Merging artificial intelligence and retinal vascular geometric parameters: A novel tool for diagnosis and prognosis prediction of diabetic nephropathy

Abstract

Background

Diabetic nephropathy (DN) is the leading cause of end-stage renal disease. The retinal microvasculature, as the only directly observable microvasculature, may reflect DN progression. This study aims to construct a non-invasive diagnostic and prognostic prediction model using the mixed effects of retinal vascular geometric parameters and clinical data.

Methods

We constructed a multimodal database including 397 patients with type 2 diabetes and chronic kidney disease from multiple centers in China. The primary cohort (374 patients) was recruited from the Department of Nephrology at the First Medical Center of the Chinese People’s Liberation Army General Hospital in Beijing between 2017 and 2022, while an external validation cohort (23 patients) was collected from five other hospitals across China from September 2022 to March 2023. Fundus images, clinical characteristics, renal biopsy diagnoses, and follow-up data were collected. Unsupervised learning and Resnet neural networks were used to segment and calculate retinal vascular geometric parameters. Weighted quantile regression (WQS), Lasso, and COX univariable regressions were employed to assess the mixed effects of retinal vascular geometric parameters and select relevant clinical characteristics. Logistic regression and Cox regression with random forest (COX-RF) were used for model construction.

Results

A multimodal database of 397 patients was constructed. A diagnostic model combining retinal parameters (WQS-diagnosis) and seven clinical characteristics achieved superior performance: AUC 0.98, accuracy 0.92 on the test set, outperforming models using only clinical data (AUC 0.93) or two retinal parameters plus clinical data (AUC 0.94). On the multi-centre validation set, the model maintained accuracy 0.91 and AUC 0.95. For prognosis, a mixed-effects parameter (WQS-prognosis) was derived; the COX-RF model achieved an AUC of 0.88.

Conclusions

Retinal microvasculature are effective biomarkers for DN. The proposed non-invasive model demonstrated high accuracy and generalizability, offering a valuable tool for optimising DN management.

Introduction

There are over 400 million people worldwide with type 2 diabetes mellitus (T2DM), approximately 50% of whom have chronic kidney disease (CKD).^[1,2] Patients with CKD and T2DM can develop diabetic nephropathy (DN), non-diabetic renal disease (NDRD) occurring alongside diabetes mellitus, or a combination of both DN and NDRD.^[3] These conditions can only be reliably distinguished through a kidney biopsy.^[4] However, in clinical practice, the invasive nature of renal biopsy and the associated risks, such as bleeding and infection, often prevent its use, particularly in medically underserved areas. Consequently, clinicians frequently diagnose DN in patients with diabetes and CKD based on the Kidney Disease Outcomes Quality Initiative (KDOQI) guidelines without performing a biopsy.^[5] This practice can result in misdiagnosis, where patients with NDRD or a combination of NDRD and DN are incorrectly identified as having DN. Misdiagnosis not only delays appropriate treatment but negatively impacts prognosis due to significant differences in treatment strategies between DN and NDRD.^[6] It also affects the accurate calculation of DN prevalence and the rational allocation of medical expenses and healthcare resources.^[7] Therefore, developing a non-invasive diagnostic model for accurately diagnosing DN and NDRD is crucial.

DN is one of the leading microvascular complications of diabetes.^[8] Changes in the morphology of glomerular capillaries, leading to hemodynamic alterations, are essential mechanisms in the onset and progression of DN.^[9] Therefore, we speculate that microvascular morphology is an important basis for the diagnosis and prognosis of DN. Our preliminary studies have shown that retinal vascular geometric parameters can assist in distinguishing between DN and NDRD.^[10] The retinal vessels, part of the body’s microcirculation system, enable the rapid and non-invasive observation of microcirculatory changes similar to those occurring in the glomeruli.^[11,12] The correlation between retinal vascular geometric parameters and renal damage, diabetic complications such as diabetic retinopathy (DR), and cardiovascular diseases further confirms our hypothesis.^[13–15] However, previous studies have analysed the retinal vascular geometric parameters independently. Although there is collinearity among these parameters, individual parameters cannot represent the entire morphology of retinal microvessels. Therefore, the mixed effects of retinal vascular geometric parameters deserve further exploration in the diagnosis and prognosis prediction of DN.

To address these issues [Figure 1], the study aims to provide a clinically applicable tool that can accurately distinguish DN from NDRD without the need for kidney biopsy, thereby facilitating timely and appropriate treatment decisions, reducing misdiagnosis, and ultimately improving the quality of life and prognosis of patients with diabetes.

Figure 1.

