Algorithms for information extraction and signal annotation on long-term biosignals using clustering techniques

• Author(s): Abreu, Rodolfo Telo Martins de
• Date: 2012
• Persistent ID: http://hdl.handle.net/10362/8250
• Origin: Repositório Institucional da UNL
• Subject(s): Biosignals; Waves; Events detection; Features extraction; Pattern recognition; k-means
• Document Type: Master thesis
• Language: English
• Advisor(s): Gamboa, Hugo
• Contributor(s): RUN

Description

Dissertação para obtenção do Grau de Mestre em Engenharia Biomédica

One of the biggest challenges when analysing data is to extract information from it, especially if we dealing with very large sized data, which brings a new set of barriers to be overcome. The extracted information can be used to aid physicians in their diagnosis since biosignals often carry vital information on the subjects. In this research work, we present a signal-independent algorithm with two main goals: perform events detection in biosignals and, with those events, extract information using a set of distance measures which will be used as input to a parallel version of the k-means clustering algorithm. The first goal is achieved by using two different approaches. Events can be found based on peaks detection through an adaptive threshold defined as the signal’s root mean square (RMS) or by morphological analysis through the computation of the signal’s meanwave. The final goal is achieved by dividing the distance measures into n parts and by performing k-means individually. In order to improve speed performance, parallel computing techniques were applied. For this study, a set of different types of signals was acquired and annotated by our algorithm. By visual inspection, the L1 and L2 Minkowski distances returned an output that allowed clustering signals’ cycles with an efficiency of 97:5% and 97:3%, respectively. Using the meanwave distance, our algorithm achieved an accuracy of 97:4%. For the downloaded ECGs from the Physionet databases, the developed algorithm detected 638 out of 644 manually annotated events provided by physicians. The fact that this algorithm can be applied to long-term raw biosignals and without requiring any prior information about them makes it an important contribution in biosignals’ information extraction and annotation.