Predicting overall survival in anaplastic thyroid cancer using machine learning approaches.

Abstract

Purpose: Anaplastic thyroid carcinoma (ATC) is a highly aggressive and lethal thyroid cancer subtype with a poor prognosis. Recent advancements in machine learning (ML) have the potential to improve survival predictions. This study aimed to develop and validate ML models using the SEER database to predict 3-month, 6-month, and 12-month (overall survival) OS in ATC patients.

Methods: Clinical and demographic data for patients with ATC from the SEER database (2004-2015) were utilized. Five ML algorithms-AdaBoost, support vector machines, gradient boosting classifiers, random forests, and naive Bayes-were evaluated. The data were split into training and testing sets (7:3 ratio), and the models were tuned using fivefold cross-validation. Model performance was assessed using the concordance index (C-index) and Brier score, with 95% confidence intervals reported.

Results: The gradient boosting model achieved the greatest performance for 3-month survival (C-index: 0.8197, 95% CI 0.7682-0.8689; Brier score: 0.1802), and the AdaBoost model achieved the greatest performance in 6-month survival (C-index: 0.8473, 95% CI 0.7979-0.8933; Brier score: 0.1775). The SVC model showed superior performance for 12-month survival (C-index: 0.8347, 95% CI 0.7866-0.8816; Brier score: 0.1476). Using SHAP with a gradient boosting model, the top five features affecting 6-month OS were identified: surgery, the presence of stage IVC, radiation, chemotherapy, and tumor size. Treatment improved survival, while higher stages reduced survival, with smaller tumors generally linked to better outcomes.

Conclusion: ML algorithms can accurately predict short-term survival in ATC patients. These models can potentially guide clinical decision-making and individualized treatment strategies.