We did a standard 80/20 training/testing trial. We found that the correct diagnosis was in the top 5 of the generated possibilities 50% of the time.

What will be helpful is to update the algorithm from user input. I'd be interested in hearing about your particular case. We use symptom/disease prevalence in the calculations, so it is usually great at getting common conditions. Maybe you can leave a comment using the feedback form on the website and we can look into why it missed your diagnosis?