DataSheet1_A deep learning mixed-data type approach for the classification of FHR signals.docx

The Cardiotocography (CTG) is a widely diffused monitoring practice, used in Ob-Gyn Clinic to assess the fetal well-being through the analysis of the Fetal Heart Rate (FHR) and the Uterine contraction signals. Due to the complex dynamics regulating the Fetal Heart Rate, a reliable visual interpretation of the signal is almost impossible and results in significant subjective inter and intra-observer variability. Also, the introduction of few parameters obtained from computer analysis did not solve the problem of a robust antenatal diagnosis. Hence, during the last decade, computer aided diagnosis systems, based on artificial intelligence (AI) machine learning techniques have been developed to assist medical decisions. The present work proposes a hybrid approach based on a neural architecture that receives heterogeneous data in input (a set of quantitative parameters and images) for classifying healthy and pathological fetuses. The quantitative regressors, which are known to represent different aspects of the correct development of the fetus, and thus are related to the fetal healthy status, are combined with features implicitly extracted from various representations of the FHR signal (images), in order to improve the classification performance. This is achieved by setting a neural model with two connected branches, consisting respectively of a Multi-Layer Perceptron (MLP) and a Convolutional Neural Network (CNN). The neural architecture was trained on a huge and balanced set of clinical data (14.000 CTG tracings, 7000 healthy and 7000 pathological) recorded during ambulatory non stress tests at the University Hospital Federico II, Napoli, Italy. After hyperparameters tuning and training, the neural network proposed has reached an overall accuracy of 80.1%, which is a promising result, as it has been obtained on a huge dataset.

胎儿电子监护（Cardiotocography, CTG）是一种广泛应用的临床监测手段，于妇产科诊所中通过分析胎心率（Fetal Heart Rate, FHR）与子宫收缩信号评估胎儿健康状态。由于调控胎心率的动态过程极为复杂，对该信号进行可靠的目视解读几乎难以实现，且会导致观察者间与观察者内存在显著的主观变异性。此外，仅通过计算机分析得到的少量参数，仍未能解决鲁棒性产前诊断的难题。因此，近十年来，基于人工智能（AI）机器学习技术的计算机辅助诊断系统被开发出来，以辅助临床决策。
本研究提出一种基于神经架构的混合方法，该架构可接收异构输入数据——包括一组定量参数与图像——用于区分健康与病理胎儿。定量回归指标可反映胎儿正常发育的不同维度，因而与胎儿健康状态密切相关；本方法将其与从胎心率信号多种表征（图像形式）中隐式提取的特征相结合，以提升分类性能。
该实现通过构建双分支连接的神经模型完成，分别为多层感知机（Multi-Layer Perceptron, MLP）与卷积神经网络（Convolutional Neural Network, CNN）。
该神经架构在意大利那不勒斯费德里科二世大学医院门诊无应激试验中记录的大规模平衡临床数据集上进行训练，该数据集包含14000份CTG描记图，其中健康与病理样本各7000份。
经过超参数调优与训练后，所提出的神经网络整体分类准确率达到80.1%；由于该结果基于大规模数据集获得，因而具有良好的应用前景。