Detail publikace

SVM Based ECG Classification Using Rhythm and Morphology Features, Cluster Analysis and Multilevel Noise Estimation

Originální název

SVM Based ECG Classification Using Rhythm and Morphology Features, Cluster Analysis and Multilevel Noise Estimation

Anglický název

SVM Based ECG Classification Using Rhythm and Morphology Features, Cluster Analysis and Multilevel Noise Estimation

Jazyk

en

Originální abstrakt

Background: Smartphone-based ECG devices comprise great potential in screening for arrhythmias. However, its feasibility is limited by poor signal quality leading to incorrect rhythm classification. In this study, advanced method for automatic classification of normal rhythm (N), atrial fibrillation (A), other rhythm (O), and noisy records (P) is introduced. Methods: Two-step SVM approach followed by simple threshold based rules was used for data classification. In the first step, various features were derived from separate beats to represent particular events (normal as well as pathological and artefacts) in more detail. Output of the first classifier was used to calculate global features describing entire ECG. These features were then used to train the second classification model. Both classifiers were evaluated on training set via cross-validation technique, and additionally on hidden testing set. Results: In the Phase II of challenge, total F1 score of the method is 0.81 and 0.84 within hidden challenge dataset and training set, respectively. Particular F1 scores within hidden challenge dataset are 0.90 (N), 0.81 (A), 0.72 (O), and 0.55 (P). Particular F1 scores within training set are 0.91 (N), 0.85 (A), 0.76 (O), and 0.73 (P).

Anglický abstrakt

Background: Smartphone-based ECG devices comprise great potential in screening for arrhythmias. However, its feasibility is limited by poor signal quality leading to incorrect rhythm classification. In this study, advanced method for automatic classification of normal rhythm (N), atrial fibrillation (A), other rhythm (O), and noisy records (P) is introduced. Methods: Two-step SVM approach followed by simple threshold based rules was used for data classification. In the first step, various features were derived from separate beats to represent particular events (normal as well as pathological and artefacts) in more detail. Output of the first classifier was used to calculate global features describing entire ECG. These features were then used to train the second classification model. Both classifiers were evaluated on training set via cross-validation technique, and additionally on hidden testing set. Results: In the Phase II of challenge, total F1 score of the method is 0.81 and 0.84 within hidden challenge dataset and training set, respectively. Particular F1 scores within hidden challenge dataset are 0.90 (N), 0.81 (A), 0.72 (O), and 0.55 (P). Particular F1 scores within training set are 0.91 (N), 0.85 (A), 0.76 (O), and 0.73 (P).

BibTex


@inproceedings{BUT143520,
  author="Radovan {Smíšek} and Jakub {Hejč} and Marina {Ronzhina} and Andrea {Němcová} and Lucie {Maršánová} and Jiří {Chmelík} and Jana {Kolářová} and Ivo {Provazník} and Lukáš {Smital} and Martin {Vítek}",
  title="SVM Based ECG Classification Using Rhythm and Morphology Features, Cluster Analysis and Multilevel Noise Estimation",
  annote="Background: Smartphone-based ECG devices comprise great potential in screening for arrhythmias. However, its feasibility is limited by poor signal quality leading to incorrect rhythm classification. In this study, advanced method for automatic classification of normal rhythm (N), atrial fibrillation (A), other rhythm (O), and noisy records (P) is introduced.
Methods: Two-step SVM approach followed by simple threshold based rules was used for data classification. In the first step, various features were derived from separate beats to represent particular events (normal as well as pathological and artefacts) in more detail. Output of the first classifier was used to calculate global features describing entire ECG. These features were then used to train the second classification model. Both classifiers were evaluated on training set via cross-validation technique, and additionally on hidden testing set.
Results: In the Phase II of challenge, total F1 score of the method is 0.81 and 0.84 within hidden challenge dataset and training set, respectively. Particular F1 scores within hidden challenge dataset are 0.90 (N), 0.81 (A), 0.72 (O), and 0.55 (P). Particular F1 scores within training set are 0.91 (N), 0.85 (A), 0.76 (O), and 0.73 (P).",
  booktitle="Computing in Cardiology 2017",
  chapter="143520",
  doi="10.22489/CinC.2017.172-200",
  howpublished="online",
  year="2017",
  month="september",
  pages="1--4",
  type="conference paper"
}