Overview of the study design and workflow. (A) Research mechanism: hypothesis that retinal microvascular geometric parameters reflect glomerular microvascular changes in DN; (B) Data collection from multiple centers, including fundus images, clinical data, renal biopsy pathology, and diagnoses; (C) Retinal microvascular segmentation and geometric parameter calculation; (D) Development and validation of non-invasive diagnostic and prognostic prediction models for DN. AVR: Arteriolar‑to‑venular diameter ratio; BA: Branching angle; CB: Central arteriolar branch; CBa: Arteriolar central branch; CBv: Venular central branch; DN: Diabetic nephropathy; DR: Diabetic retinopathy; eGFR: Estimated glomerular filtration rate; ESRD: End‑stage renal disease; FD: Fractal dimension; FDa: Arteriolar fractal dimension; FDv: Venular fractal dimension; HbA1c: Glycated haemoglobin; LDL‑C: Low‑density lipoprotein cholesterol; NDRD: Non‑diabetic renal disease; NON: Non-ESRD; PLA: Perimeter‑to‑length aspect; SBP: Systolic blood pressure; TD: Tortuosity density; TDa: Arteriolar tortuosity density; TDv: Venular tortuosity density; TOR: Tortuosity; TORa: Arteriolar tortuosity; TORv: Venular tortuosity.

Methods

Ethics approval and consent to participate

This study adhered to the Declaration of Helsinki and was approved by the Ethics Committee of the Chinese People’s Liberation Army (PLA) General Hospital (S2017-133-01). All participants signed informed consent forms.

Inclusion and exclusion criteria

All datasets were included or excluded based on the same criteria. The inclusion criteria were as follows: (1) Patients with T2DM and CKD aged over 18 years; (2) Patients diagnosed with DN, NDRD, or a combination of both through renal biopsy; (3) Availability of retinal images. The exclusion criteria were as follows: (1) Undergoing kidney replacement therapy, such as hemodialysis, peritoneal dialysis, or kidney transplantation; (2) Pregnancy or malignancy; (3) Poor-quality fundus images, including blurred images or large dark shadows; (4) Absence of crucial clinical information, such as patient demographics and biochemical results; and (5) DN without follow-up data.

Notably, this study categorised patients with a combination of DN and NDRD as having NDRD due to the need for individualised treatment.

Datasets

DKD-microvessels diagnosis dataset

According to the inclusion and exclusion criteria, 374 patients with T2DM and CKD admitted to the Department of Nephrology at the First Medical Centre of Chinese PLA General Hospital from 2017 to 2022 for kidney biopsy were included in the diabetic kidney disease (DKD)-microvessels diagnosis dataset. These patients were then divided into training and test sets at a ratio of 4:1. A flowchart illustrating the inclusion and exclusion is shown in Figure 2.

Figure 2.

Flowchart of patient inclusion and exclusion criteria. CKD: Chronic kidney disease; DN: Diabetic nephropathy; NDRD: Non‑diabetic renal disease; T2DM: Type 2 diabetes mellitus.

Multicentre diagnosis dataset

To evaluate the generalisability of our model, the algorithm was applied to a multicentre dataset. This dataset included 23 patients from five other large hospitals in China between September 2022 and March 2023. The hospitals involved were Daping Hospital in Chongqing City, The First Affiliated Hospital of Dalian Medical University in Dalian City, Beijing Tongren Hospital, The Third Medical Centre of Chinese PLA General Hospital, and The Fourth Medical Centre of Chinese PLA General Hospital.

DN-microvessels prognosis dataset

To explore the impact of retinal vascular geometric parameters on the prognosis of DN, we collected follow-up data from patients with DN from the DKD-microvessel diagnosis dataset. A total of 139 patients with DN were followed up until December 2023, constituting the DN-microvessels prognosis dataset. The dataset was subsequently divided into training and testing sets in a 4:1 ratio. The primary endpoints were death or end-stage renal disease (ESRD).

Clinical characteristics and retinal fundus image collection

The demographic and clinical data of all patients from the three datasets were collected. Baseline characteristics and clinical parameters included sex, age, body mass index (BMI), blood pressure (BP), duration of diabetes, smoking history, drinking history, and DR. Biochemical results included haemoglobin (Hb), glycated hemoglobin (HbA1c), serum creatinine (Scr), albumin (ALB), estimated glomerular filtration rate (eGFR, calculated using the CKD-EPI formula), cystatin C, total cholesterol (TC), triglycerides (TG), high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), blood uric acid (UA), 24-hour proteinuria, and renal biopsy results (DN or NDRD). Additionally, we used a KOWA retinal camera (VX-20, nonmyd 7) to capture 45° field-of-view retinal images of the patient’s left and right eyes. The fundus images included the optic disc and macular area. The right eye retinal image was preferred for measurement; if the right eye image could not be obtained, or was unclear, the left eye image was used.

