Fetal states identification in cardiotocographic tracings through discrete emissions multivariate hidden Markov models

Abstract

Background and objectivesComputerized Cardiotocography (cCTG) allows to analyze the Fetal Heart Rate (FHR) objectively and thoroughly, providing valuable insights on fetal condition. A challenging but crucial task in this context is the automatic identification of fetal activity and quiet periods within the tracings. Different neural mechanisms are involved in the regulation of the fetal heart, depending on the behavioral states. Thereby, their correct identification has the potential to increase the interpretability and diagnostic capabilities of FHR quantitative analysis. Moreover, the most common pathologies in pregnancy have been associated with variations in the alternation between quiet and activity states. MethodsWe address the problem of fetal states clustering by means of an unsupervised approach, resorting to the use of a multivariate Hidden Markov Models (HMM) with discrete emissions. A fixed length sliding window is shifted on the CTG traces and a small set of features is extracted at each slide. After an encoding procedure, these features become the emissions of a multivariate HMM in which quiet and activity are the hidden states. After an unsupervised training procedure, the model is used to automatically segment signals. ResultsThe achieved results indicate that our developed model exhibits a high degree of reliability in identifying quiet and activity states within FHR signals. A set of 35 CTG signals belonging to different pregnancies were independently annotated by an expert gynecologist and segmented using the proposed HMM. To avoid any bias, the physician was blinded to the results provided by the algorithm. The overall agreement between the HMM's predictions and the clinician's interpretations was 90%. ConclusionsThe proposed method reliably identified fetal behavioral states, the alternance of which is an important factor in the fetal development. One key strength of our approach lies in the ease of interpreting the obtained results. By utilizing a small set of parameters that are already used in cCTG and possess clear intrinsic meanings, our method provides a high level of explainability. Another significant advantage of our approach is its fully unsupervised learning process. The states identified by our model using the Baum-Welch algorithm are associated with the “Active” and “Quiet” states only after the clustering process, removing the reliance on expert annotations. By autonomously identifying the clusters based solely on the intrinsic characteristics of the signal, our method achieves a more objective evaluation that overcomes the limitations of subjective interpretations. Indeed, we believe it could be integrated in cCTG systems to obtain a more complete signal analysis.