Search

Detecting cardiac abnormalities promptly is critical for preventing unexpected and premature fatalities. In this research, four types of cardiac arrhythmias such as Ventricular Tachycardia, Premature Ventricular Contraction, Normal Sinus Rhythm and Supraventricular Tachycardia are detected from the amassed Physiobank MIT-BIH cardiac arrhythmia database. Dimensionality reduction techniques like Stochastic Neighbour Embedding (SNE), Neighbourhood Preserving Embedding, Linear Local Tangent Space Alignment and Gaussian Process Latent Variable Model are used to reduce the dimension of the ECG signals. The appropriate features of dimensionally reduced ECG signals are selected by the Elephant Search Optimization (ESO) technique. Finally, classification is performed using the relevant classifiers, such as Support Vector Machine, Adaboost, Modest Adaboost based on Ridge Regression (Modest Adaboost.RR), Extreme Gradient Boost (XGboost) and Naïve Bayes (NBC) classifiers. Multiple classifiers without and with ESO feature selection for different cardiac cases have an average classification accuracy of 62.23% and 73.61%, respectively. These multiple classifiers are defined by a set of control parameters known as hyper-parameters, which must be tuned in order to achieve optimal results. Experts have developed many approaches for detecting cardiac arrhythmias, but these multiple classifiers do not always perform well when the usual parameters for machine learning classification models are employed. In this paper, various classifiers are used in conjunction with the Stochastic Gradient Descent (SGD), Particle Swarm Optimization (PSO) and Bayesian Tree-structured Parzen Estimator (BTPE) to enhance the cardiac arrhythmia classification accuracy via hyper-parameter tuning. Multiple classifiers with SGD, PSO and BTPE hyper-parameters tuning techniques for various cardiac cases have an average classification accuracy of 80.13%, 90.67% and 94.96%, respectively. The Classifier's performance is analysed based on metrics like Classification Accuracy, F1 score, Error Rate, Matthew's correlation coefficient, Jaccard Index and Cohen's Kappa Coefficient with and without ESO features selection method and hyper-parameters tuning techniques. The analysis utilizes the MATLAB R2014a software for result evaluation. The results show that the SNE-ESO approach, along with the XGboost-BTPE, achieved the highest classification accuracy of 99.89% for detecting {Ventricular Tachycardia}-{Normal Sinus Rhythm} cases. In terms of classification benchmarks, the results exhibit that the BTPE hyper-parameter tuning technique surpasses the SGD and PSO techniques. [ABSTRACT FROM AUTHOR]

Hyper-parameter optimization is a crucial task for designing kernel-based machine learning models. Their values can be set by using various optimization algorithms. But a data-dependent objective function makes hyper-parameter's configuration changes over time in a dynamic environment. A dynamic environment is an environment where training data keep getting added continuously over time. To find the optimum values for the hyper-parameters in such an environment, one needs to run the optimization algorithm repeatedly over time. But due to the dependency of the objective function on data, the average time complexity of the optimization process increases. This paper work proposes a novel knowledge-based approach that uses particle swarm optimization (PSO) as the base optimization algorithm to optimize the hyper-parameters of support vector machine. We have introduced two major modules for designing this framework—a knowledge transfer module and a drift detection module. The knowledge transfer module in our proposed framework generates knowledge by running PSO and transfers this knowledge to the consequent time instances. On the other hand, the drift detection module is responsible for detecting changes in the objective function when new data get added to the existing data. This drift detection module helps to utilize the transferred knowledge at a particular time instance to reduce the execution time of the overall optimization process. The proposed framework has been evaluated using various standard datasets such as Adult, DNA, Nist-Digits, Segment, Splice, Mushroom and Usps for five consecutive time instances and found the average execution time as 30.72 s, which is better than the execution time 36.89 s recorded using general PSO. We have also found that our proposed framework performs the optimization of hyper-parameters much faster than the other existing approaches such as grid search, chained-PSO, dynamic model selection and quantized dynamic multi-PSO. [ABSTRACT FROM AUTHOR]