Machine learning in medicine: big pictures require small, but crucial strokes

Machine learning models are increasingly becoming more accurate and faster and they are not burdened by emotions or fatigue, for example. Nevertheless, these models are often not actually implemented in clinical practice, partly due to the trust the doctor must have in the models. This dissertation describes three crucial steps for the creation of a good and implementable model in medicine: 1) Validation of the raw data, 2) Extraction of the correct parameters and 3) Choosing the right type of model. The importance of point 1 is shown by the pulse transit time parameter, which has clinical value, but this dissertation shows that in practice it has significant and previously unknown artifacts. The importance of point 2 is demonstrated by the creation of a new parameter for the indication of whether the ductus arteriosus is open or closed in newborns. Point 3 is underlined by three published articles in which models have been trained to try to solve a clinical problem with both insightful and complex models.