Authentication based on electrocardiography signals and machine learning

Abstract

Abstract Among the information security problems involved in Telemedicine Information Systems (TMIS), the authentication area of the various entities involved has been extensively discussed in recent years and shown a wide range of possibilities. The problems caused by the application of inadequate authentication processes may lead to the death of patients who depend on Mobile Healthcare (M-Health) services. User authentication can be based on several physiological traits (e.g., iris, retina, and fingerprint) for biometric recognition, including electrocardiography (ECG) signals. Some ECG patterns are relatively robust to daily changes associated with normal heart rate variability. In fact, the relative lengths of PQ, QR, and RS intervals, as well as Q, R, and S relative amplitudes constitute individual traits. A few studies have succeeded in using ECG signals as an accurate authentication input and offered some advantages in comparison to biometrics traditional approaches. ECG-based user authentication can be built on Machine Learning (ML) models, used for classification purposes and reductions in distortions caused by misinterpretation of ECG data. Among the ML models adopted for ECG signals classification, ensembles have shown a good research opportunity. Random Under-Sampling Boosting (RUSBoost), a boosting algorithm not yet explored (to the best of our knowledge) for such a problem, can achieve comparatively high performance after a supervised training stage, even from relatively few training examples. This manuscript reports on a comparison of RUSBoost with Nearest Neighbour Search (NNS) regarding the classification of ECG signals for biometric authentication applications. The two ML techniques were compared by a random subsampling technique that considers four analysis metrics, namely accuracy, precision, sensitivity, and F1-score. The experimental results showed the better performance of RUSBoost regarding accuracy (97.4%), sensitivity (96.1%) and F1-score (97.4%). On the other hand, NNS provided better precision (99.5%).