Skip to main content
Download PDF

Predicting responsiveness to GLP-1 pathway drugs using real-world data

BMC Endocrine Disorders volume 24, Article number: 269 (2024) Cite this article

Abstract

Background

Medications targeting the glucagon-like peptide-1 (GLP-1) pathway are an important therapeutic class currently used for the treatment of Type 2 diabetes (T2D). However, there is not enough known about which subgroups of patients would receive the most benefit from these medications.

Objective

The goal of this study was to develop a predictive model for patient responsiveness to medications, here collectively called GLP-1 M, that include GLP-1 receptor agonists and dipeptidyl peptidase-4 (DPP4) inhibitors (that normally degrade endogenously-produced GLP-1). Such a model could guide clinicians to consider certain patient characteristics when prescribing second line medications for T2D.

Methods

We analyzed de-identified electronic health records of 7856 subjects with T2D treated with GLP-1 M drugs at Vanderbilt University Medical Center from 2003–2019. Using common clinical features (including commonly ordered lab tests, demographic information, other T2D medications, and diabetes-associated complications), we compared four different models: logistic regression, LightGBM, artificial neural network (ANN), and support vector classifier (SVC).

Results

Our analysis revealed that the traditional logistic regression model outperforms the other machine learning models, with an area under the Receiver Operating Characteristic curve (auROC) of 0.77.Our model showed that higher pre-treatment HbA1C is a dominant feature for predicting better response to GLP-1 M, while features such as use of thiazolidinediones or sulfonylureas is correlated with poorer response to GLP-1 M, as assessed by lowering of hemoglobin A1C (HbA1C), a standard marker of glycated hemoglobin used for assessing glycemic control in individuals with diabetes. Among female subjects under 40 taking GLP-1 M, the simultaneous use of non-steroidal anti-inflammatory drugs (NSAIDs) was associated with a greater reduction in HbA1C (0.82 ± 1.72% vs 0.28 ± 1.70%, p = 0.008).

Conclusion

These findings indicate a thorough analysis of real-world electronic health records could reveal new information to improve treatment decisions for the treatment of T2D. The predictive model developed in this study highlights the importance of considering individual patient characteristics and medication interactions when prescribing GLP-1 M drugs.

Key points

1. Patient characteristics such as poorer blood glucose control, higher body mass, and shorter duration of diabetes predict better response to medications that target the GLP-1 pathway.

2. Simultaneous use of NSAIDs (for example ibuprofen) was associated with better responsiveness in women under 40.

3. Combining GLP-1 pathway medications with some other commonly used T2D medications (for example thiazolidinediones or sulfonylureas) may not have an additional benefit.

Peer Review reports

Introduction

Over 460 million people worldwide have diabetes [1]. In the United States, 10.5% of the population has diabetes, with about 90–95% of those having T2D [2]. Metformin is the preferred first-line treatment for.

T2D. However, due to the progressive nature of diabetes, a second-line medication is commonly required at some point in the course of the disease to maintain glucose homeostasis. When combined with metformin, on average all non-insulin medications lower hemoglobin A1C (HbA1C) by about 0.7–1.0% [3]. However, at the individual level, patients are known to respond differently to different medications [4]. Moreover, many factors contribute to diabetes etiology and responsiveness to medications, including patient characteristics and environmental factors [5]. Unfortunately, there are few widely used biomarkers to aid in predicting individual patient responses [6]. Therefore, it is important to understand real-world patient responses to different combinations of medications to enhance personalized treatment for T2D.

When cost is not an issue, the ADA recommends choosing a second-line medication from one of the following: dipeptidyl peptidase-4 inhibitors (DPP4i), glucagon-like peptide-1 receptor agonists (GLP-1-RA), or sodium/glucose co-transporter inhibitors (SGLT2i) [3]. Among these options, the incretin-based drugs GLP-1-RA and DPP4i (which we will collectively call GLP-1 M) are widely used. From 2013 to 2016, 23% of T2D patients in the US received DPP4i, and 6% of patients received GLP-1-RA [7]. A recent recommendation published by a group of diabetologists suggests that GLP-1-RA are currently under-prescribed [8], and thus, use of GLP-1 M is likely to continue to increase in the future. GLP-1-RA mimic the actions of GLP-1, binding the GLP-1 receptor on the surface of β cells (and other cell types) and stimulating insulin secretion in the presence of elevated glucose. DPP4i block the activity of the enzyme that normally degrades endogenously produced GLP-1, thus, increasing its half-life in the circulation.

Importantly, despite their wide-spread usage, patient responses to GLP-1 M are heterogeneous. A study that analyzed the effects of GLP-1-RA demonstrated that patients with baseline HbA1C of 7–8% had only a 40%−60% probability of reducing HbA1C to < 7%, after 12-months of GLP-1 M treatment [9]. In addition to differences in patient genetics and pathophysiology, drug-drug interactions could contribute to the heterogenous response to GLP-1 M. GLP-1 M act through a G protein coupled receptor (GPCR) that couples to Gαs. Islets express nearly 300 GPCRs [10] and GPCRs represent 35% of all current drug targets [11]. It is highly possible that activity of other GPCRs, and medications that affect their signaling, modulate GLP-1 M activity in β cells.

In this study, we aimed to develop a predictive model for patient responsiveness to GLP-1 M treatment. We analyzed the de-identified electronic health record (EHR) database entries from 2003–2019 at Vanderbilt University Medical Center (VUMC). Subject demographics (for example: age, sex, self-identified race), lab values (for example: HbA1c, BMI, liver enzymes), and some medication information (GLP-1 M, other drugs used to treat T2D, and NSAIDs) and presence or absence of diabetes-associated complications (for example: chronic kidney disease, cardiovascular disease) were extracted as prediction features. We compared logistic regression, LightGBM, artificial neural networks (ANNs), and support vector classifier (SVC), to identify the best model for predicting GLP-1 M responsiveness. We found logistic regression to be the best model. Pre-treatment HbA1C was the most important prediction factor for responsiveness. We also found associations between usage of other T2D medications or NSAIDs with GLP-1 M responsiveness.

