Imputation of Doppler Cardiograms Using a Fine-Tuned Time-series Foundation Model

The robust and continuous measurement of Doppler Cardiograms (DCG) using radar faces challenges due to missing data segments. These missing segments can arise from various factors, such as hardware limitations, data transmission issues, and the removal of artifact-contaminated segments. The variability and non-linearity of DCGs, present significant challenges for traditional methods in imputing missing values. In this paper, we introduce DCG-Moment, a fine-tuned Time-series Foundation Model (TSFM) for DCG imputation. Experimental results demonstrate that the DCG-Moment model consistently outperforms traditional interpolation methods and other deep learning models under various conditions of data loss, achieving lower Mean Absolute Error (MAE) and higher Pearson correlation coefficient (CORR). Moreover, it shows significant improvements in Inter-beat Interval (IBI) estimation tasks. This study highlights the potential of applying large-scale pre-trained models to radar signal processing, with the learned representations offering further potential for completing additional DCG analysis tasks.