Temporal Integrative Machine Learning for Early Detection of Diabetic Retinopathy Using Fundus Imaging and Electronic...

Area of Science:

• Ophthalmology
• Medical Informatics
• Machine Learning

Background:

• Diabetic retinopathy (DR) is a leading cause of blindness, often diagnosed late due to symptom manifestation and limited expert availability.
• Early detection of DR is crucial for timely intervention and preventing vision loss.

Purpose of the Study:

• To develop and validate a novel temporal integrative machine learning system for early and enhanced DR detection.
• To leverage both fundus images and electronic health records (EHR) for improved diagnostic accuracy.

Main Methods:

• A dual-model system was developed: a temporal tabular model using historical EHR data and a deep learning multi-modal model combining EHR with fundus images.
• Models were trained using pseudo-labeling on clinical data from 5,000 patients (25,000 retinal images, up to 20 years of EHR).
• Temporal features were extracted from EHR to capture long-term patient history dynamics.

Main Results:

• The temporal-trend EHR model achieved an AUROC score of 0.881.
• The multi-modal imaging+EHR model demonstrated superior performance with an AUROC score of 0.988.
• Expert verification confirmed the models' efficacy, surpassing existing methods.

Conclusions:

• Integrating temporal EHR data with fundus imaging creates a dynamic, comprehensive system for enhanced DR detection.
• This approach provides physicians with a holistic patient view and facilitates early identification of high-risk individuals.
• The system offers new insights into DR risk factors and improves patient management strategies.