Real-world renoprotective effects of polyethylene glycol-loxenatide in type 2 diabetes: a multicenter, ambispective cohort study

Abstract

Patients with type 2 diabetes mellitus and chronic kidney disease face persistent risks despite standard care. Polyethylene glycol-loxenatide lacks real-world renoprotective evidence in high-risk Chinese populations relative to non-incretin-based standard care. This multicenter, ambispective cohort study enrolled 3305 patients with type 2 diabetes mellitus, renal impairment, and elevated cardiovascular risk at six Chinese tertiary institutions (median follow-up 36.4 months; trial registration: ChiCTR2300073001). Data were extracted via an artificial intelligence-enhanced pipeline. Co-primary analyzes comprised multivariable-adjusted Cox regression and 1:1 propensity score matching (332 per group); pre-specified landmark analyzes at 6 and 12 months addressed immortal time bias. Polyethylene glycol-loxenatide was associated with a reduced risk of the primary kidney composite outcome in the original cohort (hazard ratio [HR], 0.49; 95% CI, 0.25–0.95; P = 0.035), which was corroborated in the propensity score matched cohort (HR, 0.44; 95% CI, 0.20–0.97; P = 0.043) and progressive across 6-month (HR, 0.40; 95% CI, 0.19–0.84) and 12-month (HR, 0.33; 95% CI, 0.14–0.75) landmarks. While a reduction in 3-point major adverse cardiovascular events was observed in the unmatched cohort (HR, 0.60; 95% CI, 0.43–0.85; P = 0.004), this association was not sustained in the matched analysis (HR, 0.92; 95% CI, 0.57–1.48; P = 0.735). In conclusion, long-term therapy was associated with an attenuated risk of major renal events. Given the null findings in the matched cohort, the cardiovascular observations are exploratory and require randomized confirmation.

Introduction

The global escalation of type 2 diabetes mellitus (T2DM) has precipitated a parallel crisis in chronic kidney disease (CKD) and atherosclerotic cardiovascular disease, forming a complex cardiorenal-metabolic (CKM) syndrome that remains the primary driver of premature mortality and escalating healthcare expenditure.¹ In China, this epidemiological burden is particularly acute, where the high prevalence of T2DM is compounded by a high incidence of hypertension and late-stage renal presentation.^2,3 Despite the rigorous implementation of standard-of-care therapies, including intensive glycemic control and maximum-tolerated doses of renin-angiotensin system (RAS) inhibitors, a substantial residual risk for kidney failure and major adverse cardiovascular events (MACE) persists among high-risk phenotypes,^4,5 thereby underscoring the need for therapeutic strategies that extend beyond conventional risk factor management.

The clinical landscape of CKM management has been substantially altered by the emergence of glucagon-like peptide-1 receptor agonists (GLP-1 RAs).⁶ Conclusive evidence from landmark randomized controlled trials (RCTs), such as the LEADER and REWIND trials, and most recently the dedicated renal outcome trial FLOW, has established the cardiorenal-protective efficacy of this class.^7–10 These trials shifted the treatment paradigm from glucose-lowering alone to comprehensive organ protection, establishing GLP-1 RAs as a cornerstone in managing patients with overlapping metabolic and renal vulnerabilities. Consequently, evaluating the long-term, real-world effectiveness of these agents has become an important focus for contemporary cardiovascular and renal research.

However, an important translational gap remains because these landmark trial results primarily involve non-PEGylated analogs evaluated within the protocol-driven settings of RCTs. These trials often exclude patients with advanced comorbidities, elderly frailties, or the complex polypharmacy typical of routine clinical practice. Polyethylene glycol loxenatide (PEG-Loxe), a domestically developed once-weekly GLP-1 RA, features a third-generation 40 kDa PEGylated exenatide backbone.^11,12 This structural modification optimizes its pharmacokinetic profile to achieve an extended half-life of approximately 140 h, ensuring sustained, stable drug exposure that may confer distinct biological advantages for long-term organ protection.^12,13 Despite its clinical deployment, a critical evidence gap remains regarding the real-world effectiveness of PEG-Loxe, particularly in high-risk Chinese populations where the intersection of cardiovascular vulnerability and renal impairment is pronounced. Real-world evidence (RWE) is essential to bridge the translational gap between the idealized outcomes of RCTs and the multimorbid realities of clinical practice.⁶

To address this deficit, we conducted the FLYING 2 study, a multicenter, ambispective cohort study designed to evaluate the long-term impact of PEG-Loxe on renal and cardiovascular outcomes over a 60-month horizon.¹² The study employed an AI-enhanced screening framework that uses large language models (LLMs) to process and extract unstructured electronic health records (EHRs) across six tertiary medical centers.¹⁴ This study provides the first long-term, multi-year RWE on PEG-Loxe in a high-risk Chinese cohort, bridging the gap between controlled trial findings and multimorbid clinical realities. These findings may help inform clinical guidelines by offering high-fidelity evidence on the capacity of PEG-Loxe to attenuate the risks of progressive renal failure and MACE in patients managed outside the confines of randomized trials.

Results

Patient disposition and baseline characteristics

A total of 12,342 EHRs were initially processed through the AI-enhanced screening framework. Following the rigorous application of pre-specified inclusion and exclusion criteria, 3305 patients with T2DM and high cardiorenal risk were ultimately enrolled in the analytic cohort (Fig. 1). This cohort comprised 338 patients in the Polyethylene Glycol-Loxenatide (PEG-Loxe) group and 2967 in the non-incretin-based control group.

Fig. 1.

Flowchart of participant selection and study design

The baseline clinical profiles of the original cohort reflected distinct phenotypic disparities characteristic of real-world practice patterns. Participants in the control group were substantially older (mean 68.4 vs. 57.1 years) and exhibited a more advanced disease burden, including a higher prevalence of established cardiovascular disease (41.5% vs. 24.9%) and more frequent insulin utilization (54.5% vs. 33.1%; all SMDs > 0.1; Table 1).