Methods

Data source and ethics

This study was approved by the VUMC Institutional Review Board (IRB; #191,602) in Nashville, TN. All data are from the VUMC Synthetic Derivative (SD) database, a de-identified copy of the main hospital medical record database created for research purposes. The SD contains over 3.2 million electronic records, with no defined exclusions. No HIPAA identifiers are available in the SD. Dates, such as “January 1, 2004” have been replaced with a randomly generated date, such as “February 3, 2003”. To maintain data integrity for each subject, a single random time interval was generated, and then used to shift all the dates for the same subject. In this way, all the time intervals of subject clinic visits and treatments remain intact.

Defining the subject cohort for analysis

In this study, ICD codes were used to identify the disease. The International Classification of Diseases (ICD) 9 and ICD10 codes used to define T2D are listed in Supplemental Table 1. All other ICD and CPT (Current Procedural Terminology) codes used to identify the other diseases and procedures are listed in Supplemental Table 2. To be included, each diagnosis required that the ICD code be present a minimum of twice in the subject’s record.

Table 1 Mean values ± SD of subjects’ continuous features considered
Full size table
Table 2 Mean values ± SD of subjects’ discrete features considered
Full size table

At the time of our analysis, there were 178,340 subjects in the SD with at least one ICD code for T2D (see Supplemental Table 1). Figure 1A shows the strategy used for cohort selection. Subjects with T2D were identified using a published algorithm based on lab tests, prescriptions, and ICD codes [12]. This algorithm was previously validated across five institutions and demonstrated a positive predictive value (PPV) of 98%. We further restricted our analysis to subjects with T2D whose medical records also met the following requirements: 1. included at least two HbA1C measurements spanning at least one year; 2. included at least one ICD code for T2D (Supplemental Table 1); and 3. the first T2D code occurred after age 22 years. These criteria identified 33,428 subjects with T2D. Among these, one-third (11,604 subjects) had at least one record of GLP-1 M treatment. We considered a subject as having been treated with GLP-1 M only if there were at least two different medication entries referring to GLP-1 M treatment in the subject record, and GLP-1 M treatment occurred together with or subsequent to ICD codes for T2D (10,517 subjects were selected using these rules). Among these, 7,856 subjects had HbA1C values both before and after GLP-1 M treatment; this was the final number of subjects included in the analysis.

Fig. 1
figure 1

Cohort inclusion and exclusion strategy and response to GLP-1 M. A. Overview of the cohort building procedure. B. The number of subjects treated with GLP-1 M per year during the defined period (see Methods) in the VUMC Synthetic Derivative database. The years indicated are synthetic time (see Methods). C. Subject responsiveness to GLP-1 M is heterogeneous. Each ‘ + ’ indicates one subject. The blue line indicates an unchanged HbA1C level after GLP-1 M treatment. The blue shading contains subjects whose HbA1C levels were changed by less than ± 0.5% in the treatment period. The green shading contains subjects whose HbA1C decreased by more than 0.5% in the period after treatment. White area contains subjects whose HbA1C increased by 0.5% or more in the time period after treatment

Full size image

Each of the FDA-approved GLP-1Ms have been prescribed at VUMC (Supplemental Table 3), although the prevalence for each varied. Among this class of drugs, albiglutide (US. trade name: Tanzeum) was found to be less effective than other drugs in this class and was withdrawn from the market in 2018. We therefore excluded albiglutide use from our analysis. It should be noted that the first record of GLP-1 M usage in the VUMC database was in 2004 (Fig. 1B). However, the first GLP-1 M, ByettaTM (Exenatide), was actually approved in 2005. The discrepancy in the dates is due to the fact that for each subject, all the dates in the VUMC SD database have been randomly shifted backward by 0–364 days. All the dates in our study therefore refer to these “synthetic dates”. The prescribing of GLP-1 M In VUMC increased consistently until 2017 (Fig. 1B), and then decreased. Multiple factors may contribute to the variation of the usage of the drug. The underlying reason for this pattern of the drug application is beyond the scope of the current paper.

Table 3 Differences in feature values between GLP-1 M responders and non-responders
Full size table

Data pre-processing

For each type of lab value, the data was first converted to the same unit. A range was then used to identify and eliminate likely incorrectly entered values that fell outside known published value ranges for that particular test. For blood pressure values, when the diastolic pressure was normal, but systolic pressure was below 20, we multiplied the systolic pressure with a correction factor of 10. We developed this correction since a systolic pressure below 20 mm Hg is not compatible with life, and the presence of a normal diastolic pressure value in the record further supports that the extremely low systolic value is an error. Such an error is most likely due to the omission of a trailing zero. We did not perform corrections for other values, as it is difficult to distinguish between genuine data points and errors.

If one subject had multiple records for the same lab test, we determined the median value, as it is robust to the outlier. For the other categorical features, we used binary encoding. In our dataset, race was not mutually exclusive, e.g., a subject can have the value 1 for being both White and Asian. Race is self-reported by subjects and thus, if a subject listed two races on their intake survey, we could not exclude one or the other, since all subjects are deidentified. Only a very small percentage of our subjects have a self-reported mixed racial background and reflects real-world diversity. Excluding them from the study could lead to racial bias, so we retained them in our dataset.

For HbA1C, the SD contains HbA1C measurements obtained via the VUMC clinical laboratory as well as point of care (POC) testing. There is only an average of 0.2% difference between the average of the two values (Supplemental Fig. 1A). Moreover, no significant difference in pattern was observed when plotting these two values together (Supplemental Fig. 1B). Thus, we combined the HbA1C from these two different sources without any adjustment.

As our data span over 17 years, for each feature, after data processing, we examined the data range and distribution in each year, to ensure that the data could be combined. We did not identify any difference in data distribution, so no further adjustment was done. With the cleaned dataset, for each lab test, we used the median value within 12 months prior to initiation of GLP1-M treatment as the final value. For each diagnosis and surgery, we only considered it in our analysis if it occurred before initiation of GLP-1 M. We required any additional medications to be administered within ± 3 days of GLP-1 M treatment to be considered simultaneous.

