Related Experiment Video
Updated: Jun 13, 2026
Comparison of Predictive Performance of Three Lymph Node Staging Systems in Colorectal Signet Ring Cell Carcinoma Based on Machine Learning Model
Published on: April 18, 2025
Machine learning-driven prognostic modeling in head and neck adenoid cystic carcinoma: A retrospective cohort study
Hanan M Qasem1, Abdelrahman M Saad2, Rami Al-Fodeh3
1Faculty of Dentistry, Jordan University of Science and Technology, Irbid, Jordan.
Introduction:
Adenoid cystic carcinoma (ACC) of the head and neck is a rare malignancy characterized by perineural invasion, local recurrence, and distant metastasis. Despite its indolent growth, ACC poses significant challenges in prognosis and management. This study addresses two complementary aims to identify key prognostic factors and evaluating the predictive performance of machine learning models METHODS: A retrospective cohort of 1794 patients diagnosed with head and neck ACC between 2004 and 2015 was extracted from the SEER database. Demographic and clinical variables were analyzed, including age, sex, tumor location, AJCC staging, and treatment modalities. Feature selection involved backward elimination and Bagged CART analysis, with SHAP plots for interpretation. Multiple machine learning classifiers (e.g., neural networks, Random Forest, Support Vector Classifier (SVC)) and regression models were developed and validated using cross-validation. Survival analyses employed univariate and multivariate Cox regression.
Results:
Multivariate Cox regression identified age, T-stage, N-stage, distant metastasis, and chemotherapy as independent predictors of survival. Tumors in the maxillary sinus were associated with the poorest outcomes. Neural networks yielded the highest classification AUC (0.899), while random forest demonstrated overfitting. Logistic regression had the highest precision but lowest sensitivity. SHAP analysis highlighted age and tumor stage as the most influential features. Among regression models, SVR achieved the best performance with an R² of 0.247 and the lowest Mean Absolute Error (MAE) among all regression models evaluated indicating reliable survival duration predictions.
Conclusion:
Machine learning models effectively stratify risk and predict survival in head and neck ACC. Age, stage, and chemotherapy are critical determinants of prognosis. Neural networks and SVR emerged as optimal tools for classification and regression tasks, respectively.
