmol/mol], respectively) without any glucose-lowering medication at 1 year post-BS (5).

BS is superior to intensive medical therapy (i.e., medication and/or lifestyle interventions alone [6,7]) in inducing DR (8,9), with 60–75% of patients experiencing remission after Roux-en-Y gastric bypass (RYGB). Therefore, recent guidelines consider BS/metabolic surgery in the type 2 diabetes (T2D) treatment algorithm for obese patients with lower BMIs (i.e., BMI ≥30 kg/m² with poor glycemic control) (9,10). Experts thus anticipate a dramatic surge in the already large number of BS interventions worldwide (1).

Currently, there is a major concern in the mid- to long-term maintenance of glycemic control post-BS because observational long-term follow-up studies and randomized controlled trials (8,11,12) indicate a decreased rate of DR over time. Nearly half of the patients experiencing DR at 1 year relapse 5 years post-RYGB, with a described mean DR rate of 30%. These relapses are frequently concomitant with weight regain and the deterioration of lipid homeostasis after 1 year (13). These observations signal the need to establish useful and clinically applicable tools to predict metabolic/BS outcomes (9), both in the short and longer term post-BS. Using these tools, we inform patients about what they can expect and adapt patient follow-up according to predicted trajectories.

Scoring systems to predict 1-year DR (1y-DR) generally combine clinical variables mostly related to diabetes severity (14–17). For example, the ABCD score (combining age, BMI, C-peptide, and duration of T2D) (15,18) predicts that patients with a score >6 will not enter 1y-DR; however, our team demonstrated that the ABCD score was less accurate than the Diabetes Remission (DiaRem) score (19). The DiaRem, based on preoperative age, HbA_1c, and the use of glucose-lowering treatments, has 84% predictive accuracy for DR 1 year post-RYGB, but its predictive accuracy is limited in the middle scoring zone (8–17) (19,20). We recently improved the DiaRem predictive performance by proposing the Advanced-DiaRem (Ad-DiaRem) (21), which adds diabetes duration and the number of glucose-lowering agents, including new antidiabetic agents, and optimizes the weight of each item used in the DiaRem (21). The Ad-DiaRem, developed to predict 1y-DR post-RYGB, remains to be tested beyond 1 year of follow-up.

Only a limited number of studies have tested the DiaRem in cohorts with T2D beyond 1 year post-RYGB (22). In 400 patients with T2D, the DiaRem accurately predicted patients with nonremission after 1 year post-RYGB (i.e., all patients with a score >18), but its predictive performance was modest for longer-term DR (5–8 years). Only 50% of the patients with the lowest DiaRem score (0–2) exhibited DR after 8 years (23). When tested in another group, 20% of patients with scores >18 experienced 5-year DR (5y-DR), showing a poor predictive value even in high DiaRem scores (22). Another study with 31 patients with T2D monitored 10 years post-RYGB showed that 30% of patients with 5y-DR exhibited a DiaRem score falling into the “middle zone,” further highlighting the insufficient predictive accuracy for long-term outcomes (24).

These collective findings prompted us to 1) characterize factors involved in 5y-DR and T2D relapse post-RYGB, 2) build an improved scoring system capable of identifying 5y-DR and nonremission at 5 years (5y-NDR), and 3) examine the clinical relevance of this new established scoring system in independent cohorts.

This study included a subselection of patients from our prospective BS cohort (Bariatric Surgery Cohort of the Institute of Cardiometabolism and Nutrition [BARICAN], N = 2,229 patients) at the Pitié-Salpêtrière Hospital Nutrition Department (Paris, France), which was approved by Commission nationale de l’informatique et des libertés (No. 1222666) and the French Ministry of Research. All patients provided informed consent. They are part of several studies registered on ClinicalTrials.gov (P050318 Les Comités de Protection des Personnes approval 24 November 2006, NCT01655017, NCT01454232). These patients met standard BS recommendations (25) and are monitored according to national and international guidelines.

