Use of clustering in fetal state assessment by cardiotocography

Hypoxia is the fetal condition that occurs when there is a reduced supply of oxygen in the fetus. Some level of hypoxia occurs in all pregnancies during childbirth, and it is due to the mechanical pressure exercised during labor. The possibility of reaching critical values for the health of the fetus depends on how and for how long time there is a reduced oxygen supply. When hypoxia reaches extreme level and it results in metabolic acidosis, it is harmful to the baby. Nowadays, the cardiotocography (CTG) is the easiest, non-invasive, painless and widely used tool to simultaneously record the fetal heart rate signal and uterine contraction activity during pregnancy and childbirth. From its signal, by analyzing its characteristics it is possible to evaluate the fetal state assessment. From the systematic research of studies already present in literature, emerged that a greater attention is increasingly focused on the use of artificial intelligence for clinical decision support, particularly the use of supervised ML techniques. For this reason, this thesis has the aim to develop and implement an unsupervised ML algorithm, the CTG-GCA, able to recognize, identify and bring together cardiotocographic signals of hypoxic and non-hypoxic fetus. The proposed model was trained using the CTU-UHB database, to extract through a genetic algorithm the most relevant features of the CTG signals and subsequently, once extracted, the features were considered to create the 2 clusters using a hierarchical agglomerative clustering algorithm. Through the model, four features were selected (BW, Area ACC, Area DEC and HF) and the statistical analysis of the clusters (Cluster 1 of 289 subjects and Cluster 2 of 263) led to the conclusion that for a correct identification of hypoxic and non-hypoxic fetus the most important features are both those strictly linked to the characteristics of the CTG signal, such as accelerations, decelerations, heart rate variability and baseline, and neonatal clinical outcome information, as the 5-minute Apgar index. In conclusion, this work, thanks to the promising results obtained, can be considered a forerunner for this branch of research and a good starting point for future developments.