PhysioOmni
收藏资源简介:
PhysioOmni是一个用于多模态生理信号分析的通用基础模型,旨在处理任意缺失模态。该模型通过训练一个解耦的多模态分词器,实现模态无关和模态特定目标的掩码信号预训练,从而在保持与任意缺失模态兼容性的同时,提取通用表示。PhysioOmni在情绪识别、睡眠阶段分类、运动预测和精神负荷检测等四个下游任务上进行了广泛的实验,结果表明,PhysioOmni在保持对缺失模态的强鲁棒性的同时,实现了最先进的性能。
PhysioOmni is a general-purpose foundational model for multimodal physiological signal analysis, designed to handle arbitrarily missing modalities. This model trains a decoupled multimodal tokenizer to perform masked signal pre-training for both modality-agnostic and modality-specific objectives, thereby extracting universal representations while maintaining compatibility with arbitrarily missing modalities. Extensive experiments have been conducted on PhysioOmni across four downstream tasks, namely emotion recognition, sleep stage classification, motion prediction, and mental load detection. The results demonstrate that PhysioOmni achieves state-of-the-art performance while maintaining strong robustness against missing modalities.




