Ann: Positive Preliminary Results from SMARTCOUGH-C-2 Study, page-147

"NOTE that this is SOFTWARE that can be improved based on learnings and testing and recoding/better algorithms"

Two things wrong with this view:

1. The algorithm is trained.  To train it you tell it what the outcome is for each set of inputs.  Then the algorithm builds up knowledge about the attributes that had certain values when the condition was present, and different values when the condition is absent.  But how do you determine whether the condition is present or not?  Doctors!  Doctors who can't agree on the most important condition of all in a substantial number of instances.  So what hope have you of getting good training when your classifications are unreliable in the first place.
2. And we have exactly one algorithm, licenced from the University of Queensland, which forms the core of our IP.  If we want to drop that one and go scouting for a new one then we're going back to where we were when the company was floated.