A Deep Learning Framework for Predicting Teprotumumab Treatment Response in Thyroid Eye Disease.

Purpose: To develop and evaluate a deep learning-based framework for quantifying thyroid eye disease (TED) severity before and after teprotumumab treatment, an insulin-like growth factor-1 receptor inhibitor, and to create a predictive model for forecasting individual patient responses to therapy. Design: A retrospective cohort study was conducted at a single institution, utilizing an image-based deep learning TED severity quantification model and a regression-based prediction model for teprotumumab treatment response. Subjects: One hundred eighty-four patients with a clinical diagnosis of TED and 44 individuals with normal orbital anatomy (controls) used for model training and validation. Additionally, 19 patients with pretreatment and posttreatment imaging were analyzed for treatment response. Methods: A deep learning classification model integrating computed tomography-based orbital volumetric features and clinical data were developed to classify TED severity (normal, mild, and severe). A severity scoring model converted classification probabilities into a continuous severity metric. A regression-based model then predicted posttreatment severity using pretreatment clinical and imaging variables. Model performance was evaluated using accuracy, area under the receiver operating characteristic curve (AUC), root mean squared error, and coefficient of determination ( Main Outcome Measures: Thyroid eye disease severity scores and predicted treatment response. Results: The classification model achieved an accuracy of 0.81 and an AUC of 0.88 in the primary cohort and an accuracy of 0.81 with an AUC of 0.86 in the treatment response cohort. Within the treatment response group (n = 19), a mean improvement of 0.194 severity points ( Conclusions: This deep learning framework demonstrates promising performance for objective TED severity quantification and exploratory prediction of teprotumumab treatment response. By integrating clinical and imaging data, this approach provides a proof-of-concept for data-driven tools that may support individualized treatment planning in TED. Financial Disclosures: Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.