Quasi-periodic time series (QTS) exists widely in the real world, and it is important to detect the anomalies of QTS. In this paper, we propose an automatic QTS anomaly detection framework (AQADF) consisting of a two-level clustering-based QTS segmentation algorithm (TCQSA) and a hybrid attentional LSTM-CNN model (HALCM). TCQSA first automatically splits the QTS into quasi-periods which are then classified by HALCM into normal periods or anomalies. Notably, TCQSA integrates a hierarchical clustering and the k-means technique, making itself highly universal and noise-resistant. HALCM hybridizes LSTM and CNN to simultaneously extract the overall variation trends and local features of QTS for modeling its fluctuation pattern. Furthermore, we embed a trend attention gate (TAG) into the LSTM, a feature attention mechanism (FAM) and a location attention mechanism (LAM) into the CNN to finely tune the extracted variation trends and local features according to their true importance to yield a better representation of the fluctuation pattern of the QTS. On four public datasets, HALCM exceeds four state-of-the-art baselines and obtains at least 97.3% accuracy, TCQSA exceeds two cutting-edge QTS segmentation algorithms and can be applied to different types of QTSs.

Quasi-periodic time series (QTS) exists widely in the real world, and it is important to detect the anomalies of QTS. In this paper, we propose anautomaticQTSanomalydetectionframework (AQADF) consisting of a two-level clustering-based QTS segmentation algorithm (TCQSA) and a hybrid attentional LSTM-CNN model (HALCM). TCQSA first automatically splits the QTS into quasi-periods which are then classified by HALCM into normal periods or anomalies. Notably, TCQSA integrates a hierarchical clustering and the k-means technique, making itself highly universal and noise-resistant. HALCM hybridizes LSTM and CNN to simultaneously extract the overall variation trends and local features of QTS for modeling its fluctuation pattern. Furthermore, we embed a trend attention gate (TAG) into the LSTM, a feature attention mechanism (FAM) and a location attention mechanism (LAM) into the CNN to finely tune the extracted variation trends and local features according to their true importance to achieve a better representation of the fluctuation pattern of the QTS. On four public datasets, HALCM exceeds four state-of-the-art baselines and obtains at least 97.3 percent accuracy, TCQSA outperforms two cutting-edge QTS segmentation algorithms and can be applied to different types of QTSs. Additionally, the effectiveness of the attention mechanisms is quantitatively and qualitatively demonstrated.

ECG heartbeat classification plays a vital role in diagnosis of cardiac arrhythmia. Traditional heartbeat classification methods rely on handcrafted features and often fail to learn potentially abstract patterns, while current deep learning based methods usually consist of complex convolutional and recursive structures. In this paper, considering the time sequence property of ECG signals, we propose a novel heartbeat classification method based on Transformer, a sequence-to-sequence model with a relatively simple architecture and higher degree of parallelism. To adapt Transformer to our problem, we use only the encoder part of the model for that ECG signals do not have translation signals. We also replace the dropout layers with batch normalization layers, considering our small-size feature space and the natural differences among patients. Further, we fuse handcrafted temporal features with the features learnt by Transformer to better capture rhythmic patterns in ECG signals. We have conducted extensive experiments on the MIT-BIH arrhythmia database using both the original dataset and an augmented dataset with more balanced data. The results show that our model can achieve 99.62% accuracy, 95.09% precision and 94.12% sensitivity on the original dataset and 99.87% accuracy, 99.74% precision and 99.74% sensitivity on the augmented dataset. Besides, we have performed multiple experiments against state-of-the-art methods using their assessment strategies. Experimental results indicate that our model can achieve better performance under most circumstances.

