Application of construction and validation of logistic regression model in risk prediction of non‐proliferative diabetic retinopathy in type 2 diabetes mellitus

Abstract

Aims

Diabetic retinopathy (DR) is a leading cause of vision impairment in type 2 diabetes mellitus (T2DM), with non‐proliferative DR (NPDR) representing its most prevalent form. Early identification of high‐risk individuals remains challenging due to the complexity and poor interpretability of existing machine learning models. This study aims to develop a clinically interpretable Logistic regression model for NPDR risk prediction using routinely available clinical indicators.

Methods

A retrospective cohort of 421 T2DM patients from a single centre was divided into training (n = 295) and validation (n = 126) sets. Demographic, glycemic (fasting glucose, HbA1c), renal (UACR) and ophthalmologic (macular oedema) data were collected. Univariate and multivariate Logistic regression with stepwise selection identified independent predictors. Model performance was evaluated using area under the ROC curve (AUC), sensitivity, specificity and Hosmer–Lemeshow goodness‐of‐fit. Internal validation was performed via bootstrapping (1000 replicates), and external validation used an independent cohort.

Results

Four independent predictors were identified: macular oedema (OR = 3.247), fasting glucose (OR = 2.194), HbA1c (OR = 2.799) and UACR (OR = 1.153). The model demonstrated excellent discrimination in the training set (AUC = 0.949, sensitivity = 86.4%, specificity = 95.5%) and good calibration (H–L test, p = 0.358). Bootstrap validation confirmed stability of HbA1c and UACR. External validation yielded an AUC of 0.918, with a positive predictive value of 91.1% and maintained calibration (p = 0.282).

Conclusions

The constructed Logistic regression model accurately predicts NPDR risk using four readily available clinical variables, offering high discriminative power, interpretability and clinical utility for stratifying high‐risk T2DM patients in primary care settings.

What's new?

• A parsimonious Logistic model using only macular edema, fasting glucose, HbA1c, and UACR predicts NPDR risk with an AUC of 0.949.

• High clinical interpretability overcomes the “black box” limitation of complex algorithms, aiding primary care screening.

• External validation shows robust generalizability (AUC = 0.918, PPV = 91.1%) and excellent calibration (H‐L test P = 0.282).

1. INTRODUCTION

Diabetic retinopathy (DR) is a common microvascular complication in diabetic patients and one of the leading causes of vision loss. ¹ The incidence and prevalence of DR vary significantly across regions and populations. ² According to a systematic review and meta‐analysis, the global prevalence of DR is 27.0%, including 25.2% for non‐proliferative diabetic retinopathy (NPDR) and 1.4% for proliferative diabetic retinopathy (PDR). ³ Among Asian patients with type 2 diabetes, the prevalence of DR is 28%, with 27% for NPDR and 6% for PDR. ⁴ Additionally, a national survey in China reported a DR prevalence of 16.3% and a vision‐threatening DR prevalence of 3.2%. ⁵ These data indicate that NPDR has a relatively high overall prevalence with uneven distribution across different regions and populations. Studies have shown that prolonged diabetes duration and elevated blood glucose levels are closely associated with the occurrence and progression of DR. ⁶ Furthermore, socioeconomic status, educational level and ethnicity are among the sociodemographic factors influencing DR prevalence. ⁷ A study in the United States found a significant increase in DR prevalence among young Hispanic and African American communities, highlighting the need for more intensive screening and intervention targeting these high‐risk groups. ⁸

