Machine learning with generative mixture models for fetal monitoring

描述（由申请人提供）：多年来，一直在努力实现胎儿心率（FHR）节律分析的自动化。然而，尽管在生物医学信号分析方面取得了重大进展，但在自动化决策支持系统方面并没有任何重大改进。FHR监测现在在整个产房中无处不在，特别是使用无创多普勒监测器，但也使用胎儿头皮电极。胎儿心率模式的医生分类是已知的，因为显着的观察者间和观察者内的诊断变异性是一个不平凡的问题。这导致了 剖腹产的数量增加，从而在许多情况下增加了胎儿和母亲的风险。这进一步促使机器学习社区自动化分类过程，以提高准确性和一致性以及对噪声的鲁棒性。这方面的基本方法涉及某种类型的监督分类程序，其中将训练数据上的算法输出与“金标准”医生分类进行比较，然后在新数据集上进行测试和验证。然而，由于在存在上述诊断变异性以及显著的跟踪噪声的情况下，医生分类可能是不可靠的，因此我们提出使用无监督算法来 将胎心率数据记录聚类为临床有用的类别。我们使用非参数贝叶斯理论和马尔可夫时间依赖模型的特征序列的演变，提出的方法，将实现更高的精度。该方法涉及从FHR时间序列数据中提取特征序列，这些数据被建模为来自有限或无限Dirichlet混合模型的样本。然后，我们使用Gibbs抽样来获得每个数据集的聚类概率。聚类结果与直接的医生诊断进行比较，我们目前的结果被认为是与他们广泛一致，同时仍然提供新的信息，不同的子组的胎心率记录的字符。通过所提出的研究，将进一步提高分类性能。

DESCRIPTION (provided by applicant): For many years, there has been a concerted effort to automate the analysis of fetal heart rate (FHR) rhythms. However, despite significant advances in biomedical signal analysis, there has not been any significant improvement in automated decision support systems. FHR monitoring is now ubiquitous throughout delivery rooms, especially using the non-invasive Doppler monitor, but also using the fetal scalp electrode. Physician classification of fetal heart rate patterns is known to be a non-trivial problem because of significant inter and intra-observer variability of diagnosis. This has led to a marked increase in the number of caesarean deliveries, thereby increasing risk to the fetus and mother in many cases. This has further motivated the machine learning community to automate the classification procedure in the interest of accuracy and consistency as well as robustness with respect to noise. Usual approaches to this involve some type of supervised classification procedure, where the algorithm output on training data is compared with a "gold-standard" physician classification, followed by testing and validation on new datasets. However, since physician classification can be unreliable in the presence of the aforementioned diagnostic variability, as well as significant tracing noise, we propose the use of unsupervised algorithms to cluster FHR data records into clinically useful categories. We use nonparametric Bayes theory and Markov-time-dependence models for the evolution of feature sequences to propose methods that will achieve improved accuracy. The methods involve extraction of feature sequences from FHR time series data, which are modeled as samples from finite or infinite Dirichlet mixture models. We then use Gibbs sampling to obtain the cluster probabilities for each dataset. Clustering outcomes are compared against direct physician diagnosis and our current results are seen to be in broad agreement with them, while still giving new information on the character of different sub-groups of FHR records. With the proposed research, further gains in classification performance will be made.