LiteFDNet: A Lightweight Deep Network for Cardiotocography-Based Fetal Distress Detection

Abstract

This retrospective single-center study developed and evaluated LiteFDNet, a lightweight deep neural network for intrapartum fetal distress detection from cardiotocography (CTG). All eligible deliveries from April 2014 to December 2024 were included under IRB approval (LGECH-2023-002; consent waived). Labels were refined using umbilical artery blood gas (distress: pH < 7.20; normal: pH 7.20–7.27), with exclusions for multiple gestations, major anomalies, missing CTG/outcomes, or non-analysable signals. Distress cases were matched 1:2 to controls, yielding 4,353 dual-channel CTG segments (1,451 distress; 2,902 normal). For each delivery, 30–60 min fetal heart rate and uterine contraction traces were artefact-screened, short gaps interpolated, z-normalised, lightly smoothed, and rendered as 224×224 images. LiteFDNet was inspired by ResNet-18 and trained with AdamW, cosine annealing, and mild affine augmentation; performance was primarily assessed using stratified 10-fold cross-validation, while an independent held-out test split was retained only for supplementary transparency analyses. The primary outcome was AUC, and secondary outcomes included accuracy, sensitivity, specificity, and F1 score. Among 4,353 women (maternal age 30.3 ± 4.4 years; gestational age 38.9 ± 1.6 weeks), LiteFDNet achieved a mean AUC of 0.940 ± 0.015 across the cross-validation folds and showed favorable performance relative to the compared baselines. Ablation analyses supported the contributions of the task-specific lightweight architectural modifications (t = 2.77; p < 0.01) and squeeze-and-excitation blocks (t = 2.15; p = 0.03). LiteFDNet may provide an accurate and computationally efficient proof-of-concept approach for CTG-based fetal distress screening, but further multi-center validation and comparison with stronger direct temporal baselines are warranted.