I'm yet to see a study that convinces me of the potential of this concept. All of them when you get down to the details have at least one of:

  - An unrepresentative test set , i.e. a population with similar blood pressure at recording in which merely predicting a constant 'typical' reading would perform well in terms of accuracy.
  - A major methodology flaw, such as 'demonstrating' an ability to detect fluctuations from a baseline calibration cuff reading in situations where you would expect no significant fluctuations to occur.

I don't have access to this particular paper, but suspect it has the same issue owing to the relatively small sample size.