Temporal deep learning framework for retinopathy prediction in patients with type 1 diabetes

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Time Series Analysis

Background:

• Electronic health records provide large datasets for predicting medical complications.
• Patient data trajectories are variable, posing challenges for longitudinal data modeling.
• Deep learning for time series is advancing, but application to irregular medical time series (IMTS) is limited.

Purpose of the Study:

• To develop a deep learning framework for modeling IMTS.
• To compare sequential neural networks and time representation techniques for IMTS.
• To predict retinopathy complications in type 1 diabetes patients using longitudinal data.

Main Methods:

• Developed a generic deep-learning framework for IMTS.
• Compared transformer and long short-term memory networks.
• Utilized soft one-hot time gap representation.
• Validated on 1207 type 1 diabetes patients using glycosylated hemoglobin data.

Main Results:

• The transformer model with soft one-hot time gaps achieved an 88.65% AUC.
• Achieved 85.56% specificity and 83.33% sensitivity.
• Demonstrated an 11.7% performance improvement by including time gap information.

Conclusions:

• Modeling time gaps in medical records is crucial for improving prediction performance.
• The developed framework facilitates comparative studies of deep learning models for IMTS.
• This study is the first to predict retinopathy complications using deep learning on longitudinal data in type 1 diabetes.