Big data-driven optimal weighted fused features-based ensemble learning classifier for thyroid prediction with heuristic algorithm

Abstract

Diagnosis of thyroid disease is a most important cause in the field of medicinal research and it is a complex onset axiom. Secretion of Thyroid hormone plays a major role in the regulation of metabolism. Hence, it is very significant to predict thyroid disease in the initial stage, which is helpful for preventing more serious health complications due to thyroid cancer. The diagnostic accuracy of machine leaning-based approaches is greater but these techniques require large amounts of data for the diagnosis process. In the conventional approaches, the time needed for the prediction process is also high. Feature engineering is less investigated in conventional models and hence error produced during the prediction process is high. Hence, in this research work, a machine learning-aided thyroid disease prediction technique is designed to provide higher prediction accuracy and reliability. Initially, the thyroid data is gathered from the standard benchmark resources. Next, the data transformation process is carried out to make the data usable for analysis and visualization. After, the features are extracted using Principal Component Analysis (PCA), “One-Dimensional Convolutional Neural Network Model (1DCNN). Moreover, the statistical features are also extracted for getting more relevant information from the data. The three sets of features such as PCA-based, 1DCNN-based and statistical are concatenated and fed to the “optimal weighted feature selection” process, where the optimal features and weights are tuned by an Improved Archimedes Optimization Algorithm (IAOA). Next, the selected optimally fused features are given to the Ensemble Learning (EL) for predicting the thyroid diseases, where the EL with be suggested by incorporating stacking classifier, XGboost, and Multivariate regression classifier. Ensembling of three different classifiers provides higher thyroid disease prediction accuracy and it makes the decision about normal and abnormal classes. Here, the same IAOA is used for optimizing the parameters of every classifier. The investigational outcomes demonstrate that the proposed ensemble classifier provides higher performance than others. Experimental results prove that the thyroid prediction accuracy of the developed EL approach is 96.30%, precision is 99.67% and F1-score is 97.93%, which is more extensive than the state-of-the-art approaches.