Diabetes has a high incidence and mortality rate, thus emerging as a major global health challenge. The global prevalence of diabetes in people aged 20~79 years may increase to 12.2% by 2045 (1). Diabetic retinopathy (DR), the most common and serious ocular complication of type 2 diabetes, can result in visual impairment and even permanent vision loss (2). To date, DR is the leading cause of blindness among individuals aged 20~74, contributing to new cases of preventable blindness in developing countries (3, 4). A meta-analysis showed that the worldwide prevalence of DR is about 22.27% (5). Notably, China has the largest number of diabetic patients in the world (1), with about 27.9% of diabetes cases being DR (6), of which 34% are in rural areas. Most DR patients are asymptomatic or mild in the early stage and are only detected when great damage has occurred or at the late stage (2). Therefore, early screening of DR is crucial. However, the detection rate of DR remains suboptimal (7), particularly in certain primary hospitals and remote regions due to the constraints in medical conditions and limited social resources. Therefore, more accessible methods are needed to aid healthcare professionals in diagnosis and screening of DR.

To date, factors such as diabetes duration, age, BMI, smoking, blood pressure, HbA1c levels, and cholesterol have been identified as risk factors for diabetic retinopathy (DR) (8, 9). However, only a limited number of studies have incorporated insulin resistance (IR) into DR prediction models, despite substantial evidence (10, 11)demonstrating a strong association between IR and DR. The hyperinsulinemic-euglycemic clamp (HIEC) is the gold standard for detecting IR (12), However, it is expensive and complex, limiting its use. HOMA-IR (13) is the most commonly utilized method for estimating insulin resistance; however, it requires the measurement of fasting insulin levels, which imposes specific technical demands on laboratory capabilities. Additionally, this method is not suitable for patients undergoing insulin therapy and is not widely adopted in primary care settings or resource-limited regions. Therefore, HOMA-IR cannot be widely promoted in primary hospitals and poor areas (14, 15). A new index has been recently identified for detecting IR: Triglyceride-glucose index (TyG) (16), TyG is calculated by fasting triglyceride (TG) and fasting blood glucose (FBG), providing a simple, reliable, and cheap detection tool (17). Besides, this index tool has shown better results than HOMA-IR (18), Elsewhere, Srinivasan (19) and Yao (20) have shown that TyG and DR are closely related, while others not.

In recent years, artificial intelligence (AI) has demonstrated significant advantages in disease prediction. Machine learning (ML), encompassing both deep learning (DL) and traditional ML approaches represents the most widely adopted technique (21). While DL has proven highly effective in medical image analysis—as evidenced by studies on brain tumor classification using DCNN (22), U-Net-based MRI segmentation (23), and mammography interpretation via XAI-RACapsNet (24)—its superior performance is primarily confined to high-dimensional (imaging) data. Furthermore, DL imposes higher computational demands than traditional ML methods. Due to the inherent opacity of its multiple processing layers (the so-called "black box" problem), DL poses greater challenges to model interpretability and accountability (25). In contrast, traditional ML algorithms (e.g., SVM, LR) exhibit distinct advantages when processing tabular data: superior interpretability through SHAP, critical for clinical trust (26); reduced computational demands, enabling deployment in resource-limited settings (27); and comparable performance to DL on tabular biomedical data with limited samples (27). Given our goal of developing a primary-care-suitable tool, we prioritized lightweight, interpretable models over complex DL architectures. In the absence of imaging inputs, ML outperforms DL in tabular data prediction. The newly discovered predictors can improve the predictive value for the prediction model (28, 29), TyG, as a new index for measuring IR, has shown great value in many studies, While integrating the TyG as a novel predictor further enhances accessibility without compromising performance. However, no prediction model uses TyG to predict the DR among the DM2 population.

To bridge this critical gap, we present the first machine learning framework integrating TyG for DR prediction, which simultaneously addresses three key limitations of prior approaches: 1) Methodological Barrier: Existing models (30–33) neglect insulin resistance (IR) or rely on impractical biomarkers [e.g., HOMA-IR requiring fasting insulin (11)]; 2) Clinical Applicability: TyG leverages routine lipid/glucose tests, enabling IR assessment in resource-limited settings where specialized assays are unavailable (17, 18); 3) Model Interpretability: We employ SHAP to decode the "black box" nature of ML models, quantifying TyG's non-linear contribution to DR risk. By validating this approach across multi-tier healthcare centers (urban tertiary vs. rural primary hospitals), we establish TyG as a novel, deployable biomarker for scalable DR screening.

