User authentication from ECG Signals using 1D convolutional neural network

Different conventional techniques, including passwords, pins and biometrics, became highly prone to breaches because of the advancements in hacking and spoofing technologies. There is a pressing need for more robust and dynamic authentication methods. The proposed work implements a 1-dimensional Convolutional Neural Network (1D CNN)-based ECG authentication mechanism. We used the Global QRS (GQRS) from the Physikalisch-Technische Bundesanstalt (PTB) and Massachusetts Institute of Technology - Beth Israel Hospital (MIT-BIH) datasets. The average signal within a window created around each QRS complex is then used to construct the average QRS complex waveform. After that, the correlation coefficient between each QRS complex and the mean QRS complex waveform is determined. It generates 100 random permutations of the eight QRS complexes with the highest correlation coefficients. Then, it creates a single signal by concatenating the QRS complex waveforms, normalizing it, and adding it to a temporary data frame for each permutation. The final dataset is supplemented with this interim data frame. Later, a 1D CNN is trained using this extracted data frame. On the PTB dataset, the suggested model produced tutoring accuracy of 98%, Mean precision of 99%, Mean recall of 98%, and Mean F1-score of 98%. On the MIT dataset, the proposed method produced a training accuracy of 98.12%, Mean precision of 100%, Mean recall of 100%, and Mean F1-score of 100%. For the PTB and MIT datasets, the accuracy during authentication was 97.38% and 100%, respectively. These outcomes surpass those of conventional biometric approaches, offering an effective and safe alternative for user identification.