Unsupervised approaches for retinal vascular segmentation

We utilised unsupervised methods for the segmentation of retinal blood vessels. Unsupervised approaches do not require hand-drawn segmentations to train a classification model; instead, they rely on mathematical transformations of the raw image to identify edges corresponding to vessel walls, followed by post-processing steps such as inpainting to fill gaps in the segmented vasculature. The specific methods employed are as follows: First, we extracted the green channel image from the retinal image. Second, the green channel image was enhanced using image processing techniques, including Gaussian filtering, histogram equalisation, and adaptive histogram equalisation. Third, through adaptive thresholding, we extracted the edges of the green channel image or its enhanced version, followed by filling in the interior of these edges to obtain the segmented vascular image. Then, post-segmentation, the connected domain images were classified into single and multiple vessels. This process involved annotating single and multiple vessel images; constructing training, testing, and validation sets; configuring the ResNet152 neural network algorithm; adjusting algorithm hyperparameters; training the neural network classification model for single and multiple vessels with an accuracy of 91%; and then applying the trained classification model to classify all vessel images to distinguish between single and multiple vessel images. Lastly, for artery and vein classification after segmentation, the CLAHE algorithm was applied to enhance each of the three different channel images separately, then stacking them to create a vascular-enhanced image. Various features were extracted, including the average grayscale of the initial retinal image’s vascular region and the maximum, median, average, and minimum grayscale of the enhanced vascular image. Additional features were also obtained, such as the maximum pixel values of the hue, saturation, and value channels, median pixel values of the hue, saturation, and value channels. A dataset was built, a machine learning algorithm program was then set up, reasonable hyperparameters were configured, and machine learning classification models were trained using the obtained dataset. The accuracies achieved by different methods were as follows: random forest: 87%, XGBoost: 84%, AdaBoost: 85%, LightGBM: 89%, and ResNet152: 91%.

Calculation of retinal vessel geometric parameters and feature selection

After segmenting the retinal vascular tree, we calculated the following parameters from five dimensions [Supplementary Figure 1, http://links.lww.com/CM9/C958]: diameter (D), arteriolar diameter (Da), venular diameter (Dv), fractal dimension variable k (Fk), fractal dimension variable b (Fb), arteriolar fractal dimension variable k (Fka), arteriolar fractal dimension variable b (Fba), venular fractal dimension variable k (Fkv), venular fractal dimension variable b (Fbv), distance measure (DM), arteriolar distance measure (DMa), venular distance measure (DMv), tortuosity density (TD), arteriolar tortuosity density (TDa), venular tortuosity density (TDv), branch angle (B). The specific calculation formulas for these parameters are detailed in the published article.^[10] Based on existing research and Lasso regression, seven clinical variables that significantly impacted the diagnosis of DN were selected: LDL, diabetes duration, age, SBP, HbA1c, DR, and eGFR. For prognostic prediction of DN, using COX univariable regression, 11 clinical features were selected: Cystatin C, 24-hour urine protein, ALB, Hb, eGFR, DBP, SBP, DR, TG, WQS-prognosis, and LDL [Supplementary Tables 1–2, http://links.lww.com/CM9/C958].

WQS

Due to multicollinearity among retinal vessel geometric parameters [Supplementary Table 3, http://links.lww.com/CM9/C958], we employed WQS regression to assess the overall effect of these parameters on the diagnosis and prognosis of DN. This method is effective in characterising mixed effects. The R package gWQS (version 3.0.5, Comprehensive R Archive Network, Vienna, Austria) was used to calculate the WQS index, which is the weighted sum of individual parameters based on experience. The WQS index (ranging from 0 to 1) represents the mixed-effect level of the parameters, with non-negligible weights indicating parameters of interest. The final result was interpreted as the synchronous effect of adding one quartile to the mixed parameters on the diagnosis and prognosis of DN.

Model development and validation

Multi-dimensional diagnosis model comparison

In diagnosing DN, we used machine learning algorithms such as Support Vector Machine (SVM), random forest (RF), Logistic Regression, Lasso, ElasticNet, XGBoost, and LightGBM to develop and evaluate model performance. These models were trained and tested on the DKD-microvessels diagnosis dataset, with the model demonstrating the highest diagnostic performance being selected and validated using the Multicenter diagnosis dataset. SHAP was utilised to visualise the models.

Diagnosis model ablation study

We conducted an ablation study to validate the effectiveness of the two modalities, i.e., retinal vessel geometric parameters and clinical factors, in classifying DN and NDRD. Specifically, we removed the input of the retinal vessel geometric parameters and conducted disease classification experiments. We also used retinal vessel geometric parameters to classify DN and NDRD. The model performance was validated using the Multicentre diagnosis dataset.

Multi-dimensional prognosis model comparison

The Cox proportional hazards model is a statistical method commonly used in survival analysis. However, it has limitations when analysing high-dimensional data and dealing with collinearity. To address these challenges, we combined machine learning with the Cox proportional hazards model for the prognostic prediction of DN. We employed various methods, including COX_RF, COX_Ridge, COX_Lasso, Cox Test Estimation with Least Squares Tree (Cox Test Est LS Tree), Cox Test Estimation with Conditional Inference Tree (Cox Test Est CPH Tree), and Cox Test Estimation with ElasticNet to address high-dimensional data, multicollinearity issues, non-linear effects, and interactions. We trained and tested the models using the DKD-microvessels prognosis dataset and selected the model with the optimal prognostic performance. We used SHAP to visualise the models.

Performance evaluation for disease classification and statistical analysis

We evaluated the performance of DKD classification in terms of accuracy, sensitivity, specificity, F1-score and area under the receiver operating characteristic (ROC) curve (AUC). All analyses were conducted using R and Python (version 3, Python Software Foundation, Wilmington, USA). We compared demographic characteristics between DN and NDRD status groups using the chi-squared test for categorical variables, the t-test for normal continuous variables, and the Mann–Whitney U test for non-normal continuous variables. We used Kaplan–Meier analysis to describe the survival of patients with DN. P <0.05 was considered statistically significant.

