Fuzzy-Bayesian Classification
This research was a bit of an offshoot or something I implemented in the clinical decision support system I worked on earlier. I was dealing with continuous data read in from the patient, but the bayesian networks I was feeding the data into were discrete. Now some of the problems that can be seen here are that often times you can be on the virge of one observation, but still classified as another on a given node. For instance in our bayesian network that calculated hemorrhagic shock, if we know the rate at which a person is breathing to be almost considered hyperventilation, but still within the bounds of normal ventilation, we are really neglecting some of the data. If we have this type of problem occurring at several nodes of the bayesian network there is a good chance of incorrectly classifying the level of shock the patient may be in.
In order to add some "continuity" I first passed the data from the patient through a fuzzy classifier to find the relative memberships that the patient would be exhibiting. I could then modify the discrete bayesian network accordingly to create new a priori in order to better reflect that a person might, lets say, be exhibiting 52% normal ventilation but at the same time 48% hyperventilation. This was able to produce results more similar to that of a continuous bayesian network without the pain and hassle of creating a continous bayesian network. This technique also ensure that we utilized all of our input, not just the majority of it, as discrete bayesian networks do when fed continuous data.
|