Leibniz MMS Days 2018 - Abstract
Oswald, Marcus
In medical studies we are often faced with the situation that a certain treatment does not increase the overall success rate significantly compared to placebo. But it remains the question whether there are subgroups of patients where treatment is beneficial. We developed a machine learning method based on Mixed Integer Linear Programming that detects such subgroups and gives a treatment suggestion based only on features that are known before the treatment decision has to be done. Like in classification problems we separate the dataset into training and testing sets. In the training problem the so-called oddsratio is maximized, this ratio is the standard measurement in medicine for the effectivity of a new treatment. This method is applied to a pneumonia dataset with the task to find rules for a beneficial Makrolide treatment strategy. Computational results after cross validation are presented.