A systematic review on AI-driven noninvasive fetal ECG processing and analysis methods.

Introduction: Fetal electrocardiogram is a critical tool for fetal health monitoring, and it is obtained nowadays in clinical practice through an invasive procedure. An alternative is the use of noninvasive fetal electrocardiogram (NI-fECG) that can be obtained by placing a matrix of electrodes on the maternal abdomen. However, the extraction of NI-fECG from abdominal signals (ADSs) remains challenging due to the low amplitude of fetal signals, overlapping with maternal ECG, and because of contamination from other noise sources. Traditional signal processing methods have been widely used but are often limited by incomplete separation, signal distortion, and the need for manual parameter tuning. Artificial intelligence (AI) has emerged as a powerful solution, focusing on deep learning techniques to improve NI-fECG extraction and noise suppression, offering automated feature extraction, adaptability to signal variability, and real-time processing, and addressing the limitations of classical methods. This study presents a comprehensive systematic review of AI methods in NI-fECG processing and analysis, evaluating their effectiveness and limitations. Methods: The review follows a structured methodology, including study selection and analysis of AI techniques, following PRISMA 2020 guidelines, and covering peer-reviewed studies published between 2014 and 2025. It categorizes AI models into neural network-based, generative, hybrid, and specialized architectures. Results: Twenty-two studies met the inclusion criteria. AI-based approaches outperformed traditional signal processing in NI-fECG extraction and denoising. Generative and specialized architectures showed the strongest morphology preservation, while hybrid models added robustness. Six studies applied AI to diagnostic tasks, including fetal heart rate estimation, arrhythmia detection, and congenital heart disease classification. Limitations include scarce ground-truth datasets and a lack of standardized evaluation protocols. Discussion: No single architecture is universally superior across all pipeline stages. A key research gap is the need for AI models that conserve NI-fECG morphology for medical decision-making, alongside standardized benchmarking and explainable models.