Table 1.

Baseline characteristics of the study population before and after 1:1 propensity score matching

Characteristic	Original overall (N = 3305)	Original control (N = 2967)	Original PEG-Loxe (N = 338)	Original SMD	P-value	Matched overall (N = 664)	Matched control (N = 332)	Matched PEG-Loxe (N = 332)	Matched SMD	P-value
Demographics
Age, years	67.3 (10.5)	68.4 (9.7)	57.1 (11.6)	1.151	<0.001	57.3 (11.6)	57.4 (11.6)	57.2 (11.7)	0.016	0.838
Female, n (%)	1719 (52.0%)	1519 (51.2%)	200 (59.2%)	0.161	0.006	386 (58.1%)	190 (57.2%)	196 (59.0%)	0.037	0.694
Duration of T2DM, years	10.4 (6.2)	10.6 (6.3)	9.1 (5.1)	0.239	<0.001	9.2 (5.3)	9.2 (5.4)	9.2 (5.1)	0.009	0.906
Clinical measures
HbA1c, %	8.5 (2.9)	8.5 (3.0)	8.4 (1.1)	0.06	0.037	8.3 (2.2)	8.3 (3.0)	8.4 (1.1)	−0.004	0.962
BMI, kg/m²	24.8 (3.5)	24.6 (3.6)	26.6 (1.5)	−0.606	<0.001	26.7 (3.0)	26.7 (3.9)	26.6 (1.5)	0.048	0.539
SBP, mmHg	138.6 (16.5)	138.7 (16.5)	137.9 (16.6)	0.047	0.417	138.4 (16.5)	138.9 (16.4)	138.0 (16.6)	0.056	0.47
DBP, mmHg	81.0 (11.8)	81.0 (11.8)	81.0 (11.9)	−0.002	>0.9	81.1 (11.9)	81.3 (11.9)	80.9 (11.9)	0.032	0.68
Lipid profile (mmol/L)
Total cholesterol	4.8 [3.9, 5.8]	4.8 [3.8, 5.8]	5.0 [4.3, 5.6]	−0.061	0.006	4.8 [4.1, 5.6]	4.6 [3.7, 5.6]	5.0 [4.3, 5.6]	−0.165	<0.001
Triglycerides	1.6 [1.1, 2.4]	1.6 [1.1, 2.3]	1.9 [1.3, 2.6]	−0.035	<0.001	1.7 [1.2, 2.5]	1.5 [1.1, 2.2]	1.9 [1.3, 2.7]	−0.074	<0.001
HDL-C	1.1 (0.4)	1.1 (0.4)	1.2 (0.3)	−0.241	<0.001	1.2 (0.4)	1.1 (0.4)	1.2 (0.3)	−0.151	0.082
LDL-C	3.1 (1.1)	3.1 (1.2)	3.0 (0.9)	0.156	<0.001	3.0 (1.0)	3.0 (1.2)	3.0 (0.8)	0.042	0.63
Renal function
eGFR — ml/min/1.73 m²	56.4 (13.0)	56.1 (13.1)	59.3 (11.5)	−0.249	<0.001	59.3 (12.1)	59.3 (12.6)	59.2 (11.5)	0.006	0.941
eGFR category				0.271	<0.001				0.083	0.77
≥60	1553 (47.0%)	1366 (46.0%)	187 (55.3%)			366 (55.1%)	183 (55.1%)	183 (55.1%)
45 to <60	1051 (31.8%)	944 (31.8%)	107 (31.7%)			205 (30.9%)	99 (29.8%)	106 (31.9%)
30 to <45	567 (17.2%)	528 (17.8%)	39 (11.5%)			80 (12.0%)	42 (12.7%)	38 (11.4%)
<30	134 (4.1%)	129 (4.3%)	5 (1.5%)			13 (2.0%)	8 (2.4%)	5 (1.5%)
Comorbidities, n (%)
CVD history	1315 (39.8%)	1231 (41.5%)	84 (24.9%)	0.359	<0.001	191 (28.8%)	107 (32.2%)	84 (25.3%)	0.153	0.059
Hypertension history	2729 (82.6%)	2456 (82.8%)	273 (80.8%)	0.052	0.397	546 (82.2%)	279 (84.0%)	267 (80.4%)	0.095	0.264
Diabetic retinopathy	237 (7.2%)	186 (6.3%)	51 (15.1%)	−0.289	<0.001	66 (9.9%)	15 (4.5%)	51 (15.4%)	−0.369	<0.001
Current smoking	342 (10.3%)	289 (9.7%)	53 (15.7%)	−0.179	<0.001	114 (17.2%)	61 (18.4%)	53 (16.0%)	0.064	0.471
Medications, n (%)
Metformin	1375 (41.6%)	1215 (41.0%)	160 (47.3%)	−0.129	0.028	279 (42.0%)	121 (36.4%)	158 (47.6%)	−0.227	0.005
SU / GLN	381 (11.5%)	364 (12.3%)	17 (5.0%)	0.26	<0.001	49 (7.4%)	33 (9.9%)	16 (4.8%)	0.197	0.018
AGI	1080 (32.7%)	969 (32.7%)	111 (32.8%)	−0.004	>0.9	210 (31.6%)	101 (30.4%)	109 (32.8%)	−0.052	0.559
TZD	243 (7.4%)	215 (7.2%)	28 (8.3%)	−0.039	0.56	50 (7.5%)	23 (6.9%)	27 (8.1%)	−0.046	0.659
SGLT2i use	544 (16.5%)	493 (16.6%)	51 (15.1%)	0.042	0.522	105 (15.8%)	55 (16.6%)	50 (15.1%)	0.041	0.671
Insulin	1729 (52.3%)	1617 (54.5%)	112 (33.1%)	0.441	<0.001	295 (44.4%)	184 (55.4%)	111 (33.4%)	0.454	<0.001
ACEI / ARB	2027 (61.3%)	1820 (61.3%)	207 (61.2%)	0.002	1	414 (62.3%)	212 (63.9%)	202 (60.8%)	0.062	0.471
Beta-blocker	1210 (36.6%)	1154 (38.9%)	56 (16.6%)	0.515	<0.001	185 (27.9%)	129 (38.9%)	56 (16.9%)	0.506	<0.001
Statin	2087 (63.1%)	1855 (62.5%)	232 (68.6%)	−0.129	0.032	425 (64.0%)	197 (59.3%)	228 (68.7%)	−0.195	0.015
Fibrate	241 (7.3%)	214 (7.2%)	27 (8.0%)	−0.029	0.682	44 (6.6%)	17 (5.1%)	27 (8.1%)	−0.121	0.16
Antiplatelet	1612 (48.8%)	1516 (51.1%)	96 (28.4%)	0.477	<0.001	268 (40.4%)	172 (51.8%)	96 (28.9%)	0.48	<0.001

