Accurate and early detection of fetal heart anomalies is a critical challenge in prenatal care. This study utilizes the underexplored NInFEA Maternal and Fetal ECG Database, which provides non-invasive fetal ECG recordings captured between 21 and 27 weeks of gestation. The goal is to build a robust machine learning pipeline for fetal heart anomalies detection, starting with unsupervised learning methods and progressing to supervised classification.
Despite challenges such as weak fetal signals, high noise, and the absence of clinical labels, the NInFEA dataset offers rare and valuable insights by including simultaneous maternal and fetal ECG data for anomaly detection. The workflow begins with unsupervised learning techniques, including DBSCAN, One-Class SVM, and Isolation Forest, to explore potential anomalies in the data. Isolation Forest, with a silhouette score of 0.5868, provided the best balance between interpretability and performance.
For supervised learning, a labeling strategy was then applied to differentiate normal from anomalous fetal heart patterns. Classification models, including Logistic Regression, XGBoost, Random Forest, and Support Vector Machine, were trained using the labeled data. XGBoost achieved F1-scores of 0.96 for normal patterns and 0.92 for anomalous patterns, demonstrating strong classification performance. Data preprocessing steps, such as ECG signal filtering, feature extraction, and SMOTE for handling class imbalance, ensured the reliability of the models, and cross-validation confirmed the generalizability of the results across various fetal cases.
This work not only demonstrates the potential of machine learning to enhance fetal health monitoring but also underscores the untapped value of the NInFEA dataset. It contributes to the development of scalable, efficient, and accurate diagnostic tools in prenatal care.