Transcription

1

2 . 36.9

3

4

5

6

7

8 d = 4 p = 5

9

10

11

12

13

14

15 .8 1.5

16

17 (1 : N) 1 : N 1 : (N + 1) (N + 1) th G I (nt P ) (nt N) (nf P ) (nf N) F AR = nf P I F RR = nf N G T AR = nt P G T RR = nt N I G + I = N T AR = 1 F RR T RR = 1 F AR

18 nt P + nt N Accuracy = G + I N G << I G = 1 I = nt P = G << I G I.1 1.1

19

20 1 2

21 Amplitude 4 Vasoconstriction PPG signal with its characteristic points Systolic Peak Diastolic Peak -1 Inter Pulse Interval (IPI) Diacrotic Notch Diastolic Point Time(s)

22

23 k = k =

24 k =

25 π

26

27

28 + +

29

30

31

32 Figure 3.1: Plux pulse sensor and black Velcro strip used in this study Figure 3.2: OpenSignal Desktop Application Figure 3.3: Bitalino MCU and Placement of pulse sensor on fingertip with black Velcro strip on the go anywhere via Bluetooth. Hence, these devices indeed emulate practical recording devices such as the smart-watch. Followed protocol before each recording: An open call for participation was circulated to invite as many people as possible. Common protocol followed during each recording is as follows. Prior to recording sessions, all participants were asked to sit down on a chair in a suitable position and relax for a while to achieve a normal heart rate. Participants were briefed about objective of this study and its importance. Any questions from them were answered and an informed consent form was signed for data collection. Then, Plux sensor was placed on either their left or right index finger-tip, irrespective of which fingertip was used in earlier sessions in multi-session configurations. Proper sensor contact was ensured by adjusting Velcro strip. It was found that a tight grip with Velcro strip resulted in poor signals. Therefore in all cases, it was made sure that Velcro strip provided moderate grip. Unconstrained subject: To minimize motion artifacts, subjects were requested to avoid any hand 21

33 Amplitude Amplitude Amplitude Amplitude Filtered Relax signal for subject ID 22 Filtered Relax signal for subject ID Time (s) Filtered Relax signal for subject ID Time (s) Filtered Relax signal for subject ID Time (s) Time (s)

34 Amplitude Amplitude Amplitude Amplitude Filtered Relax signal for subject ID 39 Filtered Relax signal for subject ID Time (s) Filtered after exercise signal for subject ID Time (s) Filtered after exercise signal for subject ID Time (s) Time (s) 3

35 Amplitude Amplitude Amplitude Amplitude Amplitude Amplitude Amplitude Amplitude Amplitude Amplitude Filtered signal from session 1 for subject ID Filtered signal from session 1 for subject ID Time (s) Filtered signal from session 2 for subject ID Time (s) Filtered signal from session 2 for subject ID Time (s) Time (s) Filtered signal from sesson 1 for subject ID Time (s) Filtered signal from session 2 for subject ID Time (s) Filtered signal from session 3 for subject ID Time (s) Filtered signal from session 1 for subject ID Time (s) Filtered signal from session 2 for subject ID Time (s) Filtered signal from session 3 for subject ID Time (s)

36 Days Days Time difference between First and Second Session Subject ID Time difference between first, second and third sessions Time difference between first and second session Time difference between second and third session Time difference between first and third session Subject ID

37

38 ± ± ± ±

39

40 4 th.5 4 x norm = x min(x) max(x) min(x)

41 Data Acquisition Filtering Peak Detection Segmentation Pre-Processing Outlier Segments Removal PPG Segments PPG Segments Feature Extraction Feature Reduction PPG Feature Vector Generation PPG Feature Vectors PPG Feature Vectors System Database Train a Classifier Model Trained Classifier Model PPG Feature Vectors Predict with Trained Classifier Model Yes No Score > Threshold? Verification Phase Accept Reject Enrollment Phase

42 P1(f) P1(f) 1 2 X: Y: Frquency plot of Raw PPG signal for subjectid 67 X: Y: Frquency plot of Filtered PPG signal for subjectid 67 X: Y: X: Y: f (Hz) f (Hz)

