FROM SIGNAL PROCESSING TO DEEP LEARNING: A REVIEW OF FETAL HEALTH MONITORING TECHNIQUES USING BIOMEDICAL SIGNALS

Abstract

Abstract—Fetal health monitoring is essential for early de-tection of life-threatening conditions such as hypoxia and ar-rhythmia; however, conventional techniques including CTG and Doppler ultrasound are limited by subjectivity, operator de-pendency, and inter-observer variability. Non-invasive modalities such as fetal electrocardiography (FECG) and fetal phono-cardiography (FPCG) enable continuous monitoring but are significantly affected by maternal physiological interference and environmental noise, motivating the adoption of advanced artifi-cial intelligence approaches. This paper presents a structured comparative analysis of deep learning-based fetal signal pro-cessing frameworks, focusing on convolutional neural networks (CNNs), long short-term memory (LSTM) networks, and hybrid multimodal fusion architectures. A unified evaluation frame-work is employed using clinically relevant metrics, including accuracy, sensitivity, specificity, and signal-to-noise ratio (SNR) improvement, enabling consistent cross-method comparison. The analysis demonstrates a clear performance hierarchy, where deep learning approaches outperform traditional handcrafted meth-ods, hybrid CNN–LSTM models effectively capture spatiotem-poral dependencies, and multimodal fusion systems integrating electrical, acoustic, and physiological signals achieve state-of-the-art performance of approximately 92%–97%. Despite these advances, clinical deployment remains constrained by limited annotated datasets, computational complexity, and lack of real-time validation. The study highlights key research gaps and future directions involving IoT-enabled wearable sensing, edge AI deployment, and explainable deep learning for robust and personalized fetal healthcare.