Existing solutions for biometric recognition from electrocardiogram (ECG) signals are based on temporal and amplitude distances between detected fiducial points. Such methods rely heavily on the accuracy of fiducial detection, which is still an open problem due to the difficulty in exact localization of wave boundaries.
Figure 1 - Variability surrounding the P and T waves risk fiducial points detection.
To avoid fiducial points detection, the signal is processed holistically, using second order statistics. Our autocorrelation based method is a very simple and effective approach that does not require any waveform detection. It depends on estimating and classifying the significant coefficients of the Discrete Cosine Transform (AC/DCT) or the Linear Discriminant Analysis (AC/LDA) of the autocorrelation of heartbeat signals.
Figure 2 - Autocorrelated heartbeat signals, recorded over 3 years for the same individual.
Cardiac Irregularities
Recently, we have extended the AC/LDA method to cases of cardiac irregularities, by introducing a novel procedure for classification of healthy vs. arrhythmic ECG windows prior to ECG recognition.
We designed an identification system robust to common cardiac irregularities such as premature ventricular (PVC) and atria (APC) contractions. Criteria concerning the power distribution and complexity of the signals are used to bring to light abnormal ECG recordings, which are not employable for matching. Experimental results indicate a recognition rate of 96.2%, with misclassification taking place mostly among irregular recordings.
An attractive feature of ECG biometric systems, is the possibility of continuous authentication. Since a fresh reading can be acquired every couple of seconds, this signal is suitable for continuous subject authentication in monitoring environments (health care, field agents etc.)
However, time dependency has negative effects. In general, every reading is a nonlinear combination of physiological (anatomical), psychological factors and recording noise. While the anatomic properties of the heart do not experience drastic changes, the psychological changes (ANS effects) may cause severe destabilization of the template.
Figure 5 - In 2 hr monitoring experiment, ECG correlation with the template drops with time.
Our group developed algorithms that can automatically detect destabilization and update the biometric template without risking the identification performance. Simulation results demonstrate a boost in the recognition accuracy when such mechanisms are incorporated to the recognition pipeline.