Prediction model for metformin intolerance in Chinese elderly patients with type 2 diabetes: derivation and external validation in a cross‑sectional study

Abstract

Background

Approximately 30% of patients taking metformin experience gastrointestinal adverse events, with 5%–10% discontinuing the drug due to intolerance. The aim of our study is to develop and externally validate a prediction model for metformin intolerance in elderly patients with type 2 diabetes.

Methods

This study utilized data from three independent cohorts of Chinese elderly metformin-naïve patients with type 2 diabetes. The derivation cohort consisted of 538 patients. External validation was performed using data from a separate cohort of 319 patients. Intolerance was assessed after patients initiated metformin and maintained treatment for over four weeks. Patients who were unable to complete at least four weeks of treatment due to gastrointestinal adverse effects were also classified as intolerance.

Results

In the derivation cohort, the median age was 67 years, median HbA1c was 8.1%, and 45% were female. The external validation cohort had a median age of 70 years, median HbA1c of 7.6%, and 54.5% were female. Metformin intolerance occurred in 75 patients (13.9%) in the derivation cohort and 63 patients (19.7%) in the external validation cohort. Multivariable analysis identified ten variables: gender, age, body mass index, fasting glucose, HbA1c, aspartate aminotransferase, gamma-glutamyl transferase, antidepressants, proton pump inhibitors and clopidogrel. In the derivation cohort, the model demonstrated an AUC of 0.889, with a sensitivity of 0.813 and a specificity of 0.773. Upon external validation, the model maintained good discriminative ability with an AUC of 0.883.

Conclusions

We developed and validated a multivariable prediction model for metformin intolerance in Chinese elderly patients. This model may help physicians to be more confident when prescribing metformin, and reduce the intolerance of metformin.

Clinical trial number

Not applicable.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12902-026-02278-y.

Background

A large-scale epidemiological survey in China reveals that the prevalence of diabetes among the elderly aged 60 and above has reached 28.8% [1], imposing a substantial healthcare burden on society. Due to its multiple clinical advantages, metformin is recommended by various domestic and international guidelines as one of the first-line hypoglycemic agents for elderly patients with type 2 diabetes [2].

However, metformin has long faced two major issues in China: under-prescription and under-dosing. A nationwide large-scale epidemiological study indicates that only half of the type 2 diabetes patients in China are treated with metformin [3]. Many physicians in China advise a daily metformin dose not exceeding 1500 mg for patients. Even in the landmark MARCH study in the field of diabetes in China, the daily treatment dose in the metformin group was only 1500 mg [4]. This dosage is considerably lower than the 2000–2550 mg/day recommended by guidelines [5] and the prescribing information. The core reason for this situation is the common gastrointestinal adverse effects (GI AEs) of metformin, such as abdominal pain, diarrhea, nausea, and bloating [6]. Approximately 30% of patients taking metformin experience related symptoms, with 5%–10% discontinuing the drug due to intolerance [7]. The intolerance rate may be even higher among elderly patients [8]. Our research group previously proposed a “double-edged sword” mechanism of metformin: it exerts hypoglycemic effects by modulating the gut microbiota–short-chain fatty acids axis, but may also cause upper gastrointestinal symptoms [9].

Currently, there is no complete method to avoid such AEs. In the CONSENT trial, metformin extended release (XR) was not superior to metformin immediate release (IR) for overall AE incidence during treatment of Chinese patients [10]. Therefore, we can shift our perspective and adopt a simple method before initial medication to early predict whether elderly patients will develop metformin intolerance. This approach would help avoid treatment discontinuation due to GI AEs, which could delay treatment and undermine patients’ confidence in therapy. Thus, the main objective of this study is to develop and externally validate a prediction model for metformin intolerance in Chinese elderly patients with type 2 diabetes.

Methods

We followed the TRIPOD (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis) statement for reporting multivariable prediction model development and validation. The work described has been carried out in accordance with the Declaration of Helsinki. This study was approved by the Ethics Committee of Huadong Hospital, Fudan University, Shanghai, China (Ethics Number: 2024K214). All procedures were performed in compliance with relevant laws and institutional guidelines. Written informed consent was obtained from all subjects prior to the study.