Data are presented as Mean (SD) for normally distributed continuous variables, Median [IQR] for non-normally distributed continuous variables (TC, TG), and n (%) for categorical variables. Normality was assessed using skewness (|skew| < 1.5) and excess kurtosis (|excess kurtosis| < 3). SMD: Standardized Mean Difference calculated via pooled Cohen’s d for continuous variables, proportion difference for binary variables, and Mahalanobis distance for multi-level categorical variables (eGFR category). |SMD| < 0.1 indicates negligible imbalance; 0.1–0.2 small; > 0.2 notable

PEG-Loxe Polyethylene Glycol-Loxenatide, T2DM type 2 diabetes mellitus, BMI body mass index, HbA1c glycated hemoglobin, SBP systolic blood pressure, DBP diastolic blood pressure, TC total cholesterol, TG triglycerides, HDL-C high-density lipoprotein cholesterol, LDL-C low-density lipoprotein cholesterol, eGFR estimated glomerular filtration rate, CVD cardiovascular disease, SU/GLN sulfonylureas/glinides, AGI alpha-glucosidase inhibitors, TZD thiazolidinediones; SGLT2i sodium-glucose cotransporter-2 inhibitors, ACEI/ARB angiotensin-converting enzyme inhibitors/angiotensin receptor blockers, SMD standardized mean difference

To mitigate selection bias and ensure clinical comparability, propensity score matching (PSM) was performed at a 1:1 ratio. Notably, 332 out of 338 (98.2%) patients in the PEG-Loxe group were successfully matched to suitable controls. This exceptionally high retention rate, alongside the extensive overlap in propensity score distributions, confirms that the matched participants occupy a robust common support region, effectively addressing the potential “support problem” arising from the initial age disparity. In the resulting matched cohort (N = 664), primary continuous demographic and metabolic covariates achieved an optimal balance with SMDs below 0.10. However, residual imbalances persisted for specific background therapies—namely baseline insulin use, antiplatelet therapy, and beta-blocker use (all absolute SMDs > 0.40)—alongside a moderate imbalance in categorical cardiovascular disease history (SMD = 0.153) (Table 1). Overly restrictive caliper trimming to force balance on all background variables would have resulted in severe sample attrition and loss of external validity; therefore, the primary caliper was maintained, and these residual medication imbalances are retained as recognized limitations inherent to the observational design. The median follow-up duration for the entire population was 36.4 months (IQR, 24.1–48.6), as estimated by the reverse Kaplan-Meier method.

Primary renal composite endpoint

The primary kidney composite endpoint occurred in 11 of 338 patients (3.3%) in the PEG-Loxe group and 105 of 2967 patients (3.5%) in the original control group. In the original cohort, although the unadjusted analysis showed a favorable but non-significant trend (HR, 0.60; 95% CI, 0.32–1.13; log-rank P = 0.113), the association became statistically significant upon multivariable adjustment (Table 2). After accounting for baseline cardiorenal risk factors and stratifying by SGLT2i use to satisfy the proportional hazards assumption (Model 3), PEG-Loxe was associated with a 51.0% lower risk of the primary kidney composite outcome (stratified HR, 0.49; 95% CI, 0.2–0.95; P = 0.035). To further validate these findings, a PSM-matched cohort analysis (n = 332 per group) was conducted.¹⁵ The protective trend remained consistent in the matched cohort, with the fully adjusted model (Model 3) confirming a statistically significant association within this matched framework, yielding an effect size consistent with the primary analysis (HR, 0.44; 95% CI, 0.20–0.97; P = 0.043). Kaplan–Meier incidence curves displayed an early and widening separation between the treatment arms in both the original (Fig. 2a) and matched (Fig. 2b) cohorts. Given the differential follow-up durations, the adjusted RMST analysis corroborated these results, demonstrating a significant gain of 1.43 months (95% CI, 0.48–3.00; P = 0.003) in event-free survival over 48 months for patients receiving PEG-Loxe.

Table 2.

Association of polyethylene glycol-loxenatide with renal and cardiovascular outcomes in the original and 1:1 PSM-matched cohorts

Outcome & model	Original cohort (N = 3305)	P-value	PSM matched cohort (N = 664)	P-value
	HR (95% CI)		HR (95% CI)
Primary renal composite endpoint
Model 1	0.51 (0.27–0.96)	0.038	0.45 (0.21–0.94)	0.032
Model 2	0.51 (0.26–0.97)	0.040	0.47 (0.22–1.01)	0.053
Model 3	0.49 (0.25–0.95)	0.035	0.44 (0.20–0.97)	0.043
RMST Difference (months)	1.43 (0.48–3.00)	0.003	—	—
Secondary outcome: 3P-MACE
Model 1	0.62 (0.44–0.87)*	0.006	0.88 (0.56–1.40)	0.600
Model 2	0.62 (0.44–0.87)*	0.006	0.94 (0.58–1.51)	0.785
Model 3	0.60 (0.43–0.85)*	0.004	0.92 (0.57–1.48)	0.735
RMST difference	1.49 (0.74–2.01)	0.044	—	—

