FetalADM: a double-layer ensemble model based on SMOTE_KM and RF-RFE for fetal trisomy 21, 18, and 13 detections

Abstract Noninvasive prenatal testing (NIPT), which utilizes high-throughput sequencing technology to analyze cell-free DNA fragments from maternal peripheral plasma, has been widely adopted in clinical practice. However, accurately detecting fetal trisomy remains a challenge. To address this issue, we propose a novel double-layer ensemble model designed for detecting fetal trisomy13, 18 and 21. Firstly, we integrate the Synthetic Minority Oversampling Technique with K-Means clustering to augment positive samples, effectively balancing the training dataset. Subsequently, we implement feature selection algorithms to identify the optimal feature combination. Leveraging these enhancements, we develop a fast and accurate multi-class classification model FetalADM based on machine learning. Evaluate its performance on three independent test datasets: T54, T210, and T136. Notably, on the T54 dataset, FetalADM achieved a perfect 100% accuracy in detecting trisomy 21, 18, and 13. On the T210/T136 dataset, the model misclassified only 2/1 out of 210/136 samples (accuracy = 99.0%/99.3%), respectively, compared to 31/12 misclassifications by traditional bioinformatics methods. Specially, as a four-class classifier, FetalADM enables direct prediction of specific trisomy types, distinguishing itself from most binary-class models while maintaining high efficiency and accuracy. These results demonstrate that it outperforms conventional bioinformatics methods, underscoring its potential to improve the clinical diagnostic accuracy of fetal aneuploidies.