Home > Publications > Doctoral dissertation & Master thesis
Domestic Conference
Title Machine Learning-Based Diagnosis of Cardiac Arrest for Automated External De fibrillator
Degree Ph.D.
Author Minh Tuan Nguyen
Advisor Kiseon Kim
Graduation Date 2018.08.17 File Link icon
    Date 2019-02-25 09:58
Sudden cardiac arrest (SCA) is mainly caused by ventricular fibrillation and ventricular tachycardia (VF/VT), known as shockable (SH) rhythms. Shock advi ce algorithm (SAA), applied for the automated external defibrillator (AED), play s a vital role in detection of the SCA on electrocardiogram (ECG) and improvemen t of survival by delivering a prompt defibrillation. In the last decade, conside rable research efforts have been devoted to develop efficient SAA, which can be generally categorized as the thresholds-based and machine learning (ML)-based SA As. Basically, the common target for the SAA development is to improve the class ification performance in terms of the SCA diagnosis. This work concentrates on s pecific algorithm design and its performance analysis to comply with the America n Heart Association (AHA) recommendations.
At first, support vector machine (SVM ) classifier and a combination of 4 features is proposed for the SAA. Here, a bi nary genetic algorithm (GA)-based feature selection is used to select good featu res among 11 input candidate features (ICFs). All the combinations of 7 good fea tures are estimated for their performance on the training and the testing data u sing the SVM models to identify 6 combinations of the final feature pool. 5-fold s cross validation (CV) is then implemented carefully to validate the performanc e of the SVM classifier using final feature pool on separated and entire 5s-segm ent databases. The final combination of 4 features is addressed by the highest v alidation performance of the corresponding SVM classifier. Furthermore, existenc e of ventricular ectopic beats in the input data shows a negligible influence on the final classification performance of the proposed SAA. The improvement of th e SAA design is implemented with an expansion of ICFs using the modified variati onal mode decomposition (MVMD) technique. The SAA selects 20 features from both the preprocessed ECG segment and its NSH signal, which is reconstructed by the M VMD, and uses a SVM classifier for the binary shockable (SH)/ non-shockable (NSH ) classification.

The 20 features are identified as an efficient set of the most informativ e, among 54 ICFs, by comparing the balanced error rate of each combination, base d on two layers of feature selection. The NSH signal has been proven its effecti veness for the feature extraction. However, the SH modes corresponding to SH rhy thms among decomposed modes using MVMD have not been exploited for SAA design. T herefore, the SAA is improved with fully augmented ECG segment with its SH and N SH signal for feature extraction using MVMD. The newly proposed SAA, including a SVM classifier and a final combination of 36 features selected among 93 ICFs, w hich confirm the reliability of fully augmented ECG signals in terms of improvin g classification performance of SH/NSH rhythms.
Nowadays, DL is the state-of-the -art technique, which do not require any expertise human knowledge, feature extr action, and feature selection algorithms often used in the ML techniques. Moreov er, the strong feature learning capabilities makes the DL a suitable choice for SAA design. Therefore, a novel SAA for detection of SH/NSH rhythms on ECG signal is proposed using a combination of the ML and DL techniques. Here, the convolut ional neural network (CNN) is employed as a feature extractor (CNNE), which gene rates a feature vector used as the input of a Boosting classifier. The CNNE is s elected by a grid search with nested 5-folds cross validation (CV) procedure on the training data. Moreover, the MVMD is used to reconstruct SH and NSH signals from prepocessed ECG segment, which are then used as three input channels of the CNN. The ML and DL techniques have been proven their effectiveness when using f or the SAA design. Moreover, the augmented ECG segment with its SH and NSH signa ls is extremely effective for the feature extraction and quality improvement of the learned features extracted by the DL technique.
광주과학기술원 한·러 MT-IT 융합기술연구센터 광주과학기술원정보통신공학부