Joeky Senders

169 General discussion GENERAL DISCUSSION Due to the infiltrative nature of the disease, the median expected survival in patients with a malignant brain tumor remains dismal despite improved surgical and adjuvant treatment strategies. 1 The thin line between treatment effectiveness and patient harm underlines the importance of tailoring clinical management to the needs of the individual patient and suggests a strong potential for the emerging field of predictive analytics. Classical statistics Throughout the medico-scientific history, numerous analytical techniques have been developed to derive knowledge from experiments and observations to improve day-to- day patient care. Classical statistical methods evaluate the strength of an association between patient characteristics and outcomes within a sample population, with the aim of generalizing these conclusions to the larger population. Although these statistical techniques have become indispensable for studying treatment efficacy and identifying risk factors, their coefficients remain group-level estimates derived from the total study cohort and do not necessarily apply to the same extent in each individual patient. A clinical trial could demonstrate the efficacy of a novel neurosurgical procedure, as well as the rate of complications observed at the cohort-level. In day-to-day clinical care, however, the question remains to what extent the individual patient would benefit from this treatment and how likely he or she is to experience the dreaded adverse events. The advent of predictive analytics provides clinicians with the analytical support for personalizing treatment decisions. Regression analysis can compute patient-level predictions of the outcome by adding the population intercept and the slope coefficients pertinent to the individual patient. To develop this model, however, human experts still need to determine which variables to include, identify relevant effect modifiers, and perform data transformations to meet the underlying assumptions. This requires pre- existing human understanding to hypothesize these statistical patterns and substantial effort to define the model properties accordingly. The high level of human interference is feasible for structured data sets with a limited number of clinically interpretable variables and even provides valuable insights into the underlying relationships among variables and outcomes. But how should a human test which variables to select,