DR is closely associated with multiple risk factors, including diabetes duration, glycemic control (HbA1c), hypertension and dyslipidemia. ⁹ In terms of DR prediction and screening, recent research and technological advancements have provided new possibilities for early detection and intervention. The application of machine learning and artificial intelligence has significantly improved the efficiency and accuracy of DR screening. ¹⁰ , ¹¹ However, these approaches still face challenges such as excessive input variables and lack of interpretability. ¹² This complexity not only increases computational costs but also hinders clinicians' ability to understand and trust the model predictions. ¹³ Studies have shown that machine learning algorithms for DR prediction typically rely on a large number of input variables. For example, Wan et al. used 39 optimal variables to construct a prediction model based on the eXtreme Gradient Boosting (XGBoost) algorithm and explained the model's key features using the Shapley Additive exPlanations (SHAP) method. ¹³ Despite enhancing predictive performance, this approach increases model complexity, making it difficult for clinicians to interpret the decision‐making process. Moreover, deep learning models are often regarded as ‘black boxes’ due to their opacity in explaining how representations are learned and why specific predictions are made. ¹⁴ Furthermore, deep learning models may bias toward the majority class when handling imbalanced datasets, thereby affecting overall performance. ¹⁵ This lack of transparency limits their clinical application, as physicians need to understand the model's decision‐making process to trust and utilize these tools.

Early screening and timely treatment of DR are crucial for preventing vision loss. However, many patients fail to undergo timely retinopathy screening after a diabetes diagnosis, impacting subsequent treatment outcomes. Studies have shown that delayed screening increases the detection rate of referable DR, underscoring the importance of early screening. ¹⁶ One study analyzed the impact of delayed screening on DR detection rates, revealing that 2.3% of patients screened within 6 months of diabetes diagnosis were identified with referable retinopathy, compared to 4.2% of those screened 3 years or later after diagnosis. ¹⁶ This indicates that delayed screening not only increases retinopathy detection rates but may also exacerbate disease severity, negatively affecting patients' vision and quality of life. Additionally, for low‐risk populations, studies suggest that screening intervals can be appropriately extended. Data show a low risk of progression to vision‐threatening retinopathy within 2 years in the absence of retinopathy (4.8 cases per 1000 person‐years). ¹⁷ However, this recommendation applies only to low‐risk groups; high‐risk patients or those with existing retinopathy require more frequent screening to enable timely detection and treatment. In summary, timely DR screening is critical for preventing visual impairment. For most patients, especially newly diagnosed individuals, early screening remains key to reducing retinopathy progression and improving treatment outcomes. ¹⁶ Furthermore, compared to complex machine learning models, logistic regression prediction models offer both interpretability and clinical operability, making them easier to integrate into community health management systems. ¹⁸

Based on this, this study aims to develop a logistic regression prediction model for NPDR in patients with type 2 diabetes, which serves as a pivotal tool to fill the current gap in NPDR screening. By integrating routine indicators such as diabetes duration, HbA1c, lipid profiles and renal function, individual risk is quantified to assist primary care physicians in identifying high‐risk patients with NPDR and facilitate early screening for such patients.

2. MATERIALS AND METHODS

2.1. Study subjects

A total of 421 patients with type 2 diabetes mellitus (T2DM) who met international diagnostic criteria were retrospectively enrolled from the Affiliated Wuxi People's Hospital of Nanjing Medical University, between January 2019 and December 2022. They were divided into a modelling set (295 cases) and a validation set (126 cases) at a 7:3 ratio. The modelling set included 177 patients with uncomplicated type 2 diabetes and 118 patients with NPDR. The dataset was randomly divided into a modelling set (n = 295) and a validation set (n = 126) at a ratio of 7:3 using stratified sampling based on NPDR status to ensure a comparable distribution of the outcome between both sets.

Inclusion criteria: (1) Confirmed diagnosis of type 2 diabetes according to international standards; (2) NPDR diagnosis conforming to the Diabetic Retinopathy Preferred Practice Pattern (PPP)‐2019 Guideline; (3) Age ≥18 years; (4) Signed informed consent.

Exclusion criteria: (1) Complicated with other ocular diseases (e.g., glaucoma, cataract); (2) Severe systemic diseases (e.g., end‐stage renal disease, malignant tumours); (3) Incomplete or untraceable clinical data.

2.2. Study methods

2.2.1. Collection of basic information

Basic data including gender, age, BMI, waist circumference and hip circumference were collected. Diabetes duration (years) was calculated from the date of initial diagnosis to enrollment. A structured questionnaire was used to record history of hypertension, smoking (continuous smoking for >6 months) and alcohol consumption (≥1 time/week). Additionally, macular oedema was diagnosed by ophthalmologists via fundus fluorescein angiography (FFA).

