Diabetes is one of the leading causes of death, disability and economic loss throughout the world. Type 2 diabetes is more common (90-95% worldwide) type of diabetes. However, it can be prevented or delayed by taking the right care and interventions which indeed an early diagnosis. There has been much advancement in the field of various machine learning algorithms specifically for medical diagnosis. But due to partially complete medical data sets, accuracy often decreases, results in more number of misclassification that can lead t o harmful complications. An accurate prediction and diagnosis of a disease becomes a challenging research problem for many researchers. Therefore, aimed to improve the diagnosis accuracy we have proposed a new methodology, based on novel preprocessing techniques, and K-nearest neighbor classifier. The effectiveness of the proposed methodology is validated with the help of various quantitative metrics and a comparative analysis, with previously reported studies using the same UCI dataset focusing on pima-diabetes disease diagnosis. This is the first work of its kind, where 100% classification accuracy is achieved by feature reduction from eight to two that shows the out performance of the proposed methodology over existing methods. © 2016 IEEE.