Derivation and validation cohorts

We formed a derivation cohort using two previously published cohorts of metformin-naïve patients with type 2 diabetes (cohort A: 415 patients [11]; cohort B: 361 patients [8]). The derivation cohort consisted of 538 patients after applying inclusion and exclusion criteria from May 2013 to September 2018, from Department of Endocrinology, Huadong Hospital, Fudan University, Shanghai, China. External validation was performed using data from a separate cohort (cohort C: 333 patients) recruited from the National Metabolic Management Center, Huadong Hospital Minhang Branch, Shanghai, China, between March 2021 and November 2024 (Ethics Number: 2021K009). 319 patients were included in the external validation cohort after applying inclusion and exclusion criteria.

Inclusion criteria and exclusion criteria

Inclusion criteria were as follows: (a) diagnosed type 2 diabetes according to the 1999 World Health Organization criteria [12]; (b) metformin-naïve; (c) age ≥ 60 years; (d) body mass index ≥ 18.5 kg/m²; (e) HbA1c > 6.5% (48 mmol/mol).

Exclusion criteria included: (a) confirmed or suspected type 1 diabetes; (b) previous use of α-glucosidase inhibitors or glucagon-like peptide-1 receptor agonists; (c) history of peptic ulcer, Helicobacter pylori infection, gastrointestinal surgery, chronic gastritis, gastrointestinal tumor, or other significant gastrointestinal disorders; (d) recent (within 3 months) diabetic ketoacidosis or hyperosmolar hyperglycemic state; (e) active cardiovascular or other serious systemic disease; (f) estimated glomerular filtration rate < 45 mL/min/1.73 m²; (g) liver enzymes > 2 times the upper limit of normal at screening; (h) concomitant use of unknown or unclear medications; (i) poor medication compliance.

Study procedures and outcome

Anthropometric, serum plasma, and urine samples were collected from all patients prior to the initiation of metformin therapy. Biochemical measurements of plasma glucose, insulin, C peptide, lipids profiles, liver enzymes, creatinine, HbA1c and urine albumin were performed in a central laboratory. Organic cation transporter 1 (OCT1)-inhibiting drugs are defined as medications shown to inhibit OCT1 activity in vitro, such as antidepressants, citalopram, proton pump inhibitors (PPIs), verapamil, diltiazem, doxazosin, spironolactone, clopidogrel, rosiglitazone, quinine, tramadol and codeine [13]. Information on the concurrent use of OCT1-inhibiting drugs was collected by questionnaires [8].

All patients were prescribed metformin IR tablets at a minimum daily dose of 1000 mg for at least four weeks. The study outcome was metformin intolerance (“yes/no”). Metformin intolerance was assessed based on structured questionnaires and was defined according to established criteria from prior studies [8]. The questionnaire covered the following six digestive disturbances: (a) abdominal pain; (b) diarrhoea; (c) nausea; (d) vomiting; (e) bloating; and (f) anorexia. A scoring system of 0–4 was used to grade the severity of the symptoms, where 4 = severe (GI problem inducing ill health such as malnutrition, hospitalization), 3 = significant (GI symptoms present to the extent that daily activity is disturbed), 2 = moderate (GI symptom present with some disturbance to daily activity), 1 = mild (daily activity not affected), and 0 = no symptoms. Briefly, a patient was classified as “metformin intolerance” if they met either of the following criteria: (a) the presence of moderate-to-severe GI AEs, defined as symptoms sufficiently severe to interfere with daily activities or to lead to clinical consequences such as malnutrition or hospitalization; or (b) patient-reported intolerance, regardless of whether the symptoms met objective thresholds for severity [8].

Based on our previous studies and other published literature, GI AEs of metformin do not appear to be dose-dependent [8, 14, 15]. Therefore, metformin dose was not included as a variable in the prediction model. All patients in our cohorts received a minimum daily dose of 1000 mg, and the model is applicable to patients taking ≥ 1000 mg/day.