Model 1: adjusted for age and sex. Model 2: additionally adjusted for diabetes duration and baseline HbA1c. Model 3: further incorporated history of cardiovascular disease, baseline eGFR, BMI, baseline SGLT2i use, and smoking status. To address violations of the proportional hazards (PH) assumption, Model 3 in the original cohort was stratified by baseline SGLT2i use for the renal endpoint, and by both sex and SGLT2i use for the 3P-MACE endpoint. RMST differences represent the gain in event-free time (PEG-Loxe minus Control) over a 48-month horizon. Values > 0 favor the PEG-Loxe group. RMST differences were not computed for the PSM-matched cohort due to the reduced sample size and the corresponding imprecision of RMST estimates in smaller matched samples. *Indicates Stratified Hazard Ratios implemented to ensure validity of the proportional hazards assumption where global PH tests were violated

Fig. 2.

Cumulative incidence of primary kidney composite outcomes and 3P-MACE. a Cumulative incidence of the primary kidney composite endpoint in the original cohort (N = 3305). (b) Cumulative incidence of the primary kidney composite endpoint in the PSM 1:1 matched cohort (N = 664). c Cumulative incidence of 3P-MACE in the original cohort. d Cumulative incidence of 3P-MACE in the PSM 1:1 matched cohort. Note: The primary kidney composite endpoint included a recorded ≥50% reduction in eGFR from baseline, progression to end-stage renal disease (eGFR <15 mL/min/1.73 m² or initiation of renal replacement therapy), or death from kidney-related or cardiovascular causes. 3P-MACE included nonfatal myocardial infarction, nonfatal stroke, or cardiovascular death. P-values were derived from the log-rank test. Tick marks indicate censored observations

Secondary cardiovascular and survival outcomes

In the original cohort, 3P-MACE occurred in 42 (12.4%) patients in the PEG-Loxe group and 308 (10.4%) patients in the control group. Although the crude incidence rate appeared higher in the treatment arm, multivariable-adjusted analysis revealed a significant inverse association between PEG-Loxe treatment and 3P-MACE risk (Table 2). After adjusting for the full set of Model 3 covariates and stratifying by baseline sex and SGLT2i use to accommodate proportional hazards violations, PEG-Loxe was associated with a 40.0% reduction in the risk of 3P-MACE (stratified HR, 0.60; 95% CI, 0.43–0.85; P = 0.004). This translated to an estimated absolute risk reduction of 0.66 per 100 person-years within the study population.

The observed cardiovascular benefit was primarily driven by lower risks of nonfatal stroke (HR, 0.37; 95% CI, 0.19–0.73; P = 0.005) and nonfatal myocardial infarction (HR, 0.18; 95% CI, 0.08–0.42; P < 0.001). Consistent with the findings for renal outcomes, the association with cardiovascular death did not reach statistical significance (HR, 0.42; 95% CI, 0.08–2.09; P = 0.286), likely due to the limited number of recorded events.

In the PSM-matched cohort, the point estimate for 3P-MACE moved closer to the null and did not reach statistical significance (HR, 0.92; 95% CI, 0.57–1.48; P = 0.735). This attenuation in the matched framework likely reflects the substantial reduction in sample size (N = 664) and the corresponding loss of statistical power to detect cardiovascular differences, as well as the successful mitigation of age-related confounding that was prominent in the original cohort. The cumulative incidence of 3P-MACE for the original and matched cohorts is illustrated in Fig. 2c and Fig. 2d, respectively. To complement the Cox modeling, the adjusted RMST analysis in the original cohort confirmed a significant gain of 1.49 months (95% CI, 0.74–2.01; P = 0.044) of 3P-MACE-free survival over the 48-month horizon.

The individual ischemic components—specifically nonfatal stroke, nonfatal myocardial infarction, and coronary revascularization—demonstrated robust and statistically significant risk reductions across all analytical frameworks (all P < 0.05; Supplementary Table 1). Other clinical events, including hospitalization for heart failure and all-cause mortality, consistently favored the PEG-Loxenatide group, though these findings are considered exploratory due to limited event counts in the matched cohorts. Detailed results across multiple propensity-matched ratios and a description of the adaptive statistical modeling strategy are provided in the Supplementary Table 1.

Sensitivity analysis

To rigorously evaluate the internal validity and robustness of our findings, we conducted a comprehensive suite of pre-specified sensitivity analyzes (Table 3).

Table 3.

Sensitivity analyses for the association of PEG-loxenatide with the primary kidney composite outcome

Scenario / Analysis	Original cohort (N = 3305)	P-value	PSM matched cohort* (N = 664)	P-value
	HR (95% CI)		HR (95% CI)
Primary Analysis (Reference)	0.49 (0.25–0.95)	0.035	0.44 (0.20–0.97)	0.043
Alternative matching ratios
PSM 1:2 ratio (Model 3)	0.58 (0.27–1.24)	0.161	—	—
PSM 1:4 ratio (Model 3)	0.56 (0.27–1.16)	0.118	—	—
Addressing immortal time bias
Landmark analysis at 6 months	0.40 (0.19–0.84)	0.015	0.37 (0.16–0.85)	0.019
Landmark analysis at 12 months	0.33 (0.14–0.75)	0.008	0.33 (0.13–0.84)	0.020
Causal inference comparisons
IPTW (Full Cohort)	0.63 (0.35–1.13)	0.120	—	—
Overlap Weighting (Full Cohort)	0.66 (0.38–1.14)	0.140	—	—
Stability & model specification
Parsimonious model	0.50 (0.26–0.96)	0.036	0.41 (0.19–0.89)	0.023
Evaluating follow-up durations
Fixed follow-up at 24 months	0.69 (0.28–1.71)	0.425	0.68 (0.22–2.04)	0.486
Fixed follow-up at 36 months	0.59 (0.28–1.24)	0.165	0.456 (0.18–1.12)	0.086
Quantifying residual confounding
E-value for point estimate	3.52	—	3.97	—
E-value for confidence interval	1.29	—	1.19	—