2.2.2. Collection of clinical indicators

Morning venous blood samples were collected, and biochemical indicators such as blood glucose, lipid profiles and serum creatinine were measured using a Roche Cobas8000 automatic biochemical analyzer (Roche, Germany). Glycated haemoglobin was determined by high‐performance liquid chromatography (HPLC). Estimated glomerular filtration rate (eGFR) was calculated using the CKD‐EPI formula based on serum creatinine levels. ¹⁹ Midstream morning urine samples were collected to measure urinary microalbumin and urinary creatinine using a microalbumin ELISA kit (Abcam, Cambridge, UK) and a microplate assay kit (Abcam, Cambridge, UK), respectively, for evaluating the urinary albumin/creatinine ratio (UACR).

2.3. Statistical analysis

SPSS 27.0 software was used for statistical analysis. Normality of continuous variables was assessed by non‐parametric tests. Normally distributed continuous variables were expressed as mean ± SD, and intergroup comparisons were performed using independent‐samples t‐test. Non‐normally distributed continuous variables were presented as median (interquartile range), and intergroup comparisons were conducted using the Mann–Whitney U test. Categorical variables were compared using the chi‐squared test or Fisher's exact test. Differences in indicators between the uncomplicated diabetes group and the NPDR group were analyzed with a significance threshold of p < 0.05.

Univariate and multivariate binary logistic regression analyses were performed with NPDR occurrence as the dependent variable to identify independent risk factors for NPDR. The stepwise regression method (Forward: LR) was used to optimize the model by excluding variables with no statistical significance. Collinearity between independent variables was evaluated by variance inflation factor (VIF) and tolerance (VIF <5 and tolerance >0.1 indicated no multicollinearity). Internal validation of the model was performed using bootstrap resampling (1000 repetitions) to assess the risk of overfitting, and 95% confidence intervals (95% CI) for regression coefficients of predictors were calculated. External validation was conducted using an independent validation set (126 cases).

The discriminative ability and predictive performance of the model for NPDR were visualized using the receiver operating characteristic (ROC) curve. The calibration of the model was evaluated by the Hosmer–Lemeshow goodness‐of‐fit test (p > 0.05 indicated good calibration).

3. RESULTS

3.1. Distribution of demographic and clinical indicators in the training set

First, we analyzed the baseline clinical data of the training set population, with results shown in Table 1. Significant differences were observed between the NPDR group and the diabetes group in diabetes duration (7.5 vs. 5.5 years, p = 0.008) and the incidence of macular oedema (20.3% vs. 10.7%, p = 0.022), suggesting that NPDR occurrence may be associated with diabetes duration and macular oedema incidence. No statistically significant differences were found in other baseline data between the two groups (all p > 0.05).

TABLE 1.

Comparison of baseline characteristics between the NPDR group and non‐NPDR group in the training set (n: Number of patients).

Item	NPDR group (n = 118)	Non‐NPDR group (n = 177)	t/Z/χ 2	p
Gender (Male/Female)	66/52	95/82	0.146	0.703
Age (years)	51 (35, 64)	50 (38, 62)	0.618	0.840
BMI (kg/m2)	24.71 ± 3.87	24.55 ± 3.50	0.384	0.701
Waist circumference (cm)	91.27 ± 8.07	91.16 ± 7.31	0.119	0.906
Hip circumference (cm)	98.14 ± 5.91	97.3 ± 5.51	1.044	0.297
Diabetes duration (years)	7.5 (4.2, 11.2)	5.5 (3.2, 8.7)	1.164	0.008**
Hypertension	13 (11.02%)	9 (5.08%)	2.185	0.139
Macular oedema	24 (20.34%)	19 (10.73%)	5.245	0.022*
Smoking history	17 (14.41%)	24 (13.56%)	0.042	0.837
Alcohol consumption history	17 (14.41%)	20 (11.30%)	0.623	0.430

Note: p < 0.05 *; p < 0.01**; p < 0.001***.

