Predicting Elevated TSH Levels in the Physical Examination Population With a Machine Learning Model.

Area of Science:

• Endocrinology and Metabolism
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Elevated thyroid-stimulating hormone (TSH) levels can indicate underlying thyroid dysfunction.
• Early prediction of elevated TSH is crucial for preventing thyroid-related diseases.
• Machine learning offers a promising approach for analyzing complex clinical data to predict health outcomes.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting elevated TSH levels.
• To identify key clinical parameters that predict elevated TSH one and two years post-examination.
• To leverage large-scale physical examination data for predictive health analytics.

Main Methods:

• Utilized data from 12,735 individuals undergoing general physical examinations (2015-2019).
• Analyzed 21 clinical parameters, including demographic and laboratory values (e.g., TPO-Ab, TG-Ab).
• Trained four machine learning models (DT, LR, XGBoost, SVM) to predict elevated TSH at one and two years, assessing feature importance.

Main Results:

• XGBoost demonstrated the highest performance in predicting elevated TSH one year post-enrollment (AUC 0.87).
• FT3/FT4 ratio and TPO-Ab were identified as the most critical predictors in the XGBoost model.
• Predictive performance for elevated TSH two years post-enrollment was not significant across the tested models.

Conclusions:

• An effective XGBoost model was developed to predict elevated TSH levels one year in advance.
• Measurements of FT3 and FT4 can serve as an early warning for elevated TSH.
• This predictive capability can aid in the early prevention of thyroid diseases.