Treatment Selection Patterns and Associated Outcomes for Biguanides and SGLT2 Inhibitors in Type 2 Diabetes: A Retrospective Database Study in Japan

Abstract

Background

Type 2 diabetes mellitus has multiple treatment options and high healthcare costs. In Japan, DPP-4 inhibitors have been widely used, but since 2022 the pharmacotherapy algorithm has shifted: DPP-4 inhibitors and biguanides for BMI < 25 kg/m$_{}^2$, and biguanides and SGLT2 inhibitors for BMI $\ge$ 25 kg/m$_{}^2$. However, real-world factors influencing prescribing decisions between biguanides and SGLT2 inhibitors remain unclear.

Objective

This study investigates factors influencing the choice between biguanides and SGLT2 inhibitors in clinical practice. We also compare glycemic efficacy and medication persistence while accounting for baseline characteristics.

Methods

This retrospective cohort study used the Millennium Medical Record Database, primarily capturing care from large hospitals. Adults aged $\ge$ 18 years who received a first prescription of a biguanide or a SGLT2 inhibitor between July 2019 and April 2023 were included. Patients were categorized as 1st-line or 2nd-line therapy. Confounding was addressed using coarsened exact matching with classification tree modeling. HbA1c changes were analyzed using mixed models for repeated measures, and persistence using Kaplan–Meier methods.

Results

A total of 1925 patients were included. In the 1st-line cohort (n = 1440), key selection factors were diuretic use, lipid-modifying agents, and baseline HbA1c. At 24 weeks, HbA1c decreased in both groups (biguanide: −1.37 percentage points, 95% CI −1.48 to −1.26; SGLT2 inhibitor: −1.16 percentage points, 95% CI −1.29 to −1.02); the between-group difference was −0.21 percentage points (95% CI −0.38 to −0.04). Persistence at 48 weeks was 47.0% (95% CI 42.7–51.2) for biguanides and 56.2% (95% CI 51.2–61.2) for SGLT2 inhibitors. In the 2nd-line cohort (n = 485), selection factors were baseline HbA1c, BMI, and diuretic use. HbA1c decreased at 24 weeks (biguanide: −2.26 percentage points, 95% CI −2.47 to −2.05; SGLT2 inhibitor: −1.79 percentage points, 95% CI −2.04 to −1.54); the between-group difference was –0.47 percentage points (95% CI −0.80 to −0.15). Persistence at 48 weeks was 51.8% (95% CI 45.2–58.4) for biguanides and 49.4% (95% CI 40.3–58.6) for SGLT2 inhibitors.

Conclusions

In Japanese patients with type 2 diabetes, treatment selection was mainly associated with concomitant cardiovascular medication use. BMI influenced choice, particularly in the 2nd-line setting, but was not the predominant determinant. Biguanides were associated with greater HbA1c reductions. In intention-to-treat analyses, persistence over 48 weeks was higher with SGLT2 inhibitors in the 1st-line cohort, while persistence was comparable in the 2nd-line cohort.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s40801-026-00567-5.

Key Points

Treatment selection was primarily driven by cardiovascular risk, which contrasts with the BMI-based approach recommended in current Japanese treatment guidelines.

Biguanides showed a more pronounced reduction in HbA1c at 24 weeks.

SGLT2 inhibitors, in 1st-line cohort, demonstrated higher treatment persistence over 48 weeks.

Introduction

Background

Type 2 diabetes mellitus (T2DM), accounting for approximately 90% of all diabetes cases [1], is a complex metabolic disorder characterized by hyperglycemia resulting from insulin resistance and $\beta$-cell dysfunction [2]. The prevention of complications through appropriate glycemic control is an essential goal of diabetes management. T2DM has reached epidemic proportions globally [1], having a significant impact on healthcare finances worldwide [3]. Japan is experiencing the same trend [4, 5]. The selection of appropriate treatment methods is important from both clinical and cost-effectiveness perspectives.

Characteristics of Diabetes Treatment in Japan

Japan has implemented treatment approaches based on its unique guidelines and practices. Japan’s distinctive approach is evident in the high prescription rate of dipeptidyl peptidase-4 (DPP-4) inhibitors. Research suggests that the pathophysiology of diabetes differs among ethnic groups [6, 7], and the characteristics of impaired insulin secretion relative to insulin resistance may favor the selection of DPP-4 inhibitors. It is attributed to the safety profile in elderly populations and ease of use in patients with renal impairment [8, 9]. According to a study using the National Database of Health Insurance Claims and Specific Health Check-ups of Japan [9], DPP-4 inhibitors were the most prescribed medication from 2014 to 2017, contrasting with Western countries where biguanides have long been the first-choice treatment. A previous study reported that some non-Japan Diabetes Society-certified facilities prescribed DPP-4 inhibitors to nearly 100% of patients [9].

Biguanides and SGLT2 inhibitors have emerged as key medications in Japan’s diabetes treatment landscape, shifting away from DPP-4 inhibitors (Supplementary Fig. S1). Metformin, the most widely used biguanide, was approved in 1961 but initially saw limited use due to safety concerns but was later reconsidered after positive UK trial results and its lower economic burden [10]. Its use expanded in 2019 to include patients with mild-to-moderate renal impairment (eGFR $\ge$ 30) [11]. SGLT2 inhibitors, approved in 2014, are a relatively new glucose-lowering class and more costly but offer benefits beyond glycemic control, including cardiovascular and renal protection [12–14].

