Internal and external validation of a tumour control probability model for non-small cell lung cancer

Abstract

Tumour control probability (TCP) models are often used in radiation oncology to estimate local tumour control. However, validation of these models, while highly desirable, unfortunately remains very uncommon. Therefore, the purpose of this study is to validate a previously described TCP model for non-small cell lung cancer (NSCLC). To validate and, hence, measure the TCP model performance, two conventional approaches, internal and external validation, were used. 25 publications, which reported clinical outcomes of NSCLC, were employed in the validation procedure. The Bootstrap resampling approach was utilized to validate the TCP model internally. 1,000 bootstrap samples were randomly generated with replacement, each sample consisted of 25 data points. The TCP model outcomes were then evaluated in both the original and the bootstrap samples to determine the ‘optimism’ of the TCP model performance. For external validation, the patient cohorts were split, randomly, into a training set (70%) and a validation set (30%). Subsequently, the optimism in the model performances for both sets was determined. The bootstrap expected ‘optimism’ and ‘optimism-corrected’ performances in the coefficient of determination (R2) for two-year TCP were estimated to be 6% and 86%. Similarly, the expected ‘optimism’ and ‘optimism-corrected’ performances of the external validation samples in R2 were estimated to be 2% and 88%. These values agreed well with those observed in the internal validation samples. Nonetheless, both internal and external validation outcomes should, preferably, be the same. The TCP model in the current study demonstrates excellent performance, as suggested by the internal and external validation procedures.