FetalSleepNet: A Transfer Learning Framework with Spectral Equalisation Domain Adaptation for Fetal Sleep State Classification.

Abstract

Objective: Fetal sleep state classification is essential for identifying neurodevelopmental complications like hypoxia, but manual annotation is subjective and labor intensive, and fetal EEG (fEEG) data is extremely scarce. Method(s): We propose FetalSleepNet, the first deep learning architecture specifically developed for automated sleep staging from the fEEG. To address the scarcity of labeled data, we implement the first cross-developmental (adult-to fetal) and cross-species (human-to-sheep) transfer learning framework for this task, utilizing Spectral Equalisation (SE) to align the frequency-domain characteristics of adult human EEG with the fetal sheep target. Result(s): Our findings prove the irreplaceable effective ness of this adaptation: while direct transfer on raw EEG almost fails with only 18.7% accuracy, applying SE allows even a frozen model to reach 73.6% accuracy, effectively mitigating the cross-domain spectral mismatch. With full fine-tuning, FetalSleepNet achieves a state-of-the-art accuracy of 86.6% and a macro F1-score of 62.5%. Conclusion(s): Beyond high-accuracy classification, Fetal SleepNet establishes a robust "label engine" paradigm. By generating high-fidelity neurophysiological annotations, it facilitates a framework for training proxy sleep stagers on broader, non-invasive clinical modalities. This paves the way for scalable, real-time fetal monitoring and early risk prediction in clinical settings.Copyright © 2013 IEEE.