Japan’s pharmacotherapy algorithm for type 2 diabetes [15] represents a paradigm shift in treatment strategies, being the first specifically tailored for Japanese patients that addresses the overreliance on DPP-4 inhibitors by recommending a BMI-based approach: DPP-4 inhibitors/biguanides for patients with BMI $<$25 kg/m$_{}^2$, and biguanides/SGLT2 inhibitors for those with BMI $\ge$ 25 kg/m$_{}^2$, while maintaining flexibility for personalized medicine based on individual patient characteristics. Analysis of Japan’s National Database of Health Insurance Claims reveals a consistent trend showing decreasing prescription rates for DPP-4 inhibitors while biguanides and SGLT2 inhibitors prescriptions continue to rise, reflecting the successful implementation of these evidence-based guidelines and changing clinical practices among Japanese physicians [9].

Current Challenges

Although patients may be eligible for both biguanide and SGLT2 inhibitor treatments, the factors influencing prescription decisions remain unclear. Current guidelines lack specific numerical criteria for medication selection. Identifying key factors and decision thresholds would help nondiabetes specialists make more informed treatment choices and enable more personalized patient care. Understanding these prescription patterns also has economic significance, as biguanides provide cost advantages over newer medications, potentially helping to manage rising healthcare expenditures.

Research Objectives

This study mainly aims to investigate (1) the influential factors affecting the selection between biguanides and SGLT2 inhibitors in clinical practice. Also, we compare (2) their glycemic efficacy and (3) medication persistence rates in patients with matched baseline characteristics.

Methods

Population

This study was a retrospective cohort study using a clinical database. This retrospective cohort study included adult patients aged 18 years or older who received a first prescription of either a biguanide or an SGLT2 inhibitor between 1 July 2019 and 30 April 2023 (index date: first prescription date). We excluded patients who received concurrent prescriptions of both drug classes at the index date, and those diagnosed with type 1 diabetes or gestational diabetes. To ensure that the index prescription was the patient’s first observed prescription in the database and to ascertain baseline covariates, we required at least 90 days of available data prior to the index date, which is the baseline look-back period. We further required at least 28 days of observable postindex follow-up, regardless of initial treatment continuation, to evaluate postindex trajectories. Finally, we excluded patients with baseline eGFR below 30 mL/min/1.73 m$_{}^2$ (metformin is contraindicated), those receiving renal replacement therapy, those diagnosed with renal failure or heart failure, and those with a history of systemic steroid prescription, as these conditions could substantially restrict treatment selection and influence outcomes. The full list of inclusion and exclusion criteria is provided in the Supplementary Table S2. Details of clinical and laboratory data with corresponding codes are provided in the Supplementary Table S3.

Patients were categorized on the basis of their first prescription of either a biguanide or SGLT2 inhibitor during the study period. We divided patients into 1st-line (no previous antidiabetic medication before index prescription) or 2nd-line (with prior antidiabetic medication history) treatment groups for separate analyses. For analytical purposes, patients remained in their initially assigned treatment groups even if they switched medications during follow-up. Sample size calculations were not performed for exploratory purposes and include all data available.

Data Source

This study utilized the Millennium Medical Record Database, a comprehensive clinical data repository established under Japan’s Next Generation Medical Infrastructure Act and managed by the Life Data Initiative [16]. The database integrates electronic health records, diagnosis procedure combination data, and insurance claims from 54 medical institutions located nationwide across Japan, containing approximately 1.79 million patient records as of April 2023. About 70% of these records come from large hospitals with more than 300 beds. A key strength of this database is its collection of identifiable information, enabling linkage of multiple data sources and potential tracking of patients across participating institutions. This mitigates a common limitation of electronic health record databases regarding patient transfers. The database employs standardized coding systems: diagnoses are recorded using the International Classification of Diseases, 10th revision (ICD-10), and medications are classified according to the World Health Organization Anatomical Therapeutic Chemical (WHO-ATC) classification system.

Ethics

This study received approval from the Millennium Medical Record Database Review Committee (2024_MIL_0002) and was registered with the University of Tokyo Clinical Research Ethics Committee (2024296NIe). The study was conducted in accordance with the principles of the Declaration of Helsinki and Japan’s Next Generation Medical Infrastructure Act. Under this Act, patient consent requirements were waived as researchers received only anonymized processed information. All data were handled under strict access controls. The study involved no direct patient intervention, so no compensation was provided to participants.

Outcome Measures

The primary outcomes in this study were changes in HbA1c levels and treatment persistence. Glycemic efficacy was assessed by measuring HbA1c changes from baseline to 24 weeks. Specific measurement windows were defined as 28 days (± 14 days) for 4-week values, 84 days (± 28 days) for 12-week values, and 168 days (± 56 days) for 24-week values. When multiple measurements existed within these periods, the value closest to the reference date was selected.

Treatment persistence was also considered a key factor to evaluate the effectiveness of those medications. Treatment persistence was defined as the duration from initiation of the index drug to discontinuation of the index drug, where discontinuation was defined using a 60-day grace period such that the event day (discontinuation date) was set to 60 days after the last prescription date of the index medication. Subsequent drugs were defined as any antidiabetic treatment administered following the last prescription of the index medication.

Statistical Analysis

Objective 1: Treatment selection factors

Patient characteristics were evaluated by calculating frequencies with percentages for categorical variables and medians with interquartile ranges (IQR; 25th and 75th percentiles) for continuous variables. Between-group differences were assessed using absolute standardized mean differences (aSMD).