Data processing

The target variable was the metformin intolerance status, with 35 predictive variables a final training and validation dataset. Missing values in the training dataset were imputed using the k-Nearest Neighbors (k-NN) algorithm with “k = 10”, implemented via the “VIM” package.

Stepwise selection was employed as an exploratory approach due to the large number of candidate predictors (Table 1), with the aim of identifying a parsimonious model. Continuous variables were categorized into quartiles to balance clinical interpretability and model stability. Stepwise selection algorithm (both forward and backward) using the Akaike Information Criterion (AIC) was applied. The formula of the resulting model defined the final set of features used for subsequent model building. The continuous variables in the final feature set were discretized into quartiles. For each continuous variable, the values were binned into four categories (Q1, Q2, Q3, Q4) based on the 0%, 25%, 50%, 75%, and 100% quantiles of the training data. These categorical labels were then converted into numerical values (1 to 4) for analysis. This discretization process was applied to both the training and validation datasets. Crucially, the quartile thresholds were calculated exclusively from the training data and then applied to the validation set to prevent data leakage. Stepwise selection was again applied to this model to finalize the predictors. Feature selection was followed by retention of only significant variables (p < 0.05) to further mitigate overfitting risk.

Table 1.

Baseline characteristics of the derivation and external validation cohorts

Variable		Derivation (n = 538)	Validation (n = 319)	P Value
Female (n, %)		242 (45.1%)	174 (54.5%)	0.007*
Age (years)		67 (64–73)	70 (66–73)	0.002*
Duration of diabetes (weeks)		120 (60–180)	120 (42–201)	0.725
History of smoking (n, %)		91 (16.9%)	69 (21.6%)	0.102
History of alcohol (n, %)		48 (8.9%)	84 (26.3%)	< 0.001*
History of coronary heart disease(n, %)		76 (14.1%)	47 (14.7%)	0.840
History of stroke (n, %)		94 (17.5%)	23 (7.2%)	< 0.001*
BMI (kg/m2)		24.99 (23.2-27.24)	25.3 (23.35–27.2)	0.804
Systolic pressure (mmHg)		130 (120–140)	138 (127.5–150)	< 0.001*
Diastolic pressure (mmHg)		80 (70–80)	74 (67–81)	< 0.001*
Fasting glucose (mmol/l)		7.6 (6.1-9)	8.1 (6.75–9.85)	< 0.001*
2-hour glucose (mmol/l)		12.22 (10.3–17.5)	14 (11-17.4)	< 0.001*
Fasting C-peptide (ng/ml)		1.6 (1.2–2.5)	2.5 (1.9–3.3)	< 0.001*
Fasting insulin (uIU/ml)		7.2 (5.4-11.25)	10.5 (7.5–15.8)	< 0.001*
2-hour C-peptide (ng/ml)		3.8 (2.5–5.6)	6.9 (4.5–9.57)	< 0.001*
2-hour insulin (uIU/ml)		17.1 (11.25–25.2)	35 (19.3–65)	< 0.001*
HbA1c (%)		8.1 (7.5–8.5)	7.6 (6.7–8.55)	0.006*
(mmol/mol)		65 (59–69)	59 (49–70)
ALT (IU/l)		18 (13–26)	20 (15–27)	0.033*
AST (IU/l)		17 (14–23)	19 (16–24)	< 0.001*
GGT (IU/l)		23.7 (17–35)	21.8 (16.1–32.2)	0.236
History of gallstones (n, %)		156 (29%)	38 (11.9%)	< 0.001*
Serum creatinine (umol/l)		69 (58-84.38)	69 (59-83.15)	0.880
Urine albumin/creatinine (mg/mmol)		1.12 (0.5–3.8)	2.27 (1.69–6.82)	< 0.001*
TG (mmol/l)	1.5 (1.1-2)		1.58 (1.17–2.12)	0.053
TC (mmol/l)	4.4 (3.74–5.1)		4.9 (4.1–5.7)	< 0.001*
HDL-C (mmol/l)	1.06 (0.9–1.27)		1.16 (1.01–1.33)	< 0.001*
LDL-C (mmol/l)	2.59 (2.01–3.16)		3.08 (2.37–3.83)	< 0.001*
Diabetic retinopathy (n, %)	53 (9.9%)		44 (13.8%)	0.182
Diabetic neuropathy (n, %)	323 (60.5%)		71 (22.3%)	< 0.001*
Concomitant OCT1 inhibitors
Antidepressant (n, %)	31 (5.8%)		3 (0.9%)	< 0.001*
PPI (n, %)	11 (2%)		31 (9.7%)	< 0.001*
Doxazosin (n, %)	28 (5.2%)		1 (0.3%)	< 0.001*
Spirolactone (n, %)	24 (4.5%)		4 (1.3%)	0.009*
Clopidogrel (n, %)	56 (10.4%)		30 (9.4%)	0.724
Rosiglitazone (n, %)	10 (1.9%)		1 (0.3%)	0.062

