With the rapid production of high dimensional genetic data, one major challenge in genome-wide association studies is to develop effective and efficient statistical tools to resolve the low power problem of detecting causal SNPs with low to moderate susceptibility, whose effects are often obscured by substantial background noises. Here we present a novel method that serves as an optimal technique for reducing background noises and improving detection power in genome-wide association studies. The approach uses hidden Markov model and its derivate Markov hidden Markov model to estimate the posterior probabilities of a markers being in an associated state. We conducted extensive simulations based on the human whole genome genotype data from the GlaxoSmithKline-POPRES project to calibrate the sensitivity and specificity of our method and compared with many popular approaches for detecting positive signals including the χ^2 test for association and the Cochran-Armitage trend test. Our simulation results suggested that at very low false positive rates (<10^-6), our method reaches the power of 0.9, and is more powerful than any other approaches, when the allelic effect of the causal variant is non-additive or unknown. Application of our method to the data set generated by Welcome Trust Case Control Consortium using 14,000 cases and 3,000 controls confirmed its powerfulness and efficiency under the context of the large-scale genome-wide association studies.