Facial Expression Recognition (FER) on optomyographic (OMG) smart glasses such as OCOsense enables privacy‑preserving, camera‑free affect sensing for health and well‑being monitoring. Most existing FER systems are fully supervised: they either rely on hand‑crafted expert features with classical machine learning, or on end‑to‑end deep models trained on large labeled datasets. Collecting abundant labeled expressions is costly and hampers deployment in real‑world, personalized settings.\n\nWe investigate Self‑Supervised Learning (SSL) as a way to achieve label‑efficient FER on these glasses. A Vision Transformer is pretrained on unlabeled OCOsense recordings using a masked‑autoencoder (MAE) objective, then fine‑tuned for five‑class FER under person‑independent cross‑validation. Across label budgets ranging from 5% to 100%, pretraining consistently improves performance; with only 5% of the labels, the pretrained model outperforms a from‑scratch model trained with all labels.\n\nOur findings demonstrate that SSL pretraining is a practical approach to reduce labeling cost for smart‑eyewear FER, and we discuss its implications for future personalization and continual adaptation.\n\nReview: The self‑supervised pretraining markedly boosts recognition accuracy under scarce‑label conditions, paving the way for scalable deployment of smart‑glass FER in real‑world applications.