Results

Patients’ data

We used the Retina-DKD dataset for model training and testing, which included data from 374 patients with DKD who underwent kidney biopsy, fundus imaging, and blood tests at the Chinese PLA General Hospital. We collected clinical factors [Table 1] for each patient and selected eight factors (LDL, diabetic courses, age, SBP, HbA1c, DR, eGFR, and WQS diagnosis) for diagnosis model development. Supplementary Table 1, http://links.lww.com/CM9/C958 presents the feature importance rankings from the Lasso regression. Based on these rankings, combined with clinical expertise and variables from previous models, we selected the top eight variables as model features. As shown in Table 1, the median age of patients was 53.0 (46.7–59.0) years, with 270 (72.2%) being male. We also included 23 multicentre patients in the multi-centre validation set, and their basic information is also presented in Table 1.

Table 1.

Baseline clinical and retinal vascular geometric characteristics of the study population across the training, test, and multi-centre validation sets.

Characteristics	Total (n = 374)	Training set (n = 299)	Test set (n = 75)	Multi-centre validation set (n = 23)
Age, years	53.0 (46.7–59.0)	53.0 (46.0–59.0)	53.0 (47.0–58.0)	54.0 (47.0–63.0)
Height, cm	170.0 (164.0–174.3)	170.0 (164.0–175.0)	170.0 (160.0–174.0)	161.0 (158.0–172.0)
Weight, kg	76.0 (67.9–84.0)	75.5 (68.0–84.0)	77.2 (66.0–84.0)	70.0 (59.0–75.8)
BMI, kg/m2	26.5 (24.5–28.7)	26.5 (24.6–28.7)	26.7 (24.5–29.2)	25.0 (23.8–27.8)
SBP, mmHg	140.0 (126.0–153.0)	140.0 (125.0–153.0)	139.0 (129.0–150.0)	135.0 (122.0–156.0)
DBP, mmHg	81.0 (73.0–90.0)	80.0 (73.0–90.0)	83.0 (77.0–90.0)	79.0 (73.0–90.0)
Hypertension duration, months	24.0 (2.0–120.0)	24.0 (1.0–120.0)	24.0 (3.0–120.0)	72.0 (9.0–132.0)
Diabetes duration, months	84.0 (24.0–168.0)	96.0 (24.0–156.0)	72.0 (24.0–180.0)	96.0 (36.0–204.0)
Sex, Male, n (%)	270 (72.2)	219 (73.2)	51 (68.0)	13(56.5%)
Smoking, n (%)	168 (44.9)	134 (44.8)	34 (45.3)	–
Drinking, n (%)	152 (40.6)	122 (40.8)	30 (40.0)	–
DR, n (%)	184 (49.2)	152 (50.8)	32 (42.7)	12 (52.2%)
DN, n (%)	139 (37.2)	112 (37.5)	27 (36.0)	11 (47.8%)
Hb, g/L	123.2 ± 22.9	123.5 ± 22.6	121.8 ± 24.3	121.8 ± 25.9
HbA1c, %	6.5 (6.0–7.6)	6.6 (6.0–7.6)	6.5 (6.0–7.5)	7.2 (6.5–7.9)
Urine protein quantitation, g/L	1.8 (0.7–2.9)	1.7 (0.6–2.8)	2.1 (0.9–3.2)	–
Urine Volume, L	2.1 (1.6–2.7)	2.1 (1.6–2.7)	1.9 (1.4–2.6)	–
Proteinuria, g/24h	3.2 (1.3–5.6)	3.1 (1.3–5.6)	3.8 (1.7–6.6)	4.1 (1.1–7.8)
ALB, g/L	34.6 (27.3–39.7)	35.0 (27.4–40.3)	32.1 (26.5–38.4)	31.6 (24.4–42.3)
UN, mmol/L	7.7 (5.7–11.1)	7.6 (5.8–10.9)	7.7 (5.4–12.2)	–
Scr, μmol/L	112.2 (81.8–169.2)	112.2 (82.1–168.2)	112.2 (80.5–200.6)	108.7 (89.0–135.0)
UA, μmol/L	363.4 (301.4–422.4)	364.3 (299.4–422.6)	360.8 (312.2–420.8)	391.6 (334.2–417.0)
eGFR, mL·min–1·1.73 m–2	49.3 (27.8–75.7)	49.2 (29.7–75.3)	49.3 (24.9–78.3)	57.6 (39.6–71.7)
Cystatin C, mg/L	1.5 (1.1–2.1)	1.5 (1.1–2.1)	1.5 (1.1–2.3)	–
TC, mmol/L	4.7 (3.7–5.8)	4.6 (3.7–5.7)	5.1 (3.9–6.0)	4.8 (4.0–6.6)
TG, mmol/L	2.0 (1.4–2.8)	2.0 (1.4–2.8)	2.1 (1.4–3.0)	2.0 (1.2–3.1)
HDL-C, mmol/L	2.8 (2.1–3.9)	2.8 (2.1–3.8)	3.1 (2.2–4.3)	1.1 (1.0–1.3)
LDL-C, mmol/L	1.0 (0.8–1.2)	1.0 (0.8–1.2)	1.03 (0.8–1.2)	3.0 (2.3–4.5)
D	16.3 (13.0–19.5)	16.3 (12.9–19.4)	16.1 (13.0–21.3)	21.0 (18.7–24.9)
Da	12.9 (9.1–16.6)	12.7 (9.1–16.6)	13.6 (9.7–17.4)	20.2 (17.6–22.5)
Dv	19.0 (15.1–23.1)	19.0 (15.1–23.0)	20.1 (16.0–23.7)	25.1 (19.69–27.4)
Fk	1.6 (1.5–1.6)	1.6 (1.5–1.6)	1.6 (1.5–1.6)	1.6 (1.5–1.6)
Fb	0.7 ± 0.1	0.7 ± 0.1	0.7 ± 0.1	0.7 ± 0.1
Fka	1.4 (1.3–1.4)	1.4 (1.3–1.4)	1.4 (1.3–1.4)	1.5 (1.4–1.5)
Fba	1.0 ± 0.1	1.0 ± 0.1	1.0 ± 0.1	0.9 ± 0.1
Fkv	1.5 (1.4–1.5)	1.4 (1.4–1.5)	1.5 (1.3–1.5)	1.5 (1.5–1.6)
Fbv	0.9 ± 0.1	0.9 ± 0.1	0.9 ± 0.2	0.8 ± 0.1
DM	1.0 (1.0–1.0)	1.0 (1.0–1.0)	1.0 (1.0–1.0)	1.0 (1.0–1.1)
DMa	1.0 (1.0–1.0)	1.0 (1.0–1.0)	1.0 (1.0–1.0)	1.0 (1.0–1.1)
DMv	1.0 (1.0–1.0)	1.0 (1.0–1.0)	1.0 (1.0–1.0)	1.0 (1.0–1.0)
TD, ×10–4	5.0 (3.6–6.8)	5.1 (3.6–6.9)	4.6 (3.5–6.6)	10.7 (7.5–16.9)
TDa, ×10–4	5.2 (3.2–8.6)	5.3 (3.4–8.6)	4.7 (2.9–8.4)	13.9 (9.3–18.6)
TDv, ×10–4	4.1 (2.7–6.1)	4.1 (2.8–6.0)	4.0 (2.6–6.5)	9.5 (5.1–9.9)
B	61.7 (51.1–73.1)	62.7 (51.5–75.4)	60.7 (48.6–67.1)	62.9 (52.5–68.5)
WQS-diagnosis	1.5 (1.1–1.9)	1.5 (1.1–1.9)	1.5 (1.1–1.9)	1.4 (1.2–1.6)