Response label creation

The binary label for treatment responsiveness was created by the following method: the median HbA1C level within 12 months before GLP-1 M treatment was used as the pretreatment HbA1C. The median HbA1C recorded 12 months after treatment was initiated was used as the after-treatment HbA1C. If the difference between the two values was more than 0.5%, we considered the subject to have responded to the treatment (score of 1). If not, the subject was designated as not responding (score of 0).

Data analysis and model development

Python libraries such as Pandas, Numpy, Scipy, Sklearn, LightGBM, Keras, and Seaborn were used for data process and modeling. For tests with multiple entries, if the number of tests was relatively small, we used the Bonferroni Correction and present the adjusted p-value. If the number of tests was large, we computed the False Discovery Rate (FDR) to adjust for the multiple tests.

The percentage of missing values is shown in Supplemental Table 4. To build the model, only features with less than 30% of subject data missing were used. For logistic regression, ANN, and SVC, we applied the K-Nearest Neighbor (KNN) algorithm to fill in any missing values. Briefly, the missing value for each subject was replaced with the average value from the five most similar subjects calculated using Euclidian distance. We also used L1 regularization to generate a model with the least number of features. Since LightGBM is a gradient boost tree model that can model complex decision boundaries, it has no linear assumptions and can tolerate missing values. Thus, when using LightGBM, any missing values were not filled.

Table 4 Decrease in HbA1c in response to GLP-1 M based on subject features
Full size table

The data were divided into a training set and testing set (Supplemental Table 5). Eighty-five percent of the data was used for model training and validation; 15% of the data was used for final testing. The data distribution was similar between both sets (Fig. 2) and we ensured that the distribution of treatment years was similar between training and testing data sets (Fig. 2A). The treatment time was not used for the training set to avoid data leakage. Grid search and five-fold cross-validation were then used to search for hyperparameters. During grid search, 90% of the training data was used for the model training set and 10% was used for validation.

Table 5 Differences in responder vs non-responder subjects (binary variables)
Full size table
Fig. 2
figure 2

Histograms of training and testing data. Training data (blue) and testing data (orange) are similar for (A) treatment year and (B) HbA1C before treatment. (C) Cohort BMI. (D) Cohort age

Full size image

Results

Responses to GLP-1 M are heterogeneous

A meaningful decrease in HbA1C level is the most commonly used clinical measure for evaluating the effectiveness of any diabetes treatment. The median HbA1C level of all T2D subjects treated at VUMC was maintained at approximately 7.0 in the past 17 years, suggesting that overall, diabetes is well-controlled in this cohort.

A decrease in HbA1C of less than 0.5% has not been shown to correlate with improvements in the incidence of diabetic complications [13, 14]. Therefore, in this study, a subject was considered as a responder to GLP-1 M only when the HbA1C decreased by more than 0.5% after the treatment. We considered a subject to have a positive response to GLP-1 M if the median HbA1C within 12 months after the treatment decreased by more than 0.5% from the median HbA1C within 12 months before the treatment [13, 14]. Surprisingly, although at the population level HbA1C decreased by 0.50 ± 1.57% after GLP-1 M treatment (Table 1), a decrease in HbA1C of more than 0.5% was only observed in 43.9% of subjects (Fig. 1C, green area). In 36.7% of subjects, the HbA1C did not change by more than 0.5% in either direction, while in 17.7% of subjects, the HbA1C increased by more than 0.5% after GLP-1 M treatment (Fig. 1C, white area), potentially indicating a lack of response to drug treatment. It is important to note that GLP-1 M includes both DPP4i and GLP-1-RA, which have different mechanisms of action. DPP4i relies on endogenous GLP-1 to be effective, while GLP-1-RA does not. To investigate whether responses to these two medication categories differ, we examined patient responses to each medication separately (Supplemental Fig. 1C, D) and found a heterogeneous response to each drug, with a pattern similar to the analysis of the two treatments combined. As a result, we combined the two types of medications for the rest of the paper to increase the sample size.

Next to build a model predicting GLP-1 M treatment responsiveness, we retrieved demographic and clinical information (here, called features) for each subject. Since the data were collected as part of clinical care, features available for selection were limited to lab tests requested by the treating clinician. To have sufficient statistical power, we selected only features that were present in more than 100 subjects. Our final dataset for exploratory analysis, contained 34 continuous features (Table 1), including most of the commonly ordered lab tests, and 19 discrete features (Table 2), including demographic information, other T2D medications, and diabetes complications. The specific drug names can be found in Supplemental Table 3.

Univariate analysis of features associated with GLP-1 M treatment effectiveness

Univariate analysis was performed to explore features highly associated with GLP-1 M responsiveness without the complication of missing data. The results from the univariate analysis also aid in applying the model to other datasets in the future. This analysis revealed that pre-treatment HbA1C, fasting blood glucose, random blood glucose, and urine glucose were significantly higher in the responsive group (Table 3, Fig. 3A). Additionally, higher levels of total cholesterol, LDL, triglycerides, and lower HDL were observed in the responsive group, suggesting greater insulin resistance [15]. Elevated levels of serum glutamic pyruvic transaminase (SGPT, also as known as ALT), serum glutamic oxaloacetic transaminase (SGOT, also known as AST), urine albumin, and albumin/creatinine ratio were also noted, indicating potentially poorer kidney and liver function.

