Cardiac arrhythmia characterized by irregular heartbeats is a prevalent problem among people suffering from cardiovascular diseases (CVD). Abnormalities in the heartbeats manifested in the electrocardiogram (ECG) signal are traditionally analysed by expert cardiologists or semi-automated computer aided techniques, which can be time consuming. Recent deep neural network (DNN) methods have paved the way for fully-automated timely classification. However, low accuracy and high energy consumption are few of the common problems with the existing DNN methods. In this work, a complete solution based on hierarchical three stage DNN architecture is proposed using Resistive Random Access Memory (RRAM) based Computation in-memory (CIM) architecture. The proposed scheme using Fully Connected Network (FCN), Long Short Term Memory Network (LSTM), assisted with dimensionality reduction and online personalized partial training enables a high accuracy and energy-efficient ECG arrhythmia classification solution. Experimental study conducted with MIT-BIH arrhythmia database show that the proposed scheme achieves an accuracy of 98.73%, F1 score of 97.6%, and consumes 442.25 uJ of energy. Compared to state-of-the-art, energy improvement of up to 2.83x is achieved, with a reasonable area overhead.