[home] [Personal Program] [Help]

---

Session: The heart
Session starts: Thursday 24 January, 10:30
Presentation starts: 11:45
Room: Lecture room 559

---

Jordy Jansen (Maastricht University)
Richard Houben (2BMedical BV)
Joël Karel (Maastricht University)
Ralf Peeters (Maastricht University)

Abstract:
Sudden Cardiac Death (SCD) remains a vital health issue to this day and rapid EMS response is paramount for favourable neurological outcome [1]. Ventricular Fibrillation (VF) or Ventricular Tachycardia (VT) are the major causes of SCD. In order to detect VF and VT, many (relatively simple) single-lead ECG features have been proposed over the years that can be implemented in implantable or wearable devices. From the MIT-BIH Arrhythmia, VFDB and CUDB databases, a total of 37 of these features are extracted from electrocardiogram (ECG) segments, which were set to be 6, 12 or 18 seconds yielding three datasets. These ECGs act as a proxy for cardiograms. From a response time perspective, shorter segment lengths are preferable, but might not carry sufficient information. The research question is: what are the relevant features for VF/VT detection and what segment length ought to be used? For these three data sets individually, a Support Vector Machine (SVM) and a Gradient Boosting Tree (GBT) model are trained on the features extracted from the training data (80%) and the parameters are optimized following a 5-fold cross validation technique. The SVM and GBT models are then evaluated on the test data (20%). The SVM and GBT models with all features fitted on the 6s dataset achieved the highest AUC, as well as specificity, accuracy and positive predictive value when sensitivity is fixed to 98% or 99%. Comparing the two models (on the 6s data set), the GBT model achieves highest AUC whereas the SVM model achieves higher specificity, accuracy and positive predictive value when sensitivity is fixed to 98% or 99%. The information potential of the individual features was estimated by fitting an SVM for each feature individually. Regardless segment size, the three most informative features were time domain features, bCP [2], TCSC [3], count2 [4]. The results of this preliminary study are promising for robust VT/VF detection. REFERENCES [1] Wellens, Hein J, Lindemans, Fred W, Houben, Richard P, Gorgels, Anton P, Volders, Paul G, Ter Bekke, Rachel MA, and Crijns, Harry J. Improving survival after out-of-hospital cardiac arrest requires new tools. European Heart Journal, 37(19), 1499-1503, (2015). [2] Irusta, Unai, Ruiz, Jesús, Aramendi, Elisabete, Gauna, Sofía Ruiz de, Ayala, Unai, and Alonso, Erik. A high-temporal resolution algorithm to discriminate shockable from non-shockable rhythms in adults and children. Resuscitation, 83(9), 1090-1097, (2012). [3] Arafat, Muhammad Abdullah, Chowdhury, Abdul Wadud, and Hasan, Md Kamrul. A simple time domain algorithm for the detection of ventricular fibrillation in electrocardiogram. Signal, Image and Video Processing, 5(1), 1-10, (2011). [4] Jekova, Irena and Krasteva, Vessela. Real time detection of ventricular fibrillation and tachycardia. Physiological measurement, 25(5), 1167-1178, (2004).