Intentional user input has fairly consistent characteristics, as do each of a variety of noise/drift/interference sources. The key is to safely and accurately recognize/estimate each of those to remove or report them. I suspect now people would prefer to use some form of machine learning, but realistically there are fairly low expected power requirements, which make that difficult. There's also a battle between low latency stateless techniques and more stateful techniques that require rapid recovery when they get something wrong.