JMIR Publications

ABSTRACT

Background:

Chronic kidney disease (CKD) is a prevalent condition with substantial implications for global health. Early detection and management are crucial to prevent disease progression and complications. Deep learning models using retinal images have emerged as potential non-invasive screening tools for CKD. However, their performance can be limited, particularly in identifying individuals with proteinuria and those in specific subgroups.

Objective:

To evaluate the efficacy of integrating retinal images and urine dipstick data into deep learning models for enhanced CKD diagnosis using retinal images.

Methods:

Two deep-learning models were developed and validated: eGFR-RIDL (retinal image deep learning) and eGFR-MMDL (multimodal deep learning combining retinal images and urine dipstick data). Both models were trained to predict an estimated glomerular filtration rate (eGFR) < 60 ml/min/1.73 m², an indicator of CKD. eGFR was calculated using a 2009 CKD-EPI equation. The study employed a multicenter dataset of participants aged 20-79 years, including a development set (65,082 subjects) and an external validation set (58,284 subjects). We used a wide residual network for model development, saliency maps to visualize model attention, and sensitivity analysis to assess the impact of numerical variables.

Results:

eGFR-MMDL consistently outperformed eGFR-RIDL in both test and external validation sets, with AUCs of 0.94 vs 0.90 and 0.88 vs. 0.77 (P < 0.001 for both, DeLong test), respectively. Subgroup analysis revealed the superiority of the eGFR-MMDL in individuals aged < 65 years. Age and proteinuria were identified as crucial numerical factors influencing the model performance. The emphasis on central vessels in saliency maps underscored the importance of these features in deep learning with retinal images.

Conclusions:

Multimodal deep learning models integrating retinal images and urine dipstick data offer significant promise for noninvasive CKD screening. However, routine blood tests were recommended for individuals aged 65 years and older, as the model performance was limited in this age group.