Extracting relevant physiological information from noisy, imperfect measurements is a key challenge in many biomedical applications. In many instances, measurements contain mixtures of multiple physiological processes, such as cardiac and respiratory activity, inducing a need for algorithms to separate these different processes from one another to obtain accurate information on either of them. Moreover, signals are often contaminated by various kinds of disturbances, necessitating intelligent signal processing algorithms that are robust to outliers missing data. Finally, measurements are often indirect: the physiological quantity of interest usually cannot be measured directly, and one has to infer its state from some related quantity.

Based on current and new machine learning principles we develop solutions for real medical applications suffering from all of the above challenges, using model-based and data-driven signal processing techniques. Typical methods we employ include

• Classical signal-processing techniques, such as Kalman filters and smoothers
• Gaussian process modeling & regression
• Factor graph-based probabilistic inference
• Deep learning techniques