Exploring the impact of high rosuvastatin plasma exposure on new-onset diabetes mellitus: insights from machine learning-based prediction

Abstract

Background

The association between rosuvastatin (RST) and the risk of new-onset diabetes mellitus (NODM) is controversial. Although the link between RST and NODM is still debated, there is a lack of effective strategies to predict and prevent potential RST-induced NODM in clinical practice. This study aimed to determine the association between plasma exposure to RST and the risk of developing NODM in patients with cardiovascular disease and to establish predictive models for the early detection of RST-induced NODM.

Methods

We included 704 patients with cardiovascular disease and without diabetes who had been on RST for > 4 months. We used ultra-performance liquid chromatography–tandem mass spectrometry to detect the concentration of RST and its metabolites. We then used machine learning (ML) to analyze the impact of plasma exposure to RST and patient characteristics on NODM and established four risk prediction models for RST-induced NODM. The optimal model was determined using the receiver-operating-characteristic (ROC) curve (AUC) and the Shapley algorithm (SHAP) interpretation.

Results

Our findings indicated that high plasma exposure to RST was an independent risk factor for NODM. In terms of NODM, the Random Forest model demonstrated the highest AUC value (0.7310) in the testing set and was validated as the best performing model. The SHAP analysis identified high triglyceride levels, advanced age, and high RST plasma exposure as the top three predictors of NODM. In addition, low total cholesterol and LDL-cholesterol levels were also associated with an increased risk of NODM.

Conclusions

The study findings revealed high plasma exposure of RST as an independent risk factor for NODM in patients without diabetes who are on RST, which may relate to “on-target” effects of RST. In addition, ML interpreted using SHAP provide valuable tools for early prediction and personalized management of RST-induced NODM.

Supplementary Information

The online version contains supplementary material available at 10.1186/s40001-025-03029-w.

Introduction

Rosuvastatin (RST), a hydroxyl-methyl-glutaryl-coenzyme-A (HMG–CoA) reductase inhibitor, is effective in reducing low-density lipoprotein cholesterol (LDL-C) levels, has few drug–drug interactions, and is widely used in clinical practice [1–3]. The blood concentration of the drug is twice as high in Asians than in Caucasians at the same RST dosage, resulting in Asians taking only half the dose of Caucasians [4–6]. However, studies have shown that Asians have a notably higher risk of developing statin-induced new-onset diabetes mellitus (NODM) than Caucasians, contradicting the conclusion that high-dose statins increase the risk of NODM [7–9]. Moreover, the risk was not significant in RST users after adjusting for dosage in a population-based cohort study [10]. Investigating the impact of RST on NODM through plasma exposure may accurately reflect the influence of RST on NODM. However, no previous studies have been conducted in this aspect.

Consequently, the effect of RST on NODM remains controversial. Whether the inconsistent conclusions regarding RST-induced NODM are due to the variability in plasma exposure to RST require further confirmation. Statin-induced NODM is related to the reduction in LDL-C levels through the inhibition of HMG–CoA reductase. This phenomenon has been reported in a Mendelian randomization analysis, indicating that the increased risk of diabetes observed with statin therapy may result from the drug’s “on-target” effects—that is, the mechanism of lowering LDL-C through HMG–CoA reductase inhibition, which may increase diabetes risk [11]. However, whether RST in the plasma exhibits this on-target effect requires further investigation.

In this study, we investigated the effects of plasma exposure to RST and its metabolite Rosuvastatin lactone (RSTL), on NODM. Furthermore, we used machine learning (ML) to analyze the link between RST and RSTL plasma exposure and NODM. In addition, to construct an ML model in predicting NODM for such patients using RST plasma exposure in combination with other clinical features.

Materials and methods

Study design and population

Ethical approval was obtained from the Ethics Committees of Guangdong Provincial People’s Hospital (no.2017071H) and Guizhou Provincial People’s Hospital (no. [2020]20). Both ethics committees approved the use of secondary data derived from electronic medical records as part of this study. The study was conducted in accordance with the Declaration of Helsinki, and written informed consent was obtained from all participants. We included patients with cardiovascular disease from two hospitals who had been on RST for more than 4 months and had no prior diagnosis of diabetes and were not receiving any antidiabetic medication at the start of RST treatment. Patients with plasma levels of RST or RSTL below the detection limit were excluded. Demographic data, medical history, and medication history of all participants were collected and recorded. The inclusion criteria are shown in Fig. 1.

