Modeling and Interpreting CTG Curves from Labor Using Machine Learning and Pattern Recognition

Abstract: The main monitoring method during labor is cardiotocography, CTG, which measures the fetal heartbeat, FHR, as well as the uterine contractions, TOCO. The CTG is a valuable tool in assessing the fetal status and it is evaluated intermittently by clinicians during the progress of the labor. However, abnormalities in the CTG are often non-specific which makes the interpretation difficult and inconsistent. Therefore, a technical solution that could assist and simplify the assessment would be beneficial. By the use of different machine learning methods, this master's thesis investigates the properties and liaisons between the CTG data and the condition of the newborn baby, i.e., the Apgar score. All model-performance measures are evaluated based on 10-fold cross-validation to reduce bias. The thesis covers three linked areas. Firstly, a peak detection model was implemented to indicate peaks in the uterine contraction curve. The best model was a k-nearest neighbors that had a mean sensitivity of 98.70 (0.42) percent. This model was used for the second part; identifying different stages of labor with two approaches. For the classification of labor stage, the best model was a convolutional neural network with 23 consecutive layers. This model had a mean accuracy of 78.80 (0.80) percent. This step was conducted to enable the machine learning algorithms to be trained and evaluated on CTG data from the same stage of labor. In the last part, the data from the second part were used for predicting the fetal status. The results were discouraging, as the machine learning models did not perform considerably better than the fortuity. The best mean F1 score for classifying high and low Apgar scores was 5.26 (0.43) percent, and it was achieved by a decision tree model. The conclusion of this thesis is that additional research is required to establish the conjunction between CTG data and the status of the fetus. Furthermore, a data set with auxiliary samples of low Apgar scores would presumably be valuable in the aspiration of developing a supportive tool for fetal assessment.