Data are presented as median (interquartile range) for continuous variables and n (%) for categorical variables. Comparisons between cohorts were performed using the Wilcoxon rank-sum test for continuous variables and the chi-square test for categorical variables. P values < 0.05 were considered statistically significant

Abbreviations: BMI, body mass index; HbA1c, glycated hemoglobin; AST, aspartate aminotransferase; ALT, alanine aminotransferase; GGT, gamma-glutamyl transferase; TG, triglycerides; TC, total cholesterol; HDL-C, high-density lipoprotein cholesterol; LDL-C, low-density lipoprotein cholesterol; PPI, proton pump inhibitor

Model validation and performance evaluation

The final model’s performance was evaluated on both the cleaned training set and the independent validation set. The primary evaluation metric was the Area Under the Receiver Operating Characteristic Curve (AUC). The “pROC” package was used to compute the AUC and ROC curves. For clinical application, an optimal probability threshold of 0.12 was selected, yielding a sensitivity > 0.8 and specificity > 0.75 in the training set. The same threshold was then applied to the predictions on the validation set. The performances of prediction models were further examined through calibration metrics and plots (Figure S1) with Hosmer-Lemeshow test.

Results

Characteristics of the derivation and external validation cohorts

Based on the cohort selection flowchart (Fig. 1), patient data were derived from two independent centers. The derivation cohort was formed by merging Cohort A (n = 415) and Cohort B (n = 361), recruited from the Department of Endocrinology, Huadong Hospital, Fudan University between May 2013 and September 2018. Following the application of exclusion criteria, primarily a positive H. pylori test (n = 195) and age below 60 years (n = 43), 538 patients were included in the derivation cohort. From this cohort, 75 (13.9%) patients developed metformin intolerance. Separately, the external validation cohort consisted of Cohort C (n = 333), recruited from the National Metabolic Management Center, Huadong Hospital Minhang Branch between March 2021 and November 2024. After excluding 11 patients with incomplete data and 3 lost to follow-up, 319 patients comprised the external validation cohort, of whom 63 (19.7%) were identified as having metformin intolerance.

Fig. 1.

Cohort selection flowchart

The baseline characteristics of both cohorts are presented in Table 1. In the derivation cohort, the median age was 67 years (interquartile range [IQR], 64–73), 55.0% of patients were male, and the median body mass index (BMI) was 24.99 kg/m² (IQR, 23.2–27.24). Glycemic parameters included a median fasting plasma glucose of 7.6 mmol/L (IQR, 6.1–9.0) and a median HbA1c of 8.1% (65 mmol/mol; IQR, 7.5–8.6%). Patients in the external validation cohort differed in several respects. They were older (median age 70 years, IQR 66–73; P = 0.002), included a higher proportion of females (54.5%; P = 0.007), and had higher median fasting plasma glucose (8.1 mmol/L, IQR 6.75–9.85; P < 0.001). Significant differences were also observed in alcohol consumption, history of stroke, blood pressure, lipid profiles, and the use of OCT1-inhibiting medications such as antidepressant, PPIs, doxazosin and spironolactone (all P < 0.05). In contrast, key clinical variables including BMI, diabetes duration, and prevalence of coronary heart disease did not differ significantly between the cohorts. These differences in baseline demographics and clinical profiles reflect the expected heterogeneity between real-world patient populations and provide a robust setting for evaluating the external validity and generalizability of the prediction model.