In diabetes‐related indicators (Table 2), the NPDR group exhibited higher fasting blood glucose (8.29 vs. 8.05 mmol/L, p = 0.015), HbA1c (63 [7.9%] vs. 58 [7.5%], p < 0.001) and triglycerides (TG) (2.05 vs. 1.96 mg/dL, p = 0.017), as well as lower high‐density lipoprotein cholesterol (HDL‐C) (1.17 vs. 1.24 mmol/L, p = 0.030), indicating associations between NPDR occurrence and glycemic indicators and lipid profiles. Additionally, compared with the diabetes group, the NPDR group had significantly higher UACR (89.85 vs. 36.07 mg/g, p < 0.001) and significantly lower eGFR (115.30 vs. 126.55 mL/min/1.73 m², p = 0.001), highlighting a close correlation between renal function and DR.

TABLE 2.

Comparison of diabetes‐related indicators between the NPDR group and non‐NPDR group in the training set (n: Number of patients).

Item	NPDR group (n = 118)	Non‐NPDR group (n = 177)	t/Z/χ 2	p
Fasting blood glucose (mmol/L)	8.29 (7.73, 9.64)	8.05 (7.45, 8.59)	1.569	0.015*
HbA1c (mmol/mol (%))	63 (7.9%) (57 (7.4%), 69 (8.5%))	58 (7.5%) (51 (6.8%), 64 (8.0%))	2.115	<0.001***
TC (mg/dL)	5.04 (4.58, 5.41)	5.01 (4.56, 5.48)	1.046	0.224
TG (mg/dL)	2.05 (1.71, 2.31)	1.96 (1.49, 2.33)	1.545	0.017*
LDL‐C (mg/dL)	3.16 (2.53, 3.55)	2.87 (2.45, 3.42)	1.141	0.148
HDL‐C (mg/dL)	1.17 (0.98, 1.42)	1.24 (1.07, 1.44)	1.450	0.030*
UACR (mg/g)	89.85 (60.04, 112.74)	36.07 (29.53, 44.40)	6.703	<0.001***
eGFR (mL/min/1.73 m2)	115.30 (95.49, 133.41)	126.55 (103.85, 145.83)	1.949	0.001

Note: p < 0.05 *; p < 0.01**; p < 0.001***.

3.2. Screening and effect quantification of independent predictors for NPDR

Factors with statistically significant differences between the NPDR group and the diabetes group were further included in univariate Logistic regression analysis, with results shown in Table 3. Diabetes duration (OR = 1.099), macular oedema (OR = 2.123), fasting blood glucose (OR = 2.146), HbA1c (OR = 2.361) and UACR (OR = 1.143) significantly increased the risk of DR, while eGFR exerted a protective effect (OR = 0.980). In contrast, TG and HDL‐C showed no significant association with NPDR occurrence (both p > 0.05).

TABLE 3.

Results of univariate Logistic regression for screening NPDR risk factors.

Factors	B	SEM	Wald χ 2	p	OR	95% CI
Diabetes duration	0.094	0.031	9.400	0.002	1.099	1.034, 1.167
Macular oedema (yes)	−0.753	0.334	5.095	0.024	2.123	1.104, 4.082
Fasting blood glucose	0.764	0.194	15.562	<0.001	2.146	1.468, 3.136
HbA1c	0.859	0.170	25.496	<0.001	2.361	1.691, 3.295
TG	0.421	0.261	2.595	0.107	1.523	0.913, 2.514
HDL‐C	−0.754	0.528	2.041	0.153	0.470	0.167, 1.324
UACR	0.134	0.020	46.981	<0.001	1.143	1.100, 1.188
eGFR	−0.020	0.005	14.825	<0.001	0.980	0.970, 0.990

Abbreviations: eGFR, estimated glomerular filtration rate; UACR, urinary albumin/creatinine ratio.

