Diabetes mellitus has been identified as a keen disease which affects the human at any age. Predicting the disease at earlier would help the person in preventing and maintaining it. Towards predicting diabetes there are number of approaches available like Support Vector Machine (SVM), KNN, Genetic Algorithm (GA), Pattern mining and Decision Trees. The methods use clinical, lifestyle, professional, and other features in predicting the possibility of disease, but suffer to achieve higher prediction accuracy. This work proposes a method of Hybrid pruning tree with Manhattan Metric measures which is used to find for early diagnosis of diabetes. To start with, the method reads the PIMA data set and applies feature level normalization technique to remove noisy records and normalize the data set. Further, the features of the data set have been extracted and generate the tree according to the features extracted. Once the tree has been generated, then the method applies tree pruning and trims the tree to a simplified form. With the test sample, the features of the test sample are extracted and measures Multi Feature Manhattan Similarity (MFMS) against the tree available. According to the value of Manhattan distance, the method predicts the chance of person to get affect by diabetes and classifies the sample. The HPMMM model introduces higher prediction accuracy with less time complexity.