Mridu Sahu

Early detection of fetal cardiac abnormality facilitates timely intervention and management. Existing noninvasive fetal electrocardiogram (fECG) extraction techniques often rely on fixed or heuristic optimization parameters and segment-level evaluation strategies, limiting robustness under nonstationary maternal–fetal signal overlap and introducing potential information leakage. This work presents an adaptive optimization framework utilizing the adaptive Lévy flight-based cuckoo search (ALCS) method for improved fECG extraction and detection of fetal arrhythmias, contributing to intelligent human–machine systems in obstetric monitoring. The framework employs ALCS as an outer loop optimizer to determine an optimal initialization of the least mean square (LMS) filter coefficients for adaptive noise cancellation (ANC), where a convergence-aware Lévy flight parameter governs the search process. The Lévy parameter $\lambda$ evolves adaptively within $[1.5, 2.0]$ enabling an effective transition from exploration to exploitation. The proposed optimizer achieves approximately 1.8–$2.0\times$ faster convergence than standard LMS, reducing the mean-squared error by $\sim$74% to 0.10 within 50 iterations. A comprehensive feature set exceeding 100 descriptors is extracted from fECG signals and evaluated using ensemble classifiers. Ensuring a leakage-safe evaluation protocol, the experiments were conducted using subjectwise grouped cross-validation with foldwise feature selection on the noninvasive fetal ECG arrhythmia database available from PhysioNet. The proposed framework achieves robust arrhythmia detection, attaining 91.2% accuracy and 95.0% area under the curve across unseen subjects. To quantify systematic bias and agreement across evaluation stages, Bland–Altman analysis is employed to provide clinical interpretability beyond aggregate scores. Overall, the results demonstrate that convergence-aware Lévy-based optimization improves ANC performance for fECG analysis, while acknowledging dataset scale and offline evaluation limitations.

Foetal health classification that is accurate utilizing cardiotocography data is essential for timely medical interventions during pregnancy. The research looks into the efficacy of Machine learning methods for making use of a CTG data set to categorize three categories of fatal health: normal, suspect, and pathological. The dataset was preprocessed using the Synthetic Minority Oversampling Technique (SMOTE) and Min-Max normalization to address feature scaling and class imbalance. Machine learning evaluated five individual models Support Vector Machine, Random Forest, XGBoost, K-Nearest Neighbours, and Logistic Regression and two ensemble approaches: Blending and Stacking. The Blending Ensemble combined predictions from Random Forest, XGBoost, and SVM via majority voting, while the Stacking Ensemble used these models with Logistic Regression as a meta learner. Metrics including accuracy, precision, recall, F1-score, and AUC-ROC were accustomed to evaluate performance on a stratified 80–20 train-test split. Results indicate that Individual models were surpassed by ensemble approaches, with the Stacking Ensemble obtaining the best fidelity accuracy of 97.7% to the Blending Ensemble at 97.6%. These models also demonstrated improved detection of minority classes (suspect and pathological), which is critical for clinical reliability. Random Forest and XGBoost showed competitive individual performance, while KNN yielded the lowest accuracy. This study emphasizes how effective group learning can be, coupled with effective preprocessing, to enhance fatal health prediction, offering a robust framework for developing decision support tools in obstetrics.