Data are presented as mean ± standard deviation (SD) for normally distributed continuous variables, median (interquartile range, IQR) for non-normally distributed continuous variables, and n (%) for categorical variables. “–” denotes the absence of data for this metric. ALB: Albumin; B: Branch angle; BMI: Body mass index; D: Diameter; Da: Arteriolar diameter; DBP: Diastolic blood pressure; DM: Distance measure; DMa: Arteriolar distance measure; DMv: Venular distance measure; DN: Diabetic nephropathy; DR: Diabetic retinopathy; Dv: Venular diameter; eGFR: Estimated glomerular filtration rate; Fb: Fractal dimension variable b; Fba: Arteriolar fractal dimension variable b; Fbv: Venular fractal dimension variable b; Fk: Fractal dimension variable k; Fka: Arteriolar fractal dimension variable k; Fkv: Venular fractal dimension variable k; Hb: Hemoglobin; HbA1c: Glycated hemoglobin; HDL-C: High-density lipoprotein cholesterol; LDL-C: Low-density lipoprotein cholesterol; SBP: Systolic blood pressure; Scr: Serum creatinine; TC: Total cholesterol; TD: Tortuosity density; TDa: Arteriolar tortuosity density; TDv: Venular tortuosity density; TG: Triglycerides; UA: Uric acid; UN: Urea nitrogen.

Retinal microvascular segmentation and geometric parameter calculation

The results of the retinal microvascular segmentation are presented in Figure 3. A comparison between the blood vessels in the fundus images and the segmentation outcomes reveals that our unsupervised learning effectively segmented the retinal microvessels. The geometric parameters of the retinal vessels are illustrated in Figure 4. Figure 4A shows statistically significant differences in Fk, Fb, Fka, Fba, Fkv, and Fbv between DN and NDRD groups. Figure 4B reveals a significant difference in WQS-diagnosis values between DN and NDRD.

Figure 3.

Representative retinal microvascular segmentation results. (A) Fundus vessel segmentation outputs for DN and NDRD patients across the training, internal test, and multi-centre validation datasets; (B) Segmentation examples comparing ESRD and non-ESRD patients with diabetic kidney disease. DN: Diabetic nephropathy; ESRD: End-stage renal disease; NDRD: Non-diabetic renal disease; NON: Non-ESRD;.

Figure 4.