*The PSM-matched cohort was derived through 1:1 propensity score matching. Primary Analysis used Model 3 (adjusted for age, sex, duration, HbA1c, CVD, eGFR, BMI, and smoking), stratified by baseline SGLT2i use in the original cohort. Causal Inference Comparisons represent alternative weighting methods applied to the full cohort. These estimates focus on different target populations and statistical assumptions regarding the common support region. Parsimonious model adjusted only for age, sex, and baseline eGFR to evaluate the potential impact of model complexity. E-value indicates the minimum association strength an unmeasured confounder must have with both treatment and outcome to fully nullify the observed association

HR hazard ratio, CI confidence interval, PSM propensity score matching, IPTW inverse probability of treatment weighting

To mitigate potential immortal time bias, we performed landmark analyzes at 6 and 12 months post enrollment. In the original cohort, the renoprotective association of PEG-Loxe remained significant at the 6-month landmark (HR, 0.40; 95% CI, 0.19–0.84; P = 0.015) and demonstrated an intensified protective trend at the 12-month landmark (HR, 0.33; 95% CI, 0.14–0.75; P = 0.008). Similar trends were observed in the PSM-matched cohort, suggesting that the clinical benefit is sustained and potentially accrues over time.

We evaluated the impact of varying PSM ratios on the stability of our estimates. Compared with the primary 1:1 matching, alternative matching ratios of 1:2 and 1:4 yielded directionally consistent point estimates, with fully adjusted hazard ratios of 0.58 (95% CI, 0.27–1.24; P = 0.161) and 0.56 (95% CI, 0.27–1.16; P = 0.118), respectively (Table 3). These estimates did not reach statistical significance, likely owing to the inclusion of more heterogeneous controls from the older, higher-risk pool. Similarly, whole-cohort weighting methods (IPTW and Overlap Weighting) were constrained by limited common support and yielded non-significant associations (HR 0.63, P = 0.120; HR 0.66, P = 0.140), though their point estimates consistently favored the PEG-Loxe group. The parsimonious model produced a result consistent with the primary analysis (HR 0.50; 95% CI, 0.26–0.96; P = 0.036), while analyzes restricted to fixed follow-up durations of 24 and 36 months yielded attenuated, non-significant point estimates, consistent with the expectation of reduced statistical power at shorter observation windows (Table 3).

The E-value for the point estimate of the primary renal outcome was 3.52 in the original cohort and 3.97 in the PSM-matched cohort. These values indicate that a strong unmeasured confounder—one associated with both PEG-Loxe exposure and renal outcomes by a risk ratio exceeding 3.5—would be required to nullify the point estimate. However, the E-values for the confidence interval limits were 1.29 in the original cohort and 1.19 in the matched cohort. These lower values indicate that a confounder associated with both exposure and outcome by a risk ratio of approximately 1.2—a magnitude plausibly within the range of unmeasured socioeconomic, adherence, or health-literacy factors—could shift the confidence interval to include the null. This asymmetry between the point-estimate and CI-limit E-values warrants interpretive caution when assessing the robustness of the renal association.

Subgroup analysis

To evaluate the consistency of the treatment effect across diverse clinical phenotypes, we performed prespecified subgroup analyzes on the primary kidney composite endpoint (Supplementary Fig. 1). In the original cohort, the renal-protective association of PEG-Loxenatide remained broadly consistent across all analyzed strata. Using an RCS-derived age threshold of 68 years to account for non-linear risk trajectories, we observed a consistent protective trend in the younger stratum (HR, 0.48; 95% CI, 0.23–0.99), similar to the overall population. Formal interaction testing confirmed no significant heterogeneity in the treatment effect based on age (P_interaction = 0.436), baseline HbA1c (P_interaction = 0.620), diabetes duration (P_interaction = 0.220), or SGLT2i co-administration (P_interaction = 0.722). Consistent findings were observed in the PSM 1:1 matched cohort, where interaction P-values for all subgroups remained non-significant (all P_interaction > 0.05; supplementary Fig. 1b). The risk attenuation was notably pronounced in patients aged <68 years (HR, 0.34; 95% CI, 0.15–0.81) and current smokers (HR, 0.11; 95% CI, 0.03–0.45). Although point estimates for certain sub-populations—such as those with a longer diabetes duration (≥10 years: HR, 0.32; 95% CI, 0.08–1.32)—mirrored the primary findings, these strata demonstrated wider confidence intervals due to limited event counts. Consequently, these subgroup results are considered exploratory and primarily intended for hypothesis generation. Detailed subgroup numbers are provided in the Supplementary Material.

Discussion

In this multicenter, ambispective cohort study of high-risk Chinese patients with T2DM and concomitant renal and cardiovascular vulnerability, long-term PEG-Loxe therapy was associated with a 51.0% reduction in the risk of the primary kidney composite outcome (HR, 0.49; 95% CI, 0.25–0.95; P = 0.035) and a 40.0% attenuation in 3P-MACE (HR, 0.60; 95% CI, 0.43–0.85; P = 0.004) compared with non-incretin-based standard care. Although PSM, multiple landmark analyzes, and alternative causal inference estimators confirmed that the robust renoprotective association persisted (HR, 0.44; 95% CI, 0.20–0.97), the cardiovascular benefit was attenuated and did not retain statistical significance within the matched framework (HR, 0.92; 95% CI, 0.57–1.48). These findings directly address the translational evidence gap regarding PEGylated glucagon-like peptide-1 receptor agonists (GLP-1 RAs) in routine clinical practice, offering essential real-world insights into a phenotypically complex, multimorbid Asian population that faces an acute burden of progressive renal impairment but is frequently underrepresented in protocol-driven RCTs.

