Research And Platform Development Of Robust Classification Algorithm For Cardiovascular Disease

Cardiovascular disease seriously threatens human health,bringing heavy burdens for society and families.Early accurate diagnosis of cardiovascular disease is essential for rescue patient life.Electrocardiogram(ECG)has become the most commonly used cardiovascular diagnosis tool due to its innumerable and low-cost advantages.Under normal circumstances,cardiovascular diseases have no obvious symptoms in the early stage,and the morphological changes in ECG and the disease characteristics are not very obvious.It needs to be carefully discriminated by experts.However,at this stage,medical resources are limited,and experts' handling long-term electrocardiogram will be a very time consuming and difficult task.In this context,the automatic classification of cardiovascular disease came into being.Under this background,the research on automatic classification of cardiovascular diseases has appeared.Combined with the characteristics of ECG signals,this paper analyzes main problems that exist in the current research,and focuses the cardiovascular disease classification system on four aspects: generalization ability,anti-noise,anti-data skew and platform practicability.The research work of this paper is summarized as follows:1.The generalization ability of traditional methods is weak and the recognition accuracy is low.The traditional convolutional neural network(CNN)has the ability to extract abstract and generalized features,but its disadvantages are complex parameters and long training time.In order to solve the above problems,a recognition algorithm of myocardial infarction based on gramian angular fields and principal component analysis network is proposed.The core idea of the algorithm is to extract important features through principal component analysis network(PCANet),which has the advantages of strong generalization,convenient parameter adjustment and short training time.At the same time,in order to take full advantages of PCANet in image processing,the one-dimensional ECG signal is transformed into a single channel picture by using the technology of Gramian Angular Difference Fields(GADF).This method not only retains the amplitude information of the signal,but also retains the time dependence of the signal.After the transformation of the image through PCANet's mining specific information,using the linear support vector machine to complete the recognition and classification.Under the condition of no denoising,a cardiovascular disease classification algorithm with high accuracy is realized.2.In real life,the existing algorithms are not effective under the inter-patient paradigm,and the generalization ability needs to be improved.That is due to the difference of age,heart rate and heartbeat mode among individuals.In order to solve this kind of problem,the algorithm of myocardial infarction recognition based on discrete cosine residual network is proposed.First of all,the discrete cosine transform method is used to obtain the statistical domain information of the heartbeat,and the strong feature extraction ability of the residual network is used to further optimize the statistical features,so as to extract the key category difference features,and finally realize the cardiovascular disease classification algorithm with strong generalization ability and good anti-noise robustness.3.In order to achieve robust multi disease recognition algorithm,traditional methods usually map ECG data to time-frequency domain,abstract domain or statistical domain,and explore the difference characteristics of diseases based on this.These single mapping methods ignore the key influence of other feature domain information on cardiovascular disease recognition.To solve this problem,a multi-layer discrete wavelet dense network algorithm for cardiovascular disease recognition is proposed.The multi-layer time-frequency features are obtained by multiple twodimensional discrete wavelet transform,and each layer features are closely connected with all previous layer features.On the premise of not losing the transformation information,the multi-level time-frequency information is fused,and finally the feature channels are spliced with the abstract domain features extracted from the deep network.The deep fusion features enrich the basis for disease diagnosis.The generalization ability of robust cardiovascular disease recognition is improved.In addition,the combination of Borderline-SMOTE sampling algorithm and Focal loss function is used to improve the anti-data tilt ability of the algorithm from the perspectives of increasing representative minority samples and dynamically adjusting model loss.4.In order to realize the platform application of algorithm prototype,a big data platform is built based on deep learning.To realize the distributed computing of common classic algorithms by combining spark machine learning algorithm library and tensorflowonspark framework.The cardiovascular disease recognition algorithm is deployed in the distributed cluster to improve its operation efficiency.At the same time,the original data and running results are stored in the distributed storage file system(HDFS)of Hadoop cluster to ensure the integrity and reliability of the data.This platform reduces the difficulty of learning and developing algorithms for users,and provides users with a fast solution for algorithm implementation.