After excluding TG and HDL‐C, multivariate Logistic regression was performed to analyse the predictive effects of independent predictors for DR. As shown in Table 4, multivariate regression further confirmed macular oedema (OR = 3.448), fasting blood glucose (OR = 2.137), HbA1c (OR = 2.767) and UACR (OR = 1.150) as independent risk factors (all p < 0.05). In other words, diabetic patients with macular oedema had a significantly increased independent risk of NPDR by 2.448‐fold. Additionally, each 1‐unit increase in fasting blood glucose, HbA1c or UACR was associated with a significant independent increase in NPDR risk by 1.137‐fold, 1.767‐fold or 15% higher odds, respectively. Notably, the predictive significance of diabetes duration and eGFR disappeared in multivariate Logistic regression, suggesting that they are not independent factors for DR.

TABLE 4.

Predictive effects of independent predictors for NPDR in multivariate logistic regression.

Factors	B	SEM	Wald χ 2	p	OR	95% CI
Diabetes duration	0.047	0.057	0.685	0.408	1.048	0.937, 1.173
Macular oedema (yes)	1.238	0.599	4.264	0.039	3.448	1.065, 11.161
Fasting blood glucose	0.759	0.361	4.435	0.035	2.137	1.054, 4.331
HbA1c	1.018	0.323	9.932	0.002	2.767	1.469, 5.211
UACR	0.140	0.022	38.686	<0.001	1.150	1.101, 1.202
eGFR	−0.015	0.010	2.416	0.120	0.985	0.966, 1.004

Abbreviations: eGFR, estimated glomerular filtration rate; UACR, urinary albumin/creatinine ratio.

3.3. Construction and performance evaluation of the NPDR logistic regression prediction model

In multivariate Logistic regression, diabetes duration and eGFR showed no significance. Therefore, we further simplified the model using the stepwise regression method (Forward: LR). As shown in Table 5, stepwise regression confirmed macular oedema (OR = 3.247), fasting blood glucose (OR = 2.194), HbA1c (OR = 2.799) and UACR (OR = 1.153) as independent risk factors for NPDR (all p < 0.05). The final Logistic regression equation was: Logit(P) = −22.09 + 1.178 × macular oedema + 0.786 × fasting blood glucose + 1.029 × HbA1c + 0.143 × UACR.

TABLE 5.

Logistic regression prediction model constructed by stepwise regression.

Factor	B	SEM	Wald χ 2	p	OR	95% CI
Macular oedema (yes)	1.178	0.583	4.078	0.043	3.247	1.035, 10.183
Fasting blood glucose	0.786	0.359	4.791	0.029	2.194	1.086, 4.433
HbA1c	1.029	0.315	10.700	0.001	2.799	1.511, 5.185
UACR	0.143	0.022	41.828	<0.001	1.153	1.104, 1.204

Abbreviation: UACR, urinary albumin/creatinine ratio.

Collinearity diagnosis was performed to analyse collinearity between variables. As shown in Table S1, all independent variables had VIF values close to 1 and tolerance >0.1, indicating no multicollinearity. The predictive performance of the Logistic regression model for NPDR in diabetic patients was evaluated using the ROC curve with NPDR occurrence as the dependent variable. As shown in Figure 1a, the area under the ROC curve (AUC) was 0.949 (0.921–0.977), with a Youden index of 0.819, sensitivity of 86.4% and specificity of 95.5%, indicating good discriminative ability of the Logistic regression model for NPDR prediction.

FIGURE 1.

Predictive efficacy of the Logistic regression prediction model. (a) Receiver operating characteristic curve; (b) Calibration curve.

The calibration ability of the prediction model was evaluated using the Hosmer–Lemeshow goodness‐of‐fit test. As shown in Figure 1b, the Hosmer–Lemeshow χ ² = 8.821, p = 0.358 > 0.05, suggesting no statistically significant difference between the predicted values and actual observations, indicating good calibration of the prediction model.

