Fine-Grained Respiratory Waveform Recovery Based on Millimetre-Wave Radar

Respiratory signs, as key vital signs, play an essential role in long-term monitoring for the diagnosis of respiratory-related diseases. In respiratory sign monitoring, which recovers fine-grained respiratory waveforms rather than just predicting respiratory rate, this fine-grained respiratory volume data provides valuable insights into disease physiology. However, conventional signal processing methods can only obtain coarse respiratory waveforms, and recent methods using variational encoder-decoder architectures suffer from training instability, leading to pattern collapse. To address these issues, we propose an encoder-decoder architecture based on residual adaptive multi-scale LSTM, which enables more stable training, generates finer respiratory waveforms, and further improves the accuracy of predicting respiratory rate. Experimental results show that the method in this paper outperforms the baseline when compared to state-of-the-art methods.