In the current study, we focused solely on severely obese patients with T2D who underwent RYGB and had a mean follow-up of 5 years. We further restricted our analysis to patients with complete clinical data sets pre-BS and 3, 6, and 12 months and 5 years (48–84 months) post-RYGB (n = 175) (Supplementary Fig. 1). In addition to the BARICAN cohort, we had access to three confirmation cohorts, which were composed of patients with T2D who underwent RYGB with 5 years of follow-up and for whom the data needed for score calculation was available. These cohorts consisted of 54 patients in France (Louis Mourier Hospital), 20 patients from a randomized controlled trial in Italy (Catholic University) (8), and 50 patients from Germany (Leipzig Hospital). Patients from all cohorts provided informed consent.

The following data were collected for all patients at every time point, as described (21): the number of glucose-lowering agents; obesity-related comorbidities (hypertension, obstructive sleep apnea, and dyslipidemia); a complete list of treatments; diabetes duration (i.e., duration up to RYGB intervention); blood tests, including lipid panels, liver enzymes, and glucose control (at baseline after a 12-h overnight fast); body composition (total, trunk, and limb fat mass and fat-free mass) evaluated by body DEXA scan (Hologic Discovery, West Bedford, MA), as previously described (26); and depression and anxiety questionnaires (Hospital Anxiety and Depression Scale [27] and Beck Depression Inventory [28]). Weight loss was calculated at each time point using the following formula: (postoperative body weight − preoperative body weight)/preoperative body weight. We used the same approach to quantify body composition changes (in kg and percentages), HbA_1c, and fasting glycemia. We calculated the DiaRem (20) and Ad-DiaRem (21) for these patients.

Using previously described methods (21), we applied an original machine learning methodology (29) to establish the 5y-Ad-DiaRem. We aimed to find 1) the most relevant predictive variables, 2) optimal thresholds per selected variable (i.e., blood glucose 4.8, 5.3, and 5.8 mmol/L for the 1-year fasting glycemia) to construct bins (intervals between two thresholds), and 3) the subscores associated with each bin of patients. An optimal solution occurs if simultaneously learned interpretable binning is mapped to a class variable (5y-DR or NDR) and the weights associated with these bins contribute to the score. We applied a novel statistical machine learning approach called “Fully Corrective Binning” (29). Herein, the problem to learn a score was formulated as a feature selection task, where splitting a bin equates to adding a feature into a model and merging two bins results in deleting a feature from this model. This approach aims to compromise the trade-off between accuracy and sparsity (i.e., the number of bins). The Youden (30) method was used to calculate the thresholds.

Continuous variables are expressed as mean ± SD and categorical variables as number (percentage). Baseline comparisons were performed by ANCOVA with the effect of sex and age as covariates. Multivariate logistic regression was performed to estimate odds ratios for the potential predictors of 5-year relapse and/or nonremission. Continuous data of patients in nonremission throughout the follow-up were compared using the Wilcoxon signed rank test. Categorical data were analyzed using χ² tests for trends.

To build the 5y-Ad-DiaRem, the missing data were imputed by the median of each variable (29). Analyses were conducted using R 3.3.3 software (www.r-project.org). P values were considered significant when <0.05.

Our test cohort included 175 patients with T2D whose characteristics are summarized in Table 1. As shown in Fig. 1A, 61% of severely obese patients with T2D experienced DR (partial [PDR] or complete remission [CDR]) at 1 year post-RYGB. After a mean 5.1 ± 0.7 years of follow-up, DR prevalence decreased to 54% (5y-DR). Thus, 25% of 1y-DR patients (n = 27) relapsed at 5 years post-RYGB (5y-Relapse). In addition, 54 patients (31%) never displayed DR throughout the follow-up (5y-NDR), and 15 patients (22%) who remained with T2D at 1 year further experienced 5y-DR.

