Abstract
Several researchers have used a range of Machine Learning (ML) and a few Deep Learning (DL) approaches to predict thyroid hormonal disorders over the years. However, these researchers have recommended a need for the re-evaluation of the ML models and the use of more DL models with feature selection techniques to improve the accuracy of predicting thyroid hormonal disorders. Therefore, this study fills the identified gaps in the literature by comprehensively discussing the data understanding and pre-processing of a reconciled large-sized thyroid disease dataset obtained from Kaggle, which is a secondary source and uses the cleaned dataset with the application of an embedded method that is a features selection technique to develop three ML models, a hybrid model, and four modern DL models, to improve the accuracy of predicting thyroid hormonal disorder by using 80% of the cleaned and balanced dataset for model training, and 20% of the dataset for testing. Based on the findings attained and comparisons of the performances of the developed models using Mean Absolute Error (MAE), BiLSTM is the best-fit model because it has a minimum MAE value of 4.9202. Therefore, this study concludes and recommends BiLSTM as the DL model for the healthcare sector to adopt and be deployed to produce an intelligent medical diagnosis system for an improved prediction of thyroid hormonal disorders.