[home] [Personal Program] [Help]

---

Session: The heart
Session starts: Thursday 24 January, 10:30
Presentation starts: 11:30
Room: Lecture room 559

---

Amalia Villa Gomez (KU Leuven, imec)
Sibasankar Padhy (KU Leuven, imec)
Rik Willems (KU Leuven)
Sabine Van Huffel (KU Leuven, imec)
Carolina Varon (KU Leuven, imec)

Abstract:
Aim: Heartbeat classification is required to detect abnormalities in the electrical activity of the heart and in applications such as the analysis of Heart Rate Variability. Currently commercially available software for heartbeat classification still requires extensive manual correction. Better algorithms for automatic heartbeat classification are therefore clinically valuable. In this study, Variational Mode Decomposition (VMD) is used to describe heartbeats regarding their origin, and their classification performance is evaluated combining those with time features. A single-lead approach is considered, aiming to use the features on wearable devices. Methods: The features were evaluated on the first lead of the MIT-BIH Arrhythmia Database. Heartbeats were segmented using an asymmetric window of 0.65s, and the annotations were sorted in 3 groups following AAMI standards: Normal beats, Supra-ventricular and Ventricular extra-beats. Paced and fusion beats were not considered. Each heartbeat was decomposed in 5 modes, which correspond to the bands of frequency with the highest power. Five features were extracted from each of the bands: the power, the bandwidth, the symmetry of the wave around the R peak, the number of zero-crossings and the amplitude difference between the two lowest minima. Additionally, the distance to the previous and the next R peak, the absolute difference between them and the deviation of the previous RR interval from the average last 10 were used as time features. This set of features was fed to an LS-SVM classifier, using 10-fold cross-validation and 50% of the balanced data as training. Results: Preliminary results report an overall accuracy of 92.14%. Regarding the performance for each class, for Normal, Supraventricular and Ventricular heartbeats we obtained a Sensitivity of 92.84%, 72.56% and 91.25% respectively; and a positive predictive value of 99.59%, 25.48% and 80.48%. Conclusion: These results are in line with the state-of-the-art for heartbeat classification using single-lead, and they show promising performance for Normal and Ventricular beats. The discrimination between Normal and Supraventricular beats could be improved by including more sophisticated time features. Future work will focus on adding this kind of features, as well as on reducing the total number of features by applying a feature selection stage.