The direction of the renal risk reduction observed in FLYING 2 is consistent with the results of landmark RCTs. The dedicated renal outcome trial FLOW demonstrated a 24% relative risk reduction in its primary renal composite endpoint for semaglutide (HR, 0.76; 95% CI, 0.66–0.88), while the LEADER trial reported attenuation of secondary renal endpoints with liraglutide (HR, 0.78; 95% CI, 0.67–0.92).^7,9 Direct numerical comparison of effect sizes between FLYING 2 and these trials is not appropriate, given differences in endpoint definitions, follow-up duration, and population characteristics. Nevertheless, the phenomenon of baseline risk heterogeneity is well-characterized in observational pharmacoepidemiology and may provide a general framework for contextualizing the magnitude of relative risk reductions observed in higher-risk real-world populations.^16,17 Unlike the highly selected populations in RCTs, participants in our control group reflected the multimorbid realities of tertiary care: they were substantially older (mean age 68.4 years), with higher rates of established cardiovascular disease (41.5%) and insulin dependence (54.5%). In individuals at elevated baseline risk, the therapeutic impact of a metabolic modulator acting through multiple synergistic pathways is intrinsically magnified.

Methodologically, this study addressed immortal time bias—a frequent concern in observational designs requiring minimum treatment durations—through pre-specified landmark analyzes at 6 and 12 months post index.^18,19 These analyzes did not attenuate the renoprotective signal; rather, they revealed a progressive intensification of risk reduction (renal HR, 0.40 at 6 months; HR, 0.33 at 12 months). This temporal trajectory is inconsistent with immortal time artifact and supports a biologically driven cumulative therapeutic benefit. The 98.2% PSM success rate—wherein 332 of 338 PEG-Loxe recipients were retained—confirms that the treated cohort predominantly occupies a region of common support within the propensity score distribution. This stands in contrast to settings in which substantial proportions of real-world patients are discarded due to common support failure, and reinforces that the observed associations are not an artifact of unmatchable extreme phenotypes.²⁰

At the pharmacokinetic level, the extended 140-h half-life of PEG-Loxe may confer sustained target engagement conducive to long-term organ protection. Preclinical evidence indicates that PEG-Loxe suppresses the GRP78/PERK/eIF2α signaling axis, mitigating hyperglycemia-induced endoplasmic reticulum stress and podocyte apoptosis, while concurrent inhibition of NF-κB nuclear translocation downregulates profibrotic cytokines such as TGF-β and TNF-α.^21–25 Two exploratory subgroup findings align with these mechanisms and offer directions for future investigation. The more pronounced risk attenuation in patients aged under 68 years (HR, 0.34; 95% CI, 0.15–0.81) suggests a therapeutic window before irreversible glomerulosclerosis occurs, while the point estimate in current smokers (HR, 0.11; 95% CI, 0.03–0.45), though derived from a stratum with limited events, is directionally consistent with preclinical evidence that GLP-1 RA-mediated antioxidant pathways may attenuate oxidative stress-mediated kidney injury.^13,26 Both signals are derived from strata with limited event counts and require prospective confirmation. Future investigations should prioritize biomarker-enriched designs incorporating serial eGFR slope measurements and albuminuria kinetics to disentangle the hemodynamic, anti-inflammatory, and antifibrotic contributions of sustained PEG-Loxe exposure.

Several limitations warrant interpretive caution. First, the non-randomized design precludes causal inference; despite PSM, IPTW, landmark analyzes, and E-value quantification, residual confounding from unmeasured variables—including socioeconomic status, educational attainment, and dynamic treatment adherence—cannot be excluded.²⁷ Second, the cohort is drawn exclusively from Chinese tertiary medical centers, constraining generalizability to other populations or healthcare settings. Third, the pronounced baseline age asymmetry, while substantially mitigated by the 98.2% matching success rate, reflects the real-world prescribing preference for PEG-Loxe toward younger patients and may not be fully resolved by statistical adjustment alone. Fourth, several secondary outcomes and subgroup strata had low event counts, producing wide confidence intervals that limit precision. Fifth, the AI-enhanced extraction pipeline, although externally validated with complete inter-rater agreement in a 5% random sample, carries residual misclassification risk for complex events ascertained from narrative clinical text.^14,28 Sixth, residual immortal time bias cannot be entirely dismissed, as time-dependent covariate modeling was not feasible within the EHR data architecture. Seventh, substantial residual imbalances in baseline therapies—insulin, antiplatelet, and beta-blocker use (all absolute SMD > 0.40)—and cardiovascular disease history (SMD, 0.153) persisted after matching. These discrepancies reflect real-world channeling bias, wherein physicians preferentially prescribe GLP-1 RAs to patients requiring intensive background regimens. Because these agents are established prognostic determinants, their uneven distribution represents a source of residual confounding that warrants careful consideration when interpreting the matched-cohort estimates. Finally, the absence of serial eGFR trajectories, albuminuria kinetics, and drug adherence metrics precludes assessment of dose-response relationships.

The E-value for the renal point estimate (3.52 in the original cohort; 3.97 in the PSM cohort) indicates that a strong unmeasured confounder would be required to nullify the point estimate.²⁹ However, the confidence interval limit E-values (1.29 and 1.19, respectively) warrant particular attention: a confounder associated with both PEG-Loxe prescription and renal outcomes by a risk ratio of approximately 1.2 could shift the confidence interval to include the null. In the Chinese healthcare context, candidate confounders of this magnitude include socioeconomic status, educational attainment, health literacy, and medication adherence—factors that are plausibly associated with both access to novel incretin-based therapies and renal outcomes through pathways of diet quality, follow-up compliance, and comorbidity management. These E-value considerations frame the FLYING 2 findings as hypothesis-generating RWE.