Fig. 1.

Inclusion and exclusion criteria for patients

The primary endpoint of this study is the occurrence of NODM in patients treated with RST, which was defined as glycosylated hemoglobin (HbA1c) levels ≥ 6.5% or fasting plasma glucose (FPG) levels ≥ 7.0 (mmol/L). The thresholds for FPG and HbA1c levels were defined according to the criteria for the diagnosis of diabetes in the Standards of Care in Diabetes-2023 of the American Diabetes Association (ADA) [12].

Laboratory index measurements

The laboratory assays included measuring the levels of serum total cholesterol (TC), triglyceride (TG), high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), alanine aminotransferase (ALT), aspartate aminotransferase (AST), creatinine (CREA), creatine kinase (CK), apolipoprotein A1 (APOA), creatine kinase-MB (CKMB), lipoprotein (a) (Lpa), FPG levels, and HbA1c levels, which were retrospectively obtained from the hospital’s electronic medical records system.

Quantification of RST and its metabolite concentrations in plasma

All patients who received RST for more than 4 months were included in the study. For each participant, 5 mL of blood was collected in an ethylenediaminetetraacetic acid anticoagulant tube on an empty stomach in the morning for all recruitments. After centrifugation at 3000 ×g for 10 min at 4 °C, the supernatant was transferred to a fresh tube and stored at − 80 °C.

We used ultra-performance liquid chromatography–tandem mass spectrometry (UPLC–MS/MS) to analyze the plasma exposure to RST and RSTL in the serum samples. A rapid, sensitive, and selective UPLC–MS/MS method was developed and validated for quantifying RST and RSTL in human plasma. RST and RSTL, and the internal standard were isolated from human plasma via liquid–liquid extraction with ethyl acetate and then separated on an Acquity UPLC HSS T3 column (3.0 mm × 100 mm, 1.8 µm) (Waters, USA) with 0.1% formic acid using 0.1% (v/v) formic acid and a gradient of 30–85% acetonitrile at a flow rate of 0.30 mL/min. Mass detection was performed using a Waters Xevo TQ-S triplequadrupole mass spectrometer in positive electrospray ionization mode. The responses of RST and RSTL were optimized at m/z 482.1 → 258.1, m/z 464.1 → 270.1, m/z, respectively.

Model construction and performance evaluation

In clinical practice, missing data are common; thus, variables with more than 20% missing data were excluded from the analysis. Those with less than 20% missing data were imputed using multiple imputation by chained equations (MICE) in R, with predictive mean matching (PMM) specified as the imputation method to ensure imputed values remained within a plausible range of the observed data. Subsequently, the data were pre-processed and entered into a model-building process. We used four machine learning models—Extreme Gradient Boosting (XGBoost), Random Forest (RF), Gradient Boosting Decision Tree (GBDT), and LightGBM—to build the model. Due to the class imbalance between positive and negative events in our cohort, we addressed this issue by applying class weighting within the machine learning models. The data set was split into training (60%) and validation (40%) sets for internal validation [13, 14]. The models were evaluated using indexes such as the area under the receiver-operating-characteristic (ROC) curve (AUC), sensitivity, specificity, accuracy, and F1-score. Moreover, fivefold cross-validation was performed on the model generated using the training set. Furthermore, we used the Shapley (SHAP) method to visualize the importance of variables in the ML models and to provide indicators to explain the model’s risk-prediction mechanism.

Statistical analysis

In the analysis of baseline patient characteristics, categorical data are presented as percentages and continuous variables as mean ± standard deviation. Plasma exposures to RST and RSTL were logarithmically transformed to normalize their distribution. An independent sample t tests were used to compare the plasma concentrations of RST and RSTL between the NODM and non-NODM groups. Univariate logistic regression analysis was performed to evaluate the effects of RST and RSTL on NODM. In the results, variables with p < 0.1 were included in the multivariable model, and only variables with p < 0.05 were retained. Statistical significance was set at p < 0.05. Python (version 3.5) was used to build the ML models, R software (version 4.1.1) and Python were used for plotting, and SPSS (version 26) was used for statistical analysis.

