Poster Session 1
Category: Digital Health Technologies (DHT)
Poster Session 1
Debjyoti Karmakar, MD
Obstetrician Gynaecologist Clinician Researcher
University of Melbourne/Mercy Hospital for Women
Mercy Hospital for Women/Heidelberg, Victoria, Australia
Lochana Mendis
University of Melbourne
University of Melbourne, Victoria, Australia
Brett Manley
University of Melbourne/Mercy Hospital for Women
University of Melbourne/Mercy Hospital for Women, Victoria, Australia
Jeanie Cheong
University of Melbourne
University of Melbourne, Victoria, Australia
Emerson Keenan
University of Melbourne
University of Melbourne, Victoria, Australia
Jim Holberton
Mercy Hospital for Women
Mercy Hospital for Women, Victoria, Australia
Enes Makalic
Monash University
Monash University, Victoria, Australia
Fiona Brownfoot, MBBS, PhD (she/her/hers)
University of Melbourne/Mercy Hospital for Women
Mercy Hospital for Women/University of Melbourne, Victoria, Australia
To evaluate performance and generalizability of an interpretable AI model incorporating clinical heuristics and game theory–based feature selection, using minimally pre-processed CTG signals to predict intrapartum fetal asphyxia.
Study Design:
Diagnostic model development and external validation study. The model was trained on 36,792 labors from an Australian tertiary hospital (2010–2021) and validated on two independent cohorts: an Australian community hospital (n = 15,413; 2010–2021) and a European university hospital (n = 552; 2010–2012). Analysis followed a pre-specified protocol.
< ![if !supportLineBreakNewLine] >
< ![endif] >
Singleton labors ≥36 weeks’ gestation with ≥15 minutes of interpretable CTG in the final hour. Exclusions included major congenital anomalies and >90% signal dropout.
Fetal heart rate and uterine activity signals processed via a Python pipeline to derive physiological metrics, artifact measures, and deep learning embeddings. Waveform features were engineered using clinician expertise. Shapley Additive Explanations guided feature selection. The final model was a gradient boosting classifier integrating the features.
The primary outcome was intrapartum fetal asphyxia, defined by a Delphi-derived composite including hypoxic–ischemic encephalopathy, stillbirth, therapeutic hypothermia, low Apgar scores, prolonged resuscitation, or cord pH < 7.00. The secondary outcome was cord pH < 7.15.
Results:
Asphyxia prevalence was 3.3%, 3.2%, and 7.1% across cohorts. In the development cohort, AUROC was 0.76 (95% CI, 0.72–0.80), with 68.5% sensitivity and 74.2% specificity. At matched clinician specificity (78.5%), model sensitivity improved from 35.8% to 60.2%. External AUROCs ranged between 0.73 and 0.83. The model achieved sensitivity for pH < 7.15 of 83.9%.
Conclusion:
This interpretable AI model significantly and consistently outperformed clinician judgment and generalized across diverse clinical settings. Its integration of domain knowledge and signal integrity supports further evaluation as a real-time decision-support tool in labor.