Research

In clinical trials and social studies it is to be expected that not all participants react equally to an intervention. Goal of this body of work is to estimate the effect of the intervention on an individual basis or for subgroups of the population. To achieve this well established methods (e.g. penalised regression) or modern methods such as regression trees can be utilized. Some of the questions arising from this are:

◦    How can the predictions be validated?
◦    Which method is the best to use in a particular setting?
◦    How can groups of subjects on the basis of these predictions be defined?

---

One of the most promising applications of machine learning and AI is precision medicine. The expectation thereby is that these algorithms can yield accurate predictions about a patient‘s health on the basis of individual patient data (e.g. genetic make-up, epidemiological and environmental factors). Before a new drug can be used in patients a high level of evidence on its safety and effectiveness is required. In this research topic we investigate the question of how we can obtain a suitable level of evidence for algorithms that are sought to be used in healthcare. A particular question of interest is thereby how the dynamic nature of algorithms (i.e. the fact that they routinely get improved) can be accounted for.

---

How resources should be allocated when resources are limited is a common problem. A computer cluster needs to allocate tasks to cores, an internet provider needs to distribute bandwith and servers and in drug development the question of which intervention should be developed when or which treatment should a particular patient receive within a clinical trial. Optimal sequential decisions approaches (e.g. multi-armed bandits) are routinely used in many of these application areas. In the life sciences and in clinical trials in particular, such approaches are, however, rarely used as these methods are very good at allocating but are inefficient when one want to conclusively establish if a particular medicine is better than an alternative. In this work we develop novel approaches that seek to balance these two objectives better and develop novel statistical tests that yield improved power when comparing medicines.

---