# Objective interpretation of intrapartum cardiotocography images using attention-guided convolutional neural networks

## TL;DR

This paper introduces an automated method using deep learning to objectively analyze fetal heart rate data during childbirth, improving diagnostic consistency.

## Contribution

The study proposes an attention-guided CNN with signal preprocessing for objective CTG interpretation, reducing subjectivity in fetal monitoring.

## Key findings

- The model achieved 92.66% accuracy on the internal dataset and 95.65% on the external CTU-UHB dataset.
- The attention mechanism improved focus on clinically relevant features, aligning with expert classifications and pH-based benchmarks.

## Abstract

Automated analysis of Electronic Fetal Monitoring (EFM) is essential for the precise assessment of fetal health. However, the subjective interpretation and expertise-dependent nature of conventional cardiotocogram (CTG) analysis hinder diagnostic consistency. This study aims to develop an objective interpretation approach comprising a systematic preprocessing pipeline for signal reconstruction and an attention-guided convolutional neural network for pattern classification to mitigate the risk of missed diagnoses.

A computer vision-based deep learning approach was developed. The workflow begins with a systematic preprocessing pipeline, where raw CTG images undergo grid removal, resampling, and curve reconstruction to generate standardized signal inputs. These signals are analyzed by a classifier based on the EfficientNet-B0 architecture, enhanced with a Convolutional Block Attention Module (CBAM). This attention mechanism enables the model to focus on clinically significant morphological features. The model was trained on a private clinical dataset using clinician-labeled FIGO classifications (Normal vs. Suspicious/Abnormal) as the primary outcome. To evaluate its clinical utility and robustness, the model was externally validated on the public CTU-UHB dataset, using objective umbilical artery pH levels (pH ≥7.05 vs. pH <7.05) as the benchmark.

On the internal clinical dataset, the model achieved an accuracy of 92.66% and a macro-average F1-score of 92.14%. When tested on the external CTU-UHB dataset, the model maintained an accuracy of 95.65%. These results indicate that the proposed algorithm aligns with expert visual classification and remains consistent when validated against objective physiological outcomes (pH levels). This consistency across benchmarks supports the potential robustness and clinical relevance of the learned morphological features.

This study presents an objective method for intrapartum CTG analysis. By integrating signal standardization with automated feature learning, the proposed approach addresses the inherent subjectivity of manual interpretation. It serves as a potential clinical decision support tool to assist in the consistency of fetal status assessment.