Classification of Heart Sound Biometrics for Active User Authentication and Clinical Cardiac Applications
Classification of Heart Sound Biometrics for Active User Authentication and Clinical Cardiac Applications This research presents a method and structure for active user authentication and cardiac risk diagnosis applications using the biometric of heart sound. Active authentication can maintain confidence that the device owner is the current user without the inconvenience of requiring re-authentication. Heart sound, or photocardiogram (PCG), is an interesting possibility for active authentication because it is constantly available, hard to obtain from another person, and has been shown to be reasonably unique between individuals. Furthermore, clinical cardiology applications currently do not take advantage of the algorithms of heart sound authentication, for example to indicate a change in a patient’s heart sound on an in-home wearable smart device application. Heart sound measurements used in this study are from a public dataset created for heart sound biometric research. The data show marked inconsistencies between samples taken from the same user at different times or on a different day, a known problem with recording biometric data. This research presents two methods for preprocessing, filtering, and classifying imperfect datasets: Phase 1 using frequency-based Mel Frequency Cepstral Coefficients (MFCC) with the MATLAB Classification Learner, and Phase 2 using a custom time-based approach with a python Dichotomy-classifier. Accompanying these methods is a complete software structure to extract features and use machine learning to classify heart sound input from initial capture through classification of individual participants to actively check their heart sound for user authentication and for clinical cardiac applications. This research uses novel and existing methodologies to achieve improved authentication accuracy and to expand the analysis methods available to researchers in the field. The results of time-based feature extraction and dichotomy-classification were an authentication accuracy of 84% (EER = 16%), compared with the previously published accuracy of 63% (EER = 37%) for a structural methodology. Heart sound is different enough between individuals to be part of a multi-factor authentication methodology, and to contribute to diagnosis of cardiac patients for clinical uses.
Applied Mathematics|Computer Engineering|Computer science
Clevenger, Leigh Anne Hodges, "Classification of Heart Sound Biometrics for Active User Authentication and Clinical Cardiac Applications" (2017). ETD Collection for Pace University. AAI10745766.
Remote User: Click Here to Login (must have Pace University remote login ID and password. Once logged in, click on the View More link above)