71635 - Can prediction models aid in the process of selecting treatment for hyperthyroidism?

Abstract Introduction Hyperthyroidism (HT) is treated with antithyroid drugs, radioactive iodine, or surgery. Previous studies have demonstrated that many patients undergo repeated treatment without remission and have impaired quality of life 6-10 years after diagnosis. This study aimed to evaluate different predictive machine learning models as a means of support in selecting treatment for HT. Method The study is based on a dataset of 2916 newly diagnosed HT patients between 2003-2005, who later were invited to participate in a 6-10 year follow-up study. This dataset underwent standard preprocessing for data representation and preparation for learning algorithms. Large Language Models were assessed as an alternative to encode medical records into sentence embeddings. Various regression models were trained and evaluated to predict the probability of a successful treatment based on nine features (Age, Sex, Diagnosis, thyroid-stimulating hormone receptor antibodies, T4-levels, TRAb (reference), TRAb (corrected), TRAb (positive), Thyroid-associated Ophthalmopathy. Result Two different multi-output regression models were trained and evaluated for each pre-processed dataset: Gradient-Boosting (GB) and Random Forest (RF). Root Mean Squared Error (RMSE) and the coefficient of determination, R2 score, were measured for evaluation metrics. RF trained on the traditionally pre-processed dataset showed best result, RMSE 0.1764 and R2 score 0.1460. However, the differences between both evaluated models and the approach in data preparation were negligible. Discussion This study show promising results in using prediction models for assessing the probability of successful treatment in individual patients. Further investigation is necessary regarding the choice of prediction models and how the data are prepared and represented.