Electrocardiogram signal analysis is based on detecting a fiducial point consisting of the onset, offset, and peak of each waveform. The accurate diagnosis of arrhythmias depends on the accuracy of fiducial point detection. Detecting the onset and offset fiducial points is ambiguous because the feature values are similar to those of the surrounding sample. To improve the accuracy of this paper’s fiducial point detection, the signal is represented by a small number of vertices through a curvature-based vertex selection technique using polygonal approximation. The proposed method minimizes the number of candidate samples for fiducial point detection and emphasizes these sample’s feature values to enable reliable detection. It is also sensitive to the morphological changes of various QRS complexes by generating an accumulated signal of the amplitude change rate between vertices as an auxiliary signal. To verify the superiority of the proposed algorithm, error distribution is measured through comparison with the QT-DB annotation provided by Physionet. The mean and standard deviation of the onset and the offset were stable as $-4.02\pm 7.99$ ms and $-5.45\pm 8.04$ ms, respectively. The results show that proposed method using small number of vertices is acceptable in practical applications. We also confirmed that the proposed method is effective through the clustering of the QRS complex. Experiments on the arrhythmia data of MIT-BIH ADB confirmed reliable fiducial point detection results for various types of QRS complexes.

Recently, with an increase in the number of healthcare devices, studies measuring and diagnosing electrocardiogram (ECG) signals in daily life are emerging. ECG signal analysis is an essential study area that can diagnose fatal heart abnormalities in humans at an early stage. Conventional signal detection uses one reference beat to diagnose ECG signals; thus, the detection rate is different for each person. In this study, we design a system that can learn a reference beat and diagnose ECG signals in real-time using hardware accelerators with the approximated template-based ECG diagnosis algorithm proposed in the previous study. The proposed algorithm can easily perform personalized learning, increasing the detection rate since it has faster learning time and consumes less memory than the existing algorithm. The learning data, which occupies a small memory space, enables real-time and simultaneous diagnosis of several people. We confirmed that the proposed ECG diagnosis algorithm is suitable for hardware acceleration by accelerating the ECG signal diagnosis and measuring the parallelized result using Alveo field-programmable gate array (FPGA). The ECG diagnosis algorithm, implemented at the FPGA in real-time, can flexibly determine reference beats that vary depending on the person and diagnose each person's signal. The experimental results showed that the time required to diagnose the ECG signals of five people containing 1987 beats takes 5.70 s with software and 0.572 s with hardware accelerators, which is 89.96% shorter than software execution time.

Recently, there have been many studies on the automatic extraction of facial information using machine learning. Age estimation from frontal face images is becoming important, with various applications. Our proposed work is based on a binary classifier that only determines whether two input images are clustered in a similar class and trains a convolutional neural network (CNN) model using the deep metric learning method based on the Siamese network. To converge the results of the training Siamese network, two classes, for which age differences are below a certain level of distance, are considered as the same class, so the ratio of positive database images is increased. The deep metric learning method trains the CNN model to measure similarity based only on age data, but we found that the accumulated gender data can also be used to compare ages. Thus, we adopted a multi-task learning approach to consider the gender data for more accurate age estimation. In the experiment, we evaluated our approach using MORPH and MegaAge-Asian datasets, and compared gender classification accuracy only using age data from the training images. In addition, using gender classification, our proposed architecture, which is trained with only age data, performs age comparison using the self-generated gender feature. The accuracy enhancement by multi-task learning, i.e. simultaneously considering age and gender data, is discussed. Our approach results in the best accuracy among the methods based on deep metric learning on MORPH dataset. Additionally, our method has better results than the state of the art in terms of age estimation on MegaAge-Asian and MORPH datasets.

Recently, as interest in electrocardiogram monitoring has increased, research on real-time ECG signal analysis in daily life using lightweight embedded devices has increased. Abnormal beat detections in ECG signal analysis are an important research area to reduce processing time and cost for cardiac arrhythmia diagnosis. Abnormal beat detections can be divided into feature-based detection and shape-based detection. Feature-based detection finds it difficult to detect reliable fiducial points, and shape-based detection has difficulty detecting abnormal beats that are similar to normal beats. In this paper, we propose template cluster generation and abnormal beat detection using both detection methods. The proposed method shows robust detection of distorted normal beats by generating a template cluster rather than a single template. Moreover, abnormal beats that have normal shape can be detected using the RR interval, which is a highly reliable feature. Experiment results using the MIT-BIH arrhythmia database, provided by Physionet, showed the average processing times to generate a template cluster and detect abnormal beats for the 30-minute signal length were 1.21 seconds and 0.14 seconds, respectively. With manually adjusted thresholds, the specificity and accuracy achieved 93.00% and 97.94%, respectively. In the case of group 1 records obtained relatively stably, the specificity and accuracy achieved 99.27% and 99.44%.