Finding Order in Chaos: Learning Disentangled Features for mmWave Cardiac Sensing

Recent advances have demonstrated the great potential of millimeter-wave (mmWave) signals for contactless cardiac sensing, enabling various applications such as heart rate variability (HRV) tracking and arrhythmia detection. However, the absence of structural characterization in mmWave cardiac signals, combined with complex interference, makes it difficult to separate cardiac rhythm, cardiac pattern, and external interference. As a result, existing methods predominantly process these entangled components jointly. This missing step of feature disentanglement, which is fundamental to biosignal analysis, severely limits performance and hinders the translation of mmWave cardiac sensing into clinical practice. In this work, we propose the first disentangled feature learning framework for contactless cardiac mmWave sensing. The key innovation lies in the design of a variational feature disentanglement model, which embeds structural priors of mmWave cardiac signals to construct an inductive bias that facilitates effective disentanglement. This design enables the disentanglement of three key components from signals: cardiac-irrelevant interference features, intrinsic cardiac rhythm features, and intrinsic cardiac pattern features, thereby supporting effective cardiac sensing. We evaluate our framework on a large-scale clinical dataset comprising 7,090 outpatients. Experimental results demonstrate superior performance compared to baseline methods, achieving 27.96% and 28.42% average improvements in HRV tracking and arrhythmia detection tasks, respectively, thus highlighting its strong potential to bridge the gap between mmWave sensing technology and real-world clinical applications.