We used classification and regression tree (CART) analysis to identify and visually summarize key variables associated with treatment selection between biguanides and SGLT2 inhibitors. Trees were constructed using propensity scores estimated from a logistic regression model (after applying the inverse-logit transformation). The estimated propensity scores were used to repeatedly split the population into two subclasses at each node by selecting the split that maximized between-node heterogeneity. The minimum terminal node size was set to 5% of the total cohort, and the maximum tree depth was set to 4.

CART models were developed with a comprehensive set of covariates including demographic characteristics (age, gender, duration since diabetes diagnosis), diabetes classifications (type 2 diabetes, other specified diabetes, unspecified diabetes), diabetic complications (nephropathy, retinopathy, neuropathy), comorbidities (hypertension, ischemic heart disease, heart failure, cerebrovascular disorders, malignant neoplasm, hepatic dysfunction, renal failure, dyslipidemia, dementia, peripheral arterial disease), concomitant medications (insulin, antithrombotic agents, antihypertensives, diuretics, vasodilators, beta-blockers, calcium channel blockers, renin-angiotensin system agents, lipid-modifying agents, systemic corticosteroids, psychotropic drugs), laboratory data (estimated glomerular filtration rate (eGFR), hemoglobin, HbA1c, and low-density lipoprotein (LDL) cholesterol), as well as anthropometric and vital sign measurements (body mass index (BMI), systolic blood pressure (SBP), and diastolic blood pressure (DBP)).

The CART models described above were also used to address confounding via coarsened exact matching (CEM). Specifically, the terminal nodes of the CART were used as propensity score strata. CEM was then performed within these strata to derive matching weights for subsequent outcome analyses. In this sense, CEM+CART approach can be viewed as propensity score stratification, analogous to propensity score matching (PSM). In this study, CEM implementation was configured as a full-matching approach, so that there was no algorithmic exclusion due to matching itself. In the weighted analysis, we applied stabilized inverse probability of treatment weights (IPTW): $w_i=\frac{P(T=t_i)}{P(T=t_i\mid C_i)}$, where $w_i$ is the weight for the $i$th patient, $T$ denotes treatment assignment, and $C_i$ denotes the subclass to which the $i$th patient belonged. Here, $P(T=t_i)$ is the marginal proportion of receiving treatment $t_i$ in the overall cohort, and $P(T=t_i\mid C_i)$ is the proportion of patients receiving treatment $t_i$ within $C_i$.

Objective 2: Glycemic effectiveness

Glycemic efficacy was analyzed using a mixed model for repeated measures (MMRM) with matching weights [17]. The model included fixed effects for treatment group, time point, their interaction, and baseline HbA1c. Model-based standard errors and 95% CIs for least-squares means and between-group differences at each visit were estimated. Within-subject correlation was modeled using a first-order autoregressive covariance structure (AR(1)), prespecified a priori given the expectation that correlations are stronger for measurements taken closer in time. As a sensitivity analysis, we also fitted a compound symmetry (CS) covariance structure.

Objective 3: Treatment persistence

Treatment persistence was analyzed using Kaplan–Meier survival analysis with matched weights. The 95% CIs for the Kaplan–Meier curves were calculated using Greenwood’s formula to estimate the pointwise variance, with a normal approximation on the linear scale. The 95% CIs for the median time were calculated using the Brookmeyer–Crowley method.

For this analysis, we defined the event day (discontinuation date) as 60 days after the last prescription date of the index medication (i.e., applying a 60-day grace period). Patients were censored at the earlier of (A) the last prescription date of any glucose-lowering drug or (B) the last HbA1c measurement date, which represents the last date on which follow-up could be confirmed in the database. If follow-up ended before completion of the 60-day grace period after the last index prescription, the observation was treated as censored rather than classified as discontinuation, because the derived event day (last index prescription date + 60 days) could fall beyond the observable follow-up window. We considered this situation a potential misclassification of treatment discontinuation because patients might have continued hospital visits, but subsequent information was unavailable because it had not yet been recorded at the time of data extraction.

Addressing Missing Data

Missing covariate data were handled using Multiple Imputation by Chained Equations (MICE) in the whole analytic cohort. The imputation model included all baseline covariates used for confounding adjustment and also treatment-line. We generated m = 20 imputed datasets using MICE with predictive mean matching (PMM). PMM was selected to maintain the distributional characteristics of the original variables while addressing potential biases from complete-case analysis.

As for missing outcome data, following the methodology of previous research [18], we did not exclude patients with missing follow-up HbA1c values during the matching process. Instead, these patients were excluded from the subsequent HbA1c analysis.

For the longitudinal HbA1c analysis using MMRM, model estimates were obtained within each imputed dataset and then pooled using Rubin’s rules [19] to obtain final point estimates and standard errors.

For the treatment persistence analysis, because pooling Kaplan–Meier curves across multiple imputations is not straightforward, we constructed a single set of analysis weights based on an average of estimated propensity scores and applied these weights to the Kaplan–Meier estimation.

Sensitivity Analysis for matching procedure

We performed PSM as a sensitivity analysis to ensure robustness of our findings. Estimated propensity scores with logistic regression model induced bins of cohort. The number of subclassifications was set to 10. Same as CEM, trimming was not applied to PSM and stabilized inverse probability weight was used.

