Professors Dong Keon Yon and Sang Youl Rhee of the College of Medicine, together with their research team, have developed a multimodal artificial intelligence model capable of predicting the risk of chronic kidney disease (CKD) within five years in patients with type 2 diabetes.

The model draws on large-scale clinical and imaging datasets from both Korean and international cohorts. By combining clinical test results with retinal fundus images, the system achieved significantly higher accuracy and interpretability than conventional AI systems. Notably, the model also predicted a patient’s likelihood of developing vascular complications, offering new possibilities for precision medicine and personalized care. The study—conducted with Research Professor Selin Woo, researchers Seung Ha Hwang, Jaehyeong Cho, and Soeun Kim, and Professor Hong-Hee Won of Sungkyunkwan University—was based on large-scale clinical data from both Korea and the U.K. The findings were published in the online edition of Diabetes Care (Impact Factor: 16.6) under the title, “A Multimodal Predictive Model for Chronic Kidney Disease and Its Association With Vascular Complications in Patients With Type 2 Diabetes: Model Development and Validation Study in South Korea and the U.K.”

Multimodal AI model that combines clinical data and retinal images overcomes the limitations of single-modality tools, improving accuracy, interpretability, and real-world applicability
Diabetes is highly prevalent worldwide, and kidney disease is one of its most serious and common complications. Early prediction and prevention are essential, but conventional risk assessment tools have typically relied on a single type of data, such as clinical tests or imaging alone. This narrow approach has limited both the accuracy and the interpretability of the results. To address these limitations, the Kyung Hee research team set out to develop a multimodal artificial intelligence (AI) model capable of integrating multiple types of medical information. By combining structured clinical data with retinal fundus images, the team aimed to enhance both the precision of predictions and their practical applicability in clinical settings.

The research team developed the multimodal AI model using data from Kyung Hee University Medical Center in Korea and a diabetes cohort in the United Kingdom. By integrating structured clinical data—such as blood and urine test results and medication history—with retinal fundus images, they constructed a deep learning system capable of predicting the risk of developing chronic kidney disease within five years. The AI tool was first trained on Korean patient data and then externally validated using the UK cohort. It has demonstrated strong predictive performance, achieving an accuracy of 88.0% in the Korean cohort in the domestic dataset and 72.2% in the external validation, underscoring its potential for international clinical application.

One of the key limitations of conventional AI models in medicine is their “black-box” nature—they often produce results without revealing how those results were derived. To overcome this challenge, the researchers integrated explainable AI (XAI) techniques into the model. These methods make it possible to visually interpret the basis of the AI’s predictions, thereby enhancing transparency and increasing the model’s potential for real-world clinical adoption.

AI sheds light on key indicators of kidney disease through explainable techniques
The explainable AI analysis identified several major risk factors for chronic kidney disease, including estimated glomerular filtration rate (eGFR), the use of diabetes and hypertension medications, and the patient’s age. In the retinal imaging data, the optic disc and superior vascular arcade emerged as critical visual cues. These findings demonstrate that the algorithm not only predicts outcomes but also offers clinically meaningful insights that physicians can use as scientific evidence in patient care.

The researchers also analyzed the relationship between the model’s predicted probabilities and the actual occurrence of vascular complications. Patients with higher predicted risk scores were significantly more likely to develop major complications, such as cardiovascular and peripheral vascular disease, neuropathy, and end-stage renal disease. For instance, those in the highest tertile of model probability faced up to a 2.21 times greater risk of macrovascular complications and a 1.30 times greater risk of microvascular complications compared to those in the lowest group. These findings suggest that the AI tool could be used not only to predict CKD onset early but also to support long-term health management and the prevention of serious complications.

“This AI model enables high-accuracy predictions using only data routinely collected in clinical settings, making it a realistic tool for use in primary care,” said Research Professor Selin Woo. “It lays the groundwork for precision medicine by allowing early identification and intervention for high-risk patients.” Professor Sang Youl Rhee added, “By training and validating the model with both domestic and international data, we ensured its generalizability and reliability. This study opens new avenues for personalized patient management.”