Identification of predictors and development of a clinical scoring system

A total of 35 candidate predictors were evaluated in the initial analysis (see Table 1). Multivariable logistic regression analysis within the derivation cohort identified ten independent predictors of metformin intolerance (Table 2). Female gender (adjusted odds ratio [aOR] 3.15, 95% CI 1.69–5.87; P < 0.001) and older age (aOR 1.75 per year, 95% CI 1.34–2.28; P < 0.001) were significantly associated with an increased risk. Among clinical parameters, a higher BMI was protective (aOR 0.47, 95% CI 0.35–0.63; P < 0.001), whereas elevated aspartate aminotransferase (AST) levels constituted a notable risk factor (aOR 2.05, 95% CI 1.51–2.78; P < 0.001). Elevated fasting glucose (aOR 0.57, 95% CI 0.41–0.79; P < 0.001) and gamma-glutamyl transferase (GGT) levels (aOR 0.64, 95% CI 0.47–0.86; P < 0.001) also showed significant protective effects, while higher baseline HbA1c demonstrated a modest protective association (aOR 0.73, 95% CI 0.54–0.99; P = 0.042). Concurrent use of OCT1-inhibiting medications markedly increased intolerance risk. The use of PPIs was associated with the highest risk (aOR 6.89, 95% CI 1.27–37.46; P = 0.026), followed by clopidogrel (aOR 4.90, 95% CI 2.14–11.24; P < 0.001) and antidepressants (aOR 4.57, 95% CI 1.79–11.70; P < 0.001).

Table 2.

Multivariable logistic regression analysis for predictors of metformin intolerance in the derivation cohort

Variable	OR (95% CI)	Beta	SE	P Value
Intercept	-	-2.8163	0.8567	<0.001
Demographics
Female	3.15 (1.69–5.87)	1.1475	0.3184	<0.001
Age (years)	1.75 (1.34–2.28)	0.5576	0.1354	<0.001
Clinical Parameters
BMI (kg/m2)	0.47 (0.35–0.63)	-0.7552	0.1497	<0.001
Fasting glucose (mmol/L)	0.57 (0.41–0.79)	-0.5603	0.168	<0.001
HbA1c (%)	0.73 (0.54–0.99)	-0.3144	0.155	0.042
AST (U/L)	2.05 (1.51–2.78)	0.7189	0.1559	<0.001
GGT (U/L)	0.64 (0.47–0.86)	-0.4518	0.1531	<0.001
OCT1 inhibitors
Antidepressant	4.57 (1.79–11.70)	1.5201	0.4792	<0.001
PPIs	6.89 (1.27–37.46)	1.93	0.8674	0.026
Clopidogrel	4.90 (2.14–11.24)	1.59	0.4235	<0.001

The table presents the independent predictors retained in the final model, showing odds ratios (ORs) with 95% confidence intervals (CIs). For continuous variables, the OR represents the change in risk associated with each variable as modeled. For categorical variables, the OR represents the risk compared to the reference category (male sex for gender; non-use for medications). Corresponding regression coefficients (Beta), standard errors (SE)

Abbreviations: OR, odds ratio; CI, confidence interval; BMI, body mass index; AST, aspartate aminotransferase; GGT, gamma-glutamyl transferase; PPI, proton pump inhibitor

To translate these findings into a practical clinical tool, a points-based scoring system was developed (see Table 3). The six continuous variables were categorized into quartiles based on the derivation cohort distributions and assigned integer scores from 1 (lowest quartile) to 4 (highest quartile). The four categorical variables were assigned binary scores. For an individual patient, a linear predictor score is calculated as the sum of the product of each variable’s score and its corresponding regression coefficient (Beta from Table 2): score = Σ (Beta × variable) + Intercept. The regression coefficients (Beta) represent the weight and direction of influence for each predictor. This linear score is then transformed into a probability of intolerance (p) using the sigmoid (logistic) function: , ensuring the output is bounded between 0 and 1.