Results

Patient characteristics and their effects on NODM

In terms of NODM, 704 patients were included in the study. 99 individuals had NODM during RST treatment. The incidence rate of NODM was 14.06%. The univariate logistic regression analysis revealed that age (odds ratio (OR): 1.03; 95% confidence interval (CI) 1.01–1.05; p = 0.002), Gender (OR: 0.59; 95% CI 0.38–0.93; p = 0.022), Heart failure (OR: 5.37; 95% CI 2.58–11.20; p = 0.001), TG (OR: 1.17; 95% CI 1.00–1.36; p = 0.047), RST plasma exposure (OR: 1.53; 95% CI 1.17–1.99; p = 0.002), RSTL plasma exposure (OR: 1.32; 95% CI 1.05–1.65; p = 0.018), and use of proton pump inhibitors (OR: 1.55; 95% CI 1.01–2.40; p = 0.046) were independent risk factors for NODM. The multivariate logistic regression analysis revealed that Heart failure (OR: 4.83; 95% CI 2.25–10.35; p = 0.000), TG (OR: 1.24; 95% CI 1.05–1.47; p = 0.013), and RST plasma exposure (OR: 1.47; 95% CI 1.12–1.93; p = 0.005) were independent risk factors for NODM (Table 1).

Table 1.

Patients’ characteristics and their effects on NODM

Characteristics	Non-NODM	NODM	Univariate		Multivariate
	N (%) or mean ± SD	N (%) or mean ± SD	OR (95% CI)	P value	OR (95% CI)	P value
Demographic data
Total no.	605	99
Age	61.8 ± 10.8		1.03 (1.01–1.05)	0.002
Gender
Female	153 (25.3%)	36 (36.4%)
Male	452 (74.7%)	63 (63.6%)	0.59 (0.38–0.93)	0.022
Dosage (mg)
5	14 (2.3%)	2 (2%)
10	534 (88.3%)	85 (85.9%)
20	57 (9.4%)	12 (12.1%)	1.03 (0.96–1.10)	0.398
Medical history
Arrhythmia
No	565 (93.4%)	90 (90.9%)
Yes	40 (6.6%)	9 (9.1%)	1.41 (0.66–3.01)	0.371
Heart failure
No	587 (97%)	85 (85.9%)
Yes	18 (3%)	14 (14.1%)	5.37 (2.58–11.20)	0.001	4.83 (2.25–10.35)	0.000
Hypertension
No	353 (58.3%)	51 (51.5%)
Yes	252 (41.7%)	48 (48.5%)	1.32 (0.86–2.02)	0.203
Hyperlipidemia
No	552 (91.2%)	93 (93.9%)
Yes	53 (8.8%)	6 (6.1%)	0.67 (0.28–1.61)	0.372
Biochemical measurements
ALT, U/L	31.0 ± 26.8	34.4 ± 25.1	1.00 (1.00–1.01)	0.240
AST, U/L	30.4 ± 23.4	34.8 ± 37.1	1.01 (1.00–1.01)	0.128
CREA, μmol/L	83.0 ± 33.0	89.1 ± 44.6	1.00 (1.00–1.01)	0.113
CK, U/L	116.1 ± 140.6	141.5 ± 362.4	1.00 (1.00–1.00	0.247
APOA, g/L	1.1 ± 0.3	1.1 ± 0.3	1.03 (0.48–2.18)	0.942
TC, mmol/L	4.4 ± 1.2	4.4 ± 1.3	0.98 (0.83–1.17)	0.861
CKMB, U/L	7.2 ± 5.6	8.3 ± 12.0	1.02 (0.99–1.04)	0.166
HDL-C, mmol/L	1.0 ± 0.3	1.0 ± 0.3	0.68 (0.29–1.56)	0.360
LDL-C, mmol/L	2.7 ± 1.0	2.6 ± 0.8	0.89 (0.71–1.12)	0.334
Lpa, mg/L	273.4 ± 301.0	304.4 ± 330.2	1.00 (1.00–1.00)	0.350
TG, mmol/L	1.5 ± 1.0	1.8 ± 1.8	1.17 (1.00–1.36)	0.047	1.24 (1.05–1.47)	0.013
Plasma exposure
RST, ng/ML	3.8 ± 3.4	5.4 ± 4.6	1.53 (1.17–1.99)	0.002	1.47 (1.12–1.93)	0.005
RSTL, ng/ML	0.5 ± 0.5	0.7 ± 0.7	1.32 (1.05–1.65)	0.018
Medication
β-Blockers
No	98 (16.2%)	13 (13.1%)
Yes	507 (83.8%)	86 (86.9%)	1.28 (0.69–2.38)	0.438
ACEIs
No	298 (49.3%)	50 (50.5%)
Yes	307 (50.7%)	49 (49.5%)	0.95 (0.62–1.46)	0.818
CCBs
No	447 (73.9%)	64 (64.6%)
Yes	158 (26.1%)	35 (35.4%)	1.55 (0.99–2.43)	0.057
PPIs
No	304 (50.2%)	39 (39.4%)
Yes	301 (49.8%)	60 (60.6%)	1.55 (1.01–2.40)	0.046
Clop
No	21 (3.5%)	3 (3%)
Yes	584 (96.5%)	96 (97%)	1.15 (0.34–3.93)	0.823
Aspirin
No	31 (5.1%)	5 (5.1%)
Yes	574 (94.9%)	94 (94.9%)	1.02 (0.39–2.68)	0.976

