Gradient-Aware Data Augmentation for Federated Stress Detection under Data Incompleteness.

Area of Science:

• Health Informatics
• Machine Learning
• Wearable Technology

Background:

• Federated learning (FL) offers privacy-preserving stress detection using wearable physiological data.
• Missing data in FL settings significantly degrades model performance and creates client disparities.

Purpose of the Study:

• To investigate the impact of missing data on stress detection accuracy and training gradients.
• To develop a novel mechanism to address data missingness challenges in federated stress detection.

Main Methods:

• Analysis of missing data effects on stress detection performance and gradient magnitudes across two datasets.
• Introduction of a gradient-aware mechanism with dynamic data augmentation for FL.
• Evaluation of the proposed method's effectiveness in mitigating missing data impacts.

Main Results:

• Missing data introduces bias affecting individual client performance, correlated with gradient patterns.
• The proposed gradient-aware mechanism effectively reduces performance disparities among clients.
• Overall stress detection performance is enhanced by the dynamic data augmentation approach.

Conclusions:

• Missing data is a critical factor influencing federated stress detection, impacting both performance and fairness.
• The gradient-aware mechanism provides a robust solution for handling missing data in FL for stress detection.
• This work advances the reliability and equity of wearable-based federated learning applications.