Fig. 3
figure 3

Distribution of responders and non-responders for several key features. A. Histogram showing range of HbA1C values of the study cohort pre-GLP-1 M treatment. Blue shading: HbA1C values for subjects that did not respond to GLP-1 M treatment. Orange shading: HbA1C values for subjects that responded positively to GLP-1 M treatment. Compared to the non-responder group, the curve for the responder group is right-shifted (toward higher pretreatment HbA1C values). B. Histogram showing BMI values for the study cohort pre-GLP-1 M treatment. Blue shading: BMI values for subjects that did not respond to GLP-1 M treatment. Orange shading: BMI values for subjects that responded positively to GLP-1 M treatment. Compared to the non-responder group, the curve for the responder group has fewer people with BMI lower than 40. C. Mosaic plot showing patient response to GLP-1 M treatment stratified by race. The numbers in the plots indicate the number of the subjects in each group. The top tiles indicate the number of the responders. The bottom tiles indicate the number of non-responders. Tiles from left to right are White, African American and Asian, respectively. D. Mosaic plots showing patient response stratified by NSAIDs usage, sex and age. The numbers in the plots indicate the number of the subjects in each group. Left plot: female subjects; Right plot: male subjects. In both plots, purple tiles indicate subjects treated with both GLP-1 M and NSAIDs, pink tiles indicate subject only treated with GLP-1 M. Top tiles: subjects who started GLP-1 M before or equal to 40 years old. Bottom tiles: subjects who started GLP-1 M after 40. E. Violin plots showing subjects age of the first EHR records of NSAIDs stratified by year and sex. Blue line: median age for Female, Red line, median age for Male

Full size image

We also examined the association of other T2D medications with GLP-1 M response. Exogenous insulin administration correlated with enhanced responsiveness (Tables 4 and 5), with a significant decrease in HbA1C of 0.66 ± 1.81%. Combining GLP-1 M with metformin improved the response compared to non-insulin drugs (Tables 4 and 5). Conversely, the combination of GLP-1 M with thiazolidinediones was associated with a decreased response (Table 5), while sulfonylurea use showed no significant effect.

The ADA recommends using GLP-1 M for its additional cardiovascular and kidney benefits. Subjects with chronic kidney disease (CKD) had a reduced response to GLP-1 M treatment compared to those without CKD (Tables 4 and 5). Heart failure also correlated with a decreased response, whereas hypertension was associated with a slightly better response (Table 4). No significant differences were observed in subjects with weight loss surgeries, arthritis, or retinopathy.

In examining the potential effects of race on GLP-1 M responsiveness, subjects self-identifying as Asian showed a lower average reduction in HbA1C compared to non-Asian subjects, indicating potentially less benefit from GLP-1 M treatment for glucose control in Asian populations (Tables 4 and 5, Fig. 3C).

Association of sex, age, and NSAIDs with responsiveness to GLP1-M treatment

NSAIDs, which are among the most commonly used medications for treating various medical conditions, have a complex relationship with glucose metabolism. While aspirin and sodium salicylate were historically used to treat diabetes and cardiovascular disease [16, 17], some newer NSAIDs have been associated with increased cardiometabolic risk. Some of the beneficial effects of first-generation COX inhibitors may have been due to decreased levels of prostaglandin E2 (PGE2), which acts through GPCRs, some of which are expressed in β cells along with the GLP-1R [18] and could modulate GLP-1R effects. Recent studies suggest that some NSAIDs directly inhibit DPP4, thus elevating endogenous GLP-1 [19]. However, our study found that subjects taking NSAIDs (a list of these medications is found in Supplemental Table 3) simultaneously with GLP-1 M collectively showed an overall weaker response to GLP-1 M treatment (Table 4, FDR = 5.16E-03). A similar pattern was observed when GLP-1 M was used together with other pain relievers such as acetaminophen and hydrocodone (Table 4, FDR = 6.76E-03). It was reported that NSAID usage varies with both sex and age [20, 21], and since sex hormones can affect various physiological responses, we divided the subjects into eight groups based on sex, age, and NSAID usage. Overall, our data suggest that females show reduced glucose-lowering benefit from GLP1-M treatment Table 4). However, as shown in Fig. 3D and Table 6, among females less than 40 years old, NSAID usage was associated with a greater reduction in HbA1C in response to GLP-1 M (0.82 ± 1.72) than no NSAID usage (0.28 ± 1.70, Bonferroni adjusted p = 8E-03). Interestingly, the relationship was reversed in females over age 40 (0.52 ± 1.44 without NSAIDs vs 0.36 ± 1.45 with NSAIDs, Bonferroni adjusted p = 2E-03). We did not observe any age-associated differences in GLP-1 M responsiveness in male subjects. There is also no significant difference in the percentage of subjects responding to GLP-1 M with or without NSAID usage (Fig. 3D). To understand whether there is a difference in NSAID usage between sexes, we compared the first records of NSAID usage in VUMC SD (Fig. 3E). We did not identify any significant difference in the distribution of NSAID usage. Given the widespread use of NSAIDs in clinical practice, our findings highlight the need for further investigation into the potential interactions between NSAIDs and GLP-1 M treatment.

Table 6 Response to GLP-1 M based on the combination of sex, age and NSAIDs
Full size table

Predictive model for GLP-1 M responsiveness

While univariate analysis offers valuable insights, our primary goal was to develop a predictive model for patient responsiveness to GLP-1 M treatment, irrespective of causal relationships. We next built a machine learning model to enhance prediction and further selected from the aforementioned features with less than 30% missing datapoints to build a prediction model for GLP-1 M responsiveness. Machine learning models with high complexity face the difficulty of generalizability. Thus, we used one of the simplest available models, logistic regression, as the benchmark. Our exploratory analysis revealed both positive and negative correlations between particular features (Supplemental Fig. 2). For example, chronic kidney disease was strongly correlated with heart failure, while it was negatively correlated with metformin usage. Random blood glucose was positively correlated with urine glucose, but negatively correlated with blood sodium. Therefore, to reduce model complexity further, we applied L1 regulation (also called Lasso). With L1 regulation, when two features are highly correlated, the coefficient for one of the features will be set to zero and is not considered in the final model.

