Extending Anxiety Detection from Multimodal Wearables in Controlled Conditions to Real-World Environments.

Area of Science:

• Biomedical Engineering
• Machine Learning Applications
• Wearable Technology

Background:

• Machine learning (ML) models are increasingly used with wearable technology for health monitoring.
• Models trained in controlled settings often face challenges when applied to real-world data due to environmental noise and variability.

Purpose of the Study:

• To quantitatively assess the performance of feature-based ML models in real-world conditions using wearable sensor data.
• To investigate the robustness of these models against diverse environmental noise and analyze the impact of transfer learning.

Main Methods:

• Utilized feature-based ML models (XGBoost, Decision Trees) with real-world data from young adults.
• Analyzed transfer learning effectiveness using standard datasets (e.g., WESAD) for adaptation to complex scenarios.
• Examined feature importance across physiological signals like electrodermal activity (EDA) and electrocardiography (ECG).

Main Results:

• Feature-based models, especially XGBoost and Decision Trees, exhibited significant resilience and maintained accuracy across varying noise levels.
• Transfer learning showed potential but also limitations in adapting models to real-world complexities.
• Integrating multiple physiological data types (EDA, ECG) substantially improved model robustness.

Conclusions:

• Feature-based ML models demonstrate strong promise for practical applications with wearable technology.
• Understanding signal contributions and addressing environmental noise are crucial for enhancing model efficacy.
• Multi-modal physiological data integration is key to robust real-world ML model performance.