We examined patients’ bioclinical differences among the three different trajectories: 1) 5y-NDR, 2) 5y-Relapse, and 3) 5y-DR (PDR and CDR). Compared with 5y-NDR, 5y-Relapse patients (Table 1) and 5y-DR patients had similar baseline characteristics, except they displayed a less severe degree of T2D disease: shorter diabetes duration, fewer glucose-lowering agents, fewer insulin therapy requirements, and better T2D control (i.e., HbA_1c near the targeted goal of <7% [53 mmol/mol]). At baseline, more patients undergoing 5y-DR or later relapse were solely on metformin. By contrast, the use of glucagon-like peptide 1 analogs increased across the three trajectory groups during the follow-up; however, this reflected more severe T2D because patients received multiple glucose-lowering agents (Table 1). The 5y-DR patients were more likely to be women and heavier at baseline, with increased total fat mass but lower trunk fat mass, thus displaying a better body composition. Noteworthy, 27% of patients were taking psychotropic drugs at baseline; however, between trajectory groups, there were no differences in the number of patients treated with these drugs or in the mean prescribed treatment number. Likewise, no difference was seen in scores on the Hospital Anxiety and Depression Scale or Beck Depression Inventory questionnaires (Supplementary Fig. 2 and Supplementary Table 1). These results suggest that depression scores and treatment were not involved in diabetes outcomes post-RYGB.

The DiaRem and Ad-DiaRem (20,21) scores’ mean values were higher in patients with 5y-NDR than those of 5y-Relapse or 5y-DR. As shown on Fig. 1B and Supplementary Fig. 3A, both scores were unable to properly separate patients’ trajectories, displaying a large overlap between 5y-DR, 5y-Relapse, and 5y-NDR patients, especially for scores in the “middle zone” (8–17). Most 5y-Relapse patients had Ad-DiaRem scores <10, as expected, because they experienced 1y-DR. Importantly, whereas we confirmed that the DiaRem and Ad-DiaRem performed well at predicting 1y-DR post-RYGB, they failed to accurately predict 5y-DR. The DiaRem and the Ad-DiaRem displayed nonsignificantly different accuracies (79% and 78%, respectively) and areas under the receiver operating characteristic curve (AUROC) (81% and 84%, respectively) at predicting 5y-DR (Fig. 1C and Supplementary Table 2). Interestingly, although the Ad-DiaRem was inefficient at properly separating 5y-DR and 5y-Relapse patients, it was robust at classifying patients in 5y-NDR, because nearly all the patients with a score ≥15 remained with T2D at 5 years (Fig. 1B).

Because current scores only combining baseline variables were insufficient to accurately predict 5y-DR, we examined the variations of bioclinical variables during the 1st year postsurgery to identify the most discriminant factors between the three trajectory groups (5y-NDR, 5y-Relapse, and 5y-DR). Although 5y-DR patients were heavier at baseline, they lost significantly more weight, total fat, and trunk fat mass at 6 and 12 months post-RYGB compared with 5y-NDR patients (Fig. 1D, Table 1, and Supplementary Fig. 3B). 5y-DR patients lost more limb fat mass and gained more limb fat-free mass (relatively to body weight), suggesting improved body composition. Contrarily, 5y-Relapse patients lost less weight and less total, limb fat, and trunk fat compared with 5y-DR patients (adjusted for age and sex). Losing less weight during the 1st year was more associated with being in 5y-Relapse compared with 5y-DR (odds ratio [OR] 2.66, P < 0.05), but not compared with 5y-NDR (OR 0.63, P = 0.13). Whereas 5y-DR and 5y-NDR had limited but significant body weight regains between 1 and 5 years post-RYGB (+2.5% each, P < 0.05), 5y-Relapse patients displayed a significant and more pronounced weight regain over the same period (+8%, P < 0.001), including regaining fat mass, which resulted in patients reaching their baseline percentage fat-mass level (P < 0.05).