Statistical Software

All statistical analyses were performed using R statistical software version 4.3.3 (2024-02-29 ucrt) with the appropriate package, especially including gtsummary(1.7.2), mice(3.16.0), MatchIt(4.5.5), ranger(0.16.0), survival(3.5.8), mmrm(0.3.14), emmeans(1.10.5).

Results

Patient Selection

The analysis set comprised 1925 patients (follow-up time median 315 days, IQR 126–558). The analysis set comprised 1440 1st-line treatment recipients (861 biguanides, 579 SGLT2 inhibitors) and 485 2nd-line treatment recipients (302 biguanides, 183 SGLT2 inhibitors).

The patient flow has been explained below in Fig. 1.

Fig. 1.

Patient selection flow diagram. eGFR, estimated glomerular filtration rate; SGLT2, sodium-glucose cotransporter-2

Demographic and other baseline characteristics

In the 1st-line treatment group (Table 1, n = 1440), the SGLT2 inhibitor group had approximately 8 points more males, approximately 15 points higher rates of cardiovascular risk management medication, and a 7 mL/min/1.73 m$_{}^2$ lower median eGFR compared with the biguanide group. Median HbA1c values were 0.5 points lower in the SGLT2 inhibitor group and there was only 0.2 kg/m$_{}^2$ difference in BMI between the groups.

Table 1.

Baseline demographics for 1st-line treatment

Characteristics	Overall (n =  440)	Unadjusted			Adjusted (CEM)
		Biguanides (n = 861)	SGLT2 inhibitors (n = 579)	aSMD	Biguanides (n = 857)	SGLT2 inhibitors (n = 572)	aSMD
Demographics
Age, years, n (%)				0.14			0.13
18–34	49 (3.4)	33 (3.8)	16 (2.8)		27 (3.2)	18 (3.2)
35–44	81 (5.6)	49 (5.7)	32 (5.5)		46 (5.4)	36 (6.3)
45–54	199 (13.8)	123 (14.3)	76 (13.1)		122 (14.3)	79 (13.8)
55–64	269 (18.7)	157 (18.2)	112 (19.3)		156 (18.1)	114 (19.9)
65–74	494 (34.3)	309 (35.9)	185 (32.0)		314 (36.7)	177 (31.0)
75−	348 (24.2)	190 (22.1)	158 (27.3)		192 (22.4)	148 (25.8)
Gender, female, n (%)	489 (34.0)	321 (37.3)	168 (29.0)	0.18	302 (35.2)	185 (32.3)	0.06
Days from first diagnosis record to prescription, median (IQR)	17 (4, 56)	16 (5, 42)	21 (2, 121)	0.29	16.8 (5.0, 49.0)	17.0 (3.0, 91.0)	0.20
Anthropometric, vital sign
BMI, kg/m2, median (IQR)	25.1 (22.4, 28.4)	25.2 (22.4, 28.3)	25.0 (22.5, 28.5)	0.00	25.4 (22.5, 28.3)	25.0 (22.5, 28.9)	0.02
SBP, mmHg, median (IQR)	127 (112, 139)	128 (112, 140)	125 (113, 138)	0.09	127 (111, 140)	125 (113, 137)	0.03
Comorbidities
Diabetic nephropathy, n (%)	34 (2.4)	20 (2.3)	14 (2.4)	0.01	18 (2.1)	17 (2.9)	0.06
Diabetic retinopathy, n (%)	88 (6.1)	51 (5.9)	37 (6.4)	0.02	45 (5.3)	45 (7.9)	0.11
Diabetic neuropathy, n (%)	17 (1.2)	–$_{}^{\ast }$	–$_{}^{\ast }$	–$_{}^{\ast }$	–$_{}^{\ast }$	–$_{}^{\ast }$	–$_{}^{\ast }$
Hypertension, n (%)	242 (16.8)	141 (16.4)	101 (17.4)	0.03	143 (16.7)	102 (17.9)	0.03
Ischemic heart disease, n (%)	111 (7.7)	54 (6.3)	57 (9.8)	0.13	69 (8.0)	53 (9.3)	0.05
Cerebrovascular disease, n (%)	52 (3.6)	35 (4.1)	17 (2.9)	0.06	37 (4.3)	16 (2.8)	0.08
Concomitant medications
Insulin, n (%)	279 (19.4)	194 (22.5)	85 (14.7)	0.20	198 (23.1)	93 (16.2)	0.18
Diuretics, n (%)	110 (7.6)	20 (2.3)	90 (15.5)	0.48	66 (7.7)	44 (7.7)	0.00
Beta-blockers, n (%)	152 (10.6)	40 (4.6)	112 (19.3)	0.46	65 (7.5)	76 (13.2)	0.18
Calcium channel blockers, n (%)	186 (12.9)	91 (10.6)	95 (16.4)	0.17	112 (13.1)	77 (13.5)	0.01
Renin-angiotensin system agents, n (%)	219 (15.2)	74 (8.6)	145 (25.0)	0.45	114 (13.3)	102 (17.8)	0.12
Lipid-modifying agents, n (%)	261 (18.1)	101 (11.7)	160 (27.6)	0.41	142 (16.6)	106 (18.5)	0.05
Laboratory values
HbA1c, %, median (IQR)	7.6 (6.8, 8.8)	7.8 (7.0, 9.2)	7.3 (6.5, 8.3)	0.38	7.7 (7.0, 8.9)	7.5 (6.6, 8.6)	0.17
eGFR, mL/min/1.73 m2, median (IQR)	69.3 (56.2, 85.8)	72.4 (59.9, 88.7)	65.5 (50.2, 80.5)	0.36	70.8 (58.3, 86.4)	66.8 (52.6, 83.7)	0.20
Hb, g/dL, median (IQR)	14.0 (12.6, 15.2)	14.0 (12.6, 15.0)	14.0 (12.6, 15.3)	0.04	13.9 (12.5, 15.0)	14.2 (12.7, 15.4)	0.11
LDL, mg/dL, median (IQR)	107 (86, 133)	115 (91, 139)	99 (80, 126)	0.29	113 (90, 138)	101 (81, 128)	0.21

