Task-Space Domain Adaptation for Cross-Scene Infant Cry Detection.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Infant Health Monitoring

Background:

• Infant cry detection is crucial for care and medical diagnosis.
• Current models face performance issues due to varied deployment data distributions.
• A need exists for robust cry detection across diverse environments like homes and hospitals.

Purpose of the Study:

• To develop an effective domain adaptation method for infant cry detection.
• To enhance cry detection model performance in unlabeled target domains (e.g., hospitals) using labeled source domains (e.g., homes).
• To enable unsupervised adaptation for cry detection models.

Main Methods:

• Proposed a Task-Space Domain Adaptation (TSDA) method.
• Decomposed source domain features using cry detection labels to identify task-relevant features.
• Utilized adversarial learning with nuclear norm-based representation for unsupervised target domain adaptation.

Main Results:

• TSDA improved F1-score by approximately 30% compared to direct cross-dataset testing.
• Achieved 1-12% performance improvement over existing domain adaptation methods.
• Performance approached that of supervised learning on labeled target data.

Conclusions:

• TSDA offers an effective solution for improving infant cry detection across different deployment scenes.
• The method facilitates knowledge transfer from labeled to unlabeled domains without requiring target domain labels.
• This research promotes the practical application of infant cry detection models in diverse clinical and home settings.