Predicting thyroid cancer recurrence using supervised CatBoost: A SHAP-based explainable AI approach.

Area of Science:

• Oncology
• Machine Learning
• Medical Informatics

Background:

• Recurrence prediction in well-differentiated thyroid cancer (WDTC) is challenging.
• Accurate and interpretable models are needed for better patient management.
• Existing predictive models may lack sufficient accuracy or transparency.

Purpose of the Study:

• To develop and evaluate a supervised CatBoost classifier for predicting WDTC recurrence.
• To compare the CatBoost model's performance against other ensemble methods.
• To enhance model interpretability using Shapley Additive Explanations (SHAP).

Main Methods:

• Utilized a dataset of 383 WDTC patients with diverse clinical and pathological variables.
• Preprocessed data, handled missing values, and encoded categorical features.
• Trained and tested models using a 70:30 split, evaluating accuracy and AUC ROC.

Main Results:

• The CatBoost classifier achieved 97% accuracy and 0.99 AUC ROC, outperforming Extra Trees, LightGBM, and XGBoost.
• SHAP analysis identified treatment response, risk stratification, and lymph node involvement as key predictors.
• Local SHAP analysis revealed misclassifications stemmed from overemphasizing single factors.

Conclusions:

• The supervised CatBoost classifier offers high predictive performance and interpretability for WDTC recurrence.
• Integrating multiple predictive factors improves recurrence risk assessment.
• Further validation on larger datasets is needed for robust personalization of thyroid cancer management.