Development and validation of a web‐based calculator to predict individualized conditional risk of site‐specific recurrence in nasopharyngeal carcinoma: Analysis of 10,058 endemic cases

Abstract Background Conditional survival (CS) provides dynamic prognostic estimates by considering the patients existing survival time. Since CS for endemic nasopharyngeal carcinoma (NPC) is lacking, we aimed to assess the CS of endemic NPC and establish a web‐based calculator to predict individualized, conditional site‐specific recurrence risk. Methods Using an NPC‐specific database with a big‐data intelligence platform, 10,058 endemic patients with non‐metastatic stage I–IVA NPC receiving intensity‐modulated radiotherapy with or without chemotherapy between April 2009 and December 2015 were investigated. Crude CS estimates of conditional overall survival (COS), conditional disease‐free survival (CDFS), conditional locoregional relapse‐free survival (CLRRFS), conditional distant metastasis‐free survival (CDMFS), and conditional NPC‐specific survival (CNPC‐SS) were calculated. Covariate‐adjusted CS estimates were generated using inverse probability weighting. A prediction model was established using competing risk models and was externally validated with an independent, non‐metastatic stage I–IVA NPC cohort undergoing intensity‐modulated radiotherapy with or without chemotherapy ( n = 601) at another institution. Results The median follow‐up of the primary cohort was 67.2 months. The 5‐year COS, CDFS, CLRRFS, CDMFS, and CNPC‐SS increased from 86.2%, 78.1%, 89.8%, 87.3%, and 87.6% at diagnosis to 87.3%, 87.7%, 94.4%, 96.0%, and 90.1%, respectively, for an existing survival time of 3 years since diagnosis. Differences in CS estimates between prognostic factor subgroups of each endpoint were noticeable at diagnosis but diminished with time, whereas an ever‐increasing disparity in CS between different age subgroups was observed over time. Notably, the prognoses of patients that were poor at diagnosis improved greatly as patients survived longer. For individualized CS predictions, we developed a web‐based model to estimate the conditional risk of local (C‐index, 0.656), regional (0.667), bone (0.742), lung (0.681), and liver (0.711) recurrence, which significantly outperformed the current staging system ( P < 0.001). The performance of this web‐based model was further validated using an external validation cohort (median follow‐up, 61.3 months), with C‐indices of 0.672, 0.736, 0.754, 0.663, and 0.721, respectively. Conclusions We characterized the CS of endemic NPC in the largest cohort to date. Moreover, we established a web‐based calculator to predict the CS of site‐specific recurrence, which may help to tailor individualized, risk‐based, time‐adapted follow‐up strategies.