Early screening and identification of people at high risk of developing DR is important for prevention and treatment. However, Li et al. (39) showed that only 17.48% of T2DM undergo routine DR Screening yearly. Doctors in the diabetes department, especially in primary hospitals that lack equipments and professionals, often overlook the early signs of retinopathy in their patients (36, 37). This study introduces a paradigm shift in diabetic retinopathy (DR) risk stratification by integrating the triglyceride-glucose index (TyG) – an easily measurable, low-cost surrogate of insulin resistance (IR) – into machine learning prediction models. Unlike prior models that either overlook IR (40–42) or rely on impractical biomarkers like HOMA-IR (requiring specialized insulin assays) (11), our approach leverages routinely available lipid/glucose data, addressing a critical barrier in primary care settings. The prominent contribution of TyG in SHAP analysis (Figures 4–6) further establishes its role as a novel, interpretable predictor for scalable DR screening.

Results showed that logistic regression and SVM had the best performance among the five models, with area under ROC curve of the testing set and training set of 0.85 and 0.82 respectively. Similarly, Tsao et al. (43) showed that the SVM model has good predictive performance. Jiang et al. (37) also showed that logistic regression has good predictive performance. LR is a linear model suitable for small sample analysis, and it is sensitive to outliers. SVM finds the optimal hyperplane by optimizing the objective function, which is suitable for the analysis of complex data (37). Different algorithms have their own advantages in the modeling process, and no algorithm can be applied to all models. Therefore, the corresponding model should be selected according to the research design. This study was conducted based on the computational minimum sample size. Therefore, a study with a larger sample size is needed to compare the performance between LR and SVM. Our SVM model demonstrated test set AUC (0.85) comparable to Jiang et al. (37) (AUC = 0.89) and superior to Wang et al. (40) (AUC = 0.709). The integration of TyG provided predictive power comparable to HOMA-IR-based models (11) without requiring specialized tests. Unlike Roşu et al. (42) (AUC = 0.72) and Yang et al. (32) (AUC = 0.78) which excluded insulin resistance indicators, our TyG-incorporated model achieved higher AUC (0.85) with clinically accessible variables. This balances performance and feasibility for primary care settings.While our initial hold-out test set showed excellent performance (AUC >0.83), the additional 5-fold cross-validation provides stronger evidence of model robustness. The cross-validated AUC remained above 0.81 with standard deviations <0.02, confirming reliable performance across diverse data partitions.

Herein, the variables included in LR and SVM were consistent. The remaining variables were described in previous studies except for the significant predictive ability of TyG (9, 31, 37, 44, 45). Previous studies showed that the interpretability of ML models is challenging (46). In the present study, SHAP was used to improve model interpretability. SHAP is used to interpret a "black box" model that calculates a Shapley value for each feature in the prediction model to assess the importance of all feature combinations, reflecting their contribution to the predictive power of the overall model (26). SHAP results showed that the top 5 variables included TyG, HDL, Insulin therapy, Diabetes course, and HbA1c.

This study observed discordant results for Smoking and LDL between univariate and multivariate analyses, with a notable reversal in smoking's effect direction (univariate: lower prevalence in DR group, P = 0.007; multivariate: protective association, OR = 0.759, P = 0.008). This reversal is explained by two synergistic mechanisms: First, hypertension mediation—hypertension was more prevalent in the DR group (56.25% vs 44.32%, P < 0.001) and strongly predicted DR (OR = 1.506, P < 0.001); since smoking promotes hypertension (47), adjusting for it in multivariate models revealed smoking's residual protective effect via metabolic pathways. Second, exclusion of severe comorbidities may favor resilient smokers, consistent with the "smoker’s paradox" in chronic disease cohorts (48, 49). Regarding LDL and HDL, their reversals reflect collinearity within lipid metabolism (univariate analysis failed to disentangle HDL/LDL effects from TyG, which correlates with triglycerides) and a synergistic risk interaction where low HDL combined with high TyG increases DR risk—captured only by multivariate models. These reversals underscore that risk factors' net effects depend on comorbidities: Heavy smokers with hypertension remain high-risk despite smoking's statistical "protection," while LDL's reversal highlights complex lipid interactions in DR pathogenesis.

