Physiological Radar-Based Sedentary State Classification using Recurrent Neural Networks

While continuous-wave (CW) Doppler radar can be used to unobtrusively monitor human cardiopulmonary vital signs in uncontrolled settings, locomotion or other extraneous motion can sometimes obscure or distort such measurements. This work examines the application of recurrent neural networks (RNNs) to classify sedentary and non-sedentary states, which would facilitate monitoring long-term variation of vital signs specifically limited to sedentary periods. Both Long Short Term Memory (LSTM) and Gated Recurrent Unit (GRU) were applied to output signals from 2.4-GHz quadrature CW Doppler radar measurements of 12 subjects exhibiting sedentary and non-sedentary states. The findings indicate that applying RNNs directly to in-phase (I) and quadrature (Q) (not demodulated) radar outputs provided the most effective means of classifying subject state, acheiving accuracies greater than <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 6 \%}$</tex>. These results clearly demonstrate the value of using RNNs for classifying physiological radar data for subjects that are free to move about, as well as their potential for analyzing quadrature radar without preliminary demodulation.