43 Amplitude Amplitude Amplitude Amplitude x w f s l = 1 + k w s i i th [ ] i 1 l s i s i s 2 i s i f s <.4 f s [ ].3 f s 1.3 f s 1 length(sysp eaks) (j) j th x(z minscan) x(z) 1 Peak detection on filtered signal from session 2 1 Peak detection on filtered signal from session Time (s) Time (s) 1 Peak detection on filtered signal from session 5 1 Peak detection on filtered signal from session Time (s) Time (s)

44 Amplitude Amplitude.5 PPG Segments Waterfall diagram of PPG segments Time (s) Segments Time (s)

45 e jwt F (ω) = x(t).e jwt F (ω) ω ω ST F T (τ, ω) = x(t).w(t τ).e jwt x(t) ψ(t) a > W T x (a, b) W T x (a, b) = 1 a ( ) t b x(t). ψ a a a = 2, 2 2, τ

46 Amplitude Amplitude Daubechies 5 Mother Wavelet 1 PPG Segment Samples Time (s) 32 1 = 32

47 CWT Coefficient CWT of a PPG Segment Samples Scales

48 d = 4 p = 5 p BP (x(t)) = sign(x(t + i p d + 1) x(t)).2 i + sign(x(t + i + d) x(t)).2 i+p i= 1, x + ϵ, sign(x) =, x(t) d p x(t) p d x(t) k t + i + d > k t + i < p + d ϵ

49 n x(t) w n n n n x(t) 1s 1 k = 1 p = (4, 4) d = (1, 1) w = (5, 4) = (1, 2) p = 4 4 d = 1 x(t) 2 2p = w = 5 = 1 l = 1 + k w = w = 4 =

50 ˆR xx [m] = N m 1 i= x[i] x[i + m] ˆR xx [] x[i] x[i + m] m =, 1, 2...M 1 N M << N Z {Z j } K j=1 Z k {z ki } N k i=1 k W

51 J(W ) W = arg W J(W ) = arg W ( W T ) S b W W T S w W S b S w S b = S w = K N k (µ k µ)(µ k µ) T k=1 K N k (z ki µ k )(z ki µ k ) T k=1 i=1 µ k 1 N k Nk i=1 z ki Z k µ J(W ) W W max(w T S b W ) s.t. W T S w W = 1 L(W, λ) = W T S b W λ(w T S w W 1) L(W, λ) W = = S bw = λs w W S w W m = K 1 (S w ) 1 (S b ) S b K 1 S w (S w ) 1 (S b ) K 1 K 1 W z = W T z z W S w 32 S w = 1 S w

52 S w S b S w a S w a = S b a a T S b a/a T S w a a S b a T S b a = S b S w S b W T S b W = Λ W T S w W = I S b S b S w N N K S b = N k (µ k µ)(µ k µ) T = Φ b Φ T b (N N) k=1 Φ b = [ N 1 (µ 1 µ), N 2 (µ 2 µ), N 3 (µ 3 µ)...](n K) K N N Φ b Φ T b Φ T b Φ bv i = λ i v i Φ b Φ b Φ T b Φ bv i = λ i Φ b v i S b = Φ b Φ T b u i = Φ b v i S b u i = λ i u i S b Φ T b Φ b v i S b K K Φ T b Φ b

53 S w S w S w W = arg W J(W ) = arg W ( W T S b W ) η(w T S b W ) + (W T S w W ) η W η η > S w k(, ) = exp η σ ) ( 2 σ 2

54 y i { 1, 1} i R d i = 1...N 1,b,ξ i C N i=1, y i ( T i b) 1 ξ i i = 1, 2,..., N. ξ i ξ i ξ i 1 ξ i > 1 i ξ i C >

55 a i b N y( ) = a i y i i. + b y( ) = i=1 N a i y i k( i, ) + b k( i, ) i=1 a i a i > ( 2 ) k(, ) = exp σ σ

56 k k ν ν ν C N ξ i = C + ξ i + C i i I + i I ξ i C + C I + I C + C

57 C + n + = C n C + C = n n + n + n d d 51 GDI = 1 i p 2 i i p i i

58

59

60 S w 2 6 η σ k( i, j ) = exp ) ( i j 2 σ σ σ σ k( i, j ) i j σ k( i, j ) 1 σ σ σ σ σ η η η σ i 32 η σ σ [1 3, 1 17 ] η =.1 σ

61 [1 5, 1 7 ] σ = 1 5, 1 6, 1 7 η = 1 3, 1 4, 1 5 σ 84 C σ σ σ σ C σ σ C σ σ C + C + σ a i