TyG emerged as a key predictor with dual significance. Notably, increased TyG levels were associated with a high risk of DR due to the excessive production of mitochondrial superoxide in microvascular endothelial cells. This production is caused by pathway-specific insulin resistance, meanwhile, triggering intracellular hyperglycemia and vascular damage (50). As a biochemical indicator reflecting synergistic dysregulation of glucose/lipid metabolism (51), TyG directly links to DR pathogenesis. Our SHAP analysis revealed a “non-linear risk surge” at TyG>4, aligning with the point where dysregulated glucose/lipid metabolism triggers mitochondrial superoxide overproduction in retinal microvessels – a key driver of early vascular damage. Based on multivariate regression, each unit increase in TyG elevates DR risk by 9.1% (OR = 1.091, 95%CI:1.005–1.185). The threshold effect at TyG>4 reflects amplified risk accumulation consistent with Zhou et al.'s (52) report. This threshold is calculable from routine lipid/glucose tests. We recommend annual fundus screening for patients exceeding TyG>4. Crucially, TyG overcomes limitations of traditional IR metrics. Our study is the first to operationalize TyG within a predictive algorithm, with its threshold effect (TyG>4) providing actionable clinical guidance.Though TyG cut-offs vary by disease context [e.g., 4.65 for IR (16), 8.88 for diabetes progression (53)], our threshold specifically targets DR risk stratification in primary care settings where standardized values remain undefined. A meta-analysis (54) showed that there is no standardized threshold value for TyG and clear threshold value related to TyG and DR. Therefore, more high-quality studies are needed to determine the optimal cut-off point for TyG in the future.

Insulin therapy is another key predictor of DR. Wang, Li et al (40, 45) found that the insulin treatment is associated with a high risk of DR development, possibly because insulin therapy indicates worse islet function. While Ricard found it possibly related to the rapid reduction of blood glucose (55). Nonetheless, more basic research is needed to confirm the findings. This study, along with others, has demonstrated that the duration of diabetes mellitus is closely associated with the development of DR (30, 41, 44, 56), possibly due to the prolonged exposure of blood vessels to risk factors.

Herein, DR was linked to increased HBA1c levels. This may be due to the continuous elevation of blood glucose, which can lead to dysfunction of the retinal vascular endothelium and cause retinal ischemia and increased vascular permeability (2). Similar findings have been reported previously (11, 30, 45).

In addition, HDL was identified as a key predictor of DR. Low HDL levels may indicate a higher risk of DR, consistent with Roşu, Liu, Li et al (42, 57, 58).

By addressing three critical gaps – 1) IR assessment simplification via TyG, 2) cross-tier validation (urban tertiary/rural primary hospitals), and 3) SHAP-driven interpretability – our model enables scalable DR screening. Our model requires only 9 routinely collected variables, enabling seamless integration into electronic health records (EHR) for real-time risk stratification during diabetic outpatient visits. Based on SHAP probability thresholds (Figures 4-6), we propose a tiered screening pathway: High-risk patients (TyG >4 and other key predictors elevated): Prioritize immediate ophthalmologist referral. Low-to-intermediate-risk patients: Maintain routine annual/biannual retinal screening. This approach is particularly valuable for resource-limited settings (e.g., rural clinics), where the model can optimize specialist resource allocation by focusing on high-risk individuals identified through universally available tests like TyG (calculated from routine lipid/glucose tests). Besides, Our model exhibits inherent generalizability potential through its reliance on universal predictors requiring only routine clinical data, demonstrated demographic resilience across diverse populations, and established scalability pathways including a simplified screening protocol and ongoing multi-region validation.

This study bridges a translational gap: while TyG-DR associations were reported, we are the first to validate its clinical utility within a predictive algorithm. The integration with SHAP interpretation further distinguishes our model from conventional approaches. while retrospective data from two hospitals with different resource levels (urban tertiary vs. rural primary) may introduce selection bias, this heterogeneity enhances real-world generalizability.

Limitation: First, only internal verification was conducted, thus effective external verification is needed. Secondly, excluding high-missingness variables may omit useful information, and while mean/mode imputation is practical, advanced techniques (e.g., multiple imputation) might reduce potential bias. Last, this study did not differentiate early vs. late-stage DR in predictions due to ungraded diagnostic records. However, our model prioritizes biomarkers linked to DR progression (TyG, diabetes course). Future work should incorporate standardized severity staging to enable stage-specific optimization.

In this study, a DR prediction model was built using TyG and other easily available clinical data. LR and SVM models had the best performance. SHAP showed that the most important predictors of DR were TyG, insulin therapy, diabetes course, HbA1c, and HDL. These findings suggest that baseline characteristics of patients can be used to screen for high-risk DR to improve early diagnosis and implement timely referral for patients.