Use of an artificial neural network to predict Graves\' disease outcome within 2 years of drug withdrawal

Background Graves' disease (GD) is an autoimmune disorder characterized by hyperthyroidism, which can relapse in many patients after antithyroid drug treatment withdrawal. Several studies have been performed to predict the clinical course of GD in patients treated with antithyroid drugs, without conclusive results. The aim of this study was to define a set of easily achievable variables able to predict, as early as possible, the clinical outcome of GD after antithyroid therapy. Methods We studied 71 patients with GD treated with methimazole for 18 months: 27 of them achieved stable remission for at least 2 years after methimazole therapy withdrawal, whereas 44 patients relapsed. We used for the first time a perceptron-like artificial neural network (ANN) approach to predict remission or relapse after methimazole withdrawal. Twenty-seven variables obtained at diagnosis or during treatment were considered. Results Among different combinations, we identified an optimal set of seven variables available at the time of diagnosis, whose combination was useful to efficiently predict the outcome of the disease following therapy withdrawal in approximately 80% of cases. This set consists of the following variables: heart rate, presence of thyroid bruits, psycological symptoms requiring psychotropic drugs, serum TGAb and fT4 levels at presentation, thyroid-ultrasonography findings and cigarette smoking. Conclusions This study reveals that perceptron-like ANN is potentially a useful approach for GD-management in choosing the most appropriate therapy schedule at the time of diagnosis.