Heterogeneous Covariates-Aware Pseudo Supervised Meta-Learning for Few-Shot Diabetes Classification

Area of Science:

• Artificial Intelligence in Medicine
• Machine Learning for Healthcare

Background:

• Limited labeled data is a significant barrier to applying artificial intelligence (AI) for diabetes classification.
• Existing methods struggle with data scarcity, impacting the development of robust AI diagnostic tools.

Purpose of the Study:

• To propose a pseudo-label supervised meta-learning algorithm to address data limitations in diabetes classification.
• To enhance AI model performance in diabetes classification using heterogeneous covariates.

Main Methods:

• Clustering algorithms generate pseudo-labels for creating meta-learning tasks within a few-shot learning framework.
• Heterogeneous covariates, including dynamic glucose monitoring data (time/date) and static physiological indicators, enrich model inputs.
• A pseudo-supervised meta-learning algorithm learns features from heterogeneous covariates in a task-driven manner, followed by fine-tuning on real diabetes classification tasks.

Main Results:

• The proposed algorithm achieved a high accuracy of 95.994% on clinical data.
• An F1 score of 91.261% was obtained, demonstrating strong classification performance.
• The method shows significant effectiveness in diabetes classification tasks with limited labeled samples.

Conclusions:

• The developed pseudo-label supervised meta-learning algorithm is superior to state-of-the-art methods for diabetes classification.
• This approach provides an effective strategy for diabetes classification, particularly when dealing with incomplete or scarce labeled data.