Optimal length-constrained segmentation and subject-adaptive learning for real-time arrhythmia detection

An algorithm of data segmentation with length constraints for each segment is presented and applied in the context of arrhythmia detection. The additivity property of the cost function for each segment yields the induction proof of the exact global optimal solution. The experiments were conducted on the MIT-BIH arrhythmia dataset with the heartbeat categories recommended by the ANSI/AAMI EC57:1998 standard. The heartbeat classification task is enhanced by an adaptive learning scheme. Incremental support vector machine is used to integrate a small number of expert-annotated samples specific to the subject into the existing classifier previously learned from the dataset. The proposed segmentation scheme obtains the sensitivity of 99.89% and the positive predictivity of 99.83%. The classification sensitivities of ventricular and supraventricular detection are significantly boosted from 85.9% and 83.5% (subject-unadaptive) to 97.7% and 93.2% (subject-adaptive), respectively. Similarly the predictivities increase from 94.8% to 99.3% (ventricular), and from 67.7% to 88.0% (supraventricular) when plugging in the adaptive learning method. The signal processing framework is conducted in a simulated real-time model. As compared to the previously reported studies we achieve a competitive performance in terms of all assessment measures.