Well the only stream (to my knowledge) that could be reversible...

Well the only stream (to my knowledge) that could be reversible (has the algorithm on the device) would be the humanitarian stream. Since network connectivity is often spotty at best when they're doing field testing, so having it locally on the device would be important. Soon my way someone is going be able to reverse it is if they get it from these organisations. So I really feel like this concern about reversing isn't really what people should be concerned about. It's much more likely someone will just read the research papers, and do their own testing/create their own implementation from that (and even if that happens we still have an advantage as they will be playing catchup essentially).

But as far as your question. Basically ML (machine learning) algorithms can improve overtime when they have more data. That in and of itself doesn't make reversing the algorithm anymore difficult. Rather it's just advantageous if we have more cough data than a competitor (that is using the same technique). So if we have more data than a competitor who is using the same algorithm, then our diagnosis can be more accurate and may possibly cover more diseases.