Joeky Senders

100 Chapter 5 Discussion This manuscript and the accompanying online prediction tool provide a framework for individualized survival modeling in patients with glioblastoma that is generalizable to other cancer and neurosurgical patients. Although prior investigation in this area tends to focus on metrics of prediction performance, we advocate a multimodal assessment when constructing and implementing clinical prediction models. The online prediction tool provides interactive, online, and graphical representations of expected survival in glioblastoma patients. Few other groups have developed an online survival prediction tool for glioblastoma patients. 14–16 Gorlia et al. developed multiple nomograms based on a secondary analysis of trial data using age at diagnosis, World Health Organization performance score (WPS), extent of resection, Mini-Mental State Examination (MMSE) score, and O6-methylguanine–DNA methyltransferase (MGMT) methylation status as input features, thereby achieving a maximum C-index of 0.66. 14 Gittleman et al. developed similar nomograms including sex as an input feature and Karnofsky Performance Status (KPS) score as a measure of functional status. However, model discrimination remained similar (C-index 0.66). 15 Marko et al. developed a model in which extent of resection was modeled as a continuous covariate. This group also utilized an AFT model to account for the violated proportional hazards assumption and achieved a C-index of 0.69. 16 Higher discriminatory performance (C-index 0.63-0.77) was achieved in studies that used machine learning algorithms to analyze complex, high-dimensional data structures, such as genomic, imaging, and health-related quality of life data. 17–25 Although many machine learning algorithms are ideally suited for superior prediction performance by utilizing these high-dimensional data structures, increasing model complexity may incur other costs in terms of interpretability, ease of use, computation speed, and external generalizability. Limitations Due to the retrospective nature of the data acquisition, it cannot be excluded that adjuvant therapy was administered at outside hospitals and not corresponded back to the reporting hospital. However, because of the short survival period in this patient population, the percentage of patients with complete survival follow-up is exceptionally high. Although clinically essential features were included to mitigate the risk of confounding, the possibility of influence from unmeasured confounders cannot be excluded. Randomized data would be ideal; however, it is practically and financially infeasible to establish a cohort on this scale, and it has become ethically unjustifiable