Conference Call Summary, page-320

The Australian study wasn't a trial at all. It was aimed at training the algorithm:
1) These are the symptoms
2) This is what the cough sounded like
3) This is what (if anything) the patient had

The algorithm looks at the first two to try and find what they have in common for each condition. You have to have samples with no condition as well.

You can test how good the training is by taking one of your samples and asking the algorithm to tell you what the condition was from just the symptoms and the cough.

The key thing is that you can't use the training data as trial data because the algorithm will just be giving you back what you put in! You need to take new samples that weren't used for training and see how good the algorithm is at predicting the condition.