专为女性健康设计的可迁移穿戴模型，提升经期、症状等预测准确率。
FemWear: A Specialized Wearable Foundation Model for Women's Health
- 用低秩适配器微调预训练多模态模型，仅训练1.11%参数
- 经期阶段预测F1提升8.15%，痛经、情绪、睡眠误差下降超9%
- 适合女性健康研究者，输出一致概率结果便于临床分析
通用可穿戴基础模型未针对女性健康任务设计。本文提出FemWear，一种专用可穿戴基础模型，通过低秩残差适配器和因果任务族头，仅训练239,236参数（占2154万参数编码器的1.11%），保留原始分块投影与Transformer编码器。该模型学习共享的纵向表征，覆盖月经、症状、情绪、睡眠/恢复、自主神经、活动及妊娠相关结果。在六个队列中评估63项可比指标（含33项女性健康指标），同时保持32任务的OpenMHC能力保留基准。固定参与者划分下，周期阶段宏F1提升8.15%，痛经、情绪症状、睡眠问题的平均绝对误差分别降低9.32%、5.80%、9.43%。严格42人嵌套留一参与者审计中，24小时、72小时发病及痛经仍呈正向改善（2.87%、6.35%、2.19%）；相位、情绪、睡眠表现中性或负向，无终点具严格正向校准置信区间。容量匹配实验优于最新日多层感知机，但不及共享GRU或多门专家混合基线。仅训练校准使发病预期校准误差降低84.2%–88.2%，零时间嵌套违规。FemWear支持目标迁移与一致概率输出，但未确立普适性能优势或临床有效性。
原文摘要 · Abstract (English)
General wearable foundation models are pretrained across broad sensor streams and populations, but are not designed around women's-health tasks. We introduce FemWear, a specialized wearable foundation model that parameter-efficiently repurposes a pretrained multimodal wearable backbone. FemWear retains the patch projection and Transformer encoder, training 239,236 parameters (1.11% of a 21.54M-parameter encoder) through low-rank residual adapters and causal task-family heads. It learns one shared longitudinal representation for menstrual, symptom, affective, sleep/recovery, autonomic, activity, and pregnancy-related outcomes. We evaluate six cohorts with 63 comparable primary metrics, including 33 from women's-health cohorts, while retaining the 32-task OpenMHC ability-retention benchmark. On a fixed participant split over three seeds, FemWear improved cycle-phase macro-F1 by 8.15% and reduced mean absolute error for cramps, mood symptoms, and sleep problems by 9.32%, 5.80%, and 9.43%, respectively. In a stricter 42-participant nested leave-one-participant-out audit, 24-hour onset, 72-hour onset, and cramps retained positive changes of 2.87%, 6.35%, and 2.19%; phase, mood, and sleep were neutral or negative, and no endpoint had a strictly positive corrected confidence interval. Capacity-matched experiments outperformed a latest-day multilayer perceptron but not shared-GRU or multi-gate mixture-of-experts baselines. Train-only calibration reduced onset expected calibration error by 84.2--88.2% with zero temporal-nesting violations. FemWear enables targeted transfer and coherent probability outputs for women's-health research, but does not establish universal performance dominance or clinical validity.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。