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

Area of Science:

• Ophthalmology
• Nephrology
• Artificial Intelligence

Background:

• Chronic kidney disease (CKD) is a widespread condition requiring early detection to prevent complications.
• Deep learning (DL) models using retinal images offer a potential noninvasive screening method for CKD.
• Current DL models may have limitations in detecting proteinuria and performing in specific subgroups.

Purpose of the Study:

• To evaluate the effectiveness of combining retinal images and urine dipstick data in DL models for improved CKD diagnosis.
• To compare the performance of multimodal DL models against single-modality models.

Main Methods:

• Developed and validated three models: eGFR-RIDL (retinal images), eGFR-UDLR (urine data), and eGFR-MMDL (combined retinal images and urine data).
• All models predicted an estimated glomerular filtration rate (eGFR) <60 mL/min/1.73 m² using the 2009 CKD-EPI equation.
• Utilized a multicenter dataset with development (65,082) and external validation (58,284) cohorts; employed Wide Residual Networks and saliency maps.

Main Results:

• The eGFR-MMDL model demonstrated superior performance over eGFR-RIDL in both test and validation sets (AUCs 0.94 vs. 0.90 and 0.88 vs. 0.77).
• eGFR-UDLR showed comparable performance to eGFR-MMDL, especially in external validation, but eGFR-MMDL improved across all subgroups.
• eGFR-MMDL excelled in individuals younger than 65 or with proteinuria, indicating synergistic benefits of combined data.

Conclusions:

• The multimodal DL (MDL) model integrating retinal images and urine data shows significant promise for noninvasive CKD screening.
• MDL outperformed retinal image-only models, highlighting the value of combined data modalities.
• Routine blood tests remain recommended for individuals aged 65 and older due to the model's performance limitations in this demographic.