Paroxysmal Atrial Fibrillation (PAF) prediction viability is a line of research currently being investigated. The definition of new valid parameters for this task may generate various heterogeneous features. Genetic Algorithms (GAs) automatically find a set of parameters to maximize the diagnosis capabilities of a scheme based on the K-nearest neighbours algorithm. This is an efficient way of generating a number of possible solutions for the problem of PAF prediction. The present paper illustrates how GAs, rather than a statistical study of the database can be used to select the parameters giving the best classification rates.