Exploiting exercise electrocardiography to improve early diagnosis of atrial fibrillation with deep learning neural networks.

Atrial fibrillation (AF) is the most common type of sustained arrhythmia. It results from abnormal irregularities in the electrical performance of the atria, and may cause heart thrombosis, stroke, arterial disease, thromboembolism, and heart failure. Prior to the onset of atrial fibrillation, most people experience atrial cardiomyopathy which, if effectively managed, can be prevented from progressing to atrial fibrillation. Electrocardiogram (ECG) can show changes in the heartbeats, and is a common and painless tool to detect heart problems. P-waves in exercise ECGs change more drastically than those in regular ECG, and are more effective in the detection of atrial myocardial diseases. In this paper, we propose a deep learning system to help clinicians to early detect if a patient has atrial enlargement or fibrillation. Firstly, a Convolutional Recurrent Neural Network is employed to locate the P-waves in the patient's exercise ECGs taken in the exercise ECG test process. Relevant parameters are then calculated from the located P-waves. Then a Parallel Bi-directional Long Short-Term Memory Network is applied to analyze the obtained parameters and make a diagnosis for the patient. With our proposed deep learning system, the changes of P-waves collected in different phases in the exercise ECG test can be analyzed simultaneously to get more stable and accurate results. The system can take data of different length as input, and is also applicable to any number of ECG collections. We conduct various experiments to show the effectiveness of our proposed system. We also show that the more ECG data collected in the exercise phase are involved, the more effective our system is in diagnosis of the diseases.

Automated ECG classification based on 1D deep learning network.

The standard 12-lead electrocardiogram (ECG) records the heart's electrical activity from electrodes on the skin, and is widely used in screening and diagnosis of the cardiac conditions due to its low price and non-invasive characteristics. Manual examination of ECGs requires professional medical skills, and is strenuous and time consuming. Recently, deep learning methodologies have been successfully applied in the analysis of medical images. In this paper, we present an automated system for the identification of normal and abnormal ECG signals. A multi-channel multi-scale deep neural network (DNN) model is proposed, which is an end-to-end structure to classify the ECG signals without any feature extraction. Convolutional layers are used to extract primary features, and long short-term memory (LSTM) and attention are incorporated to improve the performance of the DNN model. The system was developed with a 12-lead ECG dataset provided by the Kaohsiung Medical University Hospital (KMUH). Experimental results show that the proposed system can yield high recognition rates in classifying normal and abnormal ECG signals.

A fast method for computing the centroid of a type-2 fuzzy set.

Type reduction does the work of computing the centroid of a type-2 fuzzy set. The result is a type-1 fuzzy set from which a corresponding crisp number can then be obtained through defuzzification. Type reduction is one of the major operations involved in type-2 fuzzy inference. Therefore, making type reduction efficient is a significant task in the application of type-2 fuzzy systems. Liu introduced a horizontal slice representation, called the α-plane representation, and proposed a type-reduction method for a type-2 fuzzy set. By exploring some useful properties of the α-plane representation and of the type reduction for interval type-2 fuzzy sets, a fast method is developed for computing the centroid of a type-2 fuzzy set. The number of computations and comparisons involved is greatly reduced. Convergence in each iteration can then speed up, and type reduction can be done much more efficiently. The effectiveness of the proposed method is analyzed mathematically and demonstrated by experimental results.

Data-based system modeling using a type-2 fuzzy neural network with a hybrid learning algorithm.

We propose a novel approach for building a type-2 neural-fuzzy system from a given set of input-output training data. A self-constructing fuzzy clustering method is used to partition the training dataset into clusters through input-similarity and output-similarity tests. The membership function associated with each cluster is defined with the mean and deviation of the data points included in the cluster. Then a type-2 fuzzy Takagi-Sugeno-Kang IF-THEN rule is derived from each cluster to form a fuzzy rule base. A fuzzy neural network is constructed accordingly and the associated parameters are refined by a hybrid learning algorithm which incorporates particle swarm optimization and a least squares estimation. For a new input, a corresponding crisp output of the system is obtained by combining the inferred results of all the rules into a type-2 fuzzy set, which is then defuzzified by applying a refined type reduction algorithm. Experimental results are presented to demonstrate the effectiveness of our proposed approach.