Table 3.

Scoring rules of Logistic Regression Model

Variable Name	Scoring Method	Scoring Rules	Individual Score	Coefficient (Beta)
Demographics
Gender	Male: 1pt, Female: 2pt		1 or 2	1.15
Age (years)	Q1(< 64):1pt, Q2(64–67):2pt, Q3(68–73):3pt, Q4(> 73):4pt	Per quartile increase	1–4 points	0.56
Clinical biomarkers
BMI (kg/m²)	Q1(< 23.2):1pt, Q2(23.2–25.0):2pt, Q3(25.0-27.4):3pt, Q4(> 27.4):4pt	Per quartile increase	1–4 points	-0.76
Fasting Glucose (mmol/L)	Q1(< 6.1):1pt, Q2(6.1–7.6):2pt, Q3(7.6-9.0):3pt, Q4(> 9.0):4pt	Per quartile increase	1–4 points	-0.56
HbA1c (%)	Q1(< 7.5):1pt, Q2(7.5–8.1):2pt, Q3(8.1–8.6):3pt, Q4(> 8.6):4pt	Per quartile increase	1–4 points	-0.31
AST (IU/L)	Q1(< 14):1pt, Q2(14–17):2pt, Q3(17–23):3pt, Q4(> 23):4pt	Per quartile increase	1–4 points	0.72
GGT (IU/L)	Q1(< 17):1pt, Q2(17–24):2pt, Q3(24–36):3pt, Q4(> 36):4pt	Per quartile increase	1–4 points	-0.45
OCT1 inhibitors
Antidepressants	0 = Not used, 1 = Used	Use: 1 point	0 or 1	1.52
PPIs	0 = Not used, 1 = Used	Use: 1 point	0 or 1	1.93
Clopidogrel	0 = Not used, 1 = Used	Use: 1 point	0 or 1	1.59

Points are assigned based on patient characteristics. Continuous variables are categorized into quartiles (Q1-Q4) using cutoff values derived from the derivation cohort distribution. Categorical variables are assigned binary points. The linear predictor score is calculated as the sum of each variable’s points multiplied by its beta coefficient (Table 2), plus the intercept of -2.82: Score = Σ(Points × Beta) + Intercept. This score is then converted to a probability via the logistic function

Abbreviations: BMI, body mass index; HbA1c, glycated hemoglobin; AST, aspartate aminotransferase; GGT, gamma-glutamyl transferase; PPI, proton pump inhibitor

Model performance in the derivation and external validation cohorts

We first evaluated the model’s performance in the derivation cohort (n = 538). As shown in Fig. 2A (ROC curve for the derivation cohort), the model demonstrated excellent discriminatory ability, with an area under the curve (AUC) of 0.889 (95% CI: 0.858–0.921). The optimal cutoff (probability threshold of 0.12) for clinical use was determined by maximizing Youden’s index. At this threshold, the model achieved a sensitivity of 0.813 and a specificity of 0.773.

Fig. 2.

Receiver operating characteristic (ROC) curves of the metformin intolerance prediction model (A) ROC curve for the derivation cohort (n = 538); (B) ROC curve for the external validation cohort (n = 319)

The model was subsequently applied to the independent external validation cohort (n = 319). Its performance in this cohort is shown in Fig. 2B (ROC curve for the external validation cohort). The model maintained good discriminative ability, with an AUC of 0.883 (95% CI: 0.842–0.925). Applying the same probability threshold of 0.12, the sensitivity was 0.857 and the specificity was 0.723. The calibration plots (Figure S1) demonstrated good agreement between predicted and observed probabilities, supporting the model’s reliability.

