Uterine Contractions Features: Evolution Throughout Pregnancy Using Electrohysterography

Tese de mestrado, Bioestatística, Universidade de Lisboa, Faculdade de Ciências, 2022Uterine Electrohysterography (EHG) is a method used to detect uterine electrical activity which pro vokes uterine contractions. This technique has been studied to detect uterine contractions, and it has been reported as an overall non-invasive better alternative compared with the gold standard IUPC (In trauterine Pressure Catheter) and the TOCOgram. The study of EHG, specifically EHG signal features, has gained more interest in preterm prediction and uterine contraction detection. Some studies also com pare features such as duration, entropies, or frequency features in different pregnancy scenarios. With the use of the longitudinal 16 electrode Iceland Database, which comprises EHG recordings from 45 women, this work aims to analyze a set of 15 of the most common EHG extracted features over the gestational period and including the effect of pregnancy and woman-specific factors, such as Age, BMI or Placental Position. Linear Mixed Models with random intercept are used to analyze this data, and given the high degree of missing values, a sensitivity analysis using multiple imputation is performed to assess the robustness of the models in order to validate the results. It was observed that some features, such as Log Energy, Log Shannon Entropy, or Log Peak Amplitude, increase towards labor. BMI ap pears to exert a low-pass filter on some features such as Log Peak Amplitude, Log Mean, and Median Frequency. Age of the woman also showed to have some effect on the features Log Sample Entropy and Crest factor. More interestingly, Placental Position was observed to significantly influence all entropy features, except for Log Shannon Entropy and Lyapunov Exponent. Some of the obtained results are consistent with the knowledge of the respective EHG features in the literature. Additionally, the sensi tivity analysis showed robust results for most of the models, confirming that Linear Mixed Models can be considered a robust technique to model correlated data, even with high levels of missingness