Retinal vascular geometric parameters and their mixed effects in DN vs. NDRD. (A) Intergroup comparison of individual fundus vessel geometric parameters between DN and NDRD patients; (B) WQS mixed effects analysis of geometric parameters in DN vs. NDRD classification. ^*P <0.05, ^†P <0.01, ^‡P <0.001. B: Branch angle; D: Diameter; Da: Arteriolar diameter; DM: Distance measure; DMa: Arteriolar distance measure; DMv: Venular distance measure; Dv: Venular diameter; DN: Diabetic nephropathy; Fb: Fractal dimension variable b; Fba: Arteriolar fractal dimension variable b; Fk: Fractal dimension variable k; Fka: Arteriolar fractal dimension variable k; Fbv: Venular fractal dimension variable b; Fkv: Venular fractal dimension variable k; NDRD: Non-diabetic renal disease; TD: Tortuosity density; TDa: Arteriolar tortuosity density; TDv: Venular tortuosity density; WQS: Weighted quantile regression.

DN diagnostic model performance and comparison

We employed various machine learning methods to integrate clinical and retinal vascular geometric parameters to construct a non-invasive diagnostic model for DN. Logistic model outperformed other models such as SVM, RF, Lasso, Elasticnet, Xgboost, and Lightboost, achieving the best performance with an AUC of 0.98 in the test set. Additionally, the model exhibited strong results across other metrics, with an accuracy of 0.92, precision of 0.92, recall of 0.85, F1-score of 0.88, specificity of 0.96, and NPV of 0.92 [Table 2, Figure 5].

Table 2.

Performance comparison of different machine learning models for DN diagnosis and results of the ablation study.

Evaluation cohorts	Factors + parameters	Accuracy	Precision	Recall	F1-score	AUC	Specificity	NPV
Test set	SVM	0.85	0.83	0.74	0.78	0.94	0.92	0.86
	Random forest	0.88	0.85	0.81	0.83	0.95	0.92	0.90
	Lasso	0.87	0.87	0.74	0.8	0.94	0.94	0.87
	Elasticnet	0.88	0.88	0.78	0.82	0.95	0.94	0.88
	Xgboost	0.87	0.95	0.67	0.78	0.97	0.98	0.84
	Lightboost	0.85	0.83	0.74	0.78	0.94	0.92	0.86
	Logistic	0.92	0.92	0.85	0.88	0.98	0.96	0.92
Ablation study	Logistic/clinical factors	0.87	0.87	0.74	0.80	0.93	0.87	0.94
	Logistic/parameters	0.76	0.76	0.48	0.59	0.67	0.76	0.92
Multi-centre validation set	Logistic	0.91	0.90	0.90	0.90	0.95	0.92	0.92

All performance metrics are reported as numerical values ranging from 0 to 1, with higher values indicating better diagnostic performance. Ablation Study: A study design to assess the contribution of individual components to model performance. AUC: Area under the receiver operating characteristic curve; DN: Diabetic nephropathy; NPV: Negative predictive value; SVM: Support vector machine.

Figure 5.

Diagnostic model performance and feature visualization. (A) ROC curves comparing different machine learning diagnostic models; (B) Multi-centre external validation ROC curves of the final logistic regression diagnostic model; (C) ROC curves of diagnostic models constructed using only clinical features; (D) ROC curves of diagnostic models constructed using only fundus vascular parameters; (E–F) Visualization maps illustrating key feature contributions to the model. AUC: Area under the receiver operating characteristic curve; DR: Diabetic retinopathy; eGFR: Estimated glomerular filtration rate; ElasticNet: Elastic net regression; HbA1c: Glycated haemoglobin a1c; LDL: Low‑density lipoprotein; LightGBM: Light gradient boosting machine; LASSO: Least absolute shrinkage and selection operator; ROC: Receiver operating characteristic; SBP: Systolic blood pressure; SHAP: Shapley additive explanations; SVM: Support vector machine; WQS: Weighted quantile sum; XGBoost: Extreme gradient boosting.

Furthermore, we validated the performance of the Logistic model on a multi-centre validation set, where it achieved an accuracy of 0.91, a precision of 0.90, a recall of 0.90, an F1-score of 0.90, a specificity of 0.92, a NPV of 0.92, and an AUC of 0.95. These results demonstrate the strong generalisation capability of our model [Table 2, Figure 5]. The calibration curve of the DN diagnostic model is presented in Supplementary Figure 2, http://links.lww.com/CM9/C958.

Model ablation study

As shown in Table 2, the performance decreased across all metrics after removing the WQS-diagnosis, with reductions of 0.05 in accuracy, 0.05 in precision, 0.11 in recall, 0.08 in F1-score, and 0.05 in AUC. Similar results for ablating the modality of clinical factors are observed in Table 2 across all validation sets. The ROCs in Figure 5A–D further demonstrate the effectiveness of jointly using both modalities.

Visualization

Visualising the Logistic model helps understand its operational principles, identify key features, and reveal the interrelationships among these features. We utilised the SHAP method to visualise the diagnostic model, as illustrated in the Figure 5. For the non-invasive diagnosis of DN and NDRD, the importance rankings are as follows: duration of diabetes, DR, age, LDL, eGFR, WQS, HbA1c, and SBP. Additionally, the hot map displayed the relationships between the two variables, such as the negative correlation between WQS-diagnosis and eGFR [Figure 5E, F].