We compared logistic regression with three different machine learning models using distinct algorithms: LightGBM, SVC, and ANNs. LightGBM is a gradient boosting framework using tree-based learning algorithms. SVC is used for classification tasks by finding the optimal hyperplane that maximizes the margin between different classes. ANNs consist of interconnected nodes (neurons) organized in layers: input, hidden, and output layers, capable of capturing complex patterns in data. The hyperparameters searched for all models are shown in Table 7. Model performance was assessed using a confusion matrix (Fig. 4A1, 4A2, Supplemental Fig. 3) and auROC (area under the Receiver Operating Characteristic curve; Fig. 4A3, Supplemental Fig. 3). As shown in Fig. 4A, all models performed similarly except SVC. Logistic regression achieved an auROC of 0.77 (Fig. 4A3), while LightGBM and ANN both achieved an auROC of 0.76 (Fig. 4A3, Supplemental Fig. 3). The performance of SVC was the worst among the models.

Table 7 Model hyperparameters and evaluation results
Full size table
Fig. 4
figure 4

Assessment of model performance. A. Confusion matrices for: (A1) logistic regression and (A2) LightGBM. A3. ROC (Receiver Operating Characteristic) curve for logistic regression (green) and LightGBM (orange). The blue dotted line indicates where the false positive rate equals the true positive rate. B. Model learning curves for (B1) Logistic regression and (B2) LightGBM. Area under the ROC (auROC) curve was used to evaluate the model performance and is shown in the y-axis. C. Logisitic regression coefficients were plotted to show feature importance

Full size image

Since none of the other models tested outperformed logistic regression, we used LightGBM to explore whether increasing the sample size would further improve either model. As indicated in the learning curves for logistic regression (Fig. 4B), the training and validation sets would likely converge at around 0.78 if more samples were added. For LightGBM (Fig. 4B2), the learning curve has already reached a plateau, suggesting that increasing the sample size would not improve model performance.

As logistic regression proved to be the best model, the model coefficients are plotted in Fig. 4C. The full set of the coefficients and the odds ratio are in the Supplemental Table 6. Pre-treatment HbA1C, subject age, ALT, AST, BMI, metformin use, and having hypertension each positively contribute to predicted GLP-1 M responsiveness. These results are consistent with the univariant analysis results (Table 2). Interestingly, the model did not identify insulin usage as being predictive of treatment responsiveness. The use of insulin in subjects might be associated with features such as age and HbA1C level, so the coefficient for insulin was set to zero when L1 regulation was applied. Sulfonylurea or thiazolidinedione use, being of Asian descent, having heart failure, retinopathy, high blood chloride, and long T2D duration all negatively contributed to the predicted value of GLP-1 M effectiveness.

Discussion

In this study, we used a de-identified patient EHR database to analyze the real-world response of T2D patients to GLP-1 M treatment and to predict which patients might benefit most from this class of T2D medications. Taken together, our model predicts that the best candidates for adding GLP-1 M as a second line T2D medication are: older age, non-Asian subjects, with short T2D duration that have elevated HbA1C, high BMI, hypertension, and no heart failure or retinopathy. A limitation of the current study is that it was performed using data from a single academic medical center with limited diversity in terms of patient population. The current results should be confirmed using datasets from different subject cohorts.

Our goal in the current study was to develop a predictive model that could guide the use of GLP-1 M medications for the treatment of T2D in the real world. To increase our power, the current study did not separately consider effects on GLP-1RA or DDP4i responsiveness. We specifically used common clinical factors that are available within current clinical practice to build our model, making it more feasible for use in real-world settings. As a starting point, we used the HbA1C 12 months before the decision point, which is typically available in clinical practice. Our model reveals clinical features that should be considered when deciding to initiate GLP-1 M combined therapy for the treatment of T2D, including age, race, T2D duration, HbA1C levels, BMI, hypertension, and the presence or absence of heart failure or retinopathy. By incorporating these factors into our model, we aim to provide clinicians with a more accurate and personalized tool for predicting which patients might benefit most from GLP-1 M treatment. Our model reveals clinical features that should be considered when deciding to initiate GLP-1 M combined therapy for the treatment of T2D. We want to emphasize that although these factors can be used to guide clinical decisions, they may not have a causal relationship with the response to GLP-1 M.

Some factors identified in this study are consistent with prior existing knowledge. Previous research showed that patients with shorter duration of T2D are more likely to benefit from GLP-1 M treatment [22]. Interestingly, univariate analysis did not identify differences in T2D duration between responder and non-responders (Table 3). However, in the final model, when considering all other factors, T2D duration was found to be negatively associated with GLP-1 M responsiveness (Fig. 4C), which is consistent with the previous study. The finding that hypertension and higher BMI positively contribute to predicting GLP-1 M responsiveness is consistent with the current approach of using GLP-1 M for protection of the cardiovascular system and promoting weight loss [3]. Consistent with these reports, recent research from one of our groups (Q.S.W.) further demonstrated that drugs targeting genes involved in the regulation of systolic blood pressure might be repurposed to treat diabetes [23].

The finding that high ALT and AST also predict a good response to GLP-1 M was not anticipated based on the literature, and seems counterintuitive. It is unlikely that impaired liver function enhances GLP-1 M effectiveness. Elevated ALT and AST might simply reflect fatty liver disease due to insulin resistance and worse severity of the underlying metabolic and physiological condition. In support of this hypothesis, we did identify a weak association between the level of C-peptide and the level of ALT (Pearson’s correlation coefficient = 0.30) and AST (Pearson’s correlation coefficient = 0.29). The popularity of GLP-1 M is partly due to its cardiovascular, kidney, and weight loss benefits. Our results suggest that although GLP-1 M have demonstrable protective effects on cardiovascular and kidney disease, for people who already have CKD and heart failure, such protection might come together with reduced benefit on lowering of blood glucose. On the other hand, for patients with hypertension, GLP-1 M have an enhanced blood glucose lowering effect.

