Automatická klasifikace zdravotního stavu plodu pomocí kardiotokografických signálů

Abstract

This thesis focuses on the classification of cardiotocographic recordings for the diagnosis of fetal health status. The aim is to apply methods of artificial intelligence as a supportive tool in decision-making during labor management, considering that the evaluation of cardiotocographic signals is influenced by the subjectivity of medical assessments. Cardiotocographic signals and clinical information from the CTU-UHB database were preprocessed, features were extracted, and classification criteria were established based on a literature review, consultations with a clinical expert, and statistical analysis. Techniques including support vector machines, random forests, multilayer perceptrons, and convolutional neural networks were applied to classify cardiotocographic recordings as normal, suspicious, or pathological. Various data divisions and hyperparameter settings were tested. The best performance was achieved using a convolutional neural network, reaching an F1 score of 0,6914, an accuracy of 0,7857, a precision of 0,6173, a recall of 0,7857, and a loss function of 0,4279. The results demonstrate the potential of artificial intelligence-based models for the classification of cardiotocographic signals. Finally, the thesis discusses the current limitations and the future prospects for the implementation of intelligent algorithms in clinical practice.