Prediction of second malignant neoplasm incidence in a large cohort of long-term survivors of childhood cancers

Background. The ability to predict adverse-event occurrences accurately in long-term survivors of childhood cancer is of high importance in late effects research, both clinically and methodologically. Procedure. This article considers a statistical prediction of future events in a cohort, taking second malignant neoplasm (SMN) incidence in a large cohort of long-term childhood cancer survivors as an example. The method consists of dividing the follow-up period of the cohort into two non-overlapping periods, using the first period as "training data," with which we model the patterns of SMN occurrences in the cohort, and the subsequent period as "testing data," with which we validate the model based on the training data. Future predictions are also applied beyond the testing-data period to. calculate the SMN incidence of the cohort in the next five years for overall and specific types of SMNs. Results. The proposed statistical prediction is shown empirically to perform well with respect to the prediction accuracy. Overall, the models were able to predict the future second cancers rates very well, with exceptions of a few cancer types that had very small observed counts in the testing period. Conclusions. Our proposed statistical method predicts future events in a cohort of long-term childhood cancer survivors and, as such, is a useful tool for late effects research on childhood cancer survivors.