In conclusion, long-term PEG-Loxe therapy was associated with an attenuated risk of major renal events in high-risk patients with T2DM and established cardiorenal vulnerability. Given the null findings in the matched cohort, the cardiovascular observations remain exploratory and do not support definitive conclusions regarding MACE protection. These findings underscore the need for prospective, randomized trials to establish the cardiorenal efficacy of PEGylated GLP-1 receptor agonists in high-risk Asian populations.

Materials and methods

Study design

The FLYING 2 study is a multicenter, ambispective, observational cohort study conducted across six tertiary medical institutions in China (see Supplementary Materials and Methods). The study comprises a retrospective phase, during which baseline identification and historical outcome ascertainment were performed for participants enrolled between July 1, 2020, and June 30, 2023, and a prospective monitoring phase extending through June 30, 2025, yielding a maximum follow-up duration of 60 months.

Time zero (the index date) for each participant was defined as the date of the first documented prescription of either polyethylene glycol loxenatide (PEG-Loxe) or a non-incretin-based antidiabetic regimen within the enrollment window. All time-to-event analyzes were anchored to this index date, thereby ensuring a uniform origin for exposure classification and outcome surveillance. To mitigate potential immortal time bias arising from the requirement of at least 6 months of continuous therapy for exposure classification, pre-specified landmark analyzes at 6 and 12 months were incorporated into the analytical framework (Statistical Analysis section).^18,30

Prospective follow-up was conducted at 3-month intervals through scheduled clinical visits or structured telephone interviews. To ensure data integrity throughout the observation period, investigators maintained a master participant list with unique identification codes. All serious adverse events, including ketoacidosis and lactic acidosis, were systematically reported and adjudicated by site investigators. The study adheres to the principles of the Declaration of Helsinki³¹ and was approved by the Institutional Review Board of the Second Affiliated Hospital of Shantou University Medical College (No.: 2023-39). Written informed consent was obtained from all participants enrolled in the prospective phase; for the retrospective component, the requirement for individual consent was waived by the ethics committee given the observational nature of the study and the use of de-identified data. The trial is registered with the Chinese Clinical Trial Registry (ChiCTR 2300073001).

Settings and participants

Participants were identified from the EHR systems of the six participating institutions (Fig. 1). Eligible patients were adults aged 18 years or older with a confirmed diagnosis of type 2 diabetes mellitus (T2DM) who concurrently met criteria for both renal impairment and elevated cardiovascular risk (detailed definitions are provided in the Supplementary Materials and Methods).

Renal impairment was defined as a clinical diagnosis of diabetic kidney disease or other forms of CKD, corresponding to either an estimated glomerular filtration rate (eGFR) below 60 mL/min/1.73 m² or persistent albuminuria (urinary albumin-to-creatinine ratio [UACR] ≥ 30 mg/g). All enrolled participants were required to have a baseline eGFR ≥ 15 mL/min/1.73 m², as progression to eGFR below this threshold constituted a component of the primary renal composite endpoint.

AI-enhanced data extraction and validation

To efficiently derive structured, analysis-ready datasets from large volumes of unstructured clinical narratives, we employed an automated extraction pipeline built upon the GPT-4 large language model (OpenAI), accessed via application programming interface.¹⁴ A domain-specific prompt engineering strategy was developed to convert free-text admission records, diagnostic reports, discharge summaries, and prescription histories into standardized JSON-formatted outputs. Regular expressions were subsequently applied to validate structural integrity prior to ingestion into a relational database. This pipeline processed the entirety of available EHR data from the six institutions, yielding an initial extraction cohort of 12,342 patients who met preliminary screening criteria. Detailed methodologies regarding the validation study, extraction fidelity, and the handling of missing data are described in the Supplementary Materials and Methods.

Drug exposure, outcomes and follow up

To ensure that the analysis captured a clinically meaningful intensity of therapeutic exposure—rather than transient or exploratory prescriptions—a minimum of 6 months of continuous therapy was required for group assignment. Participants who received once-weekly subcutaneous polyethylene glycol loxenatide throughout this qualifying period were classified into the PEG-Loxe group, while those maintained on non-incretin-based antidiabetic regimens (including metformin, sulfonylureas, alpha-glucosidase inhibitors, thiazolidinediones, insulin, or combinations thereof) comprised the control group. Drug exposure duration was calculated as the interval between the first and last documented prescriptions within the observation window. We acknowledge that this enrollment strategy introduces a period during which treated participants are, by definition, event-free—a potential source of immortal time bias. To formally address this concern, pre-specified landmark analyzes at 6 and 12 months were conducted, in which the follow-up clock was reset and only participants remaining at risk beyond the landmark were included.

The primary endpoint was a composite renal outcome, defined as the time to the first occurrence of any of the following: (1) kidney failure, operationally defined as the initiation of chronic renal replacement therapy (hemodialysis or peritoneal dialysis), kidney transplantation, or a confirmed eGFR below 15 mL/min/1.73 m²; (2) a decline of at least 50% in eGFR from the baseline value at any point during follow-up; or (3) death attributed to renal or cardiovascular causes. The inclusion of cardiovascular mortality within this renal composite was pre-specified on two grounds. First, the endpoint design was aligned with contemporary GLP-1 receptor agonist renal outcome trials—most notably the FLOW trial—to facilitate cross-study contextual comparisons. Second, and more importantly, this approach addresses the well-established competing risk framework inCKD populations: cardiovascular events represent the predominant cause of death in patients with CKD, and the omission of cardiovascular mortality would systematically underestimate the true therapeutic benefit by censoring patients who die from cardiovascular causes prior to reaching the kidney failure threshold.^9,32

Secondary outcomes included: (1) three-point major adverse cardiovascular events (3 P-MACE), defined as the composite of non-fatal myocardial infarction, non-fatal ischemic stroke, or cardiovascular death; (2) coronary revascularization (percutaneous coronary intervention or coronary artery bypass grafting); (3) hospitalization for unstable angina; (4) hospitalization for heart failure; and (5) all-cause mortality. Specific diagnostic ICD-10 codes and outcome adjudication processes are detailed in the Supplementary Materials and Methods.

