Is AI Capable of Anticipating the Progression of Diabetic Retinopathy?

In a groundbreaking research paper featured in the esteemed journal Nature Medicine, Chinese scientists have introduced a cutting-edge deep learning (DL) system called “DeepDR Plus.” This innovative system demonstrates remarkable capabilities in forecasting the advancement of diabetic retinopathy (DR) by analyzing fundus images of patients. The study reveals that DeepDR Plus exhibits a high degree of accuracy in predicting both the likelihood and timeline of DR progression over five years. This breakthrough paves the way for the development of personalized screening protocols, heralding a new era in the management of diabetic retinopathy. Stay tuned for further updates on this revolutionary advancement in medical technology.

Background

Diabetic retinopathy (DR) often arises without symptoms in adults with diabetes but can eventually lead to avoidable vision loss. The progression of DR varies among individuals and is influenced by various factors, making it difficult to predict. Consequently, patients are advised to undergo annual DR screening. However, the absence of a personalized risk model hinders physicians from extending the screening interval, even though it is cost-effective.

Artificial intelligence, particularly deep learning (DL), has shown promise in automating the detection of DR from retinal images. Nevertheless, current research lacks a prospective risk prediction for DR onset and progression beyond two years. Further studies are necessary to evaluate the impact on patient outcomes and integration into clinical workflows.

In this study, researchers expanded on their previous work to address these gaps. They developed, validated, and tested “DeepDR Plus” as a tool capable of predicting DR progression trajectories up to five years in advance. Moreover, they demonstrated the tool’s effectiveness through a real-world study conducted on patients with diabetes.

About The Study

DeepDR Plus underwent initial training on a dataset of 717,308 fundus images from 179,327 diabetic individuals participating in the Shanghai Integrated Diabetes Prevention and Care System and the Shanghai Diabetes Prevention Program. Subsequent internal development and validation utilized a dataset of 76,400 fundus images from the Diabetic Retinopathy Progression Study (DRPS) cohort, divided into developmental and internal test sets. The model’s performance was assessed using metrics such as the Concordance Index (C-index) and Integrated Brier Score (IBS).

To evaluate generalizability, external validations were performed using eight independent longitudinal cohorts with comprehensive baseline demographic, anthropometric, and biochemical data. Diabetic retinopathy grades were assigned based on the International Clinical Diabetic Retinopathy Disease Severity Scale. The study focused on three patient subgroups: (1) diabetes with no retinopathy to DR, (2) non-referable DR to referable DR and (3) non-vision-threatening DR to vision-threatening DR. Statistical analysis included the log-rank test, Cox regression analysis, and the calculation of the area under the curve, mean absolute error, and coefficient of determination.

To apply DeepDR Plus in real-world scenarios, a community-based prospective cohort study involved 2,185 Chinese adults categorized into integrated management (IM) and non-IM groups. The fundus and metadata models were employed to assess the risk of diabetic retinopathy progression. Additionally, a real-world study within an Indian prospective cohort included 992 diabetic patients with a four-year follow-up period.

Results & Discussions

the fundus model outperformed the metadata model in internal validation, displaying superior performance as gauged by C-indices, even when working with low-resolution images. Over the initial five years, the fundus model consistently exhibited strong performance across eight external datasets, accurately predicting the specific time for diabetic retinopathy (DR) progression, as evidenced by high C-indices and IBS. Subgroup analysis revealed that DeepDR Plus effectively predicted various types of DR grade deterioration over five years in three subgroups, surpassing the metadata model in C-indices and IBS.

In a real-world study based on a Chinese cohort, the fundus model demonstrated the potential to prevent 46.8% of DR progression incidents through comprehensive interventions. A similar study based on an Indian cohort showed an even more promising result, with the fundus model helping prevent 88.74% of DR progression incidents compared to the metadata model. If participants in both the IM and non-IM groups followed the personalized screening intervals recommended by the fundus model, the average screening interval could be extended from 12 months to 31.97 months. Overall, the fundus model proved adept at accurately categorizing participants, enabling personalized interventions, and reducing the frequency of diabetic retinopathy screenings while minimizing delays in detecting progression.

Detailed analysis revealed that the tool’s predictions are centered on retinal vessels and the fovea, supported by attention maps and quantitative measurements extending beyond retinal vascular geometry.

Nevertheless, it’s crucial to note limitations in our study, including the training of DeepDR Plus on a Chinese population, potential intrinsic biases, performance variations with different treatment regimens, and the absence of actual clinical application. This emphasizes the necessity for future testing and trials to validate AI-driven diabetic retinopathy screening and intervention.

Conclusion

In summary, DeepDR Plus, when using initial fundus images, can reliably forecast an individual’s risk and the time it takes for diabetic retinopathy to progress. Applying this in real-world scenarios indicates the possibility of extending screening intervals to approximately 32 months. These results highlight the encouraging integration of the tool into everyday clinical practices for tailoring diabetic retinopathy management, aiming to enhance patient outcomes.

Journal Reference:

A deep learning system for predicting the time to progression of diabetic retinopathy. Dai, L. et al., Nature Medicine (2024), DOI: https://doi.org/10.1038/s41591-023-02702-z, https://www.nature.com/articles/s41591-023-02702-z

Leave a Reply

Your email address will not be published. Required fields are marked *