aSMD absolute standardized mean difference, CEM coarsened exact matching, eGFR estimated glomerular filtration rate, Hb hemoglobin, IQR interquartile range, LDL low-density lipoprotein cholesterol, SBP systolic blood pressure, SGLT2 sodium-glucose cotransporter-2

$_{}^{\ast }$ In accordance with the rules for publishing aggregated results of anonymized information, levels with a small number of subjects are omitted

In the 2nd-line treatment group (Table 2; n = 485), similar patterns were observed. Compared with the biguanide group, the SGLT2 inhibitor group included a higher proportion of males and higher rates of prescriptions for cardiovascular risk management medication, consistent with the patterns observed in the 1st-line group. The SGLT2 inhibitor group also had a 0.4 point lower median HbA1c and a 0.9 kg/m$_{}^2$ higher median BMI.

Table 2.

Baseline demographics for 2nd-line treatment

		Unadjusted			Adjusted (CEM)
Characteristics	Overall (n = 485)	Biguanides (n = 302)	SGLT2 inhibitors (n = 183)	aSMD	Biguanides (n = 300)	SGLT2 inhibitors (n = 180)	aSMD
Demographics
Age, years, n (%)				0.12			0.16
18–34	13 (2.7)	7 (2.3)	6 (3.3)		6 (2.1)	7 (3.7)
35–44	32 (6.6)	21 (7.0)	11 (6.0)		20 (6.7)	12 (6.5)
45–54	75 (15.5)	44 (14.6)	31 (16.9)		44 (14.7)	32 (17.5)
55–64	106 (21.9)	69 (22.8)	37 (20.2)		66 (22.1)	34 (19.0)
65–74	161 (33.2)	102 (33.8)	59 (32.2)		103 (34.4)	56 (31.0)
75−	98 (20.2)	59 (19.5)	39 (21.3)		60 (20.0)	40 (22.3)
Gender, female, n (%)	150 (30.9)	104 (34.4)	46 (25.1)	0.20	96 (31.8)	53 (29.5)	0.05
Days from first diagnosis record to prescription, median (IQR)	35 (14, 140)	31 (14, 100)	42 (15, 219)	0.18	33 (14, 117)	37 (14, 192)	0.11
Anthropometric, vital sign
BMI, kg/m2, median (IQR)	24.9 (22.2, 27.9)	24.6 (21.9, 27.4)	25.5 (22.7, 28.5)	0.25	24.8 (22.1, 27.9)	25.2 (22.5, 28.4)	0.19
SBP, mmHg, median (IQR)	128 (114, 140)	126 (113, 138)	131 (114, 144)	0.21	127 (113, 140)	130 (113, 142)	0.15
Comorbidities
Diabetic nephropathy, n (%)	15 (3.1)	8 (2.6)	7 (3.8)	0.07	7 (2.5)	7 (3.6)	0.07
Diabetic retinopathy, n (%)	47 (9.7)	36 (11.9)	11 (6.0)	0.21	32 (10.6)	13 (7.5)	0.11
Diabetic neuropathy, n (%)	13 (2.7)	9 (3.0)	4 (2.2)	0.05	8 (2.8)	4 (2.0)	0.05
Hypertension, n (%)	106 (21.9)	66 (21.9)	40 (21.9)	0.00	66 (22.1)	40 (22.1)	0.00
Ischemic heart disease, n (%)	43 (8.9)	22 (7.3)	21 (11.5)	0.14	25 (8.5)	19 (10.4)	0.07
Cerebrovascular disease, n (%)	27 (5.6)	18 (6.0)	9 (4.9)	0.05	18 (6.1)	8 (4.7)	0.06
Concomitant medications
Insulin, n (%)	233 (48.0)	150 (49.7)	83 (45.4)	0.09	146 (48.7)	83 (46.1)	0.05
Sulfonylureas, n (%)	100 (20.6)	69 (22.8)	31 (16.9)	0.15	71 (23.6)	30 (16.6)	0.17
Combinations of oral blood glucose lowering drugs, n (%)	65 (13.4)	37 (12.3)	28 (15.3)	0.09	40 (13.2)	24 (13.6)	0.01
Alpha-glucosidase inhibitors, n (%)	32 (6.6)	17 (5.6)	15 (8.2)	0.10	17 (5.8)	15 (8.3)	0.10
Thiazolidinediones, n (%)	22 (4.5)	14 (4.6)	8 (4.4)	0.01	14 (4.7)	7 (3.9)	0.03
DPP-4 inhibitors, n (%)	340 (70.1)	217 (71.9)	123 (67.2)	0.10	213 (71.0)	121 (67.5)	0.08
GLP-1 receptor agonists, n (%)	39 (8.0)	18 (6.0)	21 (11.5)	0.20	17 (5.8)	22 (12.1)	0.22
Other blood glucose lowering drugs, n (%)	53 (10.9)	33 (10.9)	20 (10.9)	0.00	31 (10.3)	20 (11.1)	0.02
Diuretics, n (%)	40 (8.2)	14 (4.6)	26 (14.2)	0.33	22 (7.5)	18 (9.8)	0.08
Beta-blockers, n (%)	66 (13.6)	32 (10.6)	34 (18.6)	0.23	36 (12.0)	31 (17.2)	0.15
Calcium channel blockers, n (%)	140 (28.9)	82 (27.2)	58 (31.7)	0.10	88 (29.2)	53 (29.3)	0.00
Renin-angiotensin system agents, n (%)	139 (28.7)	74 (24.5)	65 (35.5)	0.24	80 (26.7)	61 (33.6)	0.15
Lipid-modifying agents, n (%)	190 (39.2)	108 (35.8)	82 (44.8)	0.19	113 (37.7)	77 (42.9)	0.11
Laboratory values
HbA1c, %, median (IQR)	8.5 (7.5, 10.5)	8.9 (7.6, 11.0)	8.3 (7.3, 9.7)	0.40	8.5 (7.5, 10.8)	8.5 (7.3, 10.0)	0.24
eGFR, mL/min/1.73 m2, median (IQR)	74.0 (59.9, 90.3)	75.1 (60.8, 91.4)	73.5 (58.0, 86.2)	0.20	73.9 (59.7, 90.1)	73.7 (58.5, 87.8)	0.11
Hb, g/dL, median (IQR)	14.2 (12.9, 15.2)	14.3 (12.8, 15.3)	14.1 (13.0, 15.0)	0.01	14.1 (12.8, 15.2)	14.0 (13.0, 14.9)	0.02
LDL, mg/dL, median (IQR)	116 (88, 139)	119 (91, 147)	110 (85, 133)	0.29	113 (88, 138)	112 (86, 133)	0.20