A TSK-type neurofuzzy network approach to system modeling problems.

We develop a neurofuzzy network technique to extract TSK-type fuzzy rules from a given set of input-output data for system modeling problems. Fuzzy clusters are generated incrementally from the training dataset, and similar clusters are merged dynamically together through input-similarity, output-similarity, and output-variance tests. The associated membership functions are defined with statistical means and deviations. Each cluster corresponds to a fuzzy IF-THEN rule, and the obtained rules can be further refined by a fuzzy neural network with a hybrid learning algorithm which combines a recursive singular value decomposition-based least squares estimator and the gradient descent method. The proposed technique has several advantages. The information about input and output data subspaces is considered simultaneously for cluster generation and merging. Membership functions match closely with and describe properly the real distribution of the training data points. Redundant clusters are combined, and the sensitivity to the input order of training data is reduced. Besides, generation of the whole set of clusters from the scratch can be avoided when new training data are considered.

Discovery of fuzzy temporal association rules.

We propose a data mining system for discovering interesting temporal patterns from large databases. The mined patterns are expressed in fuzzy temporal association rules which satisfy the temporal requirements specified by the user. Temporal requirements specified by human beings tend to be ill-defined or uncertain. To deal with this kind of uncertainty, a fuzzy calendar algebra is developed to allow users to describe desired temporal requirements in fuzzy calendars easily and naturally. Fuzzy operations are provided and users can define complicated fuzzy calendars to discover the knowledge in the time intervals that are of interest to them. A border-based mining algorithm is proposed to find association rules incrementally. By keeping useful information of the database in a border, candidate itemsets can be computed in an efficient way. Updating of the discovered knowledge due to addition and deletion of transactions can also be done efficiently. The kept information can be used to help save the work of counting and unnecessary scans over the updated database can be avoided. Simulation results show the effectiveness of the proposed system. A performance comparison with other systems is also given.

Entropy-based generation of supervised neural networks for classification of structured patterns.

Sperduti and Starita proposed a new type of neural network which consists of generalized recursive neurons for classification of structures. In this paper, we propose an entropy-based approach for constructing such neural networks for classification of acyclic structured patterns. Given a classification problem, the architecture, i.e., the number of hidden layers and the number of neurons in each hidden layer, and all the values of the link weights associated with the corresponding neural network are automatically determined. Experimental results have shown that the networks constructed by our method can have a better performance, with respect to network size, learning speed, or recognition accuracy, than the networks obtained by other methods.

A neuro-fuzzy approach for segmentation of human objects in image sequences.

We propose a novel approach for segmentation of human objects, including face and body, in image sequences. Object segmentation is important for achieving a high compression ratio in modern video coding techniques, e.g., MPEG-4 and MPEG-7, and human objects are usually the main parts in the video streams of multimedia applications. Existing segmentation methods apply simple criteria to detect human objects, leading to the restriction of the usage or a high segmentation error. We combine temporal and spatial information and employ a neuro-fuzzy mechanism to overcome these difficulties. A fuzzy self-clustering technique is used to divide the base frame of a video stream into a set of segments which are then categorized as foreground or background based on a combination of multiple criteria. Then, human objects in the base frame and the remaining frames of the video stream are precisely located by a fuzzy neural network constructed with the fuzzy rules previously obtained and is trained by a singular value decomposition (SVD)-based hybrid learning algorithm. The proposed approach has been tested on several different video streams, and the results have shown that the approach can produce a much better segmentation than other methods.

An ART-based construction of RBF networks.

Radial basis function (RBF) networks are widely used for modeling a function from given input-output patterns. However, two difficulties are involved with traditional RBF (TRBF) networks: The initial configuration of an RBF network needs to be determined by a trial-and-error method, and the performance suffers when the desired output has abrupt changes or constant values in certain intervals. We propose a novel approach to over. come these difficulties. New kernel functions are used for hidden nodes, and the number of nodes is determined automatically by an adaptive resonance theory (ART)-like algorithm. Parameters and weights are initialized appropriately, and then tuned and adjusted by the gradient-descent method to improve the performance of the network. Experimental results have shown that the RBF networks constructed by our method have a smaller number of nodes, a faster learning speed, and a smaller approximation error than the networks produced by other methods.