ALT, alanine transaminase; AST, aspartate transaminase; CREA, creatinine; CK, Creatine Kinase; APOA, apolipoprotein a; TC, total cholesterol; CKMB, creatine kinase-MB; HDL-C, HDL-cholesterol; LDL-C, LDL-cholesterol; Lpa, lipoprotein (a); TG, triglycerides; RST, rosuvastatin; RSTL, rosuvastatin lactone; β-blockers, beta-blockers; ACEIs, angiotensin converting enzyme inhibitors; CCBs, calcium channel blockers; PPIs, proton pump inhibitors; clop, clopidogrel

Variables with p < 0.1 were entered into the multivariable model, and only variables with p < 0.05 were retained in the model

The RST concentration of the NODM (5.4 ± 4.6 ng/mL) was significantly higher than that of the non-NODM (3.8 ± 3.4 ng/mL, OR: 1.47; 95% CI 1.12–1.93; p = 0.005) (Table 1) and (Fig. 2). The plasma exposure of RST among different individuals varied widely, from 0.1 to 34.28 ng/mL. The plasma exposure to RSTL significantly correlated with the RST (r = 0.60, p < 0.0001) (Fig. S1A). However, the plasma exposure to RSTL was much lower than that to RST. The frequency distributions of the steady-state plasma exposure for both analytes are shown in Fig. S1B.

Fig. 2.

Plasma exposure of RST and RSTL between NODM and non-NODM in all enrolled patients

Establishment of a risk prediction model on NODM

We investigated the potential for identifying NODM in patients undergoing RST therapy by analyzing RST and RSTL plasma concentrations, and relevant clinical characteristics. Within the training cohort, four predictive models were subsequently developed for NODM in this population. The RF model had the best prediction performance, with obtaining the highest AUC (73.10%) (Fig. 3A). In addition, the accuracy was highest in the fivefold cross-validation of the training set (Fig. 3B). In the internal evaluation data set, Fig. 3C provides detailed information.

Fig. 3.

A ROC curves of the best-performing ML model. B Fivefold cross-validation of the RF model training set. C Evaluation the performance of all machine learning models with other metrics

Model interpretations

In the RF model, 20 clinical features and biochemical indicators were included (Table S1). The SHAP algorithm was used to visualize the best predictive model. This explainable method provided two types of explanations: global explanation of the model at the feature level and local explanation at the individual level. As shown in the SHAP summary plots (Fig. 4A, B), the top 12 most important features contributing to NODM prediction were identified. These included high TG levels, advanced age, high RST plasma exposure, elevated ALT levels, presence of heart failure, high RSTL plasma exposure, low TC levels, high CREA levels, female sex, elevated Lp(a) levels, low LDL-C levels, and high APOA levels.

