Attention Detection by Heartbeat and Respiratory Features from Radio-Frequency Sensor

Figures

• 

  Figure 1. (a) Setup showing NCS RF sensors near the heart and diaphragm, and reference thorax and abdomen chest belts and ECG. (b) A subject wearing the setup, with wired NCS sensors. Newer version enables Bluetooth data transfer. (c) A 30 s segment showing raw respiration and heartbeat modulation on NCS signal. (d) Filtered and normalized heartbeat and respiration extracted from the NCS thorax and abdomen sensors, respectively.

• 

  Figure 2. Different clock hand status during the attention task. (a) Normal clock rotation; (b) abnormal clock rotation, after correctly detection by user; (c) incorrect user response due to missed abnormal rotation or spacebar press for a normal clock rotation.

• 

  Figure 3. (a) Frequency spectrum of NCS thorax waveforms during the relaxation routine of 5 min showing variable HR in the range [0.93,1.04] Hz, indicated by the dashed green lines. Similarly, the second harmonic of heartbeat is also distributed around 2 Hz. (b) Second harmonic NCS heartbeat waveform, where IBI is taken as the time for two cycles. (c) Instantaneous HR (60/IBI) in BPM from NCS and ECG showing high correlation. The HR shows strong variation from 50–80 BPM in a resting state.

• 

  Figure 4. NCS respiration waveform showing different features. The maxima and minima represent inspire-end (te) and inspire-begin tb, respectively. IRI = te−te−1 , II = te−tb, EI = tb−te−1, and IV = rte−rtb. Accurate inspire-begin point is estimated for a difficult case with a pause after exhalation at t=162 s.

• 

  Figure 5. Comparison of NCS and reference IBI and IRI data from all participants during relaxed–inattention and attention routines. (a,c) The IBI and IRI scatter plots with denoted Pearson’s correlation coefficients; (b,d) The Bland–Altman plots showing the mean difference m at the center dotted line and the corresponding LoA.

• 

  Figure 6. Flow charts representing our signal processing and ML system architecture. (a) Signal processing and feature extraction. * IBI is extracted from second harmonic heartbeat, requiring re-filtering of the highest SNR waveform around the dominant HR frequency; (b) ML model validation process.

• 

  Figure 7. Test accuracy distribution across the participants using the personal calibration model by (a) NCS and (b) the reference.

• 

  Figure 8. The 5-fold CV confusion matrices showing relaxed–inattention (R) and attention (A) classification by the kNN algorithm using NCS and reference HRV and RWV features in (a,b), and only HRV features in (c,d), respectively. The cells list number of epochs in each category and their overall percentage.

• 

  Figure 9. (a) Average HR change from the R to A routine as a function of RT; each circle represents a participant. (b) Trend analysis for each event across all participants. The PoCR is a function of RT for individual events. Trends of mean(HRPost/HRPre) for each event with two window sizes of 5 s and 10 s are shown in blue lines. The RT histogram across all events is shown with purple bars.

• 

  Figure 10. Box plot showing attention levels of participants as the routine progressed. Each participant is indicated by a scatter point in the 30 s time bin.