Vital Sign Monitoring Using FMCW Radar in Various Sleeping Scenarios

Remote monitoring of vital signs for studying sleep is a user-friendly alternative to monitoring with sensors attached to the skin. For instance, remote monitoring can allow unconstrained movement during sleep, whereas detectors requiring a physical contact may detach and interrupt the measurement and affect sleep itself. This study evaluates the performance of a cost-effective frequency modulated continuous wave (FMCW) radar in remote monitoring of heart rate and respiration in scenarios resembling a set of normal and abnormal physiological conditions during sleep. We evaluate the vital signs of ten subjects in different lying positions during various tasks. Specifically, we aim for a broad range of both heart and respiration rates to replicate various real-life scenarios and to test the robustness of the selected vital sign extraction methods consisting of fast Fourier transform based cepstral and autocorrelation analyses. As compared to the reference signals obtained using Embla titanium, a certified medical device, we achieved an overall relative mean absolute error of 3.6% (86% correlation) and 9.1% (91% correlation) for the heart rate and respiration rate, respectively. Our results promote radar-based clinical monitoring by showing that the proposed radar technology and signal processing methods accurately capture even such alarming vital signs as minimal respiration. Furthermore, we show that common parameters for heart rate variability can also be accurately extracted from the radar signal, enabling further sleep analyses.


1
The performance metrics that measure error were computed for each 2-minute measurement independently. When the error measures were combined to compare the total errors between entities, such as participants or different activities, each measured instance was considered equal to others. Thus, the total mean absolute error (MAE) in a set of measurements was calculated as an average MAE weighted with the measurement lengths. In contrast, to couple the MAE with a valid total standard deviation of error s over N measurements, the problem was formulated as s total = ∑ N n=1 ((w n − 1) * s 2 n + w n * (ME n − µ 2 )) ∑ N n=1 (w n − 1) , where w n is the length of the nth measurement, s n its standard deviation, ME its mean error, and µ is 2 the total mean error of the N measurements.

3
Additionally, the total root mean squared error (RMSE) over N measurements was computed with where RMSE 2 n is the mean squared error of the nth measurement.   Figure S1. Examples of interbeat intervals extracted for each subject during relaxed respiration in the supine lying position.

RMSE of the interbeat interval estimates 8
The root mean square errors for each participant and for each lying position with respect to 9 each activity are presented in Tables S1 and S2, respectively. The RMSE results agree well with the 10 corresponding results using MAE, exhibiting similar patterns with respect to participants and activities, 11 while the differences between positions remain trivial.  Reference ECG Figure S2. Example of interbeat intervals extracted from an arrhythmic sequence, extracted from both the reference ECG (visualized at the top) and the radar.

12
As opposed to MAE, the RMSE emphasizes large errors, indicating where the largest errors occur.

13
In this case, most of the largest errors took place in the measurements that also exhibited the largest 14 MAE values. This implies that the radar IBI extraction performs in a stable manner in the various 15 scenarios.

16
It is additionally noted that MAE is more robust to possible arrhythmia episodes, during which 17 there are no beats for the radar to detect. Figure S2 illustrates an arrhythmic ECG sequence together 18 with the interbeat intervals extracted from both the reference ECG and the radar. In the ECG, the 19 arrhythmia is visible as elongated gaps between the otherwise regularly periodic R peaks, whereas in 20 the IBI extracted from the reference, they are visible as peaks. The significant yet relatively rare points 21 of high error between the radar and reference IBI emphasize the root mean square error, while the

RMSE of the respiration rate estimates 28
Tables S3 and S4 present the RMSE when comparing the radar respiration rates to those given by 29 the reference devices. Similarly as with MAE, the hypopnoea simulation (shallow respiration) with 30 small respiratory motion was more difficult to detect accurately than the larger motions. As for the 31 participants, ID003 had the largest MAE. Comparing with the MAE results, this implies that ID003 32 had larger error in individual samples, or that there are more outlier-like values. In addition, the same 33 trend of increased error in lateral positions could again be observed.

34
Further analysis revealed that there were two measurements with exceptionally high RMSE. Both 35 occurred during the shallow respiration period of the hypopnoea simulation; ID003 showed an RMSE    Figure S4 and S5 depict the relationship between individual radar-measured and reference values. 41 Figure S4 illustrates the relation between interbeat interval values (IBI) and Figure S5 Figure S4. The interbeat interval derived from the radar signal presented against the corresponding reference IBI.  ID002  ID003  ID005  ID006  ID008  ID009  ID010 ID011 Ideal Figure S5. The respiration rate derived from the radar signal presented against the corresponding reference respiration rates.