Fig. 4.

A SHAP summary bar plot. B SHAP summary dot plot. The probability of NODM development increases with the SHAP value of a feature. C SHAP dependence plot. Each dependence plot shows how a single feature affects the output of the prediction model, and each dot represents a single patient

Figure 4C illustrates the relationship between the actual values of these features and their corresponding SHAP values. SHAP values above zero indicate that the model is more inclined to predict the positive class—that is, a higher risk of NODM.

Discussion

To the best of our knowledge, this is the first study to combine direct measurement of RST plasma exposure with ML to predict NODM risk in real-world patients. While prior studies have largely focused on statin dosage associations, our approach provides individual-level, pharmacokinetic evidence that supports an exposure-dependent risk mechanism. Importantly, the application of explainable ML techniques allowed us to uncover non-linear interactions and threshold effects of biochemical factors, that may not be easily detected by traditional regression models. These findings offer new insights into personalized risk stratification and the mechanistic understanding of statin-induced NODM. In addition, this study showed that during RST treatment, low TC concentrations and low LDL-C concentrations were associated with an increased risk of NODM.

Our research illustrates previously conflicting conclusions. First, a population-based cohort study indicated that the use of RST was not associated with the occurrence of NODM, contradicting other reports [10]. This study showed that the use of RST led to NODM; however, the effect of RST on NODM needs to be characterized by plasma exposure to RST. Second, high-dose statin-induced NODM; however, Asians taking low-dose RST have a higher risk of developing NODM than Caucasians. This phenomenon may be related to significant racial and individual pharmacokinetic differences owing to genetic polymorphisms. Several studies have reported that RST exposure in Caucasians is twice that in Asians at the same dose [6]. Our previous study showed that ABCG2 421 C > A (rs2231142) significantly increased the concentration of RST and RSTL in a Chinese population [15]. Although ABCG2 genotyping was not evaluated in this cohort. This variant is highly relevant to the East Asians, where the mutation frequency of ABCG2 421 C > A (allele frequency ~ 35%), notably higher than the 14% observed in Caucasians [16–18]. In addition, RSTL, a metabolite of RST, influences HbA1c levels. The effect of RSTL may also be caused by its correlation with RST concentration (Figure S1A).

The SHAP analysis in our RF model showed that lower TC and LDL-C levels were associated with a higher risk of NODM in patients treated with RST. This finding is consistent with an analysis of a Mendelian randomization study, which suggested that the increased risk of diabetes caused by statins may be related to the “on-target” effect of the drug [11]. Similarly, a large observational study of 14,120 individuals also reported that low LDL-C was associated with a higher incidence of diabetes [19]. Recent nonrandomized observational studies have identified a phenomenon known as the “cholesterol paradox” in patients with acute myocardial infarction (AMI). This paradox refers to the observation that in patients with AMI with low baseline LDL-C levels, reductions in LDL-C levels, whether due to the natural course of the disease or through lipid-lowering treatments, do not lead to improved outcomes or reduced overall mortality [20, 21]. For example, a single-center retrospective analysis in Qatar included 1808 hospitalized patients with acute ST-elevation myocardial infarction who underwent primary percutaneous coronary intervention. At an average follow-up of 40 months, the incidence of diabetes mellitus was significantly higher in the statin-naive/LDL-C 70 mg/dL group [22]. Mechanistically, statins inhibit HMG–CoA reductase, reducing cholesterol synthesis and its precursor mevalonate. Cholesterol depletion may impair β cell dysfunction and impaired insulin secretion [23, 24]. Previous animal studies also confirmed that RST causes impaired islet β cell function and decreased insulin content, accelerating the risk of elevated blood glucose [25]. Using SHAP to uncover this LDL-C–NODM association in a machine learning framework, this study offered a novel perspective that integrates complex, nonlinear relationships—an aspect not fully captured in traditional statistical models. This highlights the potential clinical importance of closely monitoring glycemic outcomes in patients undergoing aggressive LDL-C lowering.

