George B. Moody PhysioNet Challenge


Quick links for this year's Challenge:

Please post questions and comments in the forum. However, if your question reveals information about your entry, then please email info at We may post parts of our reply publicly if we feel that all Challengers should benefit from it. We will not answer emails about the Challenge to any other address.

Session S42.7

A Statistical Approach To Discrimination Of Patients At Risk Of Paroxysmal Atrial Fibrillation

C. Marchesi, M. Paoletti

Universit di Firenze

Firenze, Italy

Short term prediction of Paroxysmal Atrial Fibrillation is a critical issue because no validated rules have been so far proposed to define consistently such an arrhythmia and because PAF signs are often masked by other rhythm problems. Moreover usual difficulties of long term signal monitoring, like baseline wander, noise and artifacts, must be faced since they can heavily affect PAF detection. This study proposes a statistical approach for detecting ECG records belonging to people at risk of PAF. The signal processing procedure includes the definitions of: a signal amplitude reference, an original non linear operator to enhance wanted signal features, sequence of RR histograms to detect sudden rhythm changes (irregular RR, ectopic beats). A reliable amplitude reference is achieved by median filtering. In our implementation an hybrid FIR-median filter has provided adequate results in all the circumstances. The key operation is carried out by a simple and effective quadratic operator, capable of tracking and enhancing the significant wave fronts. An hybrid sequential-combinatorial logic net drives the quadratic operator to discriminate various events. Thus a reliable RR time series (tachogram) is computed and analyzed by building sequences of histograms. The diversity between successive beat distributions appears to be correlated with the risk of developing PAF. This approach, after tuning the statistics to the learning data set provided by Physionet, gave the following results: 35 correct answers (screening only,log number:20010427.124605).

Supported by the National Institute of Biomedical Imaging and Bioengineering (NIBIB) under NIH grant number R01EB030362.