**Background:** Diabetic kidney disease (DKD) is the leading cause of end-stage renal disease worldwide, yet screening rates remain suboptimal. Retinal photography is noninvasive and commonly used for diabetic retinopathy screening. Because the retina and kidneys share similar microvascular pathology, retinal images may contain information about kidney function. This study aimed to develop and validate a deep learning algorithm (DLA) to detect DKD from retinal photographs in patients with diabetes, and to compare image-only, risk factor (RF)-only, and hybrid models.
**Methods:** Three models were developed: (1) image-only DLA based on ResNet18 architecture, pretrained on a diabetic retinopathy dataset; (2) RF-only multivariable logistic regression (LR) model adjusted for age, sex, ethnicity, diabetes duration, HbA1c, and systolic blood pressure (SBP); (3) hybrid LR model combining RF data with standardized z-scores from the image-only model. The development dataset was the Singapore Integrated Diabetic Retinopathy Program (SiDRP), a primary-care-based program (2010–2019), including 6066 unique participants (5356 DKD cases, 7928 controls) with 26,568 retinal images. DKD was defined as eGFR <60 mL/min/1.73 m² on ≥2 consecutive visits 3 months to 2 years apart. Internal validation used 5-fold cross-validation. External testing was performed on two independent Singaporean datasets: the Singapore Epidemiology of Eye Diseases (SEED) study (1885 participants, 798 cases, 1171 controls) and the Singapore Macroangiopathy and Microvascular Reactivity in Type 2 Diabetes (SMART2D) study (439 participants, 227 cases, 485 controls). Supplementary external testing was done on two predominantly Caucasian cohorts: the Australian Eye and Heart Study (AHES, n=460) and the Northern Ireland Cohort for the Longitudinal Study of Ageing (NICOLA, n=265). Performance was evaluated using area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) at the optimal threshold defined by Youden’s J Index.
**Key Results:** In SiDRP internal validation, AUC was 0.826 (95% CI 0.818–0.833) for image-only, 0.847 (0.840–0.854) for RF-only, and 0.866 (0.859–0.872) for hybrid. In SEED external validation, AUC was 0.764 (0.743–0.785) for image-only, 0.802 (0.783–0.822) for RF-only, and 0.828 (0.810–0.846) for hybrid. In SMART2D, AUC was 0.726 (0.686–0.765) for image-only, 0.701 (0.660–0.741) for RF-only, and 0.761 (0.724–0.797) for hybrid. At optimal thresholds, the image-only model in SiDRP had sensitivity 76% and specificity 75%; in SEED, sensitivity 70% and specificity 71%; in SMART2D, sensitivity 64% and specificity 71%. NPV for the image-only model was 82% in SiDRP, 78% in SEED, and 81% in SMART2D. When using a stricter DKD definition (eGFR <45 mL/min/1.73 m²), AUC improved: image-only in SiDRP increased to 0.851 (0.843–0.860), in SEED to 0.785 (0.757–0.813), and in SMART2D to 0.759 (0.718–0.800). Hybrid model AUCs also improved: SiDRP 0.887 (0.880–0.895), SEED 0.851 (0.829–0.874), SMART2D 0.765 (0.723–0.806). In supplementary analyses on Caucasian cohorts, image-only AUC was 0.670 (0.612–0.729) in AHES and 0.638 (0.562–0.714) in NICOLA; hybrid models achieved AUCs of 0.695 (0.640–0.751) and 0.710 (0.640–0.779), respectively.
**Clinical Implications:** This study demonstrates that a DLA using retinal photographs can detect DKD with moderate accuracy in multiethnic Asian populations, and that combining retinal images with clinical risk factors improves performance. The model could serve as a noninvasive screening adjunct to existing diabetic retinopathy screening programs, potentially increasing DKD detection rates in primary care settings where retinal photography is already used. However, performance was lower in Caucasian cohorts, indicating the need for further training on diverse populations. The authors suggest that the tool could be applied in a two-stage screening approach, with high sensitivity to minimize missed cases, followed by confirmatory laboratory testing. Limitations include the lack of albuminuria data, low representation of stage G5 CKD in training, and reduced performance in non-Asian populations. Future steps include validation in younger patients with type 1 diabetes and implementation studies.