Purpose: Using a stacked ensemble-based classification scheme for predicting second primary cancer in head and neck cancer survivors, risk factors could be identified and be helpful in early screening.

Materials and Methods:
1. Databases
A hospital-based cohort of 27,455 patients diagnosed with head and neck cancer was identified from the database of the Taiwan Cancer Registry from 2009 to 2016, randomly divided in training dataset and testing dataset by 60% and 40% respectively.
2. A Stacked Ensemble-based Classification Scheme
During the training process, nice of base classifiers are trained at once. Those classifiers are modeled via the packages of “stats”, “earth”, “rpart”, “party”, “C5.0”, “evtree”, “RWeka”, “randomForest”, “nnet”, and “kernlab”.
3. Removal Processes of classifiers
A classifier removal scheme is thus designed to sweep out classifiers whose predictions are highly correlated. The inter-classifier correlations of predictions among base classifiers are calculated to form a correlation matrix. Finally, the meta-classifier learns as the base classifiers are right or wrong. It will produce an ultimate combination of base classifiers that is able to improve the perdition performance even more.

Results: The stacked ensemble-based classification had higher balanced accuracy and AUC than other classifiers in this study. In addition, the ensembled-based classifier with removal scheme can increase both balanced accuracy and AUC level in comparison with that without removal scheme.We also identified its positive and negative effect of each feature by using the information of glm coefficient signs. The finding suggested that clinical N stage, sizes of lymph nodes, lymph nodes metastasis to level I-III, combined stage, pathologic T status, dosage to CTV_L, and radiotherapy are negative risk features related to the SPCs of head and neck cancers.

Download