A feature representation learning method for temporal datasets

Predictive modeling of future health states can greatly contribute to more effective health care. Healthcare professionals can for example act in a more proactive way or predictions can drive more automated ways of therapy. However, the task is very challenging. Future developments likely depend on observations in the (recent) past, but how can we capture this history in features to generate accurate predictive models? And what length of history should we consider? We propose a framework that is able to generate patient tailored features from observations of the recent history that maximize predictive performance. For a case study in the domain of depression we find that using this method new data representations can be generated that increase the predictive performance significantly.