Different techniques used to improve the performance of a classifier of the twelve-lead electrocardiogram

Investigates the automatic classification of the 12-lead electrocardiogram (ECG) into different pathophysiological disease categories. The ECG database used in this study contained 926 recordings, with 500 records classified with 100% accuracy and 426 records classified with 75% accuracy. Each record contained a simultaneously recorded 12-lead ECG of 8-10 s duration. Each record is classed as either (i) normal; (ii) left, right or bi-ventricular hypertrophy; or (iii) anterior, inferior or combined myocardial infarction. A baseline classifier was trained using a single beat from the 500 classified recordings and resulted in a 7-way classification test-set accuracy of 55%. The following techniques were used for improving the classification performance: (1) multi-beat data, (2) regularisation of the covariance matrix, and (3) utilisation of inaccurately classified data in the training process. Combining these three techniques resulted in a classifier with a test-set accuracy of over 75%