Introduction: Anti-thyroidal drug (ATD) therapy is a treatment of choice for Graves disease because of its effectiveness and low rate of adverse effect. Response to ATD or remission rate varies among individuals, although there were few useful clues for prediction of treatment response. We applied neural network model to predict response to methimazole, based on clinical symptoms at initial clinical symptoms.
Methods & Design: We reviewed 20 new onset Graves patients and used their symptoms at initial diagnosis. A total of 12 symptoms were classified into 5 categories (general, cardiac, neuromuscular, psychologic, and gastrointestinal). In each category, the number of positive symptoms was considered the score of the category. Response to methimazole was defined as whether serum TSH normalized within 3 months. A total of six variables, 5 symptom categories and sex, were used for a 2-layer neural network model. Among 20 subjects, 15 subjects used in training set, and the other 5 subjects were use in model validation.
Results: The mean age of subjects was 49.3±12.2 years and male subjects were 6 (30%). Our neural network model with 15 subjects as a training set showed sensitivity of 1.0 and specificity of 0.545 with AUC of 0.75 to predict normalization of serum TSH within 3 months. With same model, the estimated scores of the other 5 subjects corresponded to their actual outcomes.
Conclusion: Despite the limitation of small sample size, our neural network model showed promising performance to predict response to methimazole with patients clinical symptoms in new onset Graves patients.
20 May 2017 - 23 May 2017