Prognostic prediction model performance and comparison

The baseline characteristics of the prognostic database are presented in Supplementary Table 4, http://links.lww.com/CM9/C958. The median follow-up times for the training and test sets were 36.2 and 40.0 months, respectively, with endpoint event incidence rates of 45.9% and 35.7%. A prognostic mixed-effects parameter, WQS prognosis [Supplementary Figure 3, http://links.lww.com/CM9/C958], was constructed using the WQS regression model for retinal microvascular parameters, with specific weights shown in the figure. Subsequently, six machine learning survival analysis models were employed to predict the prognosis of DN, as shown in the Figure 6A. We evaluated the performance of each model using AUC and compared it with COX-ridge, COX-Lasso, COX-Tree, COX-ElasticNet, and COX-least squares; COX-RF achieved the best performance with an AUC of 0.88. Figure 6B further illustrates that the AUC for predicting DN prognosis using clinical features and WQS-prognosis indicators are 0.81 and 0.68, respectively. The K–M survival curves for the test set are also presented in Figure 6B.

Figure 6.

Prognostic prediction model performance. (A) ROC curves comparing the prognostic prediction performance of different machine learning models; (B) ROC curves of prognostic models built separately with clinical features alone versus fundus vascular parameters alone. ROC: The receiver operating characteristic.

Discussion

In this study, we utilised unsupervised learning and deep learning to segment and quantify the microvascular structure in 397 fundus images obtained from multiple centres. We first employed the WQS model to construct a mixed-effects parameter based on retinal vascular geometric parameters. Based on these mixed-effects parameters, we used machine learning to construct a non-invasive diagnostic and prognostic prediction model for DN and performed feature visualisation.

Retinal microvasculature segmentation has witnessed significant advancements with the integration of automated computational methods. These methods have evolved from unsupervised techniques that rely on mathematical transformations to identify vascular structures,^[16,17] to more sophisticated supervised methods utilising machine learning algorithms.^[18–20] Incorporating deep neural networks has set a new benchmark, offering improved segmentation performance compared to classical techniques. Our study contributes to this evolving landscape by proposing an unsupervised learning and deep learning ensemble model for automatically segmenting retinal vasculature from fundus images. A notable advantage of our proposed model is its ability to operate without hand-drawn segmentations, reducing the dependency on expert annotations. This not only makes the model more accessible and cost-effective but also enhances its potential for scalability and broader applicability across diverse datasets. Post-processing techniques, such as inpainting, further refine the segmentation output by addressing any irregularities or discontinuities in the vascular network.

Advancements in computational methods have enabled accurate assessments of the retinal microvasculature, providing quantitative measurements such as vessel diameter, fractal dimension, and vascular density. These parameters have been widely used in research on hypertension, stroke, and coronary artery disease, demonstrating consistent associations.^[21–23] In this study, significant differences were observed in Fk, Fb, Fka, and Fba between the DN and NDRD groups, whereas no significant differences were found in DM, DMa, and DMv, among others. Fractal dimensions in retinal vessel measurements reflect the complexity and density of the vessels. Higher values indicate more complex structures. Changes in vascular fractal dimensions can be attributed to abnormal vascular changes, whether it be due to excessive vascular proliferation or insufficient angiogenesis. Previous studies have also shown that parameters related to vascular fractal dimensions are of great significance in differentiating diseases, while other indicators showed no differences.^[10,20]

Changes in retinal vascular geometric parameters have been linked to systemic diabetic complications such as DR and cardiovascular diseases.^[24,25] Recent studies have demonstrated a correlation between retinal vascular geometric parameters and kidney damage.^{[14,15,26,27]} In our study, we utilised the WQS regression to compute a mixed-effects model, accounting for individual differences and internal correlations among retinal vascular geometric parameters. This approach improves the adaptability and flexibility of the model by effectively integrating parameters such as diameter, fractal dimension, and vascular density, all of which are closely related to the development of DR. Through this comprehensive analysis of multiple parameters, the model can more accurately identify early biomarkers of DR, thereby providing earlier opportunities for intervention. Changes in the structure of retinal microvessels may reflect systemic microvascular alterations in patients with T2DM and could potentially serve as early indicators of DN. However, most current studies analyse retinal vascular geometric parameters independently, without considering the collinearity and mixed effects among parameters. This study addresses this gap by using the WQS regression model to construct a mixed-effects parameter of retinal microvasculature for the diagnosis and prognosis prediction of DN.

