Titelaufnahme

Titel
Neural network based electrocardiography anomaly detection / by Matthias Wess
Weitere Titel
Neuronale Netzwerk basierte Detektion von EKG-Anomalien
VerfasserWess, Matthias
Begutachter / BegutachterinJantsch, Axel ; Pudukotai Dinakarrao, Sai Manoj
ErschienenWien, 2017
UmfangV, 56 Blätter : Illustrationen, Diagramme
HochschulschriftTechnische Universität Wien, Diplomarbeit, 2017
Anmerkung
Abweichender Titel nach Übersetzung der Verfasserin/des Verfassers
Zusammenfassung in deutscher Sprache
SpracheEnglisch
DokumenttypDiplomarbeit
Schlagwörter (EN)Machine learning - Anomaly detection - Neural Networks
URNurn:nbn:at:at-ubtuw:1-93199 Persistent Identifier (URN)
Zugriffsbeschränkung
 Das Werk ist frei verfügbar
Dateien
Neural network based electrocardiography anomaly detection [4.74 mb]
Links
Nachweis
Klassifikation
Zusammenfassung (Englisch)

Bio-signals such as Electrocardiography and Electroencephalography are the time variant signals representing the electrical outputs from the corresponding measurement instruments and are widely used to assess the health of patients. The objective of this thesis is to implement neural network based machine learning algorithm on FPGA to detect the anomalies in ECG signals, with a better performance and accuracy, compared to statistical models. An overview of existing techniques for anomaly detection in ECG signals is presented along with their merits and demerits, implementation strategies for different aspects such as feature selection, feature reduction and classification. Based on the performed comprehensive analysis, few of the most efficient algorithms were selected for simulation in MATLAB for performance comparison. An implementation with principal component analysis for feature reduction and multi-layer perceptron for classification, proved superior to other algorithms. For implementation on FPGA, effects of several parameters and simplification on performance, accuracy and power consumption were studied. Piecewise linear approximation for activation function and fixed point implementation were effective methods to reduce the amount of the needed resources. The resulting neural network with eight inputs and four neurons in the hidden layer, achieved in spite of the simplifications the same overall accuracy as in simulation. An accuracy of 97.5% was achieved on average for 42 records in MIT-BIH database. These results suggest that the presented technique with several simplifications for FPGA implementation of neural networks is a valid technique with good performance but less power consumption, and area, and no significant loss of accuracy.