Joeky Senders

164 Chapter 9 investigate the influence of continuing prophylactic anticoagulation beyond discharge. In a multivariable analysis of 301 patients who underwent craniotomy for a high-grade glioma, the VTE rate in patients who received thromboprophylaxis for 21 days after surgery was not statistically significantly lower compared to patients who received thromboprophylaxis up to discharge. However, a significantly higher rate of intracranial hemorrhage was seen in the prolonged thromboprophylaxis group. Functional status and BMI were confirmed as predictors of VTE in this cohort. Part II: Predictive analytics in neurosurgical oncology This part constitutes a study that evaluates the utility of machine learning for outcome prediction in neuro-oncology. Chapter 5 compared 15 commonly used statistical and machine learning algorithms in their ability to predict survival in glioblastoma patients based on demographic, socio-economic, clinical, and radiological features. In a cohort of 20,281 patients, identified through the Surveillance Epidemiology and End Results (SEER) registry, the accelerated failure time model outperformed competing statistical and machine learning algorithms in terms of prediction performance (C-index = 0.70), as well as interpretability, predictive utility, and computational efficiency. This model was therefore deployed as a free, publicly-available online calculator. Part III: Natural language processing in neurosurgical oncology In this part, various natural language processing approaches were developed and evaluated for medical text analysis in brain tumor patients. In Chapter 6 , we developed an open-source natural language processing framework for automating the extraction of clinical information. To this end, we analyzed a text corpus of narratively written radiology reports of glioblastoma patients (n = 562 reports) with a regression-based algorithm (LASSO regression) as classifier. The resultant pipeline was able to extract 15 radiographic features with high to excellent performance (AUC 0.82-0.98). Model performance was correlated with the interrater agreement of the manually provided labels (ρ = 0.904; p < .001), but not with the frequency distribution of the variables of interest (ρ = 0.179; p = .52). In the subsequent Chapter 7 , we developed and compared various statistical (logistic regression, LASSO regression), classical machine learning (fully connected artificial neural networks), and deep learning (convolutional neural networks, gated recurrent unit, and long short-term memory) techniques in their ability to classify free-text radiology reports (n = 1,472) of brain metastases patients into reports that describe