Time-Series Foundation Model Embeddings as Means for Physiological Feature Extraction

Physiological time-series, such as electrocardiograms, photoplethysmograms, or respiratory waveforms are widely used for clinical assessment, prognosis and monitoring. Recently, time-series foundation models (TSFMs) have demonstrated strong performance across various downstream tasks in the medical domain. However, it remains unclear whether TSFM embeddings include physiologically meaningful information. In this study, we evaluate TSFM embeddings as robust proxies for traditional clinical biomarkers, demonstrating that they provide a significant representation premium over raw signals by mapping temporal data into a learnable space. We demonstrate that training regressors to predict clinically informative diagnostic features. Addressing the limitation that TSFM embeddings lack amplitude information due to normalization, we propose re-injecting global scale statistics to the embeddings. Experimental results show that embedding-based feature extraction achieves on average a 17.90% improvement in \textit{relative RMSE} compared to raw signal-based extraction, with particularly strong gains in real-world noisy datasets. Specifically, periodicity-based features exhibited a substantial average performance improvement of 26.02%, while amplitude-based features showed an increase of 18.28%. In comparison, morphology-based features showed comparatively limited improvements, achieving gains of 11.82% from raw signals and 2.60\% from second-derivative signals.