Umělé neuronové sítě jako přístup k extrakci plodového elektrokardiogramu a detekci R-kmitů

Abstract

This thesis deals with the extraction of fetal electrocardiogram (fECG) through methods that use Artificial Neural Networks (ANN). After careful examination of non-invasive fECG (NI-fECG) signal processing, a search of current methods using ANN for extraction of fECG signal was performed. Based on the search, methods using a Linear Adaptive Neuron (ADALINE), an Adaptive Neuro-fuzzy Inference System (ANFIS) and a Recurrent Network (RNN), the so-called Echo State Network (ESN), were selected. These methods were also used in combination with Multilayer Feedforward ANN (ANN-ADALINE, ANN-ANFIS, ANN-ESN). Testing of the chosen methods was performed on real data from the Labour dataset and Pregnancy dataset databases. R-peaks were detected using two detectors to evaluate extraction and fetal heart rate (fHR). The first detector was based on continuous wavelet transform (CWT), the second detector was based on Multilayer Feedforward ANN. For evaluation the overall probability of correct detection (ACC), sensitivity (SE), positive predictive value (PPV) and the harmonic mean of SE and PPV (F1) were determined. The functionality of chosen methods was verified by comparison to reference anotations. In comparison to methods ADALINE, ANFIS, ANN-ADALINE, ANN-ANFIS a ANN-ESN, the ESN method achieved the best results. For data from the Labor dataset, the ACC value reached 78.65 %, for data from the Pregnancy dataset, the ACC value was over 80 %. A graphical user interface (GUI) was created for processing, analysis and evaluation in MATLAB.