Heartbeat Signal Imputation in Doppler Radar Using Bidirectional Long Short-Term Memory for Random Body Movement Mitigation

Continuous heart rate (HR) monitoring is vital for early diagnosis and the prevention of patient deterioration. Although clinical gold-standard methods, such as electrocardiography (ECG), offer high accuracy, their contact-dependent nature causes discomfort and restricts mobility during prolonged observation. Non-contact sensing via continuous-wave (CW) Doppler radar provides a compelling alternative by remotely capturing cardiopulmonary micro-movements. However, random body movement (RBM) presents a critical challenge. Because cardiac-induced chest displacements are on the sub-millimeter scale, the inherently weak heartbeat signal is easily submerged by unpredictable, non-periodic transient motions (e.g., spontaneous postural shifts). To overcome this interference, this study proposes a signal imputation framework utilizing a lightweight Bidirectional Long Short-Term Memory (Bi-LSTM) network. The pipeline comprises: 1) RBM detection via a missing data model to isolate motion-corrupted segments; 2) baseline heartbeat extraction using an adaptive narrowband filter bank; 3) predictive imputation using the Bi-LSTM architecture with subject-specific fine-tuning to adapt to physiological trends; and 4) zero-crossing detection for precise beat-to-beat interval (BBI) estimation. As a proof-of-concept evaluated on 15 healthy young adults performing cued spontaneous seated transient motions, the imputed signals achieve a Mean Absolute Error (MAE) of 27.24 ms and a Root Mean Square Error (RMSE) of 40.52 ms, a mean per-subject relative improvement of 38.66% and 58.83%, respectively, over the uncorrected filter bank baseline, and compare favorably with unidirectional LSTM, EEMD, and VMD baselines under matched beat-to-beat evaluation. These results demonstrate the feasibility of Bi-LSTM-based imputation for improving beat-to-beat accuracy of non-contact radar monitoring during transient seated motion.