Advancing diabetes prediction with a progressive self-transfer learning framework for discrete time series data.

Area of Science:

• Computational Medicine
• Data Science in Healthcare
• Biostatistics

Background:

• Diabetes mellitus is a complex disease often studied using limited data features.
• Existing research frequently overlooks multivariate, multi-instance, and time-series data complexities.
• Challenges in medical datasets include irregular data collection and sparsity.

Purpose of the Study:

• To develop a novel diabetes prediction model integrating complex data types.
• To address data sparsity and irregularity using self-supervised learning and transfer learning.
• To enhance the accuracy of diabetes diagnosis and disease management.

Main Methods:

• Applied bidirectional recurrent imputation for time series (BRITS) and least absolute shrinkage and selection operator (LASSO) for data imputation and feature selection.
• Utilized self-supervised algorithms and transfer learning to handle medical data challenges.
• Developed a novel discrete time-series data preprocessing approach with shifting/rolling windows and modified time resolution.

Main Results:

• Evaluated a progressive self-transfer network for diabetes prediction.
• Demonstrated significant improvements in AUC, recall, and F1 score compared to non-progressive and single self-transfer methods.
• The proposed approach effectively mitigates accumulated errors and captures temporal information.

Conclusions:

• The novel model effectively handles complex multivariate, multi-instance, and time-series data for diabetes prediction.
• The progressive self-transfer network offers a powerful tool for accurate diabetes diagnosis and management.
• Considering data complexities is crucial for advancing predictive modeling in diabetes.