Furthermore, our RF model demonstrated superior predictive performance for NODM compared to traditional single markers. Consistent with previous studies, we identified several key clinical factors associated with the development of NODM, including high TG levels, advanced age, high ALT levels, heart-failure, high CREA levels, female, high Lpa levels, and high APOA levels [26–34]. While earlier studies have primarily relied on conventional statistical approaches to explore the relationships between individual predictors and diabetes risk, this study used ML and SHAP dependence plots. These methods helped us see more complex and non-linear relationships between the risk factors and NODM. Therefore, combining these clinical features in a machine learning model may predict NODM more accurately than using single markers.

In this study, we used ML to build a predictive model for RST use leading to NODM. ML models can comprehensively analyze all types of clinical data to provide a more comprehensive data interpretation. In this study, the RF model outperformed the other three models, achieving an overall AUC of 73.10% in the test set, also exhibiting better accuracy and F1-score. With the development of ML technology, ML has been explored for diseases, such as cancer, infectious diseases, and cardiovascular disease [35–37]. ML can automatically select and extract important features related to target variables and eliminate the limitations of traditional statistical methods, which are subject to overfitting and multicollinearity. Our results showed that ML models may serve as potential tools in clinical decision-making by identifying patient-specific patterns based on individual characteristics and medical history, thus contributing to more personalized treatment planning.

This study has a few limitations. First, although this study employed an ML model (RF) to predict NODM and achieved a high AUC (73.10%), the “black-box” nature of ML models remains a challenge. While the SHAP method enhances model interpretability to some extent, the clinical feasibility and practicality of ML-based predictions require validation in larger data sets. Therefore, the model's generalizability remains to be further evaluated. Second, this study found that lower LDL-C levels are associated with an increased risk of NODM, which may be linked to the"on-target"effect of statins. However, the underlying mechanism remains unclear, as current evidence primarily stems from Mendelian randomization analyses and observational studies. Further basic and clinical research is required to determine whether LDL-C reduction directly leads to β-cell dysfunction or increased insulin resistance.

Conclusion

In this study, the effects of RST and RSTL plasma exposure on the risk of NODM were systematically analyzed. Our findings demonstrated that high RST plasma exposure is an independent risk factor for NODM, and this association was validated by ML models. Furthermore, low LDL-C levels were also identified as an important predictor of NODM risk during RST therapy, further supporting the impact of the"targeted"effect of statins on glucose metabolism.

Author contributions

Y.C. and F.Y.M. was responsible for conceptualization, data curation, formal analysis, and writing the original draft. FS contributed to the acquisition and interpretation of data and revision of the manuscript. C.L.T. B.L.C. S.H.Z. Z.Q.G. L.L. X.Z.L. Q.C. C.C.S. Q.M. and J.E.L. contributed to the data analysis and revision of the manuscript. S.L.Z. contributed to conceptualization, supervision, resources, and writing review and editing. X.B. contributed to conceptualization, supervision, and project administration, secured funding.

Funding

This research was supported by the National Natural Science Foundation of China (No.81960681), 2024 Drug Safety Research Project (No.ADR2024MS15), the Guizhou Provincial People’s Hospital 2019 National Natural Science Foundation of China post-subsidy fund (No. GPPH-NSFC-2019–26[2019]GPPH-NSFC-D-2019–25), the Hospital Pharmacy Innovation Project for the youth established by Chinese Pharmaceutical Association and Servier (Tianjin) Pharmaceutical Co., Ltd. (No. CPA-B04-ZC-2024–001), the Science and Technology fund project of Health and Family Planning Commission of Guizhou province (NO. gzwjkj2018-1–002).

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

Ethical approval was obtained from the Ethics Committees of Guangdong Provincial People’s Hospital (no.2017071H) and Guizhou Provincial People’s Hospital (no. [2020]20).

Consent for publication

The study was conducted in accordance with the Declaration of Helsinki, and written informed consent was obtained from all participants.

Competing interests

The authors declare no competing interests.