Uncertainty Estimation with Data Augmentation for Active Learning Tasks on Health Data.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Data Science

Background:

• Supervised machine learning (ML) in healthcare faces challenges with manual, costly, and time-consuming data labeling for medical sensor signals.
• Active learning strategies aim to reduce labeling effort by querying uncertain samples, but current methods are inefficient.

Purpose of the Study:

• To develop a novel framework for estimating uncertainty in sensor signals using data augmentation.
• To improve the efficiency and effectiveness of active learning in healthcare ML applications.

Main Methods:

• Exploited data augmentation techniques to estimate uncertainty in medical sensor signals.
• Developed a new framework for selecting informative, unlabeled samples for active learning.
• Compared the proposed framework against baseline methods in medical signal classification tasks.

Main Results:

• The proposed framework selects more diverse and informative samples compared to existing methods, achieving up to 50% greater diversity.
• The framework demonstrates superior sample selection by picking unlabeled samples with higher average point distances early in the querying process.
• Augmentation-based uncertainty estimation leads to better decision-making by prioritizing informative signals and diverse features.

Conclusions:

• The novel framework enhances the selection of informative samples for active learning in healthcare ML.
• Improved sample selection through data augmentation-based uncertainty estimation accelerates model training and improves accuracy.
• This approach has the potential to expand the application of ML in real-world healthcare scenarios by reducing data labeling burdens.