Joeky Senders

170 Chapter 10 interactions to include, or transformation to perform if the variables are composed of individual words in a text document, pixels in a picture, genes on a chromosome, or voxels in an MRI scan, let alone specify all these model properties by hand? Machine learning This is wheremachine learning comes into play. In contrast to classical statistics, modern machine learning prioritizes prediction over inference, even if it is achieved at the cost of its interpretability. 2 Compared to regression analysis, the modeling process (e.g., the inclusion of nonlinear associations and interaction terms) occurs rather automatically in many machine learning algorithms. Furthermore, they are less concerned with providing interpretable coefficients but rather oriented towards computing accurate predictions. Because they require less human guidance, these algorithms can model complex patterns automatically, even those that are potentially undetectable or meaningless for humans. Similar to regression analysis, however, classical machine learning algorithms, such as fully connected artificial neural networks, random forest, and support vector machines, are limited to the analysis of structured data (i.e., data in tabular, two-dimensional format in which observations are represented by rows and variables by columns). As a result, a neuroradiologist still has to measure the size of a brain tumor manually and insert this value into a data collection sheet to allow for the construction of classic machine learning models. This poses a significant burden on the clinician or researcher and introduces human subjectivity with regard to the generation and selection of input features. Furthermore, it ignores the potentially relevant hierarchical relationship between individual data points. Voxels close to each other in the scan might have a different, yet relevant relationship compared to voxels far away from each other. This spatial, temporal hierarchy would be missed if the data is shoehorned into a tabular format. Deep learning has emerged as a family of techniques that were designed to develop models directly from the raw, unstructured data itself. 3 It allows the computer to ingest and analyze high-dimensional data formats (e.g., free text, pictures, MRI scan) and identify meaningful representations within the data. Considering the same neuro- imaging example, nodes in the lower layers of a computer vision model might be susceptible for detecting simple straight lines in the brain MRI, subsequent hidden layers can learn how to detect shapes by recognizing combinations of lines, and the top layers utilize this condensed knowledge to produce clinically meaningful estimates, such as diagnostic classifications, volumetric segmentations, or outcome predictions. This process of condensing high-dimensional data to meaningful features within the model is called feature extraction and allows the raw data to speak for itself.