Joeky Senders

171 General discussion Instead of engaging into the futile efforts of defining when an algorithm becomes machine or deep learning, they can be considered as an extension of traditional statistical approaches. Machine learning algorithms exist along a continuum, determined by how much is specified by humans and how much is learned by the machine, referred to as the machine learning spectrum. 4 The current thesis describes several studies along the continuum of the machine learning spectrum as it applies to neurosurgical oncology (Figure 1). Thirty day outcomes after craniotomy Deep learning Classic machine learning Regression analysis Human decision making Part I - Outcomes and Risk Factors Part II - Predictive Analysis Part III - Natural Language Processing Venous thromboembolism and intracranial hemorrhage Length of thromboprofylaxis Online survival calculator Expert opinion Automating clinical chart review Classifying brain metastasis 2 3 4 2 3 4 5 5 1 1 6 6 7 8 7 8 Comparison of learning curves Relative Human-to-Machine Decision Making FIGURE 1. The machine learning spectrum as it applies to the current thesis. Numbers 2 to 8 correspond to the chapters in the current thesis. Part I: Outcomes and risk factors in neurosurgical oncology In Chapters 2 and 3 , the inferential utility of regression-based algorithms was used to identify risk factors associated with 30-day outcomes in patients operated for a malignant brain tumor. Among patients undergoing craniotomy for a primary malignant brain tumor, 12.9% experienced a major complication within 30 days after surgery, in particular elderly patients and patients with worse functional status or more comorbidity. The increased risk of adverse events should be considered and balanced against the expected survival benefit in this particular patient population. Reoperation and venous thromboembolismwere identified as the two most common postoperative