aSMD absolute standardized mean difference, CEM coarsened exact matching, DPP-4 dipeptidyl peptidase-4, GLP-1 glucagon-like peptide-1, eGFR estimated glomerular filtration rate, Hb hemoglobin, IQR interquartile range, LDL low-density lipoprotein cholesterol, SBP systolic blood pressure, SGLT2 sodium-glucose cotransporter-2

After matching, covariate balance improved in both the 1st-line and 2nd-line cohorts, with aSMD below 0.25 [20].

The proportion of non-missing baseline data for each variable is reported in Supplementary Table S7. In addition, diagnostics for multiple imputation are provided in Supplementary Fig. S4.

Treatment Selection

While 20 distinct datasets were created through MICE to address missing covariates and each generated different CART models, the CARTs presented below were constructed on the basis of the aggregation of all imputed datasets similar to the previous study [18]. This can be interpreted as an average of the 20 CART models. Distribution of the estimated propensity scores 1st-line and 2nd-line therapy are shown in Supplementary Figs. S2 and S3.

Figure 2 displays the CART model illustrating prescription drug selection patterns. The most influential variables for classification, in order of relative importance, were diuretics (65.7%), followed by lipid-modifying agents (15.6%) and HbA1c (13.1%).

Fig. 2.

Decision tree for diabetes treatment selection and estimated probability of SGLT2 inhibitor prescription 1st-line treatment. eGFR estimated glomerular filtration rate, SGLT2 sodium-glucose cotransporter-2

For 2nd-line users (Fig. 3), the influential variables for classification were HbA1c (35.9%) and BMI (26.5%), followed by diuretics (17.9%).

Fig. 3.

Decision tree for diabetes treatment selection and estimated probability of SGLT2 inhibitor prescription (2nd-line treatment). eGFR estimated glomerular filtration rate, SGLT2 sodium-glucose cotransporter-2

Change in HbA1c

A total of 1378 patients were included in the change in HbA1c analysis (exclusions due to missing follow-up HbA1c data). Among 1st-line treatment recipients, 592 (68.8%) biguanide users and 405 (69.9%) SGLT2 inhibitor users had follow-up HbA1c data. Among 2nd-line treatment recipients, 237 (78.5%) biguanide users and 144 (78.7%) SGLT2 inhibitor users had follow-up HbA1c data. The unadjusted longitudinal trend in HbA1c changes can be seen in Figs. 4 and 5.

Fig. 4.

Change in HbA1c from the initiation of the index drug for 1st-line therapy (n = 997). SGLT2 sodium-glucose cotransporter-2

Fig. 5.

Change in HbA1c from the initiation of the index drug for 2nd-line therapy (n = 381). SGLT2 sodium-glucose cotransporter-2

For 1st-line therapy (Table 3), weighted MMRM analysis showed HbA1c reductions were −1.37 points (95% CI −1.48 to −1.26) for biguanides and −1.16 points (95% CI −1.29 to −1.02) for SGLT2 inhibitors, with a between-group difference of −0.21 points (95% CI −0.38 to −0.04) favoring biguanides.

Table 3.