We also observed that glucose control during the 1st year was associated with 5y-DR post-RYGB. Although the three groups improved their T2D, 1-year HbA_1c and 1-year FPG values were associated with 5y-NDR (OR 3.63 [P < 0.001] and 5.64 [P < 0.001]) compared with 5y-DR, respectively) (Table 2 and Supplementary Table 3). However, these clinical variables were not associated with T2D relapse. As expected, 1y-DR status also influenced the 5-year status: being in PDR at 1 year was a risk factor for T2D relapse (OR 5.26, 95% CI 2.1–14.2, P < 0.001); likewise, most of the 1y-NDR patients remained NDR at 5 years (OR 9.80, 95% CI 4.8–21.1, P < 0.001).

To develop our new clinical score, for subsequent use in routine care, we selected the most easily accessible clinical variables among those that significantly differed between the three groups at baseline or 1 year post-RYGB, based on their OR values for being placed in the 5y-Relapse or 5y-NDR groups with 5y-DR as reference (Table 2 and Supplementary Table 3). Hence, the following variables were included: baseline diabetes duration (OR 1.48 [5y-Relapse vs. 5y-DR] and OR 6.73 [5y-NDR vs. 5y-DR]), baseline number of glucose-lowering agents (OR 1.85 and 4.98), baseline HbA_1c (OR 1.19 and 6.69), 1-year remission status, 1-year FPG (OR 1.16 and 1.95), 1-year number of glucose-lowering agents, and weight loss percentage from baseline to 1 year. Furthermore, random forest analysis evaluated the effect of each of those variables in 5y-DR (Supplementary Fig. 4), the most important of which were diabetes duration and weight loss amount.

We then applied machine learning methods to optimize the score (31). As previously described (21), we applied the Fully Corrective Binning approach to attribute the optimal intervals and weight for each variable. This approach led us to identify the optimal score for predicting 5y-DR, called “5y-Ad-DiaRem,” that is to be calculated at 1-year postsurgery (Table 3). An online tool allowing easy calculation of the score can be found at http://5y-Ad-DiaRem.nutriomics.org. Patients’ repartition according to their scoring is shown in Fig. 1E. The 5y-Ad-DiaRem displayed a better negative predictive value (90% vs. 82%) and positive predictive value (80% vs. 67%) than the DiaRem and reassigned 13 of 39 patients misclassified with the DiaRem (i.e., score ≥7 while undergoing 5y-DR or <7 while being in 5y-NDR). Thus, the 5y-Ad-DiaRem displayed increased accuracy (85% vs. 79%) and sensitivity (90% vs. 72%), but the specificity remained unchanged compared with the DiaRem (Supplementary Table 2). Importantly, the 5y-Ad-DiaRem had very good accuracy (>90%) for scores ≤11 to predict long-term remission and >18 to predict long-term nonremission.

To evaluate the relevance of the 5y-Ad-DiaRem in a wide range of patients reflecting routine clinical care in different European countries, we tested the score’s validity in three independent cohorts whose characteristics are summarized in Supplementary Tables 4–6. Whereas the French cohort was similar to the original BARICAN cohort, the German cohort was older and more obese, and the Italian cohort was younger yet consisted of patients with more severe T2D who were mostly men. In the French confirmation cohort, the 5y-Ad-DiaRem performed better than the DiaRem and Ad-DiaRem in predicting 5y-DR (accuracy 89%, specificity 88%, sensitivity 89%, and AUROC 90%) (Fig. 1F, Supplementary Fig. 5A, and Supplementary Table 2). In the other two independent cohorts, the accuracy and AUROC (70% and 85% in the Italian cohort and 90% and 91% in the German cohort, respectively) to predict long-term remission were excellent and confirmed the robustness of our new score in a wide range of BMI and T2D severity stages (Supplementary Fig. 5B and C and Supplementary Table 2).

