Joeky Senders

176 Chapter 10 Lastly, by extracting and analyzing patient characteristics and outcomes automatically, natural language processing could facilitate a health care system that continuously learns from clinically derived data, thereby narrowing the gap between research and patient care. This resultant collective learning curve can be used to inform and optimize clinical decision-making in the individual patient at point of care. The machine learning spectrum in neurosurgical oncology It is the increasingavailabilityof high-dimensional clinical informationandcomputational power that has propelled the use and popularity of algorithms on the high end of the machine learning spectrum (i.e., deep learning). However, these ‘black box’ algorithms do not constitute the computational panacea to all medico-scientific problems due to the lack of interpretability and need for enormous amounts of data to grasp the full complexity of the data without overfitting. 3 This thesis confirms that placement on the high end of the spectrum does not necessarily imply superiority over other algorithms. Regression analysis, as demonstrated in Part I , and other methods for statistical inference will remain pivotal for clarifying clinically relevant associations at the group- level. It performs well and consistent, even on relatively small data sets. Predictive analytics has the potential to personalize these estimates after collection of sufficient amounts of training data. Even in the predictive realm, however, machine learning does not necessarily outperform classical statistical algorithms, as shown in a recent systematic review as well. 7 In Part II , for example, we deployed an algorithm on the low end of the machine learning spectrum (i.e., the accelerated failure time) because of its superior predictive performance and interpretability. In Part III , however, we gravitated towards the high end of the machine learning spectrum (i.e., deep learning). In these natural language processing studies, the input consisted of unstructured high- dimensional data, namely free-text clinical reports. Manual specification of the almost infinite number of associations, interactions terms, and data transformations would be virtually impossible and meaningless for humans. Algorithms on the high-end of the machine learning spectrum, on the other hand, allowed for automated analysis of the hierarchical and semantic relationships among words, without the need for manual specification. Future research Instead of focusing merely on novel and complex algorithms on the high end of the machine learning spectrum, future research should focus on tailoring the modeling approach to the computational and clinical problem at hand. After all, different problems require different levels of human involvement.