The Use of Shape Factors for Heart Beats Classification in Holter Recordings

The heart-beat classification in Holter ECG processors can be succesfully done by comparing the shape coefficients of the around-QRS fragments instead of comparing the signals themselfs. The additional computation time is compensated during the classification procedure, since up to 10 coefficients are stored for each class and compared instead of 30 signal samples (for 3-channels, 128 Hz-sampled ECG, QRS duration about 100 ms). The most discriminative shape coefficients are choosen with regard to the particular heart-diseases probability supported by the MIT-BIH database. The newly proposed classification method has been verified and the reached misclassification error is of about 0.5%. The other advantage of our method is the insensibility to the amplitude changes nor to the small desynchronization between the compared QRS, thus the simplest and fastest fiducial point detectors can be used without the loss of 2. MATERIALS AND METHODS The aim of our investigation was to find the most specific shape factors for the frequently observed heart-beats types. As a medical reference we used the MIT-BIH standard database (directory: MITDB) containing 44 half-hour recordings [9]. Due to the poor signal quality, the records 104, 105, 208, 213, 223 and 228 were excluded. Among all annotated beat types the 9 most frequent (i. e. 99.3% of whole beat number) were choosen as shown in table 1. Table 1. Heart beat types used to the classifiers adjustment and tests MIT-BIH code abbreviation MIT-BIH contribution c [%] description 1 NORMAL 64,5 normal beat 2 LBBB 7.98 left bundle branch block beat 3 RBBB 9.02 right bundle branch block beat 4 ABERR 0.05 aberrated atrial premature beat 5 PVC 4.26 premature ventricular contraction 6 FUSION 1.23 fusion of ventricular and normal beat 7 NPC 0.53 nodal (junctional) premature beat 8 APC 0.97 atrial premature contraction 12 PACE 9.45 paced beat The learning set consisted of 10 randomly choosen examples for each considered beat type without regard to their contribution to the MIT-BIH database. Initially we have proposed 10 different shape factors computed on the constant-length windowed signal. Since the applied QRS detector produces his positive response (fiducial point) in the initial sector of QRS, the window was assymetrical to the QRS fiducial point, that means the fiducial point is allways in 1/4 of window length. The window lengths were: 60, 80, 100, 120 and 140 ms. Having do this, the set of 50 shape factors was tested in order to discriminate the choosen 9 beat types. The best shape factor should meet both of the following criteria: maximize the average distance d between classes, minimize the average class size e. All "geometry" values like "distance" or "size" are expressed in absolute logarithmic units regardless to their physical units. Initlialy we tried to separate all 9 classes by a single shape coefficient, but while the results were unsatisfactory we increasing the argument space (number of shape coefficients considered simultaneously) by 2, and then by 3. In order to express the discriminating capability by a single value, the quotient: dn = δ1...n ε1...n n = 2...9 (1) is computed for a subset of 2, 3, 4 ... 9 most frequent classes, and the obtained values were cumulated with regard to the class contribution c in the MIT-BIH database (see tab. 1.). and the maximum value of D is interpreted as the best class discrimination.