HbA1c reduction (percent point) at 24 weeks estimated with CEM-weighted MMRM (n = 1378)

	1st-line (n = 997)		2nd-line (n = 381)
Treatment group	Estimate	95% CI	Estimate	95% CI
Biguanide	−1.37	−1.48 to −1.26	−2.26	−2.47 to −2.05
SGLT2 inhibitor	−1.16	−1.29 to −1.02	−1.79	−2.04 to −1.54
Between-group difference	−0.21	−0.38 to −0.04	−0.47	−0.80 to −0.15

CI confidence interval, CEM coarsened exact matching, MMRM mixed model for repeated measures, SGLT2 sodium-glucose cotransporter-2

For 2nd-line therapy (Table 3), weighted MMRM analysis showed HbA1c reductions of −2.26 points (95% CI −2.47 to −2.05) for biguanides and −1.79 point (95% CI −2.04 to −1.54) for SGLT2 inhibitors, with a between-group difference of −0.47 points (95% CI −0.80 to −0.15) favoring biguanides.

Sensitivity analyses using PSM and other covariance structures are reported in Supplementary Table S8.

Treatment Persistence

For 1st-line therapy, adjusted treatment persistence proportions at 48 weeks (Fig. 6) were 47.0% (95% CI 42.7–51.2) in biguanides and 56.2% (95% CI 51.2–61.2) in SGLT2 inhibitors. For 2nd-line therapy, adjusted treatment persistence proportions at 48 weeks were (Fig. 7) 51.8% (95% CI 45.2–58.4) in biguanides and 49.4% (95% CI 40.3–58.6) in SGLT2 inhibitors.

Fig. 6.

Treatment persistence estimated with CEM-weighted Kaplan–Meier method for 1st-line. CEM coarsened exact matching, SGLT2 sodium-glucose cotransporter-2

Fig. 7.

Treatment persistence estimated with CEM-weighted Kaplan–Meier method for 2nd-line. *Masked for anonymization. CEM coarsened exact matching, SGLT2 sodium-glucose cotransporter-2

As a sensitivity analysis, treatment persistence estimated with PSM-weighted Kaplan–Meier methods is shown in Supplementary Figs.  S5 and S6. Censoring status and reasons are summarized in Supplementary Table S9, and subsequent antidiabetic medications are seen in Supplementary Table S10.

Discussion

Principal Findings

This real-world study explored trends in treatment selection. Consistent with current treatment algorithms, patients at higher cardiovascular risk were more likely to receive SGLT2 inhibitors, as reflected in both the baseline characteristics and the CART analysis. In contrast, overall BMI distributions were broadly similar between treatment groups, although BMI contributed to treatment selection, particularly in the 1st-line setting. Taken together, these findings suggest that BMI plays a role in real-world treatment selection, but its relative importance may vary by treatment line and may be less dominant than would be expected under a strictly BMI-driven algorithm.

We also compared the effectiveness of biguanides and SGLT2 inhibitors in a Japanese clinical setting. The results suggested that both drug classes achieved clinically meaningful glycemic control in both 1st-line and 2nd-line therapy settings. In an intention-to-treat analysis, biguanides showed a more pronounced reduction in HbA1c compared with SGLT2 inhibitors, particularly in the 2nd-line cohort. In contrast, long-term treatment persistence over 48 weeks was higher with SGLT2 inhibitors in the 1st-line cohort, whereas persistence was comparable between groups in the 2nd-line cohort, as defined by prescription records in this database.

Interpretation and Context

This study explored shifts in medication selection. In Japan, DPP-4 inhibitors have historically been the most commonly prescribed class, but biguanides and SGLT2 inhibitors have been increasingly used in certain patient populations. In addition, the uptake of glucagon-like peptide-1 receptor agonists (GLP-1RAs), which target the incretin system and are generally used as alternatives to DPP-4 inhibitors, may have contributed to the declining use of DPP-4 inhibitors in Japan [21].

Metformin is often used with caution in older adults aged $\ge$75 years because of heterogeneous renal function and comorbidity profiles [11]. The inclusion of these patients in our analytic cohort should be taken into account when assessing the generalizability of our findings to middle-aged patient populations.

Time from diabetes diagnosis to prescription was defined as the number of days between the earliest diabetes diagnosis record observed in the database and the index prescription date. Because the earliest recorded diagnosis may occur after referral or transfer from a prior clinic, this measure may not reflect the patient’s true disease duration and should be interpreted with caution. Nevertheless, examining the association between time from the first recorded diagnosis at the participating institution and subsequent treatment selection may still provide meaningful insights into real-world prescribing patterns.

In real-world clinical databases, hypertension defined by diagnosis codes may be under-recorded [22, 23]. In our cohort, the prevalence of coded hypertension was lower than the prescription rates of antihypertensive drug classes, suggesting under-coding in structured ICD-10 fields. Therefore, we included both diagnosis-based comorbidities and concomitant cardiovascular medications as candidate covariates. In the CART models, medication variables contributed to treatment classification, supporting the use of prescriptions as proxies for cardiovascular risk management in this setting.

Comparison with Prior Work

Regarding glycemic control efficacy, biguanides appear to yield greater HbA1c reductions, particularly with dose escalation [24, 25]. In our data, the estimated between-group difference was also in the direction of greater HbA1c reductions with biguanides; the difference was modest in the 1st-line setting but was numerically larger in the 2nd-line setting, including among patients already treated with other antidiabetic medications.

The 1-year continuation rates were slightly lower than those reported in previous studies [26, 27]. This discrepancy may be attributed to our database primarily containing information from large hospitals, where clinicians might more actively adjust medications based on patients’ glycemic control and clinical factors compared with smaller facilities. It is also important to note that definitions of treatment persistence vary across studies, potentially interfering direct comparisons of these results.