Whereas the 5y-Ad-DiaRem accurately predicts longer-term DR post-RYGB, we examined its clinical relevance for patients predicted to remain with T2D or to relapse at 5 years postsurgery. As displayed in Fig. 1E, most of the 5y-Relapse patients were correctly classified, with a score >11 (i.e., the fixed 5y-DR threshold). Yet, eight patients who further relapsed had a score <11 and thus were misclassified. Interestingly, compared with 5y-DR patients, these eight patients had well-controlled diabetes at baseline (i.e., 100% with HbA_1c levels on target at <7% [<53 mmol/mol] vs. 65% for the 5y-DR), but they lost less weight than 5y-DR patients during the 1st year (−25% vs. −30%), approaching statistical significance (P = 0.09). Although classified as 5y-Relapse based on biological threshold, five of these eight patients (62%) were free of antidiabetic medications at 5 years post-RYGB. By contrast, the 19 well-classified 5y-Relapse patients (5y-Ad-DiaRem >11) displayed more severe diabetes at baseline (i.e., mean Ad-DiaRem value [9.4 vs. 3.9], P < 0.001) and a reduced, but statistically nonsignificant, weight loss (−21% vs. −25%) compared with the 8 relapse patients with a score <11. At 1 year, those 19 patients were mostly PDR (84%), thus with higher fasting glucose and HbA_1c compared with relapse patients with a score ≤11 who were mostly (75%) in complete DR. Importantly, although they experienced relapse, most patients received oral glucose-lowering agents compared with their previous regimens involving insulin therapy, showing the beneficial effects of RYGB. However, insulin therapy was restarted in two patients in this group.

Examining the 5y-NDR group (n = 54), 81% of the patients were well classified with a 5y-Ad-DiaRem score ≥18, and 100% were >11 (i.e., the fixed 5y-DR threshold) (Fig. 1E). T2D worsened in 30% (n = 16) of the 5y-NDR group at 5 years compared with baseline (as documented with higher HbA_1c values). These patients had a high mean 5y-Ad-DiaRem of 20. Importantly, the other patients of the NDR group (n = 38) improved their metabolic condition, although they remained obese and with T2D (Supplementary Table 3). Their BMI decreased by 10 kg/m² along with improved body composition at 1 year. They subsequently regained a nonsignificant amount of fat mass during the rest of the follow-up. Their diabetes control improved, as seen by reaching a mean HbA_1c of 7.6% (60 mmol/mol) concomitantly with a significant reduction in the number of glucose-lowering agents and a reduction in insulin therapy (i.e., 35 patients were on insulin therapy at baseline vs. 19 at 1 year and 15 at 5 years). Interestingly, the type and number of glucose-lowering agents was not significantly different from 1 to 5 years post-RYGB, suggesting that improvement occurred early and further stabilized. The HbA_1c goal of <7% (53 mmol/mol) was achieved in 34 of 54 patients 1 year post-RYGB, which later decreased to 18 patients at 5 years. These NDR patients also improved their liver functions (alanine aminotransferase and γ-glutamyl transferase), lipid values (triglycerides and HDL cholesterol) despite a reduction in antilipid drugs, and blood pressure as seen by a reduction in antihypertensive drugs.

Interestingly, 15 patients (21%) classified as NDR at 1 year experienced DR at 5 years. These patients were well classified by the Ad-DiaRem, with a score of 8 (thus, below the cutoff of 10 predicting 1y-DR) (21). They exhibited a remarkable weight loss response, because they lost the same amount as 5y-DR patients both at 1 and 5 years.

