Non-inferiority analysis of a fetal heart rate artificial intelligence algorithm to registered nurse assessment

Cardiotocography (CTG) is widely used for monitoring fetal heart rate (FHR) and uterine activity (UA) during pregnancy and labor. Clinician interpretation of CTG tracings may be subjective, leading to significant inter- and intra-observer variability and inconsistent clinical decision-making. An artificial intelligence (AI) algorithm using deep learning and rule-based techniques was developed to label FHR and UA data during labor. Further development of such algorithms may add future objective monitor outputs to support clinician interpretation. This study assessed the non-inferiority of the algorithm's performance compared to registered nurse (RN) reader assessment.

This retrospective non-inferiority assessment utilized CTG data across five US hospital facilities (2013-2024). CTG labeling assessment by the AI algorithm was compared with that of RN readers for primary endpoints (baseline value, accelerations, decelerations, contraction frequency, and contraction duration) and secondary endpoints (variability and deceleration types), using a panel of qualified expert RNs who accepted or rejected the assessment. Exploratory endpoints included frequency of true and false positives/negatives and sensitivity/specificity of accelerations and decelerations. A non-inferiority threshold of 15% was required between the RN reader and algorithm.

Of 2639 de-identified fetal tracings, 800 were randomly selected for this study. The most frequently reported maternal characteristics were full-term gestational age (n = 355; 44.38%) and non-Hispanic White (n = 409; 51.13%). The algorithm performed comparably to the RN reader for all endpoints except contraction duration, with percent differences (95% confidence interval [CI]) in acceptance rates (ARs) of 3.38% (-1.5%, 8.25%) for baseline value, 3.03% (0.72%, 5.34%) for acceleration, 7.14% (3.68%, 10.61%) for deceleration, 8.25% (3.57%, 12.93%) for contraction frequency, 22.13% (17.39%, 26.86%) for contraction duration, -1.75% (-4.10%, 0.60%) for variability, and 9.13% (5.38%, 12.88%) for deceleration type. For contraction duration, AR decreased to 13.75% (9.89%, 17.61%) when allowing for a 20-s range of variability to account for visual approximations.

This study demonstrated the non-inferiority of a novel AI algorithm compared to RN readers in analyzing and assessing fetal tracings. These findings highlight AI's potential to serve as a decision-making support tool in obstetric care. By providing consistent analyses, the algorithm may help reduce observer variability and enhance clinicians' confidence in identifying fetal parameters.