Automatically classifying ECG recordings for arrhythmia is difficult since even normal ECG signals exhibit irregularities, and learning algorithms suffer from class imbalance. We propose a hybrid SVM to combat class imbalance rampant in biomedical signals. Consequently, we significantly reduce the number patients falsely classified as normal. The Hybrid SVM is suitable for a variety of multiclass problems; here, we used the MIT-BIH Arrhythmia database, and the position and magnitude of local singularities as features. We enhance relevant singularity-driven Hölder features proposed earlier; while the use of these features results in higher accuracy, using the Hybrid SVM shows even more gains. Copyright © 2010 Inderscience Enterprises Ltd.