Here, we propose the 5y-Ad-DiaRem, a new score easily usable in routine care, to predict longer-term (5-year) diabetes status post-RYGB. This score displays a 90% accuracy to predict 5y-DR (threshold ≤11). The 5y-Ad-DiaRem accurately reclassified 13 patients compared with the DiaRem (out of 39 errors) and 12 compared with the Ad-DiaRem (out of 38). The 5y-Ad-DiaRem predictive relevance was confirmed in three independent cohorts from different countries, including a wide range of age and stages of T2D severity before RYGB, thus suggesting its robustness. The use of machine learning methods together with the selection of features by physicians has again (21) displayed its superiority compared with clinical experience (20). Although RYGB enables long-term T2D improvement in most patients, some undergo T2D relapse after 1 year, which is associated with lower weight loss during the 1st year and weight regain afterward, a clinical condition that could be targeted for intervention.

Whereas an abundant literature (8,10,11) demonstrates the efficiency of BS in inducing 1y-DR, studies now mention T2D relapse (8,13,32–34) afterward. Therefore, it becomes crucial to provide accurate mid- to long-term predictors to precisely adapt personalized patient pathways (35). As such, patients predicted to remain with T2D would probably benefit from a rigorous follow-up with a specialized endocrinologist or nutritionist. Alternatively, patients with predicted remission could benefit from general practitioner or specialized nurse counseling (36). This personalized strategy might impact health care costs and patients’ quality of life. Previous studies examining DR and relapse mostly focused on bariatric cohorts with a relatively small sample size in RCTs (<30 operated-on patients) (8,37) or similar sample sizes to ours but on specific subgroups (13). For instance, Arterburn et al. (38) focused on a population of veterans who were mostly males with a higher prevalence of poor glycemic control than the usual bariatric population. Population characteristic differences could eventually explain the wide range of observed relapse rates, going from 17% (13) to 35% of patients undergoing 1y-DR (8,33). Another recent study with 2–12 years post-RYGB follow-up assessed T2D remission rates and showed a decrease from 75% to 51% between years 2 and 12, thus showing decreased beneficial effects over time (39).

Our interest in developing this new 5y-DR prediction score lies in the modest accuracy to predict longer-term DR with previously described scores (15,20), which were constructed to predict 1y-DR (0.78 for both the DiaRem and the Ad-DiaRem) (20,21). Indeed, whereas DiaRem correctly predicted 5y-DR, patients with the higher scores still underwent 5y-DR, suggesting a poor specificity (22). In another cohort examined at 2 and 10 years post-RYGB, low DiaRem scores reasonably predicted 5y-DR. However, a third of their patients fell into the score middle zone (8–17) where the DiaRem displays its poorest accuracy. This results in a 50% chance of patients being misclassified (24). Our current study is in agreement with this limited accuracy of the DiaRem for 5-year prediction.

The ABCD score proved interesting to predict 5y-DR (40), although the predictive accuracy was not calculated. In addition, C-peptide measurements used in this algorithm are expensive and not widely performed in routine care, preventing its generalization. Compared with our 5y-Ad-DiaRem, the lack of accuracy of the DiaRem might originate from the weights that were optimized locally and separately for each variable, whereas here, machine learning methods found a global optimum weighting (29).

Interestingly, the Ad-DiaRem (21) precisely predicted 5y-NDR (i.e., all patients with a score ≥15), yet did not perform well to predict 5y-Relapse. By contrast, the 5y-Ad-DiaRem provided an optimal separation between patients with 5y-DR, 5y-NDR, and 5y-Relapse. The 5y-Ad-DiaRem partly relies on well-known presurgical parameters associated with DR. It is confirmed that the severity of T2D (i.e., HbA_1c [33], the number and type of glucose-lowering agents) and its duration (15,21,33), which are key components of the DiaRem (20), ABCD score (15), Ad-DiaRem (21), and DiaBetter (41), are critical to predict 1y-DR. Diabetes duration is also associated with 2-year DR (15,40,41). However, the use of weight loss and remission status at 1 year improved the predictive value of 5y-DR.

