Endocrine Abstracts

University Hospitals Dorset (UHD) has over 1,000 thyroid patient contacts annually. These are primarily patients with autoimmune hyperthyroidism and are treated by titration of Carbimazole. Dose adjustments are made by a healthcare professional (HCP) based on the results of thyroid function tests. Once the test results are available, the HCP decides on a prescription dose and communicates this to the patient, which is time-consuming and introduces delays. This project aims to replace some of the time-intensive manual dose-adjustment with a technological solution, in the form of a mobile app available to the patients themselves. Data of 421 hyperthyroidism patients at UHD was manually extracted and anonymised from patient records. These data were subjected to processing and cleaning stages and a total of 353 (83.85%) were included (of those 79% were female). A wide range of machine learning classification algorithms was tested under different data processing regimes in an iterative approach consisting of an initial model selection followed by a feature selection method to further improve the model performance. All models were assessed using weighted F1 scores (1=best) and Brier scores (smaller is better) to select the best performing model with the highest confidence. Preliminary findings show the best performance is achieved by using a Random Forest approach resulting in good average F1 scores of 0.731. Based on a balanced assessment considering the prediction accuracy (F1=0.755) as well as model confidence (Brier score= 0.366) a model was selected to be initially deployed to the app. This initial model will be further assessed under supervision of experienced clinicians to ensure its safety. It is estimated that with this new patient-centred technology, the number of patient contacts could be reduced by as much as 50-70% and will have a significant positive impact on the HCP workload.