Recent studies in the context of medical big data show promising results with respect to accuracy of predictions. Thus, analysing medical data appears as a useful addition to research in health services.

In multiple work packages, the subproject "Big Data" researches how analysing complex patient data can help to improve the care of patients.

During the evaluation, we also address privacy and ethical questions of our approach.

From a computer science perspective, state of the art regarding big data analytics shows that it is hardly possible to craft algorithms by hand. Therefore, functions are derived, i.e., learned, automatically from training data. The derived functions, e.g., classification or regression functions, can be characterised by different performance measures such as loss ratio, error ratio, precision and recall, or false positive and false negative rates, to name a few.

When learning functions, a large amount of data is needed to ensure independence of training data. For example, to achieve a predefined confidence interval, it is well known that a sufficient amount of (independent) data is necessary. To use (small) samples of the overall data set as a basis to learn functions (for example with cross validation), the function often does not exceed a certain performance measure due to missing independence in training data. Hence, often the performance of learned functions using sampling does not suffice for pratical use.

The learning process of functions is based on a data distribution, which is available while learning. An assumption often used is that the function is deployed on data with the same or at least a similar distribution to achieve a low error ratio with a high probability. Unfortunately, the assumption does not always apply. In many cases, the function needs to be adjusted while in use. An adjustment can be incrementally performed either by supervised or unsupervised learning.