3.4. Evaluation of overfitting risk in the logistic regression prediction model by bootstrap resampling

Bootstrap resampling (1000 times) was further used to evaluate the internal validation performance of the model. As shown in Table 6, the 95% CIs of regression coefficients for HbA1c (Bootstrap 95% CI: 0.377–1.893) and UACR (Bootstrap 95% CI: 0.114–0.200) did not include 0, indicating that these two variables are robust predictors. However, although macular oedema and fasting blood glucose had small Bootstrap biases and significant p‐values, their 95% CIs included 0 (−0.043 to 2.463 and −0.011 to 1.706, respectively), suggesting unstable effects and potential sampling variation risks. The overall performance attenuation of the model was only 5.4% (original AUC 0.949 → Bootstrap mean 0.898), indicating controllable overfitting risk.

TABLE 6.

Validation results of predictor stability by Bootstrap resampling.

Factor	B	Bia	Bootstrap SEM	p	95% CI
Macular oedema (yes)	1.178	0.011	0.650	0.036	−0.043, 2.463
Fasting blood glucose	0.786	0.034	0.420	0.046	−0.011, 1.706
HbA1c	1.029	0.053	0.374	0.003	0.377, 1.893
UACR	0.143	0.006	0.021	0.001	0.114, 0.200

3.5. External validation of discriminative ability and calibration of the logistic regression prediction model in an independent cohort

A total of 126 samples from an independent validation set for the Logistic regression prediction model were collected according to the same inclusion and exclusion criteria. The performance of the Logistic regression model for NPDR prediction in the validation set at the same cutoff value (0.392) was analyzed using the ROC curve. As shown in Figure 2a, the Logistic prediction model exhibited excellent generalizability and clinical utility. The ROC curve of the validation set showed a steep upward trend, with an AUC slightly lower than that of the training set (0.918 vs. 0.949).

FIGURE 2.

Validation set validation of the Logistic regression prediction model. (a) Receiver operating characteristic curve; (b) Calibration curve

The confusion matrix in Table 7 showed that the overall accuracy of the Logistic regression prediction model was 86.5% (95% CI: 0.793–0.919). The sensitivity was 87.8% (95% CI: 0.807–0.949), indicating a strong ability to detect DR‐positive cases. Additionally, the specificity was 84.1% (95% CI: 0.733–0.949), showing reliable performance in excluding non‐NPDR populations. Notably, the positive predictive value (PPV) was as high as 91.1% (95% CI: 0.849–0.974), meaning that over 90% of patients identified as high‐risk by the model were truly affected.

TABLE 7.

Confusion matrix of the model in the training set and validation set.

Data	Accuracy (95% CI)	Sensitivity (95% CI)	Specificity (95% CI)	PPV (95% CI)	NPV (95% CI)
Training	0.919 (0.881–0.947)	0.955 (0.924–0.985)	0.864 (0.803–0.926)	0.914 (0.873–0.954)	0.927 (0.879–0.976)
Validation	0.865 (0.793–0.919)	0.878 (0.807–0.949)	0.841 (0.733–0.949)	0.911 (0.849–0.974)	0.787 (0.670–0.904)

Abbreviations: NPV, negative predictive value; PPV, positive predictive value.

The calibration ability of the prediction model was evaluated using the Hosmer–Lemeshow goodness‐of‐fit test. As shown in Figure 2b, the Hosmer–Lemeshow test results (χ ² = 9.769, p = 0.282 > 0.05) confirmed high consistency between the predicted probabilities and actual NPDR incidence. No statistically significant difference was found between the predicted values and actual observations, indicating good calibration of the prediction model.

4. DISCUSSION

This study successfully identified core independent risk factors for DR using a Logistic regression model and constructed a high‐performance predictive tool. In the univariate analysis of the modelling set (295 cases), significant differences were observed between the NPDR group and non‐NPDR group in indicators such as diabetes duration, macular oedema, fasting blood glucose, HbA1c, UACR and eGFR (all p < 0.05). Multivariate Logistic regression further confirmed macular oedema (OR = 3.247), fasting blood glucose (OR = 2.194), HbA1c (OR = 2.799) and UACR (OR = 1.153) as independent risk factors. These results quantify the biological mechanisms underlying NPDR risk: for instance, each 1‐unit increase in HbA1c elevates NPDR risk by nearly 2.8‐fold, while each 1 mg/g increase in UACR raises the risk by 15.3%. For model construction, the stepwise regression method optimized the equation as Logit(P) = −22.09 + 1.178 × macular oedema + 0.786 × fasting blood glucose + 1.029 × HbA1c + 0.143 × UACR. Collinearity diagnosis (with VIF values close to 1) ensured the independence of variables. The model exhibited excellent discriminative ability in the training set, with an AUC of 0.949, sensitivity of 86.4% and specificity of 95.5%, indicating its effectiveness in distinguishing high‐risk from low‐risk patients.

