<p>Severe fetal arrhythmias may lead to fetal heart failure and even death, with fetal electrocardiogram (FECG) signals being the pivotal means for monitoring fetal cardiac health and identifying potential arrhythmias. Existing fetal arrhythmia detection methods are notably less developed compared to those for adults, and are further hindered by the scarcity of datasets and the limited variety of arrhythmia types represented, posing challenges for model generalization. Addressing these issues, this study proposes a novel QCNN-BiLSTM based fetal arrhythmia detection method to precisely categorize non-invasively acquired FECG signals into arrhythmia (ARR) and normal rhythm (NR) groups. The Quadratic Convolutional Neural Network (QCNN), with its Quadratic neurons, outperforms traditional neurons in feature extraction, effectively discerning both shared and unique features. However, the temporal limitations of QCNN due to its fixed sliding window size are overcome by integrating Bidirectional Long Short-Term Memory (BiLSTM) networks, which excel at capturing long-term dependencies in sequential data, thereby enhancing feature extraction accuracy. Experiments conducted on the PhysioNet NIFEADB dataset validate the proposed model, achieving high performance metrics: accuracy (97.25%), recall (97.50%), F1 score (97.50%), and specificity (97.72%). To enhance the interpretability and clinical trustworthiness of the model, we incorporate SHAP values to identify the most influential features and signal segments contributing to the model’s predictions. This explainability analysis not only improves the transparency of the model but also enhances its practical utility in clinical settings. Overall, this work not only provides a robust solution for early fetal arrhythmia detection, but also highlights the need for more diverse datasets in this critical area of healthcare to expand the scope and improve the generalizability of future models.</p>