Classifying the type of delivery from cardiotocographic signals: A machine learning approach. (November 2020)

• Record Type: Journal Article
• Authors: Ricciardi, C.; Improta, G.; Amato, F.; Cesarelli, G.; Romano, M.
• Is Part Of: Computer methods and programs in biomedicine. Volume 196(2020)
• Publication Date: 2020-11
• Subjects: Foetal heart rate -- Cardiotocography -- Machine learning -- Caesarean section; Medicine -- Computer programs -- Periodicals; Biology -- Computer programs -- Periodicals; Computers -- Periodicals; Medicine -- Periodicals; Médecine -- Logiciels -- Périodiques; Biologie -- Logiciels -- Périodiques; Biology -- Computer programs; Medicine -- Computer programs; Periodicals; Electronic journals; 610.28
• DOI: 10.1016/j.cmpb.2020.105712
• Languages: English
• ISSNs: 0169-2607

Abstract

Highlights: A dataset of 370 CTG signals has been collected in clinical environment. A custom-made automatic software was used to extract 17 features from the CTGs. J48, RF, GBT, ADA-B of J48 and decorate were used to classify patients according to their delivery. The procedures of 10 folds cross-validation and SMOTE were also applied. Several evaluation metrics were computed for comparing the algorithms. Abstract: Background and objective: Cardiotocography (CTG) is the most employed methodology to monitor the foetus in the prenatal phase. Since the evaluation of CTG is often visual, and hence qualitative and too subjective, some automated methods have been introduced for its assessment. Methods: In this paper, a custom-made software is exploited to extract 17 features from the available CTG. A preliminary univariate statistical analysis is performed; then, five machine learning algorithms, exploiting ensemble learning, were implemented (J48, Random Forests (RF), Ada-boosting of decision tree (ADA-B), Gradient Boosting and Decorate) through Knime analytics platform to classify patients according to their delivery: vaginal or caesarean section. The dataset is composed by 370 signals collected between 2000 and 2009 in both public and private hospitals. The performance of the algorithms was evaluated using 10 folds cross validation with different evaluation metrics: accuracy, precision, sensitivity, specificity, area under the curve receiver operating characteristic (AUCROC). Results: While only two features were significantly different (gestation week and power expressed by the high frequency band of FHR power spectrum), from the statistical point of view, machine learning results were great. The RF obtained the best results: accuracy (91.1%), sensitivity (90.0%) and AUCROC (96.7%). The ADA-B achieved the highest precision (92.6%) and specificity (93.1%). As expected, the lowest scores were obtained by J48 that was the base classifier employed in all the others empowered implementations. Excluding the J48 results, the AUCROC of all the algorithms was greater than 94.9%. Conclusion: In the light of the obtained results, that are greater than those ones found in the literature from comparable researches, it can be stated that the machine learning approach can actually help the physicians in their decision process when evaluating the foetal well-being.