Artificial intelligence may offer insight into factors determining individual TSH level

Area of Science:

• Endocrinology
• Computational Biology
• Medical Informatics

Background:

• Serum Thyrotropin (TSH) levels are critical indicators of thyroid function.
• Previous studies have explored TSH determinants using various statistical methods.
• Machine learning applications in large population databases like NHANES for TSH prediction remain underexplored.

Purpose of the Study:

• To comparatively analyze machine learning methods for predicting TSH levels.
• To evaluate the performance of different models in classifying TSH into normal, low, and high categories.
• To identify key predictor variables influencing TSH levels within a large population dataset.

Main Methods:

• Comparative analysis of Linear Regression, Random Forest, Support Vector Machine, Multilayer Perceptron, and Stacking Regression.
• Prediction and classification of TSH levels using Free T4, Anti-TPO antibodies, T3, BMI, Age, and Ethnicity as covariates.
• Evaluation of model performance using coefficient of determination (r2) and Area Under the Curve (AUC).

Main Results:

• Random Forest, Gradient Boosting, and Stacking Regression achieved the highest r2 value of 0.13 for TSH prediction, with a mean absolute error of 0.78.
• Anti-TPO antibodies emerged as the most significant predictor of TSH levels, followed by Age, BMI, T3, and Free-T4.
• Random Forest demonstrated superior performance in classifying TSH levels, achieving AUC values of 0.61 for low TSH, 0.61 for normal TSH, and 0.69 for elevated TSH.

Conclusions:

• Machine learning models, particularly Random Forest, are effective tools for predicting and classifying TSH levels.
• Anti-TPO antibodies are a crucial factor in TSH level determination.
• AI and machine learning can provide valuable insights into the hypothalamic-pituitary-thyroid axis and guide therapeutic decisions for thyroid hormone dosing.