Biomedical advances achieved by reducing class noise

Abstract Class noise, also known as class blending, occurs when classes of different types of objects are erroneously grouped together, often as the result of trying to classify by similarity. For example, if we create a class of individuals who have pneumonia and we mix together cases of bacterial pneumonia, viral pneumonia, and smoking-related pulmonary damage, then a clinical trial conducted on a population of such individuals may fail to demonstrate the efficacy of a potentially useful antibiotic (because the nonbacterial conditions will be nonresponsive to treatment). To draw any sensible conclusion from the clinical trial, we would need to begin with a class whose members all have a defined relationship, in this case bacterial pneumonia. Class noise is quite possibly the most significant source of error in studies that compare the response of different classes to a particular agent or operation. Finding the sources of “noise” in classified data is another job that is best done by the deductive data analyst.