The aim of this work is to develop an efficient data-driven method for automatic medical decision making, especially for cardiac arrhythmia diagnosis. To achieve this goal, we have targeted the most common arrhythmia worldwide -atrial fibrillation (AF). Most of reported studies are dealing with inter-beat interval time series analysis coupled with univariate and/or multivariate data-driven methods. The state of the art of this subject revealed that although satisfactory detection findings have been achieved for long AF durations, there is still scope for improvement which needs to be addressed for brief episodes which is highly desired by healthcare professionals. Relevance vector machine (RVM) has been developed to address this issue. Several kernel functions and parameters have been tested to optimize RVM. Five geometrical and nonlinear features were extracted from 30s inter-beat time series. The RVM classifier was trained on 3000 randomly selected samples from four publicly-accessible sets of clinical data and tested on 1000 samples. The performance of the diagnosis model was evaluated by 10-fold cross-validation method. The results showed that the RVM model performed better than do existing algorithms, with 96.58% success rate. The automatic diagnosis on another dataset of 118985 samples of AF and Normal Sinus Rhythm (NSR) has yield 96.64% of classification accuracy. This automated data-driven decision making approach can be exploited for medical diagnosis of other arrhythmias.