The observation that differences in sex and age resulted in some subjects having better responsiveness to combined treatment with NSAIDs and GLP-1 M was also unexpected. One caveat of these studies is that many NSAIDs do not require a prescription and are one of the most commonly used over the counter medications. The current dataset is from the hospital records, which include self-reported information. Thus, we cannot rule out the possibility that some NSAID usage was not captured in the EHR/SD. Most likely, subjects listed as receiving NSAID treatment were experiencing chronic pain or inflammation. Subjects without NSAID treatment listing in the SD might include subjects that used NSAIDs only for a short time and did not inform their attending physician. Further research is required to validate the effects of NSAID usage and to understand the underlying mechanisms behind the interaction between NSAIDs and GLP-1 M. Several features, including NSAIDs and insulin usage, were ultimately not included in the final models. This was due to our decision to apply L1 regulation in order to obtain the simplest model for the purpose of helping clinicians determine which of their patients with T2D would most likely benefit from adding GLP-1 M as the second-line medication.

Our current analysis cannot exclude the possibility that the patient criteria for considering GLP-1 M treatment are generalizable for other T2D medications since we only considered GLP-1 M and did not compare it with responsiveness to other treatments. Future studies applying multi-label classification to a larger data set can help address this question. However, it has traditionally been believed that T2D drugs are more efficacious for subjects with higher HbA1C, partly due to the relief of the stress from glucotoxicity [24], and that is what we observed as well. In contrast, higher HbA1C has recently been linked to poor metformin responsiveness [25]. In our study, pretreatment HbA1C level was the dominant factor that affected GLP-1 M treatment outcome in the current analysis. Thus, for patients who have higher HbA1c at the time of diagnosis, it might be beneficial to use GLP1-M earlier in their course of treatment.

Our model also suggests that GLP-1 M treatment combined with sulfonylureas or thiazolidinediones has no added benefit and may actually decrease treatment effectiveness. In fact, both sulfonylureas and thiazolidinediones are negative predictors for GLP-1 M responsiveness. These associations are independent of the pre-treatment HbA1C for two reasons: 1. HbA1C is included in the model; 2. we did not observe any obvious differences in the pre-treatment HbA1C between subjects taking sulfonylureas and thiazolidinediones and those who were not taking these medications Previous studies have reported that combining sulfonylureas or thiazolidinediones with GLP-1 M is an effective treatment for T2D [26,27,28,29]. However, those studies did not compare the effectiveness of this drug combination with the use of GLP-1 M alone. For example, in a previous study, the effectiveness of dulaglutide combined with pioglitazone was compared with that of pioglitazone and metformin [29]. In our analysis, we compared the effectiveness of GLP-1 M treatment alone with that of the combination of GLP-1 M treatment with other medications. Interestingly, we only observed the negative effects of thiazolidinediones and sulfonylureas on GLP-1 M effectiveness in our final predictive model, which was built using all the features shown in Supplemental Fig. 2. Our univariate analysis did not detect any association. Therefore, further study is required to fully understand the combined effect of these drugs on improving HbA1C.

Machine learning and artificial intelligence have been successfully used in various fields. Although machine learning models often appear to be a black box, having such a prediction model would be extremely beneficial for real-world applications and patient treatment decisions. Our best prediction model was the logistic regression model with a 0.77 auROC. Complex models such as LightGBM and artificial neural network all performed similarly to logistic regression. This result is not unprecedented; in a recent study predicting patient survival after heart transplantation, all models tested demonstrated similar performance to logistic regression [30]. The authors hypothesized that the quality of the clinical datasets might limit the application of the machine learning techniques. In addition to data quality, in our experience one limitation comes from the inherent bias within healthcare data sets. A typical patient has a limited number of standard-of-care lab tests commonly ordered by their attending physician. Thus, the models used here had no opportunity to explore more rare features which might have as yet un-identified relationships with GLP-1 M treatment. Since it is impossible to test all the lab values for a patient, a combination of genomic information with clinical information might be the best way to identify the individuals that will benefit most from GLP-1 M treatment. In the future, approaches such as this may help to define physiologic and mechanistic causes for the differences in GLP-1 M responsiveness.

In conclusion, this study demonstrates that a predictive model based on commonly available clinical features can guide the use of GLP-1 M medications in T2D treatment. By identifying key factors such as age, race, T2D duration, HbA1C levels, BMI, hypertension, and the presence of other conditions like heart failure or retinopathy, our model provides a personalized tool for clinicians to make informed treatment decisions. Future research incorporating patients’ genomic information are expected to build better model.

Data availability

All the data in this study are from the VUMC SD database. Restriction and institutional approval are required for data access. Therefore, the data for this study are not publicly available. The original code is in a Jupyter notebook with the data embedded. Thus, they are not publicly available to protect the data used in this study.

References

  1. Saeedi P, Petersohn I, Salpea P, Malanda B, Karuranga S, Unwin N, Colagiuri S, Guariguata L, Motala AA, Ogurtsova K, Shaw JE, Bright D, Williams R. Global and regional diabetes prevalence estimates for 2019 and projections for 2030 and 2045: Results from the International Diabetes Federation Diabetes Atlas, 9(th) edition. Diabetes Res Clin Pract. 2019;157:107843.

    Article  PubMed  Google Scholar 

  2. Centers for Disease Control and Prevention USDoHaHS (Ed.). Centers for Disease Control and Prevention. National Diabetes Statistics Report, 2020. 2020. https://www.cdc.gov/diabetes/php/dataresearch/index.html.

  3. Association AD: 9. Pharmacologic Approaches to Glycemic Treatment: <em>Standards of Medical Care in Diabetes—2021</em>. Diabetes care 2021;44:S111-S124

  4. Dennis JM. Precision Medicine in Type 2 Diabetes: Using Individualized Prediction Models to Optimize Selection of Treatment. Diabetes. 2020;69:2075–85.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  5. Hill J, Nielsen M, Fox MH. Understanding the social factors that contribute to diabetes: a means to informing health care and social policies for the chronically ill. The Permanente journal. 2013;17:67–72.

    Article  PubMed  PubMed Central  Google Scholar 

  6. Taylor SI, Yazdi ZS, Beitelshees AL. Pharmacological treatment of hyperglycemia in type 2 diabetes. J Clin Invest. 2021;131:e142243.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  7. Koye DN, Montvida O, Paul SK. Third-Line Antidiabetic Therapy Intensification Patterns and Glycaemic Control in Patients with Type 2 Diabetes in the USA: A Real-World Study. Drugs. 2020;80:477–87.

