Feature extraction from ventricular late potentials with clinical validations using the adaptive orthogonal projection method

Non-invasive techniques to record the activation wave of ventricular late potentials have so far depended on signal averaging. Although this approach produces representative signals, any beat-to-beat variations are removed by the averaging process. These beat-to-beat variations are important in the diagnosis of many heart anomalies, particularly arrhythmia. This paper describes an experimental system which can detect ventricular late potentials at the body surface while preserving beat-to-beat variations. The proposed method can be used to validate the ventricular late potentials as reliable features for clinical diagnosis. The study was achieved on six subjects, three of them healthy, the others selected for their pathological signs. The first patient had trouble with ventricular tachycardia, the second patient suffered from cardiac dysplasia and the last presented extrasysoles. The system uses a different technique, but an important feature is the use of blind separation sources using higher-order statistics for the removal of mains interference, and multidimensional adaptive orthogonal projection followed by adaptive filtering to cancel random noises such as electrode noise, electronic noise, motion artefacts and electromyographic (EMG) signals. This technique was evaluated by using real cardiac signals recorded by an array of four small surface electrode pairs