Development and Validation of a Frequency-based Quality Index for Contactless Respiratory Volume Analysis

In contactless continuous monitoring of vital signs, assessing the quality of extracted signals is essential to increase the fidelity of the estimation. For this reason, a set of frequency-based quality indices is studied and analysed on a public dataset of torso displacements related to respiratory volume and recorded remotely under free breathing conditions. All the indices are based on the ratio between the power extracted in a frequency band associated with respiratory activity, against the complementary frequency band associated with noise. Parameters affecting the proposed indices, such as the number of harmonics and the frequency bandwidth around each harmonic, are evaluated to choose the most appropriate frequency band. Moreover, the indices are evaluated theoretically to understand in what range of noise they could properly work. Three experiments are performed: in the first one, white Gaussian noise is added to the considered signals to generate respiratory signals with a signal-to-noise ratio (SNR) ranging from -30 to 27 dB; in the second one, the temporal noise extracted from a depth camera at distances less than 4 meters from a wall is added to the considered signals to simulate a scenario closer to the real application. Mean absolute error (MAE) of less than 2 dB, root mean square error (RMSE) of less than 3 dB, and Pearson’s correlation between 0.97 to 0.99 in the defined range of applicability are the obtained results for the evaluated quality indices. In the third one, an experimental scenario with recordings from Intel RealSense D455 and Kinect V2, the proposed metrics demonstrated to discriminate adequately between lower and higher quality signals determined by subjects at increasing distances from the camera. The metrics showed to recognise the presence of occlusions and motion artefacts; a Respiration-to-Noise Ratio (RNR) with frequency interval of 0.4 Hz showed a great sensitivity to occlusions with a median worsening of the signal quality of 6 to 10 dB; on the other hand, an RNR with frequency interval of 0.05 Hz showed greater sensitivity to low-frequency body movements when comparing the breathing signals extracted from sitting against standing subjects.