Joeky Senders

101 Glioblastoma survival calculator to randomize newly diagnosed patients to a placebo arm now that a proven, effective adjuvant treatment for glioblastoma has emerged. 26 Predictive modeling on this scale remains therefore bound to observational data, thereby highlighting the need for exploring analytical approaches to mitigate confounding. On average, 3.3% of all data points were missing in the total data set, which was multiply imputed by means of a random forest algorithm to mitigate the risk of systematic bias associated with a complete-case analysis. Nonetheless, survival performance in the current study is limited by the type and number of features included in the SEER registry. As a result, KPS score, isocitrate dehydrogenase 1 (IDH1) mutation, 1p/19q co-deletion, and MGMT methylation status were not included in the current iteration of the prediction model. Despite these limitations, the performance of the current proposed prediction tool exceeds that of the currently available prediction tools and even approximates the performance of many complex radiogenomic models, 17–25 yet with the ease, speed, accessibility, interpretability, and generalizability of clinical prediction tools. Furthermore, this study presents a framework that can be updated and reiterated when novel variables are added to the SEER registry or when novel large-scale multicenter glioblastoma registries are assembled. Because these models are trained on data from thousands of patients from numerous hospitals across the U.S., we expect the fitted models to be less prone to overfitting to data from a single institution and plausibly more generalizable to patients from diverse geographic regions undergoing a variety of clinical treatments. Implications Survival prognostication is critical for clinical and personal decision-making in glioblastoma patients. Although our current prediction tool provides an interactive interface for survival modeling with potential clinical utility, it is designed as a research tool and should not be implemented in clinical practice prior to prospective validation on multiple heterogenous cohorts. Using a population-based registry might be more representative of the typical glioblastoma patient in the US; however, testing the current model on single institutional or multicenter data might be essential to confirm its prognostic value at point-of-care. Furthermore, predictive models should inform rather than direct clinical decision-making. We advocate a multidimensional approach for survival prognostication, in which model predictions are adjusted and balanced against complementary information that is available including clinical experience, neuropsychological testing, imaging data, and genomic information. Many statistical and machine learning algorithms allow for the analysis of historical patient cohorts to predict survival in new patients. However, prediction performance,