Multi-Branching Temporal Convolutional Network With Tensor Data Completion for Diabetic Retinopathy Prediction.

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Ophthalmology

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss in working-aged adults.
• Low screening compliance and high costs hinder early detection, leading to irreversible vision loss.
• Electronic Health Records (EHRs) offer a cost-effective data source for developing predictive tools.

Purpose of the Study:

• To propose a novel machine learning model, MB-TCN-TC, for predicting diabetic retinopathy using longitudinal EHR data.
• To address challenges in EHR data such as imbalanced classes and missing values.
• To improve the accuracy and efficiency of DR prediction compared to existing methods.

Main Methods:

• Development of a Multi-branching Temporal Convolutional Network with Tensor Data Completion (MB-TCN-TC) model.
• Analysis of longitudinal EHR data from diabetic patients.
• Evaluation of model performance using metrics like AUROC, AUPRC, and F1 score.

Main Results:

• The MB-TCN-TC model effectively handles imbalanced data and missing values in EHRs.
• The model successfully captures temporal correlations and complex interactions within clinical data.
• MB-TCN-TC achieved superior prediction performance, with AUROC of 0.949 and AUPRC of 0.793, significantly outperforming traditional TCN models.

Conclusions:

• The MB-TCN-TC model demonstrates significant potential for cost-effective and accurate diabetic retinopathy prediction.
• Leveraging EHR data with advanced AI can improve patient outcomes by enabling earlier detection and intervention.
• This approach offers a scalable solution to mitigate the public health burden of diabetic retinopathy-related vision loss.