Diagnosis of Chronic Kidney Disease Using Retinal Imaging and Urine Dipstick Data: Multimodal Deep Learning Approach

Abstract

Background: Chronic kidney disease (CKD) is a prevalent condition with significant global health implications. Early detection and management are critical to prevent disease progression and complications. Deep learning (DL) models using retinal images have emerged as potential noninvasive screening tools for CKD, though their performance may be limited, especially in identifying individuals with proteinuria and in specific subgroups.

Objective: We aim to evaluate the efficacy of integrating retinal images and urine dipstick data into DL models for enhanced CKD diagnosis.

Methods: The 3 models were developed and validated: eGFR-RIDL (estimated glomerular filtration rate-retinal image deep learning), eGFR-UDLR (logistic regression using urine dipstick data), and eGFR-MMDL (multimodal deep learning combining retinal images and urine dipstick data). All models were trained to predict an eGFR<60 mL/min/1.73 m², a key indicator of CKD, calculated using the 2009 CKD-EPI (Chronic Kidney Disease Epidemiology Collaboration) equation. This study used a multicenter dataset of participants aged 20-79 years, including a development set (65,082 people) and an external validation set (58,284 people). Wide Residual Networks were used for DL, and saliency maps were used to visualize model attention. Sensitivity analyses assessed the impact of numerical variables.

Results: eGFR-MMDL outperformed eGFR-RIDL in both the test and external validation sets, with area under the curves of 0.94 versus 0.90 and 0.88 versus 0.77 (P<.001 for both, DeLong test). eGFR-UDLR outperformed eGFR-RIDL and was comparable to eGFR-MMDL, particularly in the external validation. However, in the subgroup analysis, eGFR-MMDL showed improvement across all subgroups, while eGFR-UDLR demonstrated no such gains. This suggested that the enhanced performance of eGFR-MMDL was not due to urine data alone, but rather from the synergistic integration of both retinal images and urine data. The eGFR-MMDL model demonstrated the best performance in individuals younger than 65 years or those with proteinuria. Age and proteinuria were identified as critical factors influencing model performance. Saliency maps indicated that urine data and retinal images provide complementary information, with urine offering insights into retinal abnormalities and retinal images, particularly the arcade vessels, being key for predicting kidney function.

Conclusions: The MMDL model integrating retinal images and urine dipstick data show significant promise for noninvasive CKD screening, outperforming the retinal image-only model. However, routine blood tests are still recommended for individuals aged 65 years and older due to the model's limited performance in this age group.

Keywords: chronic kidney disease; fundus image; multimodal deep learning; saliency map; urine dipstick.

Figure 1.. Study flowchart from data preparation to deep learning modeling and validation. Images with abnormal brightness or haziness were excluded. CHA: CHA Bundang Medical Center; eGFR: estimated glomerular filtration rate; eGFR-MMDL: multimodal deep learning model for estimated glomerular filtration rate<60 mL/min/1.73 m²; eGFR-RIDL: retinal image deep learning model for estimated glomerular filtration rate<60 mL/min/1.73 m²; SCHPC: Severance Checkup Health Promotion Center.

Figure 2.. Wide Residual Network architecture for detecting kidney function decline using retinal images. For the multimodal model, the urine dipstick measurements, age, and sex were concatenated with the feature vector from the image. Conv: convolutional layer.

Figure 3.. ROC curves and AUC of low-eGFR detection models. The AUCs of the eGFR-MMDL and eGFR-UDLR models were compared with the AUC of eGFR-RIDL in (A) the test set and (B) the external validation set (DeLong test with Bonferroni correction). *P=.002; ***P<.001. AUC: area under the curve; eGFR: estimated glomerular filtration rate; MMDL: multimodal deep learning model; RIDL: retinal image deep learning model; ROC: receiver operating characteristic; UDLR: urine dipstick logistic regression.

Figure 4.. ROC curves and AUC of low-eGFR detection models in subgroups of the test set. (A) Age <65 years (solid line) and ≥65 years (dotted line); (B) nondiabetes (solid line) and diabetes (dotted line); (C) no hypertension (solid line) and hypertension (dotted line); and (D) no proteinuria (solid line) and proteinuria (dotted line). The AUCs of the eGFR-MMDL and eGFR-UDLR models are compared with the AUC of the eGFR-RIDL (DeLong test with Bonferroni correction; P value indicators next to the AUC values); *P=.013, **P=.001, ***P<.001; NS: not significant. P value indicators are added to the upper left corner of the eGFR-MMDL model curve to denote significant differences in AUCs between the 2 subgroups (DeLong test); *P=.011, **P=.001, ***P<.001; NS: not significant. AUC: area under the curve; eGFR: estimated glomerular filtration rate; MMDL: multimodal deep learning model; RIDL: retinal image deep learning model; ROC: receiver operating characteristic; UDLR: urine dipstick logistic regression.

Figure 5.. Saliency map results from the eGFR-RIDL and eGFR-MMDL models for eGFR decline detection. (A) A nondiabetes male aged 60 years with proteinuria level 3 (2+). (B) A male with diabetes and proteinuria level 4 (3+) aged 54 years. (C) A male with diabetes and proteinuria level 4 (3+) aged 60 years. (D) A male without hypertension or diabetes aged 60 years, with a proteinuria level of 0 (negative). From left to right: original retinal image, retinal image postpreprocessing via CLAHE and color normalization, eGFR-RIDL–generated saliency map, and eGFR-MMDL–generated saliency map. CLAHE: Contrast Limited Adaptive Histogram Equalization; eGFR: estimated glomerular filtration rate; MMDL: multimodal deep learning; RIDL: retinal image deep learning.

Figure 6.. Sensitivity analysis: feature importance of age, sex, and 10 urine measurements in predicting the probability of eGFR decline by eGFR-MMDL using a test set. The sensitivity analysis scores indicate the relative impact of each variable on the predicted probability. Higher sensitivity analysis scores correspond to a greater influence. The error bars on the right end represent 95% CIs. 1e-08: 0.00000001; eGFR: estimated glomerular filtration rate; MMDL: multimodal deep learning.