This paper puts forward a LSTM and CNN based assemble neural network framework to distinguish different types of arrhythmias by integrating stacked bidirectional long shot-term memory (SB-LSTM) network and two-dimensional convolutional neural network (TD-CNN). Particularly, SB-LSTM is used to mine the long-term dependencies contained in electrocardiogram (ECG) from two directions to model the overall variation trends of ECG, while TD-CNN aims at extracting local information of ECG to characterize the local features of ECG. Moreover, we design an ensemble empirical mode decomposition (EEMD) based signal decomposition layer and a support vector machine based intermediate result fusion layer, by which ECG can be analyzed more effectively, and the final classification results can be more accurate and robust. Experimental results on public INCART arrhythmia database show that our model surpasses three state-of-the-art methods, and obtains 99.1% of accuracy, 99.3% of sensitivity and 98.5% of specificity.

Electrocardiogram (ECG) signal based arrhythmias classification is an important task in healthcare field. Based on domain knowledge and observation results from large scale data, we find that accurately classifying different types of arrhythmias relies on three key characteristics of ECG: overall variation trends, local variation features and their relative location. However, these key factors are not yet well studied by existing methods. To tackle this problem, we design an attention-based hybrid LSTM-CNN model which is comprised of a stacked bidirectional LSTM (SB-LSTM) and a two-dimensional CNN (TD-CNN). Specifically, SB-LSTM and TD-CNN are utilized to extract the overall variation trends and local features of ECG, respectively. Furthermore, we add a trend attention gate (TAG) to SB-LSTM, meanwhile, add a feature attention mechanism (FAM) and a location attention mechanism (LAM) to TD-CNN. Thus, the effects of important trends and features at key locations in ECG can be enhanced, which is conducive to obtaining a better understanding of the fluctuation pattern of ECG. Experimental results on the MIT-BIH arrhythmias dataset indicate that our model outperforms three state-of-the-art methods, and achieve 99.3% of accuracy, 99.6% of sensitivity and 98.1% of specificity, respectively.

Hypertension is a common cardiovascular disease, which will lead to severe complications without timely treatment. Accurate hypertension identification is essential to preventing the condition deteriorated. However, the state of art hypertension identification methods only extract features from very few aspects, and hence have limited identification accuracy. Furthermore, they only can judge whether the subjects are hypertensive or not, more meaningful information (such as, why the subjects suffer from hypertension) that can help doctors to improve their diagnosis level are absent. In this paper, we propose a class association rules-based method to identify hypertension. Particularly, its key idea is to utilize the relationship existing in multi-dimensional features to characterize hypertension pattern more effectively, in order to improve the identification performance. In addition, it can also generate a set of class association rules (CARs), which can reflect the subjects' physiological status and are proved to be useful for doctors to analyze subject's condition deeply. Experiments based on 128 subjects (61 hypertension patients and 67 healthy subjects) shows that our method outperforms the baseline methods and the accuracy, precision and recall reach 85.2%, 85.0%, and 83.6%, respectively. Additionally, a user study based on five clinicians demonstrates the utility of the generated CARs.

Obstructive sleep apnea (OSA) is regarded as one of the most common sleep-related breathing disorders, which causes various diseases and affects people's daily life severely. Up to now, massive efforts have been devoted to identifying OSA events during sleep based on different signals (e.g., PSG, ECG, nasal airflow and EMG, etc.). However, there still are more or less shortcomings in current studies. In this paper, we propose a novel framework to improve the performance of identifying OSA events. Particularly, the key idea of our framework is to divide each potential event segment (i.e., a data segment that may or may not contain an OSA event) into different phases, from which we further extract fine-grained features to characterize respiratory pattern comprehensively. Concretely, we first automatically locate potential event segments from raw ballistocardiography (BCG) data by identifying arousals. Afterwards, each potential event segment is divided into three phases (i.e., Apnea Phase, Respiratory Effort Phase and Arousal Phase) by an adaptive threshold-based division algorithm. Based on these phases, we further extract and select efficient features that can characterize respiratory pattern from different aspects. Finally, these potential event segments are classified into OSA events or non-OSA events using BP neural network. Experimental results based on a real BCG dataset that contains 3,790 OSA events and 2,556 non-OSA events show that our framework outperforms the baselines and the precision, recall and AUC reach 94.6%, 93.1%, and 0.951, respectively.