Compressive Wireless Pulse Sensing

Transcription

1 Compressive Wireless Pulse Sensing CTS 205 Internet of Things Harvard University Kevin Chen Harnek Gulati HT Kung Surat Teerapittayanon Tracking reliable pulse waves for long term health diagnostics

2 Motivation Classification of Heart Health Classification of heart conditions derived from heart rate over time [] Peng, Chung-Kang. "Toward a General Principle of Health and Disease." Toward a General Principle of Health and Disease. Harvard Medical School, Cambridge. 26 Mar Lecture.

3 Motivation Diagnostics based on pulse Heart Rate Apnea Apnea Sleep apnea diagnosis based on changes in heart rate Time (Minutes) Blood Volume Time (Minutes) - Wrist - Finger Blood pressure calibration from phase change of PPG signals in two locations Time (Seconds) 3

4 Message With the recent availability of low-power wireless chips, for the first time, we can monitor pulse waves over a long period of time for applications such as measuring heartrate variability. However, we are still limited by the power budget available on wearables. In this paper, we will show how we can use compressive sensing to reduce power consumption. 4

5 Problem to Solve Power Consumption of Wearables Battery life of heartrate watches Apple Watch Mio Link Mio Alpha Garmin Forerunner Lifetime (hours) Battery consumption of wearables restricts its ability to continuously monitor pulse wave With new low-power wireless chips like BLE and additional power-saving compressive sensing techniques of this paper, it is now feasible for batterypowered wearables to monitor pulse wave continuously for days or even weeks. 5

6 Overview of System Tracking reliable pulse waves for long term health diagnostics Signal Acquisition 6

7 Video Demo of Pulse Wave Reconstruction 7

8 Outline of Presentation. Signal Acquisition Compressive sensing for pulse waves 2. Wireless Data Transmission Forward error correction by interleaving and randomization Adaptations in response to channel quality 3. Signal Recovery Reconstruction of pulse wave through sparse coding Noise removal 8

9 Part One: Signal Acquisition Compressive sensing for pulse waves 9

10 Compressive sensing formulation. Compression by linear projection y S x Obtain = Sensing Matrix y 2. Finding sparse representation of x S D = Sensing Matrix Dictionary z Givens Givens Trained Solve x 3. Reconstruction of x = D Dictionary z Trained Given 8

11 Uniform subsampling to reduce sensor wake-up time y = U We use uniform subsampling as the sensing matrix x The measurements y is a linear projection of x Obtain subsampling

12 Finding the sparse representation of x Givens x y = U D z subsampling Uniform subsampling matrix Trained Dictionary 2

13 Reconstructing the signal from sparse representation x = D Dictionary z Trained Solved Simple Matrix Multiplication 3

14 Experimental Results With a dictionary trained on pulse waves, uniform subsampling performs better than classic compressive sensing methods. Low construction error and very efficient to implement 4

15 Wireless Data Transmission Forward error correction by interleaving and randomization Adaptations in response to channel quality 5

16 Naïve transmission scheme Putting a whole signal segment in one packet is not ideal, because there is no way to recover information without resending Batch of packets # #2 #3 #4 #p A whole segment of signal is lost 6

17 Packet interleaving By interleaving packets, we can recover the information of lost packet from neighboring received packets. Batch of packets # #2 #3 #4 #p Packet Packet 3 Packet n n segment n+ n+3 2n segment 2 segment 3 segment p 7

18 Problems with burst packet loss However, consecutive packet loss still results in consecutive sample loss in each segment Batch of packets # #2 #3 #4 #p Packet Packet 3 Packet n n segment n+ n+3 2n segment 2 segment 3 segment p 8

19 Randomizing packet sending order We can avoid consecutive sample loss by sending packets in randomized order Batch of packets # #2 #3 #4 #p Randomized sending order #0 #3 # #2 #2 st Pkt Sent 2nd Pkt Sent 3rd Pkt Sent 4th Pkt Sent p-th Pkt Sent 9

20 Reconstruction with updated packet transmission scheme We can represent the packet interleaving as a projection Before y = U Uniform subsampling matrix D Dictionary z After w C U D z Matrix that represents how we interleave packets = Channel matrix Uniform subsampling matrix Dictionary 20

21 Reconstruction error with varying packet loss rates Transmission rate is adaptive to packet loss Channel is bad, but we get loss tolerance by simply increasing sampling rate Channel is good, so we can sample at a very low rate. 2

22 Signal Recovery Reconstruction of pulse wave through sparse coding Noise Removal 22

23 Reconstructing the signal w. Use sparse coding to recover z = C Channel matrix U Uniform subsampling matrix Known D Dictionary z Solve 2. Reconstruct x x = D z Dictionary 23

24 Cleaning the signal from outliers There can be outliers caused by movements, sensor voltage change, etc. 24

25 Augmenting the dictionary for noise removal With a little tweak, we can even tolerate corrupted measurements Corrupted measurement w = Dictionary CUD Identity matrix z 25

26 Reconstruction error at different noise levels We can deal with corrupted samples by increasing sampling rate 26

27 Implications of Our Results Pulse Diagnostics Readily Available Long term health monitoring made possible Classification of heart conditions derived from heart rate over time Heart Rate Apnea Apnea Sleep Apnea diagnosis based on changes in heart rate Blood Volume Time (Minutes) Time (Seconds) - Wrist - Finger Blood Pressure Calibration from phase change of PPG signals in two locations 27

28 Summary With new BLE chips, continuous health monitoring is possible for the first time Lower wakeup frequency Signal Acquisition 28

29 Conclusion Due to the recent availability of pulse sensing chips, and low-power wireless chips, for the first time we can monitor pulse waves over along period for applications such as measuring heartrate variations. But we have a challenge of coping with limited power budget available on wearables. We have shown in this paper that we can use compressive sensing to reduce power consumption. 29

30 Training a dictionary with pulses* x = (remove?) D Trained on earlier samples z Dictionary 30

31 3

32 Naïve transmission scheme Putting a whole signal segment in one packet is not ideal, because there is no way to recover information without resending Batch of packets # #2 #3 #4 #p c A whole segment of signal is lost 32

33 33

34 Compressive Wireless Pulse Sensing Kevin Chen Harnek Gulati HT Kung Surat Teerapittayanon Signal Acquisition 34