Considering the risks associated with renal biopsies, such as infection and bleeding, high cost, and limited accessibility in underdeveloped regions,^[28] there has been a growing interest in developing multiple non-invasive diagnostic models for DN based on clinical parameters both domestically and internationally.^[29–36] However, previous models typically relied on single-dimensional clinical data, often focusing on common indicators like blood glucose and blood pressure while neglecting other factors that might influence the diagnosis of diabetic kidney disease, such as retinal vascular geometric parameters. This study demonstrates that a comprehensive model combining clinical data and retinal vascular geometric parameters offers significant advantages in diagnosing DN. Compared to traditional models based solely on clinical data, the comprehensive model more accurately reflects the renal pathology of patients with diabetes, thereby improving diagnostic accuracy and specificity. The results of our ablation experiments further confirmed the effectiveness of multi-dimensional data modelling. The diagnostic performance of the model decreased when the mixed effects of retinal vascular geometric parameters were excluded. This finding provides new insights and methods for the non-invasive diagnosis of diabetic kidney disease. Moreover, compared to previous studies conducted by our team,^[10] the diagnostic model constructed using mixed effects of retinal vascular geometric parameters significantly enhances the model’s diagnostic performance. The generalisation capability of our model has been validated using a multicentre dataset.

Our study constructed a prognostic prediction model for DN based on clinical data and retinal microvascular geometric parameters. Previous studies on DN prognosis have identified several factors that influence the prognosis of DN, including cystatin C, ALB, Hb, 24-hour urine total protein, eGFR, duration of diabetes, fundus lesions, BMI, hypertension grade, CKD stage, glomerular hyalinosis, and increased extracellular cells in capillaries.^[37–40] Our study corroborates these findings, identifying cystatin C, 24-hour urinary protein, ALB, Hb, eGFR, and blood pressure as prognostic factors for DN. The prognostic prediction model developed in this study achieved an AUC of 0.88. However, the 3–4 year follow-up period is relatively short for patients with diabetic nephropathy. In this study, all the DN patients included were clearly diagnosed by renal biopsy and strictly followed up for 3–4 years, which could basically reflect the process of disease progression and exacerbation. Our study confirmed that even with 3–4 years of follow-up, 45.9% of the patients reached the endpoint, so it could reflect the natural course of the disease. The follow-up period of several studies currently reported is also 3–4 years.^[29,39,41] Moreover, our findings indicate that while retinal vascular geometric parameters do enhance the diagnostic performance of models, clinical factors play a much more significant role in both diagnosis and prognosis. This is consistent with previous results, which showed that using only clinical features can achieve high performance in the diagnosis and prognosis prediction of DN.^[29,35,37] Although the performance based solely on WQS for the diagnosis and prognosis prediction of DN is lower, the addition of WQS still significantly improves the performance due to the additional dimension of data it provides.^[10]

There are some limitations in the study. Despite the fact that patients who have both kidney biopsy results and retinal photographs are rare, the study involved a relatively small sample. To confirm these results, research with a larger and more diverse population is necessary. However, we are actively collecting more data and plan to expand our sample size in the future to further validate our findings.

In conclusion, this study constructed a non-invasive intelligent diagnostic and prognostic prediction model that combines mixed-effect retinal vascular geometric parameters (WQS) with clinical characteristics, significantly enhancing the accuracy of early diagnosis and prognosis of DN. Additionally, retinal vascular mixed-effect parameters identified in this study can serve as valuable biomarkers for DN diagnosis and prognosis. The results of this research reduce dependence on renal puncture biopsy, thereby lowering patient risk and medical costs. Furthermore, the model offers a feasible diagnostic method for remote areas, significantly improving early intervention and management of patients with DN.

Acknowledgements

We also thank Jijun Li (Department of Nephrology, Fourth Medical Center of Chinese PLA General Hospital, Beijing, China), Yanqiu Geng (Department of Nephrology, Third Medical Center of Chinese PLA General Hospital, Beijing, China), Jinkui Yang, Fangyuan Yang (Beijing Key Laboratory of Diabetes Research and Care, Beijing Diabetes Institute, Beijing Tongren Hospital, Capital Medical University), Yani He, Kehong Chen, Jie Yang (Department of nephrology, Daping Hospital, Army Medical University), Fenghua Yao (Department of Nephrology, Fifth Medical Center of Chinese PLA General Hospital, Beijing, China), Hongli Lin, Ming Fang (The First Affiliated Hospital of Dalian Medical University), Lin Zhang (Dept. of Endocrinology, Dept. of Medical Record; Beijing Tongren Hospital, Capital Medical University), Jie Zhao (Senior Department of Ophthalmology, the Third Medical Center of PLA General Hospital), Yan Yu, Xiang Ma, Jingying Zhou (The First Affiliated Hospital of Dalian Medical University) for their efforts in data collection.

Funding

This work was supported by grants from Science Technology Project of Beijing, China (No. Z221100007422121), Natural Science Foundation of China (Nos. 32141005, 32000530 and 32371035), Beijing Nova Program (No. 20230484246), the National Key R&D Program of China (No. 2022YFA1105502), Young Elite Scientists Sponsorship Program by CAST (No. 2023QNRC001, YESS20240790), the national key R&D projects (No. 32141005), Beijing Natural Science Foundation (Nos. L223021, L232122, 7244305, L222133, and 7252018), Capital’s Funds for Health Improvement and Research (No. CFH 2024-1-5021), and Noncommunicable Chronic Diseases-National Science and Technology Major Project (No. 2026ZD0557100, 2026ZD0557105).

Conflicts of Interest

None.

Supplementary Material

cm9-139-2975-s001.docx ^{(887.1KB, docx)}