On signal variability of ECG-based biometric system under practical considerations

This article presents a study of an ECG-based biometric system with hybrid extraction of 10 features in the face of three challenges: variability intra- and inter-subject, ECG acquisition length and population length, arising from practical scenarios. To analyze its results, an ECG database with 1400 records of 10 s each one for 140 healthy subjects with scenarios of ECG variability was generated. The authentication performance with 5 matching metrics was studied, obtaining for 140 subjects an EER = 4.88%, 6.15% and 8.51% with records of 10 s, 5 s and 3 s respectively by means of Mahalanobis distance. Also, the identification performance was analyzed through 3 classifiers, obtaining for the same number of records an IR = 94.63%, 88.07% and 76.75% with records of 10 s, 5 s and 3 s respectively by means of k-NN (k = 1) with Mahalanobis distance.