Applying machine learning in motor activity time series of depressed bipolar and unipolar patients compared to healthy controls