[Paper Review] Detection of preventable fetal distress during labor from scanned cardiotocogram tracings using deep learning

Martin G. Frasch, Shadrian B. Strong+3 more|arXiv (Cornell University)|Jun 1, 2021

Neonatal and fetal brain pathology25 references19 citations

TL;DR

This study proposes a deep learning system that detects preventable fetal distress from scanned cardiotocogram (CTG) tracings using SSD-based object detection, achieving 93.6% accuracy in identifying 'Point A'—a critical early sign of fetal injury—enabling timely conservative interventions to prevent adverse outcomes during labor.

ABSTRACT

Despite broad application during labor and delivery, there remains considerable debate about the value of electronic fetal monitoring (EFM). EFM includes the surveillance of the fetal heart rate (FHR) patterns in conjunction with the maternal uterine contractions providing a wealth of data about fetal behavior and the threat of diminished oxygenation and perfusion. Adverse outcomes universally associate a fetal injury with the failure to timely respond to FHR pattern information. Historically, the EFM data, stored digitally, are available only as rasterized pdf images for contemporary or historical discussion and examination. In reality, however, they are rarely reviewed systematically. Using a unique archive of EFM collected over 50 years of practice in conjunction with adverse outcomes, we present a deep learning framework for training and detection of incipient or past fetal injury. We report 94% accuracy in identifying early, preventable fetal injury intrapartum. This framework is suited for automating an early warning and decision support system for maintaining fetal well-being during the stresses of labor. Ultimately, such a system could enable a physician to timely respond during labor and prevent adverse outcomes. When adverse outcomes cannot be avoided, they can provide guidance to the early neuroprotective treatment of the newborn.

Motivation & Objective

• To develop an automated system for detecting early, preventable fetal injury during labor using deep learning on scanned CTG tracings.
• To identify 'Point A'—a critical early warning sign of fetal compromise—before acidosis or irreversible injury occurs.
• To reduce unnecessary cesarean deliveries by enabling timely, conservative interventions based on real-time fetal heart rate and contraction pattern analysis.
• To support clinical decision-making and reduce obstetrical malpractice risk by providing objective, AI-driven fetal surveillance.

Proposed method

• Trained a single-shot detector (SSD) on rasterized PDF images of cardiotocogram tracings to detect 'Point A'—a clinical indicator of impending fetal injury.
• Used a unique historical archive of 5,000 CTG tracings with known adverse outcomes to train and validate the model.
• Defined specific EFM pattern features (e.g., deceleration recovery, variability changes, checkmark pattern) using modified clinical criteria for consistency with expert interpretation.
• Applied transfer learning and lightweight model architecture (MobileNet) for efficient deployment on mobile or clinical devices.
• Focused on detecting 'Point A' as a proactive intervention point, not on post-injury rescue.
• Validated model performance on a pilot dataset of 36 tracings with expert-annotated 'Point A' markers.

Experimental results

Research questions

• RQ1 Can deep learning accurately detect 'Point A'—an early, preventable sign of fetal injury—directly from scanned PDF cardiotocogram tracings?
• RQ2 Does the model’s detection of 'Point A' correlate with clinically significant fetal compromise and preventable injury?
• RQ3 Can the system reduce the need for emergency cesarean deliveries by enabling earlier, conservative management?
• RQ4 How does the performance of the deep learning model compare to expert human interpretation of CTG tracings in identifying early fetal distress?

Key findings

• The deep learning model achieved 93.6% accuracy in detecting 'Point A'—a critical early sign of preventable fetal injury—on a pilot dataset of 36 scanned CTG tracings.
• 20% of initially normal tracings reached 'Point A', with 25% of these reverting to normal with conservative management, indicating a window for intervention.
• 'Point B'—a more severe stage of fetal compromise—was reached in approximately 0.5% of cases, suggesting a low but critical threshold for emergency delivery.
• The model was trained to detect clinical decision points rather than outcomes like low Apgar score or acidosis, focusing on actionable, pre-acidotic signals.
• The system enables early detection without requiring digitized signal data, making it deployable on existing scanned medical records.
• The approach is generalizable and can be adapted to other formats (e.g., digitized signals, tabular data) post-training, enhancing clinical integration potential.