* simple version should threshold to a range `[0, a]` where a is calculated from the training data (say `b*IQR(signal)`). * Should check that the multiplier is ok on upswings / downswings * complicated version would use the outlier detection functions in epiprocess