Federated learning enables privacy-preserving stress detection when leveraging wearable devices to monitor physiological indicators without transmitting raw data. However, missing data in federated settings remains a critical challenge, disrupting model training, introducing disparities, and leading to degraded performances among clients. In this work, we explore the impact of missing data on stress detection performances and the gradient magnitudes observed during training across two datasets. Our analysis reveals that the bias induced by missing data directly impacts client performance and is closely related to patterns of gradients during training. To mitigate the effects of missing data in FL setting, we introduce a flexible gradient-aware mechanism that dynamically adjusts data augmentation. Our results show the efficacy of our approach in improving overall stress detection performance while reducing performance disparities among clients.