The Influences of Different Window Functions and Lengths on Image-based Time-Frequency Features of Fetal Heart Rate Signals

CÖMERT Z., Boopathi A. M. , Velappan S., Yang Z., KOCAMAZ A. F.

26th IEEE Signal Processing and Communications Applications Conference (SIU), İzmir, Türkiye, 2 - 05 Mayıs 2018 identifier identifier


In the clinical practice, the fetal distress conditions such as hypoxia are detected routinely during antepartum and even intrapartum periods with the help of electronic fetal monitoring device, often called Cardiotocography (CTG). Due to the noticeable advances in signal processing, pattern recognition, machine learning techniques and the introduction of the quantitative diagnostic indices, the automated CTG analysis has become a quite essential tool. In this study, we come up with a new investigation on the influences of different window functions on image-based time-frequency (IBTF) features of fetal heart rate (FHR) signals for fetal hypoxia detection. In addition to the traditionally used morphological features, the spectrogram images provided by Short Time Fourier Transform (STFT) were taken into account with different windows functions such as Hamming, Hann, Kaiser, and Blackman as well as different windows lengths. Then, the spectrogram images were converted into 8-bits gray-scale images and IBTF features were obtained using Gray Level Co-occurrence Matrix (GLCM). At the end of the feature extraction stage for signal representation, we achieved a quite large feature set, and we employed genetic algorithm (GA) and support vector machine (SVM) classifier in order to reveal the most relevant features. The whole experiments were performed on an open CTU-UHB intrapartum CTG database. The experimental results show that the IBTF features have relatively increased the classification performance. All window functions ensured encouraging results. Furthermore, the GA ensured the determination of the 7 most relevant features. Thus, the dimension of feature space was reduced from 28 to 7. Moreover, the classification success increased. Consequently, the most efficient performances (Quality Index = 73.45%) were achieved with Hamming and Kaiser window functions.