Joeky Senders

173 General discussion Surgery, and neurosurgery in particular, is characterized by balancing outcome probabilities. In the decision-making process, the surgeon has to weigh the chances of a favorable outcome against the risks of surgery, keeping in mind the natural course of the disease. Large cohort studies allowed us to estimate the expected outcomes in the total population and even differentiate between various risk strata. These strata, however, comprise clusters within the total population, ranked on a single or few cardinal features. As such, the physician still relies on group-level statistics complemented with their own clinical experience. The lack of personalized outcome and risk assessment can result in informed consent procedures that are ambiguous and biased towards the mean (e.g., “Trials have shown a median increase in survival of …”, “Generally, X out of 100 will develop …”). Predictive models in contrast intend to quantify the estimated outcomes in the individual patient. As such, a personalized overview of the estimated outcomes can be provided when communicating different surgical strategies with patients and their families. This not only improves the patient selection and surgical decision-making but also enhances the patient’s autonomy throughout the decision-making process. To facilitate its transparency, reproducibility, and utility, we deployed the model developed in Chapter 5 as an online calculator for survival through a free, publicly available software. This prediction tool provides an online and interactive interface for survival modeling with the potential to inform clinical and personal decision- making in the individual glioblastoma patients. External and prospective validation on heterogenous cohorts from multiple institutions remains necessary, however, to confirm its prognostic value at point-of-care prior to clinical implementation. Furthermore, the online calculator, as well as clinical prediction tools in general, should be considered as dynamic rather than static products developed on the best available evidence available at that point. Continuousmodel evaluation and optimization remains therefore mandatory to improve its accuracy and precision based on supplementary patient data and novel insights. Currently, we are working on the first model update utilizing the recently published SEER data of glioblastoma patients diagnosed in 2016 as well. This update has improved model performance according to the Harrell’s C-index from 0.70 (95%CI 0.70 – 0.70) to 0.73 (95%CI 0.73 – 0.73). In the future, we aim to re- iterate the analysis and further optimize model performance. Collection of information on functional status and molecular markers in the SEER registry could be a valuable first step towards optimizing the model again in the near future.