    Article  CAS  PubMed  Google Scholar 

  8. Schernthaner G, Shehadeh N, Ametov AS, Bazarova AV, Ebrahimi F, Fasching P, Janež A, Kempler P, Konrāde I, Lalić NM, Mankovsky B, Martinka E, Rahelić D, Serafinceanu C, Škrha J, Tankova T, Visockienė Ž. Worldwide inertia to the use of cardiorenal protective glucose-lowering drugs (SGLT2i and GLP-1 RA) in high-risk patients with type 2 diabetes. Cardiovasc Diabetol. 2020;19:185.

    Article  PubMed  PubMed Central  Google Scholar 

  9. Peng XV, McCrimmon RJ, Shepherd L, Boss A, Lubwama R, Dex T, Skolnik N, Ji L, Avogaro A, Blonde L. Glycemic Control Following GLP-1 RA or Basal Insulin Initiation in Real-World Practice: A Retrospective, Observational, Longitudinal Cohort Study. Diabetes therapy : research, treatment and education of diabetes and related disorders. 2020;11:2629–45.

    Article  CAS  PubMed  Google Scholar 

  10. Röthe J, Kraft R, Schöneberg T, Thor D. Exploring G Protein-Coupled Receptor Signaling in Primary Pancreatic Islets. Biol Proced Online. 2020;22:4.

    Article  PubMed  PubMed Central  Google Scholar 

  11. Sriram K, Insel PA. G Protein-Coupled Receptors as Targets for Approved Drugs: How Many Targets and How Many Drugs? Mol Pharmacol. 2018;93:251–8.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  12. Kho AN, Hayes MG, Rasmussen-Torvik L, Pacheco JA, Thompson WK, Armstrong LL, Denny JC, Peissig PL, Miller AW, Wei WQ, Bielinski SJ, Chute CG, Leibson CL, Jarvik GP, Crosslin DR, Carlson CS, Newton KM, Wolf WA, Chisholm RL, Lowe WL. Use of diverse electronic medical record systems to identify genetic risk for type 2 diabetes within a genome-wide association study. Journal of the American Medical Informatics Association : JAMIA. 2012;19:212–8.

    Article  PubMed  Google Scholar 

  13. Hameed UA, Manzar D, Raza S, Shareef MY, Hussain ME. Resistance Training Leads to Clinically Meaningful Improvements in Control of Glycemia and Muscular Strength in Untrained Middle-aged Patients with type 2 Diabetes Mellitus. N Am J Med Sci. 2012;4:336–43.

    Article  PubMed  PubMed Central  Google Scholar 

  14. Selvin E, Marinopoulos S, Berkenblit G, Rami T, Brancati FL, Powe NR, Golden SH. Meta-analysis: glycosylated hemoglobin and cardiovascular disease in diabetes mellitus. Ann Intern Med. 2004;141:421–31.

    Article  CAS  PubMed  Google Scholar 

  15. Fox CS. Cardiovascular disease risk factors, type 2 diabetes mellitus, and the Framingham Heart Study. Trends Cardiovasc Med. 2010;20:90–5.

    Article  PubMed  PubMed Central  Google Scholar 

  16. Robertson RP. Prostaglandins, glucose homeostasis, and diabetes mellitus. Annu Rev Med. 1983;34:1–12.

    Article  CAS  PubMed  Google Scholar 

  17. Raber I, McCarthy CP, Vaduganathan M, Bhatt DL, Wood DA, Cleland JGF, Blumenthal RS, McEvoy JW. The rise and fall of aspirin in the primary prevention of cardiovascular disease. Lancet (London, England). 2019;393:2155–67.

    Article  PubMed  Google Scholar 

  18. Ahrén B. Islet G protein-coupled receptors as potential targets for treatment of type 2 diabetes. Nat Rev Drug Discov. 2009;8:369–85.

    Article  PubMed  Google Scholar 

  19. Chittepu VCSR, Kalhotra P, Osorio-Gallardo T, Gallardo-Velázquez T, Osorio-Revilla G. Repurposing of FDA-Approved NSAIDs for DPP-4 Inhibition as an Alternative for Diabetes Mellitus Treatment: Computational and in Vitro Study. Pharmaceutics. 2019;11:238.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  20. Demyanets S, Wojta J. Sex differences in effects and use of anti-inflammatory drugs. Handb Exp Pharmacol. 2012;(214):443–72.

  21. McCarberg BH. NSAIDs in the older patient: balancing benefits and harms. Pain medicine (Malden, Mass). 2013;14(Suppl 1):S43-44.

    Article  PubMed  Google Scholar 

  22. Freeman JS: Optimizing outcomes for GLP-1 agonists. The Journal of the American Osteopathic Association 2011;111:eS15–20

  23. Shuey MM, Lee KM, Keaton J, Khankari NK, Breeyear JH, Walker VM, Miller DR, Heberer KR, Reaven PD, Clarke SL, Lee J, Lynch JA, Vujkovic M, Edwards TL. A genetically supported drug repurposing pipeline for diabetes treatment using electronic health records. EBioMedicine. 2023;94:104674.

  24. Chan JCN, Lim LL, Wareham NJ, Shaw JE, Orchard TJ, Zhang P, Lau ESH, Eliasson B, Kong APS, Ezzati M, Aguilar-Salinas CA, McGill M, Levitt NS, Ning G, So WY, Adams J, Bracco P, Forouhi NG, Gregory GA, Guo J, Hua X, Klatman EL, Magliano DJ, Ng BP, Ogilvie D, Panter J, Pavkov M, Shao H, Unwin N, White M, Wou C, Ma RCW, Schmidt MI, Ramachandran A, Seino Y, Bennett PH, Oldenburg B, Gagliardino JJ, Luk AOY, Clarke PM, Ogle GD, Davies MJ, Holman RR, Gregg EW. The Lancet Commission on diabetes: using data to transform diabetes care and patient lives. Lancet (London, England). 2021;396:2019–82.

