Deep Learning Approaches for Monitoring Fetal Acidosis

Lack of oxygen being shared with a fetus (Intrapartum asphyxiation) can lead to fetal acidosis, a life-threating condition that can cause organ complications and brain injury in neonates. Detection of acidosis early is done using cardiotocography (CTG); medical practitioners assess key features in CTG signals to evaluate fetal health. Despite standardized CTG interpretation guidelines, high false-positive and false-negative rates have prompted efforts to make CTG interpretation more objective. Deep learning can offer a promising solution by automating the process of feature extraction and acidosis classification. This study investigates using convolutional neural networks and attention mechanisms to enhance acidosis prediction by identifying relevant features. Traditional models based on raw Fetal Heart Rate and Uterine Contraction signals were less effective, while models incorporating waveform mixing, CNNs, and attention mechanisms significantly improved precision, recall, and F1-scores. Spectral mixture models trained with Gaussian noise showed exceptional classification metrics, capturing complex patterns associated with acidosis.