personal authentication by SINGLE- CHANNEL ecg

Hanoi University of Science and Technology Telecommunication and Electronics Department personal authentication by SINGLE- CHANNEL ecg 4/28/23 Students: Class: Vu th i m in h BME K58 Instructor: DR Nguyen viet dung Purposes Research about biometric using ECG signal To authenticate a person lead to identify person in the future 4/28/23 4/28/23 Background information • Block diagram • ECG acquisition • Pre-processing • Feature extraction • Classifcation • Results • Conclusion • outline Background information Biometric authentication: "Are you indeed Mr or Mrs A?" Biometric Biometrics identifcation •    False Acceptance Rate (FAR)  "Who are you?" FAR =   False Reject Rate (FRR)  FRR =  4/28/23 Background information Bio- signal A material carrier of the information about the state of the analyzed biological systems Give more detailed characteristics about the system ECG Non- electric bio- signals 4/28/23 Electrical bio-signals Background information ELECTROCARDIOGRAM (ECG)  f: 0.05 Hz- 100 Hz  A: 1- 10 mV ( dynamic range)  peaks and valleys: P, Q, R, S, T 4/28/23 Background information ELECTROCARDIOGRAM (ECG)  PR: 0.12- 0.25s  QRS: 0.08- 0.12s  QT: 0.35-0.44s  ST: 0.05- 0.15s 4/28/23 Block diagram Classifcation Feature extraction Pre- processing Classify ‘A’ or ‘not A’ Find features ECG Record P, Q, R, S, T peaks Record from Kardia Filter to remove noises Statistical data mobile Digitize data 4/28/23 ECG Acquisition  Kardia mobile o To converts electrical impulses from fngertips into ultrasound signals transmitted to the mobile device’s microphone o Specifcations 4/28/23 ECG Channels Single Channel Input Dynamic Range 10 mV Frequency Response 0.5Hz - 40 Hz A/D Sampling Rate 300 Hz Resolution 16 bit Heart Rate Range 30- 300 bpm Battery Type 3V Coin Cell Battery life 12 months typical use ECG Acquisition  Kardia mobile 4/28/23 10 Pre-processing Filter: Band-pass flter • High-pass flter: 0.05 and 0.5 Hz (low-frequency cutoff ) • Low-pass flters : 40, 100, and 150 Hz (high-frequency cutoff) Filter confguration 0.05–40 Hz 0.5–40 Hz 0.05–100 Hz 0.5–100 Hz 0.05–150 Hz 0.5–150 Hz Cutting segments: from 20s to 50s Choose 4/28/23 14 Pre- processing  Pre-processing : Chebychev band-pass flter 0.5- 40 Hz The power spectrum of original signal 4/28/23 The power spectrum of fltered signal 15 Feature extraction Scaling function of Daubechies Wavelet: A progression {αk; kϵZ} satisfying the following four conditions for all integer N≥2: The expression relating the mother wavelet to the scaling function is: 4/28/23 16 Feature extraction Daubechies 4/28/23 Daubechies 17 Feature extraction Decomposition level • • • • S: original signal/ time series data Ai: approximation low frequency content Di: Detail high frequency content Level decomposition: S = A1+ D1 • Level decomposition 2: S = A2 + D2 +D1 • Level decomposition n- level: S = An + Dn +Dn-1 + Dn-2+…+ D1 4/28/23 18 Feature extraction • Find R peak in decomposition level From R peaks, Find peak >= Find peaks fnd the other 60% max inversely on peaks based value original signal on duration of them Let frst peak be R peak Find the other peaks based on the minimum and maximum 4/28/23 Mean values of P, Q, R, S, T peaks 19 Feature extraction o Mean o Median absolute deviation (MAD) o Standard deviation (SD) 4/28/23 o Skewness: a measure for the degree of symmetry in the variable distribution o Kurtosis: a measure for the degree of tailedness in the variable distribution 20 Classif cation Classifcation Unsupervised Supervised classifcation classifcation Support vector machine (SVM) 4/28/23 K- Nearest Neighbor 21 Results o Total data: 150 samples / 60 samples of authenticated person  120 training data : 60 data of authenticated person, data/ each other ( total 60)  32 testing data: the ratio 16 /16data 4/28/23 o Feature extraction: 11 features  MeanP, MeanQ, MeanR, MeanS, MeanT  Mean, Median, SD, MAD, Skewness, Kurtosis 22 Results Classification Train Validate (Test/ train ratio) Accuracy of train and different validate 10/90 20/80 30/70 40/60 100% 90.9% 91.3% 94.3% 93.5% 99.1% 90.9% 91.3% 97.1% 91.3% KNN SVM ⇒ ⇒ Test/ train ration: 30/ 70 Weighted KNN, Medium Gausisan SVM 4/28/23 23 Results Trained model: Medium Gaussian SVM 4/28/23 Trained model: Weighted KNN 24 Results Results of test 30 samples with 15 samples of authenticated person Trial SVM KNN 87.50 % 78.13% 0.125 0.125 0.000 0.094 Accuracy (%) FAR FRR 4/28/23 25 Conclusion Classifcation accuracy of SVM is higher than one of KNN Base on biometric criteria: KNN is better Some causes affect to results: 4/28/23  Small number of samples  Reconstruction  Timer for cutting segments  Feature extraction method 26 Thank you 4/28/23 27 References o o o o Joseph D B.( n.d ) Biomedical engineering fundamental (3rd edition) Trinity college, Hartfort, Conecticuit ,U.S.A o o o https://heimdalsecurity.com/blog/biometric-authentication/ o o http://matlab.izmiran.ru/help/toolbox/wavelet/ch01_i15.html o Englehart, K.; Hudgins, B.; Parker, P.A.; Stevenson, M Classifcation of the myoelectric signal using time-frequency based representations Med Eng Phys 1999, 21, 431–438 o Phinyomark, A.; Phukpattaranont, P.; Limsakul, C Feature reduction and selection for EMG signal classifcation Expert Syst Appl 2012, 39, 7420–7431 https://www.gemalto.com/govt/inspired/biometrics http://cimss.ssec.wisc.edu/wxwise/class/aos340/spr00/whatismatlab.htm Ruben J.M.D , Neson E.V.P, Erika U.,Wavelet Daubechies (db4) Transform Assessment for WorldView-2 Images Fusion, Distrital University Francisco José de Caldas, Carrera No 40B – 53, Bogotá D.C., Colombia https://www.mathworks.com/discovery/pattern-recognition.html Hargrove, L.; Zhou, P.; Englehart, K.; Kuiken, T.A The effect of ECG interference on pattern-recognition-based myoelectric control for targeted muscle reinnervated patients IEEE Trans Bio Med Eng 2009, 56, 2197–2201 Zardoshti-Kermani, M.; Wheeler, B.C.; Badie, K.; Hashemi, R.M EMG feature evaluation for movement control of upper extremity prostheses IEEE Trans Rehabil Eng 1995, 3, 324–333 4/28/23 28