Data mining algorithms are extensively used to classify gene expression data, in which prediction of disease plays a vital role. This paper aims to develop a new classification algorithm for cancer gene expression data using minimal number of gene combinations i.e. minimum gene subsets. The model uses classical statistical technique for gene ranking and two different classifiers for gene selection and prediction. The proposed method proves the capability of producing very high accuracy with very minimum number of genes. The methodology was tried with three publicly available cancer databases and the results were compared with the earlier approaches and proven better and promising prediction strength with less computational burden. This paper also focuses on the importance of applying an efficient gene selection method prior to classification can lead to good performance and the results are proven to be the best.