The findings of this study are highly consistent with global research on NPDR epidemiology and risk factors, while highlighting the advantage of clinical interpretability offered by the Logistic regression model. The global prevalence of NPDR is approximately 27%, and it is closely associated with metabolic indicators such as diabetes duration and HbA1c, ³ which aligns with our conclusion that fasting blood glucose and HbA1c serve as independent predictors. A study noted that HbA1c variability is an independent risk factor for DR, particularly in patients with type 2 diabetes, where greater HbA1c variability correlates positively with the incidence of retinopathy. ²⁰ Additionally, Arnqvist et al. emphasized that long‐term HbA1c control is closely linked to the severity of DR, recommending maintaining HbA1c at a low level to reduce NPDR occurrence. ²¹ As the ‘gold standard’ for long‐term glycemic control, elevated HbA1c (OR = 2.799) directly reflects the glycotoxic damage of chronic hyperglycemia to microvascular endothelial cells, serving as a core driver of NPDR pathogenesis. Moreover, fasting blood glucose (OR = 2.194) may reflect immediate glycemic fluctuations or poor control. Studies have shown that fasting blood glucose variability is significantly associated with the occurrence and progression of DR, especially in patients with type 2 diabetes, where high variability in fasting blood glucose substantially increases the risk of retinopathy. ²² These findings underscore the importance of monitoring not only average blood glucose levels but also glycemic stability in diabetes management to better predict and prevent DR.

DR is a common microvascular complication in diabetic patients, and macular oedema is one of the primary causes of vision loss. The results of this study indicate that macular oedema (OR = 3.247), as a severe complication or early manifestation of DR, intuitively reflects that local pathological changes in NPDR itself are important risk signals for its progression. Consistently, previous studies have demonstrated that diabetic macular oedema (DME) is closely associated with the severity of NPDR and can serve as an indicator for predicting NPDR progression. ²² Interestingly, research has found that the severity of diabetic nephropathy (DN) is significantly correlated with the development of NPDR and macular oedema, and declining renal function can predict the occurrence of NPDR, ²³ which aligns with our findings. In this study, UACR (OR = 1.153), as a marker of glomerular endothelial injury, showed a significant association, highlighting the existence of a diabetic kidney‐retinopathy axis and suggesting a key role of systemic vascular endothelial dysfunction in NPDR pathogenesis. Song et al. revealed that elevated urinary microalbumin levels are positively correlated with the severity of NPDR. ²⁴ Furthermore, the combination of urinary microalbumin and low haemoglobin levels has been identified as a strong predictor of NPDR and its complications, ²⁵ consistent with our results. Notably, although univariate analysis showed that elevated TG and reduced HDL‐C were associated with NPDR (p < 0.05), their predictive significance disappeared in the multivariate model. This differs from other studies exploring the relationship between atherosclerotic risk scores and NPDR, ⁹ suggesting that the independent contribution of blood lipids may be weakened or involve more complex interactions when considering blood glucose, renal function and macular oedema status.

