HSF-IBI: A Universal Framework for Extracting Inter-Beat Interval from Heterogeneous Unobtrusive Sensors

Figures

• 

  Figure 1. Overview of proposed HSF-IBI method. Multi-channel CVSs are first processed through a set of bandpass filters, followed by another preprocessing step: squaring and low-pass filtering to shift the frequency to a lower position for calculating HarSum spectrums in the HarSum-HR. The HarSum-HR computes the mean heart rate and signal quality of a signal segment (SQIHarsum), and the obtained mean heart rate and signal quality are used as references for selecting the signal channel and optimal band selection (OBS), calculating the template group, determining and selecting the beat-by-beat IBIs. The template group, correlation coefficient function (CCF), and beat-by-beat IBIs are calculated from the signals in the optimal frequency band of the selected satisfactory signal channels. Finally, the ultimate beat-by-beat IBIs are chosen based on SQIbeat.

• 

  Figure 2. Calculating mean HR from FFT spectrum using HarSum-HR. (a) The FFT spectrum of an 8 s BCG signal from a piezoelectric film and the kernel function Kf;f′ at f′=1.67 Hz. (b) The FFT spectrum of the BCG signal and the result after HarSum transformation. (c,d) show the short-time Fourier transform (STFT) spectrum and the time–frequency spectrum result after HarSum-HR of a 200 s BCG signal from a force sensor.

• 

  Figure 3. Schematic diagram of inter-beat interval (IBI) fusion across channels based on SQIbeat (signal quality index for each heartbeat). Panels (a) and (b) display examples from Dataset A and Dataset B, respectively. The last row of signals represents the reference ECG signal, with red dashed lines indicating the reference heartbeats divided by the R-peaks of the ECG. The signals in the rows above are the cardiac vibration signal (CVS) channels selected in Section 3.3. PEC, Radar0, and Radar1 represent CVSs from piezoelectric sensors and two radar channels in Dataset A. Film0, Film1, LC0, and LC1 correspond to electromechanical films and load cells. All IBI values shown in the figure are labeled in milliseconds. The parameter β is the weighting coefficient applied to the SQIbeat of each channel, as defined by the filtering parameters in Equation (2). The β-weighted SQIbeat values are displayed below each channel, with each SQI value corresponding to a single heartbeat. Overall, the final IBI is selected as the heartbeat with the highest SQIbeat that also meets the threshold requirement (θ=0.75) for the preceding heartbeat.

• 

  Figure 4. Dataset description. (a) Sensor placement diagram for Dataset A: Radar mounted above the subject’s head and PEC sensors positioned beneath the mattress. (b) Sensor placement diagram for Dataset B: Four EMFi sensors arrayed on the mattress and four load cells situated under the four bed legs [40]. (c) CVSs samples from Dataset A. (d) CVS sample from Dataset B.

• 

  Figure 5. Histogram of IBI distribution. (a) Distribution of IBIs for a total of 12,481 heartbeats in Dataset A. (b) Distribution of IBIs for a total of 16,375 heartbeats in Dataset B.

• 

  Figure 6. Comparison of the performance of HarSum-HR and the baseline method in calculating mean heart rate. The baseline method employs discrete wavelet transform (DWT) to extract heartbeat components followed by FFT. However, it experienced frequency doubling errors when postural changes altered the signal morphology.

• 

  Figure 7. Bland−Altman plots of IBI estimation errors for two exemplary subjects under different scenarios, along with their corresponding DetIBI and MAEIBI values. (a) Results for Subject #1 in Dataset A. (b) Results for Subject #4 in Dataset B.

• 

  Figure 8. Diagram of the effect of the introduction of OBS on the IBI error. (a) The result of a segment of BCG signal processed by bandpass filters with different cutoff frequencies and the detected heartbeat location. The position on the filtered signal is marked with the same mark on the original signal. (b) IBI sequence line diagram of signal in (a) after processing by different filters.

• 

  Figure 9. Comparison of the proposed framework with other methods. (a) MAEIBI (b) DetIBI.

• 

  Figure 10. Variation in IBI MAE with downsampling ratio from 1000 Hz. (a) Results for Dataset A. (b) Results for Dataset B. Left: Data downsampled from 1 kHz by integer factors. Right: Data downsampled from 1 kHz by integer factors and then interpolated back to 1 kHz using quadratic spline interpolation. Standard deviation between subjects is also shown at each point on the line chart. Downsampling ratios and their corresponding frequencies: 1 (1000 Hz), 2 (500 Hz), 4 (250 Hz), 5 (200 Hz), 10 (100 Hz), 15 (67 Hz), 20 (50 Hz), 25 (40 Hz).