Exploring the utility of artificial intelligence of intrapartum cardiotocography: a systematic review

Abstract Background: Cardiotocography (CTG) interpretation is complex and highly subjective. Misinterpretation can result unnecessary, late, or inadequate intervention; potentially harming the mother or fetus. Artificial intelligence (AI) could play a role in reducing the likelihood of these incidents. Purpose: To identify the current state-of-the-art in AI models for CTG interpretation and provide clinicians and AI developers alike with an overview of this landscape and guide the development of future models. Methods: We searched PubMed, EMBASE, Ovid Medline, and IEEE Xplore for studies published from 01/06/2005 to 07/06/2020. Studies focused on AI applications to CTG were included, with the performance metrics (accuracy, sensitivity, and specificity) being extracted for analysis. PROSPERO registration: CRD42021250394. Results: 38 articles were eligible for inclusion, though all studies were focused on pre-clinical performance evaluation. The types of AI techniques employed included support vector machines (11/38), decision trees (10/38), random forests (8/38), neural networks (23/38), and novel custom algorithms (11/38). Each model demonstrated promise in a pre-clinical setting although true clinical value is presently uncertain. Other issues included the classification systems used by AI developers, as well as the limited scope of these classification systems when compared to the more comprehensive clinical AI systems already implemented clinically in similar medical applications. Conclusion: AI shows promise as an adjunct surveillance tool in CTG interpretation. However, currently, it is too early to conclusively determine its implementation value in a clinical setting. To do so, these AIs need to be developed for and validated in high quality prospective clinical evaluations.