Yalçın Emre, Aslan Serpil, Toğaçar Mesut, Demir Süleyman Cansun
Department of Obstetrics and Gynecology, Division of Perinatology, Cukurova University School of Medicine, 01330 Adana, Turkey.
Department of Computer Engineering, Faculty of Engineering and Natural Sciences, Malatya Turgut Ozal University, 44210 Malatya, Turkey.
Diagnostics (Basel). 2025 Jun 6;15(12):1444. doi: 10.3390/diagnostics15121444.
The aim of this study is to develop a hybrid artificial intelligence (AI) approach to improve the accuracy, efficiency, and reliability of Down Syndrome (DS) risk prediction during first trimester prenatal screening. The proposed method transforms one-dimensional (1D) patient data-including features such as nuchal translucency (NT), human chorionic gonadotropin (hCG), and pregnancy-associated plasma protein A (PAPP-A)-into two-dimensional (2D) Aztec barcode images, enabling advanced feature extraction using transformer-based deep learning models. : The dataset consists of 958 anonymous patient records. Each record includes four first trimester screening markers, hCG, PAPP-A, and NT, expressed as multiples of the median. The DS risk outcome was categorized into three classes: high, medium, and low. Three transformer architectures-DeiT3, MaxViT, and Swin-are employed to extract high-level features from the generated barcodes. The extracted features are combined into a unified set, and dimensionality reduction is performed using two feature selection techniques: minimum Redundancy Maximum Relevance (mRMR) and RelieF. Intersecting features from both selectors are retained to form a compact and informative feature subset. The final features are classified using machine learning algorithms, including Bagged Trees and Naive Bayes. : The proposed approach achieved up to 100% classification accuracy using the Naive Bayes classifier with 1250 features selected by RelieF and 527 intersecting features from mRMR. By selecting a smaller but more informative subset of features, the system significantly reduced hardware and processing demands while maintaining strong predictive performance. : The results suggest that the proposed hybrid AI method offers a promising and resource-efficient solution for DS risk assessment in first trimester screening. However, further comparative studies are recommended to validate its performance in broader clinical contexts.
本研究的目的是开发一种混合人工智能（AI）方法，以提高孕早期产前筛查中唐氏综合征（DS）风险预测的准确性、效率和可靠性。所提出的方法将一维（1D）患者数据（包括颈项透明层（NT）、人绒毛膜促性腺激素（hCG）和妊娠相关血浆蛋白A（PAPP-A）等特征）转换为二维（2D）阿兹特克条形码图像，从而能够使用基于Transformer的深度学习模型进行高级特征提取。数据集由958份匿名患者记录组成。每份记录包括四个孕早期筛查标志物，即hCG、PAPP-A和NT，以中位数倍数表示。DS风险结果分为三类：高、中、低。采用三种Transformer架构——DeiT3、MaxViT和Swin——从生成的条形码中提取高级特征。将提取的特征合并成一个统一的集合，并使用两种特征选择技术进行降维：最小冗余最大相关性（mRMR）和RelieF。保留两个选择器的相交特征，以形成一个紧凑且信息丰富的特征子集。使用包括Bagged Trees和朴素贝叶斯在内的机器学习算法对最终特征进行分类。所提出的方法使用朴素贝叶斯分类器，在通过RelieF选择的1250个特征和来自mRMR的527个相交特征的情况下，实现了高达100%的分类准确率。通过选择一个更小但更具信息性的特征子集，该系统在保持强大预测性能的同时，显著降低了硬件和处理需求。结果表明，所提出的混合AI方法为孕早期筛查中的DS风险评估提供了一种有前景且资源高效的解决方案。然而，建议进行进一步的比较研究，以验证其在更广泛临床背景下的性能。
