To lessen the rate of false critical arrhythmia alarms, we used robust heart rate estimation and cost-sensitive support vector machines. The PhysioNet MIMIC II database and the 2015 PhysioNet/CinC Challenge public database were used as the training dataset; the 2015 Challenge hidden dataset was for testing. Each record had an alarm labeled with asystole, extreme bradycardia, extreme tachycardia, ventricular tachycardia or ventricular flutter/fibrillation. Before alarm onsets, 300 s multimodal data was provided, including electrocardiogram, arterial blood pressure and/or photoplethysmogram. A signal quality modified Kalman filter achieved robust heart rate estimation. Based on this, we extracted heart rate variability features and statistical ECG features. Next, we applied a genetic algorithm (GA) to select the optimal feature combination. Finally, considering the high cost of classifying a true arrhythmia as false, we selected cost-sensitive support vector machines (CSSVMs) to classify alarms. Evaluation on the test dataset showed the overall true positive rate was 95%, and the true negative rate was 85%.