Discussion

Our study successfully developed and externally validated a multivariable prediction model for metformin intolerance in elderly, metformin-naïve patients with type 2 diabetes for the first time. Utilizing clinical data from a derivation cohort of 538 patients, we identified ten clinical parameters: gender, age, BMI, fasting glucose, HbA1c, AST, GGT level, the use of antidepressants, PPIs and clopidogrel, which were independently associated with intolerance risk. The model demonstrated excellent discriminative ability in the derivation cohort (AUC 0.889) and maintained good performance upon external validation in an independent cohort of 319 patients (AUC 0.883). The derived linear prediction model provides a direct tool for pre-emptive identification of high-risk individuals, aiming to prevent treatment interruption of metformin in elderly patients. Although our model includes laboratory measures, these tests are routinely performed in clinical practice and are widely accessible for diabetic patients in China, supporting the practical applicability of the scoring system.

The efficacy of metformin was dose-dependent, but the GI AEs did not seem to have this feature [8, 14, 15]. Therefore, metformin dose was not included as a predictor variable in the final model. All patients in our cohorts received a minimum daily metformin dose of 1000 mg, and the developed model is therefore applicable to patients taking ≥ 1000 mg/day. Despite this clinical importance, the underlying mechanisms of the GI intolerance associated with metformin are poorly known. According to previous studies, the reported AEs may be caused by alteration of gut microbiota [16–18], altered metformin uptake from the intestine [19], OCT1 and serotonin transporter polymorphism [13, 20–22], enhancement of glucose uptake and glucagon like peptide-1 [9], and reduction of bile acid absorption [6, 23]. However, there was no significant difference in metformin pharmacokinetics or systemic measures of lactate, serotonin or bile acids in severe intolerance cohort [24]. Metformin XR tablets stay in the stomach longer and release more slowly in the small intestine than metformin IR tablets. In theory, metformin XR has the advantages of only needing to be taken once a day and having better GI tolerance. However, the use of metformin XR has not achieved ideal results in some clinical trials [10, 25]. Although starting at a low dose and titrating slowly may help prevent some GI AEs associated with metformin [5], some patients are unable to tolerate metformin at all and it may also be difficult to convince patients to start metformin again after a bout of AEs [26]. The core finding of our study is that metformin intolerance in the elderly can be predicted with good accuracy using common clinical data before treatment. This addresses a significant pragmatic gap in diabetes management. The under-prescription and under-dosing of metformin in China, largely driven by fear of GI AEs, may lead to suboptimal glycemic control. Our model offers a proactive strategy to mitigate this issue by enabling personalized risk assessment. There are multiple glucose-lowering drugs available as alternative initial monotherapy for type 2 diabetes patients with intolerance to metformin, such as sulfonylureas (SUs), thiazolidinediones (TZDs), glinides (NIDEs), α-glucosidase inhibitors (AGIs), dipeptidyl peptidase-4 inhibitors (DPP-4is), sodium-glucose cotransporter-2 inhibitors (SGLT2is), insulins (INSs), and glucagon-like peptide-1 receptor agonists (GLP-1RAs) [27].

