Research On ECG Arrhythmia Intelligent Recognition Based On Improved Broad Learning System

Arrhythmias are the main cause of sudden cardiac death,not only posing a serious life threat to patients,but also imposing a heavy burden of diagnosis and treatment.The early and accurate diagnosis of arrhythmia is of great significance.In clinical practice,the electrocardiogram(ECG)is the most reliable method for diagnosing arrhythmia.Since most arrhythmias are characterized by intermittent episodes,the capture of abnormal ECG waveforms relies on the analysis of long-term ECG recordings by professional physicians,which is undoubtedly time-consuming and labor-intensive.In addition,prolonged observation tends to make physicians fatigued and eventually leads to misdiagnosis and missed diagnosis.Therefore,the development of high-performance arrhythmia computer-aided analysis systems to improve the accuracy and efficiency of diagnosis has become a research hotspot.In recent years,there has been a proliferation of automatic arrhythmia analysis algorithms based on traditional machine learning and deep learning.Among them,deep learning-based methods output recognition results directly through end-to-end networks,which are not only simple to operate but also easy to obtain higher recognition accuracy.Nevertheless,deep networks need to train complex structures to build recognition models,which consumes a lot of learning time and make it difficult to achieve real-time diagnosis of cardiac arrhythmias.In addition,the severe class imbalance problem and morphological differences in ECG waveforms among different individuals prevent existing studies from achieving satisfactory results in real-world scenarios.Therefore,research on fast algorithms with high resistance to data imbalance and strong adaptability for automatic arrhythmia analysis still faces significant challenges.With its simple structure,real-time speed,and advanced performance,broad learning system(BLS)has received extensive attention in many fields.BLS can directly solve the connection weights of the network by ridge regression theory,thus avoiding the complex iterative learning process faced by the deep structure,which is well suited to meet the requirements of rapid arrhythmia diagnosis.However,the conventional BLS is difficult to solve the above-mentioned problems.To this end,this paper improves BLS from the perspective of improving the resistance to data imbalance and adaptability to achieve fast and accurate recognition of various arrhythmias in different individuals.The main research work of this paper is as follows:1.Most existing algorithms and conventional BLS are designed for balanced datasets,and when facing imbalanced ECG classification scenarios,there exists a model performance that is severely skewed to the normal category with a larger number of samples.To solve the above problem,an arrhythmia recognition algorithm based on the training subset active selection-modified broad learning system(TSAS-MBLS)is proposed.TSAS-MBLS transforms the imbalanced classification task into multiple relatively balanced learning tasks in an iterative manner,and combines multiple weak learners to form a stable classification model.To improve the resistance to data imbalance,a TSAS strategy is designed,in which a certain percentage of high-value samples from each class of the training set are selected to form a training subset based on marginal sampling criteria.The MBLS with output corrected by Sigmoid function is proposed as the base learner for each iteration and is used to identify test heartbeats and provide posterior probabilities for the TSAS process while saving running time.The maximum number of iterations of TSAS-MBLS is proportional to the degree of imbalance in the dataset.When the iteration stops,the final class labels are obtained through a voting strategy.Experimental results demonstrate that compared with conventional BLS,the proposed TSAS-MBLS not only achieves higher overall recognition performance,but also has higher recognition sensitivity for each arrhythmia class with a small number of samples,significantly improving the resistance to data imbalance.2.Although TSAS-MBLS has good performance on imbalanced datasets,it requires additional training subset selection operations,which significantly increases the running time.To improve the real-time performance,an arrhythmia recognition algorithm based on class-specific weighted broad learning system(CSWBLS)is further proposed.The idea of CSWBLS is to optimize the objective function of traditional BLS,and constrain the contribution of each class to the model by constructing the least squares error term by class and assigning weight coefficients.To improve the resistance to data imbalance,a weighting strategy based on the class distribution is designed,in which the weight coefficients are calculated according to the sample size and fine-tuned with a preset scale factor,and classes with smaller sample sizes receive higher weighting coefficients.In addition,the incremental learning algorithms under the weighting strategy are also proposed to guarantee the flexibility of CSWBLS in structure expansion.CSWBLS can still rebuild the model quickly by incremental learning without the need to train from scratch when additional nodes are added.The experimental results show that CSWBLS significantly improves the resistance to data imbalance while retaining essentially all the advantages of conventional BLS.Compared with TSAS-MBLS,CSWBLS not only achieves better performance on the imbalanced arrhythmia classification task,but also run faster.3.In real-world application scenarios,training beats and test beats are usually from different patients(domains),and there are significant morphological differences even for the same class of heartbeats.Although CSWBLS can demonstrate superior performance on imbalanced heartbeat classification tasks without differentiating patients,its recognition performance is severely degraded when the training sets and test sets are from completely different domains.To further improve the adaptability,two CSWBLS-based domain adaptation algorithms(CSWBLS-DA)are proposed,namely,one-step CSWBLS-DA(OCSWBLS-DA)and two-step CSWBLS-DA(TCSWBLS-DA).The design idea of OCSWBLS-DA is to construct weighted least squares error terms of the source and target domains simultaneously to optimize the objective function of CSWBLS for one-step adaptation between these two domains.TCSWBLS transfers knowledge learned from source domain to target domain in two steps.The first step trains a CSWBLS with the common dataset from source domain.The second step considers the output weights of CSWBLS obtained in the first step as a regularization term and constructs the objective function together with the weighted least squares error terms of target domain.Experimental results demonstrate that both algorithms not only inherit the high learning efficiency and resistance to data imbalance of CSWBLS,but also have stronger generalization performance.