The importance of weight loss during the 1st year has already been associated with 1y-DR (42) and, interestingly, with 5y-DR (22,32,33,41), yet was never included in a predictive score. Furthermore, our current random forest analysis emphasizes the importance of weight loss in long-term remission (Supplementary Fig. 4). A previous study (13) found a 17% prevalence of relapse in a cohort similar to ours. However, their patients lost more weight (despite having similar baseline BMI), which could eventually explain their lower rate of diabetes relapse than that observed in our study. Here, we found not only that 1-year weight loss was significantly different among our three trajectory groups but also that the long-term weight regain between 1 and 5 years (+8%) was associated with higher risk of relapse, as previously seen (13,22,33). Patients who had stable weight in the long term remained in their original status (i.e., 5y-DR or 5y-NDR), and relapse was associated with increased fat mass at 5 years compared with both 5y-DR and 5y-NDR. These observations underline that weight management is key to maintaining T2D remission.

Noteworthy, our results are in line with several studies that observed a weight regain in their cohort (39,43), yet the link with diabetes relapse was not assessed in those studies. By contrast, another RCT (44) monitoring patients for 2 years post-RYGB confirmed that weight stabilization after the 1-year mark enabled the maintenance of the proportion of patients achieving DR. Importantly, diabetes control (HbA_1c, DR status, and the number of antidiabetic drugs) at 1 year was a major predictive factor of longer-term outcomes in our study. Patients who further relapsed mostly exhibited partial DR at 1 year, confirming previous results (33) and suggesting the need to optimize T2D control during the 1st year.

Using the Youden method (30), the 5y-Ad-DiaRem threshold for 5y-DR was determined as a score of ≤11, whereas the cutoff for persistent T2D in the longer term was ≥18. Importantly, 96% of patients with a score ≤11 were free of antidiabetic medication at 5 years compared with only 47% of patients with a score >11 (P < 0.001). Whereas some patients remained with T2D, we show that they generally had improved diabetes control post-RYGB (8,13). However, a small portion of our cohort only exhibited relatively little improvement or even worsened after RYGB.

The strength of our study lies in the development of this new score evaluated in obese T2D patients, representative of usual bariatric cohorts with a myriad of phenotypes, whom were monitored for 5 years. The relevance of the 5y-Ad-DiaRem was confirmed in three independent RYGB cohorts from different countries and, most importantly, with different stages of T2D severity, yet all operated on with RYGB, suggesting its robustness in every clinical setting. Although the number of patients exhibiting diabetes relapse is low, it is concordant with published data (8,13,32–34). The 5y-Ad-DiaRem should now be validated in patients with lower BMI (30–35 kg/m²), because literature demonstrates that BS also enables DR in this population and is associated with weight loss (45).

We acknowledge a few limitations. For score development purposes, the test cohort included 175 patients who had not missed any of the follow-up time points during the 1st year. The 5y-Ad-DiaRem should be further evaluated in an increased number of patients lacking a full 1-year follow-up. Loss of patients to follow-up remains a major issue in the bariatric setting, which can reach 50% of patients at 5 years (46), and has been associated with poor weight loss outcomes. Despite this limitation, the 5y-Ad-DiaRem appeared robust when tested in three independent cohorts from different European countries, yet these were all Caucasian populations. Further validation in other ethnic groups will be of interest. Despite this, the current literature, including Africans Americans, Caucasians, or different Asian ethnic groups, displays no general effect of ethnicity in T2D remission after BS (47,48).

