FemWear: A Specialized Wearable Foundation Model for Women's Health

FemWear utilizes low-rank residual adapters to repurpose a 21.54M-parameter multimodal wearable backbone for women's health tasks. Training only 1.11% of the encoder parameters, the model improves cycle-phase macro-F1 by 8.15% across six cohorts while maintaining performance on the 32-task OpenMHC benchmark.