62 51

63 1 : 4 1 : 8 1 : 16 1 : 32 1 : 64 1 : : 2 : 16 1 : 2 : 32 1 : 2 : 64 1 : 2 : 128, 1 : 4 : 16 1 : 4 : 32 1 : 4 : 64 1 : 4 : : 4 : 136

64 σ η =.1

65 EER (%) DLDA & R-KDA Performance comparison with CWT CWT with DLDA CWT with R-KDA CWT without reduction Single Session Short Time Lapse Across Exercise Two Sessions Long Time Lapse Three Sessions, Train on 1st, Test on 2nd BioSec.Lab PPG Database Configurations Three Sessions, Train on 1st, Test on 3rd Three Sessions, Train on 2nd, Test on 3rd σ = 1 5, 1 6, 1 7 η = 1 3, 1 4, 1 5 η =

66 ϵ

67 1 st 2 nd 1 st 3 rd 2 nd 3 rd

68 η =.1 σ

69 EER (%) LBP 4 points with DLDA LBP 4 points with R-KDA LBP 4 points without reduction LBP 6 points with DLDA LBP 6 points with R-KDA LBP 6 points without reduction DLDA & R-KDA Performance Comparison with 1DMRLBP Single Session Short Time Lapse Across Exercise Two Sessions Long Time Lapse Three sessions, Train on 1st, Test on 2nd Three Sessions, Train on 1st, Test on 2nd Three Sessions, Train on 2nd, Test on 3rd BioSec.Lab PPG Database Configurations p = (4, 4), d = (1, 2) p = 4, d = 1 p = 4, d = 2 n n 1 ϵ =.1 ϵ =.1 p = (4, 4, 4, 4) p = (4, 4) p = (4, 4, 4, 4) p = (4, 4, 4, 4, 2, 2) η = 1 3, 1 4, 1 5 σ = 1 5, 1 6, 1 7 η = 1 4 η =.1 σ 4 6

70 p = (4, 4, 4, 4) p = (4, 4, 4, 4, 2, 2) , 36.9,

71 σ η = 1e 5

72 EER (%) 5 AC with DLDA AC with R-KDA 4 AC without reduction DLDA & R-KDA Performance Comparison with Auto-Correlation Single Session Short Time Lapse Across Exercise Two Sessions Long Time Lapse Three Sessions Train on 1st, Test on 2nd BioSec.Lab PPG Database Configurations Three Sessions Train on 1st, Test on 3rd Three Sessions Train on 2nd, Test on 3rd η = 1 5 η

73 EER (%) 5 Performance Comparison of Best Results from all Features CWT 1DMRLBP AC DWT Single Session Short Time Lapse Across Exercise Two Sessions Long Time Lapse Three Sessions Long Time Lapse, Train on 1st, Test on 2nd BioSec.Lab Database Configurations Three Sessions Long Time Lapse, Train on 1st, Test on 3rd Three Sessions Long Time Lapse, Train on 2nd, Test on 3rd 1

74 FRR FRR FRR FRR FRR FRR FRR ROC Curves Comparison for Single Session CWT LBP AC DWT ROC Curves Comparison for Two Sessions Long Time Lapse CWT LBP AC DWT FAR FAR ROC Curves Comparison for Across Exercise CWT LBP AC DWT ROC Curves Comparison for Short Time Lapse CWT LBP AC DWT FAR ROC Curves Comparison for Three Sessions Long Time Lapse, Train on 1st, Test on 2nd CWT LBP AC DWT ROC Curves Comparison for Three Sessions Long Time Lapse, Train on 1st, Test on 3rd FAR CWT LBP AC DWT ROC Curves Comparison for Three Sessions Long Time Lapse, Train on 2nd, Test on 3rd CWT LBP AC DWT FAR FAR FAR

75

76 p = (4, 4, 4, 4) p = (4, 4, 4, 4, 2, 2)

77 EER (%) 4 Capnobase Database Results with different Feature 35 3 DLDA + Random Forest DLDA + SVM No Reduction + Random Forest No Reduction + SVM CWT LBP 4 LBP 6 AC Features

78 FRR FRR ROC Curves Comparison for Capnobase Database.5 CWT LBP4 LBP6.4 AC DWT ROC Curves Comparison for DEAP Database CWT LBP4 LBP6 AC DWT FAR FAR =