The 5y-Ad-DiaRem, which combines preoperative data with weight loss trajectory and diabetes reoccurrence variables obtained 1 year post-RYGB, is relevant to predict 5y-DR in routine care. Although some patients remained in NDR, they still improved their clinical conditions, as previously seen (8,13). We propose to use the 5y-Ad-DiaRem score as an easily integrated tool to identify patients at risk for longer-term relapse during their 1-year follow-up assessments to propose strategies targeting optimized weight reduction and maintenance to maximize the length of their remission. For example, increasing physical activity should be tested to evaluate whether it could reduce T2D relapse as already seen in the reduced incidence of T2D in glucose-intolerant patients (49). Eventually, for patients identified with the 5y-Ad-DiaRem to remain in long-term NDR add-on strategies could also be proposed to further improve their diabetes control and prevent long-term complications. For example, restrictive dietary interventions could be proposed because they were shown to induce T2D remission along with antidiabetic drugs cessation outside the bariatric setting (50). These add-on therapies should be evaluated in future studies, and in this context dedicated prospective lifestyle interventions are necessary to assess the beneficial impact of these add-on therapies on patients’ metabolic health.

Acknowledgments. The authors thank Dr. Raluca Pais and Dr. Mihaela Docan for seeing some of the patients, Valentine Lemoine (Institute of Cardiometabolism and Nutrition) for patient follow-up, Dr. Florence Marchelli (Centre de Recherche en Nutrition Humaine, Ile-de-France, France) for data management, and Rohia Alili (Centre de Recherche en Nutrition Humaine, Ile-de-France, France) for her contribution to biobanking, as well as the Institute of Cardiometabolism and Nutrition Biology Resource Center members. Flavien Jacques (INSERM) developed the website for easy 5y-Ad-DiaRem score calculation in clinical routine. Tim Swartz (Institute of Cardiometabolism and Nutrition) performed manuscript English editing.

Funding. The clinical study was supported by Assistance Publique Hôpitaux de Paris (APHP) and Centre de Recherche Clinique (CRC) for clinical investigation (CRC P050318 to C.P., PHRC 02076 to K.C., and CRC-FIBROTA to K.C. and J.A.-W.). Agence Nationale de la Recherche supported the clinical study (ANR Adipofib and the national program “Investissements d’Avenir” with the reference ANR-10-IAHU-05) and methodological work (ANR JCJC DiagnoLearn). J.-D.Z. received a Contrat d’Interface grant from APHP and K.C. received an award from Fondation de France. J.A.-W. received a grant from Institut Benjamin Delessert.

The sponsors of the study had no role in study design, data collection, analysis and interpretation, or writing of the manuscript.

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. J.D., N.S., M.C., S.P., R.C., L.G., G.d.T., J.-L.B., C.P., J.-M.O., M.B., M.S., G.M., S.L., J.-D.Z., K.C., and J.A.-W. reviewed the manuscript. J.D., G.d.T., J.-D.Z., K.C., and J.A.-W. contributed to data presentation and the manuscript. J.D., G.d.T., and J.A.-W. analyzed data. J.D., K.C., and J.A.-W. wrote the manuscript. N.S. and J.-D.Z. developed the 5y-Ad-DiaRem. M.C. and S.L. recruited and selected patients in the validation cohort from the Louis Mourier Hospital. S.P. participated in data management of the validation cohort recruited from Italy. R.C., M.B., and M.S. recruited and selected patients in the validation cohort from Germany. L.G. and J.-L.B. operated on patients. C.P., J.-M.O., K.C., and J.A.-W. contributed to patient recruitment and coordinated clinical investigation, patient phenotyping, and sample collection. G.M. recruited and selected patients in the validation cohort from Italy. J.-D.Z., K.C., and J.A.-W. designed the study. K.C. and J.A.-W. are the guarantors of this work and, as such, had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.

Prior Presentation. Parts of this study were presented orally in abstract form at the 34ème Réunion Scientifique Annuelle de l’Association Française d’Etude et de Recherche sur Obésite (AFERO), Lyon, France, 23–24 January 2018; at Congrès 2018 de la Société Francophone du Diabète, Nantes, France, 20–23 March 2018; at Congrès Annuel de la Société Française et Francophone de Chirurgie de l'Obésité et des Maladies Métaboliques (French bariatric national meeting), Nancy, France, 24–26 May 2018; and as a poster at the 25th European Congress on Obesity, Vienna, Austria, 23–26 May 2018.