Few-Shot User-Adaptable Radar-Based Breath Signal Sensing

Abstract

Vital signs estimation provides valuable information about an individual’s overall health status. Gathering such information usually requires wearable devices or privacy-invasive settings. In this work, we propose a radar-based user-adaptable solution for respiratory signal prediction while sitting at an office desk. Such an approach leads to a contact-free, privacy-friendly, and easily adaptable system with little reference training data. Data from 24 subjects are preprocessed to extract respiration information using a 60 GHz frequency-modulated continuous wave radar. With few training examples, episodic optimization-based learning allows for generalization to new individuals. Episodically, a convolutional variational autoencoder learns how to map the processed radar data to a reference signal, generating a constrained latent space to the central respiration frequency. Moreover, autocorrelation over recorded radar data time assesses the information corruption due to subject motions. The model learning procedure and breathing prediction are adjusted by exploiting the motion corruption level. Thanks to the episodic acquired knowledge, the model requires an adaptation time of less than one and two seconds for one to five training examples, respectively. The suggested approach represents a novel, quickly adaptable, non-contact alternative for office settings with little user motion.

Machine Learning in RADAR-based Physiological Signals Sensing: A Scoping Review of the Models, Datasets and Metrics

“…Continuous vital signs monitoring can prevent life-threatening scenarios by providing early warnings before a decline of the patient's status [104]. In this sub-category of papers on value estimation we include papers dealing with the estimation of vital signs, such as HR (n = 21) [17], [22], [25], [28], [29], [52], [55]- [57], [62], [64], [74], [76], [82], [87], [91], [93], [96]- [98], BR [58], [66], [88], [89], [99], [102] (n = 6), both HR and BR [31], [34], [40], [53], [54], [92], [103] (n = 7), BP [65], [67], [75], [83], [85], [100], [101] (n = 7), oxygen saturation (SpO2) [84] (n = 1), and spyrometric indices [61] (n = 1) in a given windows of observation.…”

“…From Table V, we can notice that the authors made an effort to make the dataset balanced between the two sex. The synchronisation between the RADAR recording and the reference is solved by different methods, such as Precision Time Protocol [30], [68], [94], custom LABVIEW program [73], [100], post-processing manual synchronisation based on clearly defined time events [107], binary shared sequences [78], [108] or by matching of the time stamps [17], [40], [65], [66], [72], [87]. Moreover, we find two methods used to deal with synchronisation errors.…”

“…value estimation (n=13) [17], [25], [28], [29], [34], [57], [58], [62], [82], [84], [89], [101], [103]; physiological event detection (n=3) [59], [78], [79] MSE= N i (predictioni − truevaluei) 2 value estimation (n=5) [25], [56], [66], [84], [100]; physiological event detection (n=1) [59] RMSE=…”

“…The filtering row lists the filtering procedures adopted to remove the clutter and noise in the maps which can be further processed by locating the subject in the target location row. Filtering the radar data or locating the subject is a step that appears in 57% of the included studies; typical filters are band pass filters (often Butterworth) in the adequate cutoff frequencies for the wanted vital sign (n = 15) [28], [29], [40], [44], [58], [72], [74]- [76], [79], [81], [91], [93], [95], [99], [103] and derivative filters (n = 4) [36], [44], [81], [91]; typical algorithm for target location are threshold based detection (n = 13) [29], [30], [34], [41], [44], [55], [57], [63], [66], [68], [69], [94], [96]. The augmentation row shows the techniques to increase the dataset size in a synthetic way.…”

“…Augmentation is a necessary steps when there is a lack of real data and the complexity of the algorithms used for the regression or classification task necessitate an increased number of examples. The most typical augmentation procedures include segmentation of the recordings (n = 19) [17], [22], [28]- [30], [32], [34]- [36], [40], [41], [44], [55]- [57], [59], [63], [65], [66], [68], [69], [72]- [76], [78], [79], [81], [82], [87], [89], [91]- [96], [98]- [100], [102], [103], additive noise (n = 3) [29], [44], [96] and rotation of the I/Q data (n = 2) [69], [94]. The last row of reference processing refers to the elaboration of the reference signal needed to reduce the amount of information used in the training process.…”