AI Developed to Predict Dementia Risk Using Fundus Images

Seoul National University Hospital Analyzes 36,000 Participants and 108,000 Fundus Images
"Outperforms Conventional Risk Assessment Tools"

A team of South Korean researchers has developed an artificial intelligence (AI) model capable of detecting existing dementia and predicting future dementia risk using only retinal fundus photographs. The team also applied explainable AI technologies that visualize the retinal areas used as a basis for the AI’s judgment, assessing the model’s clinical applicability.

On August 28, Seoul National University Hospital announced that a joint research team—comprising Professor Sangmin Park from the Department of Family Medicine at Seoul National University Hospital, Dr. Jooyoung Jang from Zimed Co., Ltd., Assistant Research Professor Changho Han from the Medical Big Data Research Center at Seoul National University College of Medicine, and Dr. Jaewon Kim from the Department of Medical Science at Seoul National University College of Medicine—developed and evaluated a dementia screening and prediction AI model. The study used 108,008 fundus images collected from 36,322 participants who underwent medical checkups at Seoul National University Hospital between 2004 and 2016.

From the left, Sangmin Park, Professor of Family Medicine at Seoul National University Hospital; Jooyoung Jang, Ph.D. at Zimed Co., Ltd.; Changho Han, Research Assistant Professor at the Medical Big Data Research Center, Seoul National University College of Medicine; Jaewon Kim, Ph.D. in Medical Science at Seoul National University College of Medicine. Seoul National University Hospital

Dementia is typically diagnosed based on clinical assessment, cognitive tests, and brain magnetic resonance imaging (MRI), but these methods have limitations in terms of cost and accessibility. The retina, as an extension of the central nervous system, is known to reflect vascular and neural changes in the brain, making it an attractive biomarker for assessing dementia risk.

The research team defined 1,001 patients (with 2,868 images) who were diagnosed with dementia within two years before or after their fundus images were taken as the dementia group, and matched them at a 1:4 ratio with a dementia-free control group to build the development dataset. They then developed 15 different AI models by combining five foundation models with three fine-tuning strategies, and compared their performance.

The results showed that the model based on the foundation model 'RETFound-MAE,' which had been pre-trained on millions of fundus images and then underwent partial fine-tuning, achieved the best performance. This model recorded an AUROC of 0.750 for current dementia detection and a C-index of 0.812 for predicting future dementia risk—both higher than the scores of the standard clinical tool CAIDE (AUROC 0.624, C-index 0.689).

Even after adjusting for established dementia risk factors such as age and gender, higher AI scores were still significantly associated with both current and future dementia risk. This suggests that fundus images contain additional dementia-related information not captured by conventional risk factors.

In decision curve analysis, the AI model outperformed both the CAIDE score and the blanket screening/non-screening strategies in terms of net clinical benefit. When combining CAIDE with the AI score, the AUROC improved from 0.749 to 0.774, indicating the AI could complement existing dementia risk evaluation approaches.

The researchers also analyzed which parts of the fundus image the AI relied on for its decisions. The model showed high attention levels in the optic disc and surrounding areas for both dementia screening and risk prediction.

The overall attention score for the optic disc was 5.72 times higher than that for other peripheral areas, with the highest attention observed in the temporal side of the optic disc. Elevated attention was also seen in the papillomacular bundle, which connects the optic disc and the macula, as well as the peripapillary region.

These results correspond to retinal regions previously identified in research as being associated with dementia, supporting the hypothesis that the AI is using biologically valid retinal signals in its decision-making, according to the research team.

Professor Park stated, "We hope that this AI technology, which leverages routine checkup data without requiring additional costly tests, will become a valuable supplementary tool in dementia prevention and early intervention strategies." Dr. Jang added, "This study not only assessed the predictive performance of AI, but also its explainability and clinical utility, pointing to a meaningful direction for the development of AI-based fundus image biomarkers." The research findings were published in the latest issue of the international journal 'npj Digital Medicine.'

This content was produced with the assistance of AI translation services.