For this analysis, we utilized the Millennium Medical Record Database, which contains patient records from multiple Japanese medical institutions. Crucially, this database collects identifiable data, enabling linkage of information from multiple data sources through patient identification. This capability allows for more comprehensive clinical laboratory information compared with standard claims or health checkup data, thereby enabling more thorough analysis of clinical outcomes. Additionally, our analysis of 2nd-line therapy contexts is relatively uncommon in existing literature.

Limitations

Our study has several limitations.

First, regarding data availability in the database, we lacked structured data elements to accurately define lifestyle modification plans and glycemic control targets. Although blood glucose levels are crucial for diabetes management alongside HbA1c, we could not incorporate them into our analysis because laboratory test labels varied across institutions, making it difficult to reliably distinguish fasting plasma glucose from other glucose measurements. In addition, missing laboratory values in structured fields may reflect results documented in referral letters or other unstructured records rather than the absence of testing. Because such unstructured information was unavailable for analysis, uncertainty may remain regarding baseline clinical status. Finally, several comorbidities were defined using broad ICD-10 groupings; however, these categories may combine clinically heterogeneous conditions (e.g., malignant neoplasms), which could reduce clinical interpretability.

Second, several important confounders were unavailable or difficult to incorporate in our real-world data analysis (Supplementary Table S7). For example, qualitative dipstick protein results were not captured, and quantitative urinary albumin-to-creatinine ratio (UACR) measurements were available only for a limited subset of patients. Accordingly, proteinuria/albuminuria status was not included as a covariate in the treatment-selection model. Because SGLT2 inhibitors are more likely to be prescribed to patients with proteinuria/albuminuria [28], the inability to adjust for this factor could bias the comparative effectiveness estimates toward smaller observed HbA1c reductions in the SGLT2 inhibitor group. In addition, some variables were missing partly because we used a shorter baseline period of 3 months than is typical in conventional studies. This shorter window was chosen to maximize sample size while maintaining data quality and completeness.

Third, in constructing the CART for implementing coarsened exact matching, we examined several scenarios regarding the impact of changing hyperparameters such as tree depth and the number of cutoff value candidates; however, we did not conduct a systematic search for optimal hyperparameters, and therefore, the procedure for determining these parameters could be improved. In real-world practice, early insulin therapy may be selected for marked hyperglycemia [28]. Insulin initiation remained slightly more common in the biguanide group even after matching, which may have contributed to the larger early HbA1c reduction observed in the biguanide group at 12 weeks. We implemented PSM as a sensitivity analysis for matching procedure

Fourth, the generalizability of our findings is limited by both the characteristics of the data source and the cohort construction. Because the database predominantly captures care delivered in large hospitals, the findings may not directly generalize to primary care clinics. In addition, a substantial proportion of initiators were excluded due to eligibility criteria and data-availability requirements (Supplementary Table S4). These exclusion rules and observability requirements may have preferentially retained patients with more stable follow-up and more complete recorded data, which could have influenced both HbA1c trajectories and persistence estimates. Therefore, caution is warranted when extrapolating these results to all initiators in routine practice. Nevertheless, because one objective of this study was to describe contemporary treatment selection patterns in real-world practice, our results may still be informative for nonspecialist physicians managing type 2 diabetes.

Future Directions

The Millennium Medical Record Database contains unstructured text data that may capture clinically rich information not available in structured fields. Leveraging these data in future research could enable more detailed analyses and more nuanced interpretation of prescribing patterns, including factors such as diet and exercise counseling, individualized HbA1c targets, and treatment history at prior clinics.

In addition, access to prescribing department or specialty information could provide more clinically actionable insights into real-world treatment selection. These kinds of subgroup analyses may help clarify whether the observed selection factors reflect differences in treatment trajectories, or therapeutic priorities across disciplines.

Conclusions

In Japanese patients with type 2 diabetes, treatment selection appeared to be influenced primarily by concomitant cardiovascular medication use. BMI was also associated with treatment choice, especially in 2nd-line therapy, but did not predominate overall. HbA1c decreased more with biguanides. Biguanides were associated with larger HbA1c reductions. Treatment persistence over 48 weeks differed by treatment line in intention-to-treat analyses: it was higher with SGLT2 inhibitors in 1st-line therapy but comparable between groups in 2nd-line therapy.

Supplementary Information

Below is the link to the electronic supplementary material.

Author Contributions

Conceptualization: RM (lead), DK (supporting). Data curation: RM. Formal analysis: RM. Funding acquisition: DK. Investigation: RM (lead), DK (supporting). Methodology: RM (lead), DK (supporting). Project administration: RM (lead), DK (supporting). Resources: DK. Supervision: DK. Writing—original draft: RM. Writing—review & editing: RM (lead), DK (supporting). All authors read and approved the final version.

Funding

This research was supported by Japan Agency for Medical Research and Development (AMED) under Grant Number: JP266k0237001. The funder had no involvement in the study design, data collection, analysis, interpretation, or the writing of the manuscript.

Data Availability

Data used in this study are available from the Millennium Medical Record Database through a commercial agreement.

Code Availability

The statistical analysis code used in this study is available from the corresponding author upon reasonable request.

Declarations

Conflict of Interest

None declared.

Ethics approval

This study received approval from the Millennium Medical Record Database Review Committee (2024_MIL_0002) and was registered on the University of Tokyo Clinical Research Ethics Committee (2024296NIe).

Consent to Participate

Not applicable.

Consent for Publication

Not applicable.