Platform to Assist Medical Experts in Training, Application, and Control of Machine Learning Models Using Patient Data from a Clinical Information System

Abstract

In recent years, clinical data scientists achieved major breakthroughs advancing machine learning models for the medical domain, which have great potential assisting medical experts. Machine learning models can be leveraged to assist medical experts in tasks such as analyzing and diagnosing patient data, for example, from computed tomography scans. However, it is a challenge to translate the latest advancements in academic research fields such as computer sciences and physics into clinical practice. For this purpose, clinical data scientists and medical experts need to closely collaborate. This thesis tackles challenges of accessibility and usability of state-of-the-art machine learning models as well as designing a scalable computing architecture. Hence, conceptual ideas of possible strategies, as well as a prototype of such a machine learning platform, are presented. A systematic literature review was conducted on the current approaches to create medical machine learning platforms, the management of machine learning models, and the version management of large data sets. Afterward, the functional and nonfunctional requirements of the new machine learning platform were elicited as part of the requirements analysis. Two streamlined workflows for clinical data scientists and medical experts were derived from the requirement analysis. The workflow for the clinical data scientists includes steps to define, train, and share machine learning methods, including pre- and postprocessing modules, and management of data sets. Medical experts are able to analyze patient data using pre-defined machine learning methods. Building on the result of these analyses, the architecture of the platform was derived. The architecture consists of a scalable infrastructure stack, a lightweight and easy-to-use web interface, as well as a backend component to provide the required functionalities. The final design decisions solve the issue of efficiently standardizing, parallelizing, and applying machine learning workflows within a scalable computing infrastructure. The proposed platform was evaluated with 22 participants, consisting of clinical data scientists (N=12) and medical experts (N=10). Both groups were asked to rate specific workflows of the platform, as well as the platform as a whole, and to provide additional ideas and feedback. 92% of the medical experts and 90% of the clinical data scientists rated their overall impression of the platform as very good. Furthermore, medical experts and clinical data scientists strongly agreed that the platform facilitates method development and collaborations with 92% and 90%, respectively. The conducted expert survey suggests that the here proposed platform could be used to develop, optimize, and apply machine learning methods in the medical domain and beyond, thereby easing the collaboration between medical experts and clinical data scientists.