An Improved Framework for Detecting Thyroid Disease Using Filter-Based Feature Selection and Stacking Ensemble

Abstract

Machine learning (ML) had become quite significant in determining what type of thyroid disease an individual has and in informing individuals that they may possess one in the last several years.  Single ML models have been examined but their capability to produce accurate forecasts is usually impaired by information imbalance and model-specific constraints.  These problems were addressed using a number of ML techniques applied to a thyroid dataset primarily to increase the accuracy of the predictions.  The most useful features that could make a correct evaluation were found by means of feature selection methods.  Some of the models used to determine thyroid disease were support vector machines (SVM), decision trees (DT), k-nearest neighbors (KNN), logistic regression (LR), and artificial neural networks (ANN).  There was also the ensemble approach of stacking and voting a classifier to combine multiple models to improve performance.  There were significant improvements in the stacking ensemble including LightGBM, SVM, DT, KNN, LR, ANN and vote classifier including Boosted Decision Tree and ExtraTree.  The best of these was the voting predictor whose accuracy was 98% reliable. This demonstrates that ensemble techniques may be useful in addressing imbalance of data and achieve improved outcomes in the prediction of what thyroid disease a patient has.