Rigorous internal and external validation of the Logistic regression prediction model was conducted in this study. Internal validation via Bootstrap resampling (1000 iterations) showed that HbA1c and UACR are highly robust factors in the model. Despite potential sampling variability in macular oedema and fasting blood glucose, the overall performance attenuation of the model was limited (original AUC 0.949 → Bootstrap mean 0.898), indicating controllable overfitting risk. More importantly, the model maintained excellent generalizability in the independent validation set (126 cases). The Logistic regression model achieved an AUC of 0.918 in the validation set, with an overall accuracy of 86.5%, sensitivity of 87.8% and specificity of 84.1%. Additionally, the high PPV (91.1%) in the independent cohort is particularly noteworthy, meaning over 90% of patients identified as high‐risk by the model were indeed affected. This performance, superior to many studies that only report AUC or sensitivity, holds significant value for guiding clinical decisions (e.g., referral for specialized screening). ¹³ Another core advantage of the model lies in its calibration (predicted risk ≈ actual risk). The Hosmer–Lemeshow test (χ ² = 9.769, p = 0.282) reaffirmed the model's good calibration, with predicted probabilities highly consistent with actual risks. This feature is crucial for clinical stratified intervention, as it avoids the risks of over‐medicalization or missed diagnoses. By accurately quantifying individual risks, the model facilitates risk‐stratified management and optimizes the allocation of screening resources. Furthermore, compared to complex machine learning models, this study achieved comparable predictive performance (AUC >0.9) with a simplified set of variables (only four core indicators), avoiding the ‘black box’ issue and enabling more intuitive quantification of risk effects.

Despite constructing a high‐performance Logistic regression prediction model, this study has several key limitations that need to be addressed in future research: ① Limitations in sample representativeness and generalizability: The study is based on retrospective data from a single medical centre with a limited sample size, which may restrict the model's generalizability to other regions or populations and reduce its applicability in diverse social contexts. ② Macular oedema and fasting blood glucose exhibited potential sampling variability in internal Bootstrap resampling. Although overall overfitting was controllable (AUC attenuation of only 5.4%), the instability of these two variables may affect the reliability of clinical applications. Specifically, the bootstrap 95% CIs for their regression coefficients included zero, indicating that their effects are less stable than those of HbA1c and UACR, which should be considered when interpreting the model's reliance on these specific variables. ③ While the paper emphasizes the ‘black box’ issue of machine learning models, it did not compare performance with such models in the same cohort, making it difficult to fully demonstrate the relative advantages of the simplified model. ④ Although the model's high PPV (91.1%) facilitates the identification of high‐risk patients, the NPV in the validation set was only 78.7%, indicating a potential risk of missed diagnoses. Future studies should expand sample diversity, prospectively include multi‐centre, multi‐regional cohorts, integrate sociodemographic and comprehensive metabolic indicators, and further improve the prediction model.

5. CONCLUSION

In conclusion, this study successfully constructed and validated a Logistic regression‐based NPDR risk prediction model. By integrating four core clinical indicators—macular oedema, fasting blood glucose, HbA1c and UACR—the model achieves high‐precision prediction and clinical interpretability. It effectively addresses the ‘black box’ problem of complex algorithms, providing a highly efficient and intuitive tool for primary medical institutions to screen high‐risk NPDR patients. Future work could focus on integrating this model's formula into electronic health record systems to generate automated risk flags for clinicians, thereby facilitating point‐of‐care decision support. This study can serve as a foundational framework for subsequent research, which may incorporate more comprehensive clinical data to build a more holistic risk prediction system, including the development of mobile risk assessment tools or integration with AI screening systems to achieve precise prevention of DR.

AUTHOR CONTRIBUTIONS

Jiaoyan Zhang and Xiaodong Cao conducted the experiments, analyzed the data and drafted the manuscript. Xin Yu, Minfeng Jiang, Tianhua Xie, Yangningzhi Wang provided the resource and made critical revisions on the manuscript. Zhengyuan Tang supervised and funded the project. All authors reviewed and approved the final version of the manuscript.

FUNDING INFORMATION

Youth Project of Wuxi Nursing Association (Q202301).

CONFLICT OF INTEREST STATEMENT

The authors declare no conflict of interest.

ETHICS STATEMENT

The research related to human use has been complied with all the relevant national regulations, institutional policies and adheres to the Declaration of Helsinki, and has been approved by the ethical committee of The Affiliated Wuxi People's Hospital of Nanjing Medical University.

Supporting information

Table S1.

DATA AVAILABILITY STATEMENT

Data included in article/supporting information/referenced in article.