@conference{
author = "Ivanović, M. D. and Ring, M. and Hannink, J. and Baronio, F. and Hadžievski, Ljupčo and Eskofier, B.",
year = "2019",
abstract = "Ventricular fibrillation (VF) represents the most frequent initial rhythm in out-of-hospital cardiac arrest (OHCA). It is characterized by rapid and disorganized contraction of the heart muscle cells which can lead to a sudden cardiac death. Optimizing defibrillation strategy (immediate defibrillation versus cardiopulmonary resuscitation) by evaluating the probability of the successful outcome could significantly enhance resuscitation. Over the past few decades, different classification strategies were applied to predict the defibrillation outcome of OHCA patients, but none have achieved superior performance to be widely accepted and implemented in automated external defibrillators. All these reported strategies utilized conventional machine learning (ML) approach with feature engineering. Here, we compare the performances of 7 ML algorithms (Logistic Regression (LR), Naïve Bayes (NB), Decision tree (C4.5), AdaBoost M1(AB), Support Vector Machine (SVM), k Nearest Neighbour (kNN) and Random Forest (RF)) [1] with a novel approach based on convolutional neural networks (CNN). For conventional ML approach we engineered 28 “hand-crafted” features using time domain, frequency domain, time-frequency domain and non-linear dynamical analysis of the 4s pre-shock VF signal. The best performing feature combination was chosen using the wrapper feature selection method, which utilizes the classifier in evaluating selected feature subset. In deep learning approach, the CNN was capable of learning useful features from the raw VF signals. We used 3-stage CNN feature extractor, which contained convolution, rectified linear unit activation, dropout (only in training) and max-pooling and 2 layer perceptron for classification. Our results show that the SVM, kNN and RF outperformed other conventional ML algorithms. The mean accuracy obtained over 10 fold cross-validation of these 3 ML algorithms were: 81.5%, 81.8 % and 82.8 %, respectively. On the other hand deep learning approach demonstrated the superiority over the conventional ML approach with engineered features. Obtained averaged accuracy of 93.6 %, along with sensitivity of 98.8 % and specificity of 88.2 %, which satisfy the condition of at least 50 % specificity at 95 % sensitivity for being considered safe, indicate that the proposed CNN model can be considered as a safe and useful predictor for defibrillation decision.",
publisher = "Belgrade : Vinča Institute of Nuclear Sciences, University of Belgrade",
journal = "PHOTONICA2019 : 7th International School and Conference on Photonics & Machine Learning with Photonics Symposium : Book of abstracts",
title = "Defibrillation outcome prediction as a potential guide to resuscitation",
pages = "51-51",
url = "https://hdl.handle.net/21.15107/rcub_vinar_11867"
}
