Prediction of Thyroid Disease using Machine Learning Approaches and Featurewiz Selection

Abstract

Thyroid disease is one of the most disturbing hormonal disorders faced by the global population. To help the healthcare industry to diagnose the disorder rapidly and accurately, supervised machine learning algorithms and feature selection were introduced to play an essential role in predicting whether a patient has developed thyroid disease from his/her various characteristics. Therefore, in this work, a new feature selection library was introduced, which was the Featurewiz in the Python library. The goals were to present the performance of the Featurewiz library and to decide on a remarkable model for thyroid disease prediction among several machine learning models, such as Decision Tree, K-Nearest Neighbor, Logistic Regression, Naïve Bayes, Support Vector Classifier, and ensembled machine learning algorithms (Random Forest and Extreme Gradient Boost). A data set consisting of records of thyroid patients in Australia was used to develop the machine-learning models. After the data set was cleaned, exploratory data analysis was carried out. The models were then built in two ways: without feature selection and with feature selection. The feature selection process was conducted by using a new Python library called Featurewiz. The performances of the models from the two operations were evaluated using three performance metrics, including accuracy, F1-score, and AUC (Area Under Curve) value from ROC (Receiver Operating Characteristics Curve). From the two operations, the results are similar in the way that tree-based models, especially those formed by the ensemble method, outperform the statistical models. Initially, in the process without feature selection, the champion model is XGBoost with 99.23% accuracy, while Random Forest ranks second with 98.79% accuracy. However, after the feature selection, the result reveals that the champion model is Random Forest. This model achieves an improvement of 0.66% in accuracy (99.45%), making it the best model from both operations. The model also scores 0.99 and 0.97 in F1-score and AUC values, respectively. The valuable insights gained from this study can serve as a comprehensive framework for machine learning applications in predicting thyroid illness. Additionally, the study highlights the advantageous utilization of the Python feature selection library, Featurewiz. With the combination of Featurewiz and machine learning applications, medical authorities can save time and reduce the risk of misdiagnosis when identifying patients with thyroid disease.