Domain Generalization in Biosignal Classification.

Area of Science:

• Machine Learning
• Biosignal Processing
• Artificial Intelligence

Background:

• Machine learning models typically assume training and evaluation data share the same distribution.
• Domain shift, where evaluation data differs from training data distribution, poses a challenge for model generalization.
• Domain generalization aims to develop models robust to unseen data distributions without direct access during training.

Purpose of the Study:

• To propose and evaluate a novel domain generalization method for biosignal data.
• To address the challenge of domain shift in machine learning models applied to biosignals.
• To improve model performance on inaccessible, domain-shifted datasets.

Main Methods:

• A two-stage domain generalization method is proposed.
• Unseen domains are represented using a set of known basis domains.
• Classifier fusion is employed for classifying data from unseen domains.
• The method is demonstrated using a collection of heart sound databases.

Main Results:

• The proposed classifier fusion method achieved accuracy gains of up to 16% across four completely unseen domains.
• The method effectively simplifies the domain generalization process for biosignal data.
• Demonstrated good performance on both unseen and adopted basis domains.

Conclusions:

• The study presents the first investigation into domain generalization for biosignal data.
• The proposed learning strategy effectively learns domain-relevant features while considering class differences.
• The method offers a simplified yet effective approach to domain generalization in biosignals.