FemWear: Specialized Wearable Foundation Model for Women's Health
FemWear is a novel wearable foundation model aimed at women's health, detailed in arXiv:2608.08244. It efficiently repurposes a pretrained multimodal wearable backbone, maintaining the patch projection and Transformer encoder while optimizing only 239,236 parameters (1.11% of a 21.54M-parameter encoder) via low-rank residual adapters and causal task-family heads. This model captures a unified longitudinal representation for various outcomes, including menstrual, symptom, affective, sleep/recovery, autonomic, activity, and pregnancy-related metrics. Evaluations were performed across six cohorts with 63 comparable primary metrics, 33 of which were from women's health cohorts, while also upholding the 32-task OpenMHC ability-retention benchmark. FemWear enhanced cycle-phase macro-F1 by 8.15% and decreased mean absolute error for cramps and mood symptoms on a fixed participant split across three seeds. The research is accessible on arXiv.
Key facts
- FemWear is a specialized wearable foundation model for women's health.
- It repurposes a pretrained multimodal wearable backbone.
- Only 239,236 parameters are trained, which is 1.11% of a 21.54M-parameter encoder.
- Training uses low-rank residual adapters and causal task-family heads.
- It learns a shared longitudinal representation for menstrual, symptom, affective, sleep/recovery, autonomic, activity, and pregnancy-related outcomes.
- Evaluation includes six cohorts with 63 comparable primary metrics, 33 from women's health cohorts.
- Retains the 32-task OpenMHC ability-retention benchmark.
- Improved cycle-phase macro-F1 by 8.15% and reduced mean absolute error for cramps and mood symptoms.
Entities
Institutions
- arXiv