Follow-up commenced on the index date and continued until the occurrence of a primary or secondary endpoint event, death from any cause, loss to follow-up, or the administrative censoring date (June 30, 2025), whichever occurred first. The median follow-up time was estimated using the reverse Kaplan-Meier method to account for the differential censoring patterns between groups.

Statistical analysis

Statistical analyzes were performed using a landmark-indexed approach, in which participants were required to complete at least 6 months of continuous therapy to ensure a stable exposure window for observing renal outcomes. This design explicitly acknowledges the potential for immortal time bias during the initial exposure qualification period, which was formally addressed through pre-specified landmark analyzes at 6 and 12 months (detailed below). Two complementary analytical strategies were employed as co-primary approaches: (1) multivariable-adjusted Cox proportional hazards regression applied to the entire unmatched cohort (N = 3305), and (2) Cox regression applied to a 1:1 propensity score-matched cohort to address baseline imbalances and enhance causal inference.¹⁵ Continuous variables were expressed as mean ± standard deviation (SD) or median (interquartile range, IQR) depending on distributional characteristics, and compared using independent-samples t-tests or Mann–Whitney U tests. Categorical variables were summarized as frequencies and percentages, with between-group differences assessed using chi-square tests or Fisher’s exact tests where appropriate. Detailed definitions of all covariates utilized for multivariable adjustment and propensity score estimation are provided in the Supplementary Materials and Methods.

For the primary analysis in the unmatched cohort, we constructed three nested Cox models to evaluate the stability of the treatment effect estimate: Model 1 adjusted for age and sex; Model 2 additionally adjusted for diabetes duration and baseline HbA1c; and Model 3 (the fully adjusted model) further incorporated history of cardiovascular disease, baseline eGFR, BMI, baseline SGLT 2 i use, and smoking status. Hazard ratios (HRs) and 95% confidence intervals (CIs) were estimated for the association between PEG-Loxe exposure and each outcome. The proportional hazards assumption was formally tested using Schoenfeld residuals. For covariates exhibiting significant violations of the proportional hazards assumption (defined as P < 0.05 in the Schoenfeld test), a stratified Cox model was implemented, allowing the baseline hazard function to vary across strata of the violating covariate while maintaining a common treatment effect estimate.

To mitigate confounding by indication and selection bias arising from the substantial baseline imbalances between treatment groups, PSM was performed. The propensity score—representing the conditional probability of receiving PEG-Loxe given observed baseline covariates—was estimated for each participant using a multivariable logistic regression model. The model included the following pre-specified covariates: age, sex, diabetes duration, BMI, history of cardiovascular disease, baseline eGFR category, baseline HbA1c, smoking status, and baseline use of SGLT2i. Participants in the PEG-Loxe group were matched to control participants using a 1:1 nearest-neighbor matching algorithm without replacement, with a caliper width set at 0.2 times the standard deviation of the logit of the propensity score to ensure high-quality matches.³³ The adequacy of covariate balance after matching was assessed using standardized mean differences (SMD), with an SMD < 0.1 considered indicative of negligible residual imbalance. Cox proportional hazards regression was subsequently applied to the matched cohort, with robust standard errors accounting for the matched-pair structure.

Kaplan-Meier survival curves were constructed to estimate cumulative incidence functions for time-to-event outcomes, and between-group differences were assessed using the log-rank test. Given the differential follow-up durations between treatment groups, the median follow-up time was estimated using the reverse Kaplan-Meier method, which treats the censoring distribution as the event of interest and provides an unbiased estimate of potential follow-up time.

All statistical analyzes were performed using R software (version 4.3.3) ³⁴ with the following packages: “survival” ³⁵ for Cox regression and Kaplan–Meier estimation, “MatchIt” ³⁶ for PSM, “survminer” ³⁷ for visualization, and “E-Value” ²⁹ for sensitivity analysis to unmeasured confounding. A two-sided P-value < 0.05 was considered statistically significant for all hypothesis tests.

Author contributions

Conceptualization: Z.-S.C., K.H. Methodology: J.Y., Q.C., Y.T. Software: J.Y., F.C., Q.C. Validation: F.C., D.Z., C.Z., Y.T. Formal analysis: J.Y., Q.C., Y.T. Investigation: F.C., D.Z., C.Z., S.Y., W.X., B.X. Resources: J.L., Y.C., S.Y., W.X., K.H. Data curation: F.C., D.Z., C.Z., Q.C., Y.T., S.Y., W.X., B.X. Writing - original draft: J.Y., J.L., Y.C. Writing - review & editing: J.Y., Z.-S.C., K.H. Visualization: J.Y., F.C., Q.C., B.X. Supervision: Z.-S.C., K.H. Project administration: J.L., Y.C., K.H. J.Y., J.L., and Y.C. contributed equally to this work. Z.-S.C. and K.H. jointly supervised this work and share corresponding authorship. All authors have read and approved the article.

Data availability

The datasets generated and analyzed during the current study are not publicly available due to institutional restrictions regarding the confidentiality of electronic health records and patient privacy. However, de-identified datasets that support the primary findings of this study are available from the corresponding author on reasonable request for non-commercial research purposes. The trial protocol and statistical analysis plan are also available upon request. The trial is registered with the Chinese Clinical Trial Registry (ChiCTR2300073001).

Code availability

The customized R scripts and analytical protocols used for the multivariable-adjusted Cox proportional hazards models and propensity score matching in the FLYING 2 study are available from the corresponding author upon reasonable request.

Competing interests

The authors declare no competing interests. While the primary baseline study (FLYING 1) received support from Jiangsu Hansoh Pharmaceutical Group Co., Ltd., the sponsor played no role in the design, execution, analysis, or funding of this secondary extended study (FLYING 2).

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41392-026-02945-5.