This present work introduces an automatic diagnosis system of cardiac arrhythmias by employing three stage feature selection and classification approach. For discriminating different types of arrhythmias, DWT decomposition of ECG signals has been performed. The bi-orthogonal wavelet transform is used up to 5 level decomposition and various non-linear and statistical features are extracted from the sub band coefficients. In every stage of classification model different set of features are employed for efficient classification. In the first stage of classification, nonlinear characteristics of DWT coefficient extracted from RR signal efficiently distinguish normal and arrhythmic conditions of heart. In the later stages, the crucial morphological variations of ECG signals are employed for classification of different types of arrythmia. In the proposed work, significance of selected feature set is evaluated in each stage of classification mode. Two alternative approach of feature ranking scheme based on statistical p-value and conventional Euclidean distances are introduced to study the clinical significances. Hence the adaptive weighted factors of individual features are estimated for construction of fused feature set called Integrated Feature Index to improve the classification results with less computational burden. Experimental results of each stage show superior classification performances using the Integrated Feature Index compare to the selected feature combinations. The computational simplicity, reduction of feature dimension and encouraging classification accuracies make the model effective for mobile based cardiac heath monitoring system.