Application of Machine Learning for Multicenter Learning

Abstract

Advancements in radiation oncology are driving more specific, and thus improved, treatment opportunities. This creates challenges on the assessment of treatment options, as more information is needed to make an informed decision. One of the methods is to use machine-learning techniques to develop predictive models. Although prediction models, embedded in clinical decision support systems (CDSSs), are the foreseen solution, developing/training such prediction models requires large amounts of detailed patient information to reach decisive power. The amount of patients needed to train a reliable prediction model rapidly outgrows the numbers available in a single institution, hence the need for multicenter machinelearning. To be able to learn over multiple centers, several infrastructural prerequisites need to be addressed. First, data needs to be extracted from multiple source systems and represented using standardized terminologies, preferably including the semantics (the actual description) of the represented data. For research and model training purposes, this means that value representations (e.g. “m” or “f” indicating gender) need to be converted into standardized terms (the NCI Thesaurus codes C20197 or C16576, respectively), and that patient-identifiable information (e.g. name, institutional ID, address, etc.) needs to be removed or changed in a non-identifiable way. If datasets from different institutions use the same standardized terminology and data structure, data can be merged. Finally, after merging, prediction models can be learned on the complete dataset, in this chapter known as centralized learning.