Comparing supervised and semi-supervised machine learning approaches in NTCP modeling to predict complications in head and neck cancer (HNC) patients

Abstract

Head and neck cancer patients (HNC) treated with radiotherapy often suffer from radiation-induced toxicities, most often xerostomia (dry mouth) and dysphagia (difficulty swallowing). As to reduce the risk of toxicities in these patients, Normal Tissue Complication Probability (NTCP) modeling is used to determine the probability to develop toxicities based on patient and treatment characteristics. Currently, most often supervised logistic regression methods are used in NTCP modeling. However, as the toxicity outcomes that are used in NTCP modeling of HNC are often recorded long after treatment started, ‘unlabeled’ data are also available. In these data the patient and treatment characteristics are recorded, but not yet the toxicity outcomes. Semi-supervised methods are able to incorporate unlabeled data in model development and may thereby gain in predictive performance compared to the current supervised logistic regression models. Here, it will be evaluated how current regression models compare to the semi-supervised method of self-training, and to regression models after multivariate imputation by chain equation (MICE) of the unlabeled data. The models were developed for the most common toxicity outcomes in HNC patients, xerostomia and dysphagia, measured at six months after treatment, in a development cohort of 750 HNC patients. The models were externally validated in a validation cohort of 395 HNC patients. It was found that MICE and self-training did not have a gain in performance in terms of discrimination or calibration at external validation compared to current regression models. Therefore, the addition of unlabeled patient data by using the semi-supervised method of self-training or MICE would not be preferred in current NTCP modeling for HNC patients.