For the detection of genomic loci associated with polygenic diseases, an alternative to traditional genome-wide association studies is machine learning with feature ranking according to importance contribution to predictive model performance. To implement this approach, it is necessary to address the class imbalance problem caused by differences in the size of case and control samples, and learn how to select features based on their importance metric, which unlike p-values does not have a threshold. This work presents a bioinformatic approach that solves both problems simultaneously. It is based on training a random forest algorithm on randomized case-control samples of similar size, followed by feature ranking according to decreasing importance score and selection based on frequency among top-ranked values, as well as stability of importance scores. The approach has been tested on simulated genotypephenotype data containing single nucleotide polymorphisms. Two types of synthetic datasets were applied. The first one contained the genomic loci associated with polygenic disease. The second one did not have such loci.