LDA pathology features lead thyroid cancer recurrence model

A retrospective study of 310 adult patients with differentiated thyroid cancer tested a machine-learning model that combines structured medical-record variables with features extracted from surgical and cytopathology reports’ free-text. The best-performing method used an XGBoost classifier augmented with latent Dirichlet allocation (LDA) topic features. After hyperparameter tuning, it achieved an F1 score of 0.631 and an area under the receiver operating characteristic curve (AUC) of 0.741 on a held-out test set. Although the work suggests predictive information can be extracted from unstructured pathology text, limitations include poor calibration, a small number of recurrence cases, and no external validation.