    Article  PubMed  Google Scholar 

  25. Bielinski SJ, Yanes Cardozo LL, Takahashi PY, Larson NB, Castillo A, Podwika A, De Filippis E, Hernandez V, Mahajan GJ, Gonzalez C, Shubhangi, Decker PA, Killian JM, Olson JE, St Sauver JL, Shah P, Vella A, Ryu E, Liu H, Marshall GD, Cerhan JR, Singh D, Summers RL. Predictors of Metformin Failure: Repurposing Electronic Health Record Data to Identify High-Risk Patients. J Clin Endocrinol Metab. 2023;108(7):1740-1746

  26. Bailey T. Options for combination therapy in type 2 diabetes: comparison of the ADA/EASD position statement and AACE/ACE algorithm. Am J Med. 2013;126:S10-20.

    Article  PubMed  Google Scholar 

  27. Kim SW. Triple Combination Therapy Using Metformin, Thiazolidinedione, and a GLP-1 Analog or DPP-IV Inhibitor in Patients with Type 2 Diabetes Mellitus. Korean diabetes journal. 2010;34:331–7.

    Article  PubMed  PubMed Central  Google Scholar 

  28. Udiawar MV, Bain SC. Liraglutide in Combination with Metformin or Sulfonylurea for the Treatment of Type 2 Diabetes in Adult patients. Clinical Medicine Insights: Therapeutics. 2012;4:CMT.S7283.

    Google Scholar 

  29. Wysham C, Blevins T, Arakaki R, Colon G, Garcia P, Atisso C, Kuhstoss D, Lakshmanan M. Efficacy and safety of dulaglutide added onto pioglitazone and metformin versus exenatide in type 2 diabetes in a randomized controlled trial (AWARD-1). Diabetes Care. 2014;37:2159–67.

    Article  CAS  PubMed  Google Scholar 

  30. Miller PE, Pawar S, Vaccaro B, McCullough M, Rao P, Ghosh R, Warier P, Desai NR, Ahmad T. Predictive Abilities of Machine Learning Techniques May Be Limited by Dataset Characteristics: Insights From the UNOS Database. J Cardiac Fail. 2019;25:479–83.

    Article  Google Scholar 

Download references

Acknowledgements

ChatGPT was used in some places in the manuscript to suggest edits to the text and proper word usage and grammar. Text suggested by ChatGPT was not used verbatim, but was modified by the authors.

Funding

M.G. was supported by Merit Review Awards from the Department of Veterans Affairs (I01-BX003744 and I01- BX005399) and an R01 from the NIH (DK120626). J.M.S. was supported by a Merit Review Award from the Department of Veterans Affairs (I01-BX005459) and R01s from the NIH (DK109102 and HL144846). Q.S.W. was supported by awards from the American Heart Association (17SFRN33520017) and the NIH (1R01HL140074).

Author information

Authors and Affiliations

  1. Department of Veterans Affairs, Tennessee Valley Health Authority, Nashville, TN, USA

    Xiaodong Zhu, John M. Stafford & Maureen Gannon

  2. Department of Medicine, Vanderbilt University Medical Center, Nashville, TN, USA

    Xiaodong Zhu, Michael J. Fowler, Quinn S. Wells, John M. Stafford & Maureen Gannon

  3. Department of Biomedical Informatics, Vanderbilt University Medical Center, Nashville, TN, USA

    Quinn S. Wells

  4. Department of Molecular Physiology and Biophysics, Vanderbilt University, Nashville, TN, USA

    John M. Stafford & Maureen Gannon

  5. Department of Cell and Developmental Biology, Vanderbilt University, Nashville, TN, USA

    Maureen Gannon

Authors
  1. Xiaodong Zhu
    View author publications

    You can also search for this author inPubMed Google Scholar

  2. Michael J. Fowler
    View author publications

    You can also search for this author inPubMed Google Scholar

  3. Quinn S. Wells
    View author publications

    You can also search for this author inPubMed Google Scholar

  4. John M. Stafford
    View author publications

    You can also search for this author inPubMed Google Scholar

  5. Maureen Gannon
    View author publications

    You can also search for this author inPubMed Google Scholar

Contributions

The authors have nothing to disclose and have no competing interests. Author contributions: Xiaodong Zhu, Michael J. Fowler, and Maureen Gannon conceived of the study. Xiaodong Zhu, Michael J. Fowler, Quinn S. Wells, and Maureen Gannon developed the study design. Xiaodong Zhu conducted the analysis, developed the model, and made all figures and tables. Xiaodong Zhu, Quinn S. Wells, John M. Stafford, and Maureen Gannon interpreted the data. Xiaodong Zhu wrote the main draft. Xiaodong Zhu, Quinn S. Wells, John M. Stafford, and Maureen Gannon edited the manuscript.

Authors’ information

Not applicable.

Corresponding author

Correspondence to Maureen Gannon.

Ethics declarations

Ethics approval and consent to participate

All methods were carried out in accordance with relevant guidelines and regulations. All experimental protocols were approved by the Vanderbilt University Medical Center IRB Committee/Committee on Human Research Protection. All of the data used in the current study was obtained from a fully deidentified synthetic derivative of the electronic health record from Vanderbilt University Medical Center. Individual informed consent for participation in the current study was waived since subject information is deidentified and specific patients cannot be identified. However, at the time of their visit, patients can opt out of being included in the database. Thus, all subjects in the database chose to be included.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.

Additional information

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Zhu, X., Fowler, M.J., Wells, Q.S. et al. Predicting responsiveness to GLP-1 pathway drugs using real-world data. BMC Endocr Disord 24, 269 (2024). https://doiorg.publicaciones.saludcastillayleon.es/10.1186/s12902-024-01798-9

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doiorg.publicaciones.saludcastillayleon.es/10.1186/s12902-024-01798-9

Keywords

BMC Endocrine Disorders

ISSN: 1472-6823

Contact us