The identified predictors align with and extend existing pathophysiological understanding. Female gender, older age and lower BMI are established risk factors for metformin-related GI AEs, possibly due to differences in pharmacokinetics and pharmacodynamics [11, 22]. In addition, a previous study has found that patients with diabetes with lower BMI or HbA1c are less likely to prescribe metformin, but it cannot be confirmed whether it is related to metformin intolerance [28]. Elderly patients with diabetes usually have several complications and tend to take a variety of concomitant medications, such as antidepressants, PPIs, clopidogrel and verapamil——all of which belong to OCT1 inhibitors. In previous studies, reduced OCT1 transport is an important determinant of metformin intolerance [13, 20, 22]. Intolerance was over four times more likely to develop in individuals with two reduced-function OCT1 alleles who were treated with OCT1 inhibitors [13]. In our cohort, the inclusion of concomitant OCT1 inhibitors such as antidepressants, PPIs and clopidogrel independently associated with metformin intolerance. Notably, elevated levels of both AST and GGT were identified as independent predictors of metformin intolerance, yet they exhibited opposite predictive values. However, current pharmacokinetic evidence indicates no clinically significant alteration in metformin metabolism in patients with chronic liver disease [29], ruling out a direct path of drug accumulation. Elevated AST predominantly signals hepatocellular injury and is a common marker of underlying metabolic dysfunction, such as insulin resistance and hepatic steatosis. In contrast, elevated GGT is more closely associated with conditions like cholestasis, alcohol consumption, or altered drug metabolism [30]. Elevated AST or GGT likely serves as a marker for an underlying systemic metabolic disturbance, such as steatosis liver disease and insulin resistance. This disturbed state may synergize with metformin’s actions on the gut-liver axis, collectively lowering the GI tolerance threshold through multiple pathways [6]. Further studies are needed to confirm this hypothesis.

This study has several limitations. First, the external validation cohort lacked patients using antidepressants, preventing validation of this specific predictor in a new population. Second, although the derivation and validation cohorts were from independent centers, they were geographically similar. The relatively small number of events and the limited sample size of the derivation cohort introduce a potential risk of overfitting, which could affect the model’s application. When stepwise regression is used for feature selection, it may not only lead to overfitting but also cause information loss from categorization. Multicenter validation across diverse regions and populations is needed to improve the generalizability of the model. Third, the model does not include genetic variants of OCT1 or detailed gut microbiome data, which could improve precision but reduce clinical practicality. Our model prioritizes immediate clinical feasibility. Further research should focus on interventional trials to assess the clinical utility and impact of implementing this prediction model. A prospective study could evaluate whether using the model to guide therapy could actually reduce intolerance rates, improve glycemic outcomes, and increase clinician confidence in prescribing metformin.

Conclusion

In conclusion, we have developed and validated a practical prediction model for metformin intolerance in Chinese elderly patients with type 2 diabetes. It leverages easily obtainable clinical variables, including gender, age, BMI, fasting glucose, HbA1c, AST, GGT level, the use of antidepressants, PPIs and clopidogrel. This model may help physicians to be more confident when prescribing metformin, and reduce the intolerance of metformin in elderly patients.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

BMI

Body mass index

HbA1c

Glycated hemoglobin

AST

Aspartate aminotransferase

ALT

Alanine aminotransferase

GGT

Gamma-glutamyl transferase

TG

Triglycerides

TC

Total cholesterol

HDL-C

High-density lipoprotein cholesterol

LDL-C

Low-density lipoprotein cholesterol

PPI

Proton pump inhibitor

Author contributions

WT: Formal analysis, Investigation, Methodology, Writing; YL: Data curation, Investigation, Resources; BL: Project administration, Supervision; JZ: Writing - review & editing, Software, Supervision, Funding acquisition; YH: Funding acquisition, Project administration, Validation, Visualization.

Funding

This work was supported by Noncommunicable Chronic Diseases-National Science and Technology Major Project (2023ZD0507203), Emerging Talent Program of Huadong Hospital, Fudan University (XXRC2217), Yangfan Project, Science and Technology Commission of Shanghai Municipality (23YF1411000), and Young Scientists Fund of National Natural Science Foundation of China (82304452).

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

The work described has been carried out in accordance with the Declaration of Helsinki. This study was approved by the Ethics Committee of Huadong Hospital, Fudan University, Shanghai, China (Ethics Number: 2024K214). All procedures were performed in compliance with relevant laws and institutional guidelines. Written informed consent was obtained from all patients prior to the study.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.

Abbreviations

OR, odds ratio; CI, confidence interval; BMI, body mass index; AST, aspartate aminotransferase; GGT, gamma-glutamyl transferase; PPI, proton pump inhibitor.

Abbreviations

BMI, body mass index; HbA1c, glycated hemoglobin; AST, aspartate aminotransferase; GGT, gamma-glutamyl transferase; PPI, proton pump inhibitor.