We present a methodology for identifying patients who have experienced Paroxysmal Atrial Fibrillation (PAF) among a given subject population. Our work is intended as an initial step towards the design of an unobtrusive portable system for concurrent detection and monitoring of chronic cardiac conditions. The methodology comprises two stages: off-line training and on-line analysis. During training the most significant features are selected using machine learning methods, without relying on a manual selection based on previous knowledge. Analysis is done in two phases: feature extraction and detection of PAF patients. Light-weight algorithms are employed in the feature extraction phase, allowing the on-line implementation of this step on wearable sensor nodes. The detection phase employs techniques borrowed from the field of failure prediction. While these algorithms have found extensive application in diverse scenarios, their application to automated cardiac analysis has not been sufficiently investigated to date. The proposed methodology is able to correctly classify 68% of the test records in the PAF Prediction Challenge database [1], performing comparably to state of the art off-line algorithms. Nonetheless, the proposed method employs embedded signal processing for the critical feature extraction step, which is executed on resource-constrained body sensor nodes. This allows for a real-time and energy-efficient implementation.