Data-Driven Strategies for Carbimazole Titration: Exploring Machine Learning Solutions in Hyperthyroidism Control

Abstract Background University Hospitals Dorset (UHD) has over 1,000 thyroid patient contacts annually. These are primarily patients with autoimmune hyperthyroidism treated with Carbimazole titration. Dose adjustments are made by a healthcare professional (HCP) based on the results of thyroid function tests, who then prescribes a dose and communicates this to the patient via letter. This is time-consuming and introduces treatment delays. This study aimed to replace some time-intensive manual dose adjustments with a machine learning model to determine Carbimazole dosing. This can in the future serve patients with rapid and safe dose determination and ease the pressures on HCPs. Methods Data from 421 hyperthyroidism patients at UHD were extracted and anonymised. A total of 353 patients (83.85%) were included in the study. Different machine-learning classification algorithms were tested under several data processing regimes. Using an iterative approach, consisting of an initial model selection followed by a feature selection method the performance was improved. Models were evaluated using weighted F1 scores and Brier scores to select the best model with the highest confidence. Results The best performance is achieved using a random forest (RF) approach, resulting in good average F1 scores of 0.731. A model was selected based on a balanced assessment considering the accuracy of the prediction (F1 = 0.751) and the confidence of the model (Brier score = 0.38). Conclusion To simulate a use-case, the accumulation of the prediction error over time was assessed. It was determined that an improvement in accuracy is expected if this model was to be deployed in practice.