79 EER (%) 35 3 DLDA + Random Forest DLDA + SVM No Reduction + Random Forest No Reduction + SVM DEAP Database Results with different Features CWT LBP 4 LBP 6 AC Features

80 44

81

82

83

84 FRR FRR FRR FRR FRR FRR FRR ROC Curve Comparison for Single Session with CWT.3 DLDA R-KDA.25 Without Reduction ROC Curve Comparison for Two Sessions Long Time Lapse with CWT DLDA R-KDA Without Reduction FAR ROC Curve Comparison for Across Exercise with CWT 1 DLDA.9 R-KDA Without Reduction FAR ROC Curve Comparison for Short Time Lapse with CWT 1 DLDA.9 R-KDA Without Reduction FAR ROC Curve Comparison for Three Sessions Long Time Lapse Train on 1st, Test on 2nd with CWT 1 DLDA.9 R-KDA Without Reduction.8 ROC Curve Comparison for Three Sessions Long Time Lapse Train on 1st, Test on 3rd with CWT 1 DLDA.9 R-KDA Without Reduction FAR ROC Curve Comparison for Three Sessions Long Time Lapse Train on 2nd, Test on 3rd with CWT 1 DLDA.9 R-KDA Without Reduction FAR FAR FAR

85 FRR FRR FRR FRR FRR FRR FRR ROC Curve Comparison for Single Session with 1DMRLBP.3 LBP4 DLDA LBP4 R-KDA.25 LBP4 without Reduction LBP6 DLDA LBP6 R-KDA LBP6 without reduction ROC Curve Comparison for Two Sessions Long Time Lapse with 1DMRLBP LBP4 DLDA LBP4 R-KDA LBP4 without Reduction LBP6 DLDA LBP6 R-KDA LBP6 without reduction FAR ROC Curve Comparison for Across Exercise with 1DMRLBP 1 LBP4 DLDA.9 LBP4 R-KDA LBP4 without Reduction.8 LBP6 DLDA LBP6 R-KDA.7 LBP6 without reduction FAR ROC Curve Comparison for shorttimelapse with 1DMRLBP LBP4 DLDA LBP4 R-KDA LBP4 without Reduction LBP6 DLDA LBP6 R-KDA LBP6 without reduction FAR ROC Curve Comparison for Three Sessions Long Time Lapse, Train on 1st, Test on 2nd with 1DMRLBP 1 LBP4 DLDA.9 LBP4 R-KDA LBP4 without Reduction.8 LBP6 DLDA LBP6 R-KDA.7 LBP6 without reduction ROC Curve Comparison for Three Sessions Long Time Lapse, Train on 1st, Test on 3rd with 1DMRLBP 1 LBP4 DLDA.9 LBP4 R-KDA LBP4 without Reduction.8 LBP6 DLDA LBP6 R-KDA.7 LBP6 without reduction FAR ROC Curve Comparison for Three Sessions Long Time Lapse, Train on 2nd, Test on 3rd with 1DMRLBP 1 LBP4 DLDA.9 LBP4 R-KDA LBP4 without Reduction.8 LBP6 DLDA LBP6 R-KDA.7 LBP6 without reduction FAR FAR FAR

86 FRR FRR FRR FRR FRR FRR FRR ROC Curve Comparison for Single Session with AC 1 DLDA.9 R-KDA Without Reduction ROC Curve Comparison for Two Sessions Long Time Lapse with AC DLDA R-KDA Without Reduction FAR ROC Curve Comparison for Across Exercise with AC 1 DLDA.9 R-KDA Without Reduction FAR ROC Curve Comparison for Short Time Lapse with AC 1 DLDA.9 R-KDA Without Reduction FAR ROC Curve Comparison for Three Sessions Long Time Lapse, Train on 1st, Test on 2nd with AC 1 DLDA.9 R-KDA Without Reduction.8 ROC Curve Comparison for Three Sessions Long Time Lapse, Train on 1st, Test on 3rd with AC 1 DLDA.9 R-KDA Without Reduction FAR ROC Curve Comparison for Three Sessions Long Time Lapse, Train on 2nd, Test on 3rd with AC 1 DLDA.9 R-KDA Without Reduction FAR FAR FAR

87

88

89

90

91

92

93

94