The Impact of Domain Shift on Predicting Perceived Sleep Quality from Wearables

Area of Science:

• Machine Learning
• Personal Informatics
• Wearable Technology

Background:

• Machine learning models for personal informatics are typically trained on specific user populations, leading to population models.
• These models face performance degradation due to domain shift, caused by variations in data distributions across users and contexts.
• Domain adaptation techniques, such as personalization, can mitigate this issue but often require user-specific data or labels.

Purpose of the Study:

• To quantify the impact of domain shift on population and personalized models in sleep quality recognition.
• To introduce a novel unsupervised domain adaptation approach to address the limitations of existing models.
• To make the new BiheartS dataset available to the research community.

Main Methods:

• The study quantifies domain shift impact on sleep quality recognition models using the new BiheartS dataset.
• Performance of population models and personalized models are compared under domain shift conditions.
• A novel unsupervised domain adaptation method, the cluster-based population model (CBPM), is proposed.

Main Results:

• Domain shift decreases the accuracy of population models by up to 18.54 percentage points on new data.
• Personalized models demonstrate robust performance across datasets but require user-specific data or labels.
• The proposed CBPM achieves accuracy improvements of up to 13.45 percentage points over population models without user data or labels.

Conclusions:

• Domain shift poses a significant challenge for population-based machine learning models in personal informatics.
• Personalized models offer robustness but have practical limitations regarding data requirements.
• The unsupervised CBPM approach effectively mitigates domain shift issues, enhancing model applicability in real-world scenarios.