Physiological interpretation of cardiotocograph (CTG): Inter-observer and “Observer-AI” agreements in recognising types of fetal hypoxic stress

Abstract

Listen

Translate article

Introduction: Interobserver reliability in interpreting cardiotocographs (CTGs) using traditional categorization into "normal," "suspicious," and "pathological" is typically very low ranging from Kappa 0.3 to 0.6. Physiological CTG interpretation focuses on identifying specific features of different types of fetal hypoxic stress and a combination of features which are associated with adverse perinatal outcomes. Objective: To evaluate the agreement among members of the Editorial Board (EBM) of the international expert consensus statement on physiological CTG interpretation, members of the international expert consensus panel (ICP), and the Tweris Mini App (TMA), which is an AI-based CTG interpretation tool developed based on international expert consensus statement. Materials & Methods: Thirty 10–15-minute CTG trace segments, representing different types of fetal hypoxic stress (chronic, gradually evolving compensated, gradually evolving decompensated, subacute, and acute) and abnormal CTG patterns (atypical sinusoidal or the “Poole Shark Teeth”, typical sinusoidal and the ZigZag patterns), were independently reviewed by 3 editorial board members and 3 international expert consensus panel members. An orthopedic surgeon independently analyzed the same traces using the TMA. Fleiss' Kappa and Z-scores were used for statistical analysis. Results: Inter-observer agreement was 0.8 (95% CI: 0.72-0.87, p < .001) among EBM and 0.68 (95% CI: 0.60-0.76, p < .001) among ICP, with a statistically significant difference between these groups (p < .05). Agreement between EBM and the Tweris Mini App was higher than between ICP and the Tweris Mini App (0.81 vs 0.73, p = .06). Conclusion: The inter-observer agreement when using physiological CTG interpretation surpasses that of the inter-observer agreement reported with traditional systems of CTG classification, with higher interobserver agreement among editorial board members compared to international expert consensus panel members. There was a substantial agreement between editorial board members and the Tweris Mini app which was higher than between ICP and the Tweris Mini App. These findings highlight the potential of AI-assisted tools, such as the Tweris Mini App, based on physiological CTG interpretation, to provide expert-level diagnostic accuracy in clinical practice. The Tweris Mini App was found to be superior in consistently recognising rare fetal heart rate patterns.