RF-HeartSSL demonstrated substantial performance improvements in heart rhythm monitoring and arrhythmia diagnosis compared with state-of-the-art RF-based cardiac sensing.
Does the RF-HeartSSL self-supervised learning framework improve performance in RF-based cardiac monitoring tasks compared to state-of-the-art methods?
A novel self-supervised learning framework for RF-based cardiac sensing improves heart rhythm monitoring and arrhythmia diagnosis without requiring large annotated datasets.
Radio-frequency (RF) sensing has emerged as a promising technique for cardiac monitoring, enabling fully contactless and operation-free measurements. Although recent machine learning approaches have achieved remarkable improvements over traditional signal processing methods, their reliance on supervised training is constrained by the scarcity of annotated RF data, as RF signals are inherently difficult to interpret and hard to annotate manually. This data-scale bottleneck limits the scalability and generalization of existing methods, motivating the need for self-supervised learning (SSL) that can exploit large volumes of unlabeled RF data. However, existing SSL frameworks cannot be directly applied to RF-based cardiac sensing, as the high interference nature of RF signals and the lack of fine-grained cardiac dynamics modeling together can lead to false self-supervision. In this paper, we propose RF-HeartSSL, a self-supervised learning framework that leverages unlabeled RF data to pre-train radio representations for efficient learning in downstream cardiac monitoring tasks. RF-HeartSSL exploits the inherent consistency within radio signals to formulate self-supervised objectives. At the signal level, it models waveform variations driven by signal interference, enforcing representation consistency through contrastive learning. At the physiological level, it constructs a self-temporal alignment strategy that enforces consistent temporal feature extraction along the cardiac progression, enabling the model to capture fine-grained cardiac dynamics. Together, these two objectives enable the model to learn high-fidelity cardiac representations directly from unlabeled RF data. We implement RF-HeartSSL with 3,147 hours of unlabeled RF data for pretraining and evaluate its effectiveness on a large-scale cohort of 7,338 outpatients. The results demonstrate substantial performance improvements in downstream cardiac monitoring tasks, including heart rhythm monitoring and arrhythmia diagnosis, compared with state-of-the-art RF-based cardiac sensing. Moreover, cross-domain evaluations across diverse environments and external clinical validations further confirm the superior robustness and generalization capability of the proposed framework.
Cai et al. (Mon,) conducted a other in Arrhythmia (n=7,338). RF-HeartSSL vs. State-of-the-art RF-based cardiac sensing was evaluated on Performance in downstream cardiac monitoring tasks including heart rhythm monitoring and arrhythmia diagnosis. RF-HeartSSL demonstrated substantial performance improvements in heart rhythm monitoring and arrhythmia diagnosis compared with state-of-the-art RF-based cardiac sensing.