S. S. Chen2, N. Zaborek2, A. R. Doubleday2, K. L. Long2, S. C. Pitt2, R. S. Sippel2, D. F. Schneider2 2University Of Wisconsin,Madison, WI, USA
Introduction: After thyroidectomy, patients require Levothyroxine (LT4), and it may take years of dose adjustments to achieve euthyroidism. During this time, patients encounter undesirable symptoms associated with hypo- or hyperthyroidism. Currently, providers adjust LT4 dose by clinical estimation, and no algorithm exists. The objective of this study was to build a decision tree that estimates LT4 dose adjustments and reduces the time to euthyroidism.
Methods: We performed a retrospective cohort analysis on 320 patients who underwent total or completion thyroidectomies at our institution between 2008 and 2016 and required one or more dose adjustments from their initial postop LT4 dose before attaining euthyroidism. Using the Classification and Regression Tree (CART) algorithm, we built various decision trees from patient characteristics that estimated the dose adjustment to reach euthyroidism. We evaluated tree accuracy with repeated 10-fold cross validation. The most accurate decision tree was developed on a training set of 214 patients, with the remaining 106 patients making up the evaluation set. We compared the accuracy of the decision tree to the actual dose adjustments made by an expert provider and to a naïve system that increased or decreased the dose by 12.5 mcg based on patient TSH.
Results: In our study cohort, an expert provider adjusted LT4 doses, and achieved euthyroidism after one dose adjustment for 156 patients (48.8%), two dose adjustments in 90 patients (28.1%), and three or more dose adjustments in 74 patients (23.1%). Figure 1 shows the most accurate decision tree using TSH values at first dose change (mean absolute error = 13.0 mcg). In comparison, the naïve system had an absolute error of 17.2 mcg, and the expert provider had an absolute error of 11.7 mcg. In the evaluation dataset (106 patients), the decision tree correctly predicted the dose adjustment within the smallest LT4 dose increment (12.5 mcg) 79 of 106 times (75%, CI = 65% – 82%). In comparison, expert provider estimation correctly predicted the dose adjustment 76 of 106 times (72%, CI = 62% – 80%). When measuring dose error within two LT4 dose increments (25 mcg), the decision tree was correct 97 of 106 times (92%, CI = 84% – 96%), whereas the expert provider was correct 93 of 106 times (88%, CI = 80% – 93%).
Conclusion: A decision tree predicts the correct LT4 dose adjustment with an accuracy exceeding that of a completely naïve system and comparable to that of an expert provider. Since this tree-based algorithm approximates an expert provider’s accuracy in adjusting LT4 doses, it can assist providers inexperienced with LT4 dose adjustment.
