Incorporating natural language processing-derived features from pathology reports into machine learning models to predict differentiated thyroid cancer recurrence

To evaluate the performance of machine learning (ML) models tasked with predicting thyroid cancer recurrence using natural language processing (NLP) and large language model (LLM) derived features.

This was a retrospective study of adult thyroid cancer patients treated at an academic medical center between January 1, 2010 through December 31, 2022. Patients were classified as having cancer recurrence or no recurrence. Novel NLP features were extracted from cytopathology and surgical pathology reports using Term Frequency-Inverse Document Frequency (TF-IDF), Latent Dirichlet Allocation (LDA), and a zero-shot large language model (LLM) classification. Five learning algorithms (Decision Tree, Gradient Boosting, Logistic Regression, Random Forest, and XGBoost) were trained to predict cancer recurrence utilizing a combination of NLP-derived features and structured electronic medical record (EMR) variables. All models were evaluated using stratified 5-fold cross-validation. Model performance was assessed using F1 score and area under the receiver operating characteristic curve (ROC-AUC). The top performing model was hyperparameter-tuned, and feature importance was evaluated using SHapley Additive exPlanations (SHAP) values and gain-based importance.

310 patients with differentiated thyroid cancer diagnosed on final surgical pathology were included in this study. The optimized XGBoost model incorporating EMR variables and LDA-derived features achieved the best performance, with a cross-validated F1 score of 0.647 on the training set and a held-out test F1-score of 0.631 (95% CI 0.509-0.758) and ROC-AUC of 0.741 (95% CI 0.572-0.892). SHAP analysis and gain-based feature importance identified surgical pathology topic-cervical and lateral neck lymph node dissection-as the dominant predictor of cancer recurrence.

Various classifiers including machine learning models utilizing novel NLP features derived from unstructured pathology reports may enable improved prediction of recurrence in patients with differentiated thyroid cancer.