Machine learning techniques for detection of fetal hypoxia and prevention of birth injury

分娩期间子宫内缺氧（也称为胎儿缺氧）对新生儿健康造成不利影响，如代谢性酸中毒、神经发育问题或死亡。因此，在分娩过程中监测胎儿是至关重要的，以防止胎儿缺氧对婴儿和家庭的破坏性影响。心脏造影（CTG）用于监测胎儿心率（FHR），它可以在分娩期间检测胎儿子宫内缺氧，在分娩期间使用高达90%。因此，CTG可以帮助识别受损胎儿（通过心率变异性、加速、减速和基线），并将其分为正常、可疑和病理。这使临床医生能够干预分娩过程，如剖腹产，以减少对新生儿的不利影响，同时确保母亲的安全。然而，视觉CTG受到观察者之间解释不一致的影响，这可能会延迟适当的干预，对婴儿造成伤害。此外，一些决策可能是直观的，具有一定程度的不确定性，这可能导致CTG解释的差异。此外，视觉CTG易被误诊为缺氧胎儿。假阴性病例对婴儿和家庭造成破坏性影响，而假阳性病例则导致不必要的干预。一项系统综述显示CTG监测下剖宫产率显著增加。这增加了住院时间，增加了医疗费用，并给母亲带来了感染和婴儿受伤等风险。为了解决视觉CTG的缺点，引入了计算机CTG，通过标准化解释和减少“灰色地带”特征来帮助异常FHR的决策。这样可以更快地对胎儿做出反应。一项随机对照试验表明，计算机化CTG提高了口译的质量，同时最大限度地减少了决策时间。然而，对六项研究的荟萃分析显示，视觉CTG和计算机CTG对胎儿健康没有显著改善。最近的一项随机对照试验，婴儿试验，调查了计算机化CTG作为支持决策工具减少新生儿不良结局的能力，结果显示视觉CTG和计算机化CTG之间没有显著差异，计算机化CTG无法准确检测异常FHR。从这里开始，没有证据表明计算机化CTG可以改善结果。因此，研究人员已经探索了实现机器学习（ML）来对FHR进行分类。ML是人工智能中一个强大的工具，由于其出色的决策结果，在医学研究中受到越来越多的关注。机器学习可以从以前的数据中学习和识别模式，以获得信息和经验，对新数据进行预测。虽然机器学习技术显示出有希望的结果，但迄今为止的研究有限。例如，研究使用了来自特定临床试验条件的有限数量的“病例”和“对照”，依赖于任意选择的心率特征，和/或使用替代结果，如专家对正常/异常CTG的分类。在将方法转化为临床实践之前，需要使用大规模真实世界数据和具有临床意义的新生儿结局来严格开发和验证预测模型。因此，本项目旨在开发和验证机器学习算法，利用CTG分析和从电子健康记录中获得的临床危险因素来预测胎儿缺氧。这个与医疗软件公司Clevermed的合作项目为分析EFM和详细的临床数据提供了新的机会。Clevermed生产用于新生儿和产妇记录的BadgerNet平台，包括EFM痕迹的存储。该系统在英国、澳大利亚和新西兰的26个产科单位使用，每年捕获超过15万名孕妇和新生儿。

Lack of oxygen supply in the womb (also known as fetal hypoxia) during labour causes adverse effects on newborns' health, such as metabolic acidosis, neurodevelopmental issues or death. Therefore, fetal monitoring during labour is crucial to prevent the devastating impact of fetal hypoxia on babies and families.Cardiotocography (CTG) is used to monitor fetal heart rate (FHR), which can detect fetal hypoxia in the womb during labour is used up to 90% during labour. Therefore, CTG can help identify compromised fetal (through heart rate variability, acceleration, deceleration and baseline) and categorise this into normal, suspicious and pathological. This allows clinicians to intervene in the labour process such as caesarean sections to reduce the adverse effect on newborns while ensuring the mother's safety. However, visual CTG suffers from interpretation inconsistencies between observers which can delay appropriate intervention, causing harm to babies. Furthermore, some decision making can be intuitive and with some level of uncertainty which may contribute towards discrepancy in CTG interpretation. In addition, visual CTG is vulnerable to misidentifications of hypoxic fetus. False-negative cases cause damaging effects on babies and families, while false-positive cases lead to unnecessary interventions. A systematic review demonstrated a significant increase in caesarean section with CTG monitoring. This increases healthcare costs by increasing hospital admissions length and imposes a risk for the mother, such as infection and injuries towards babies.To tackle the shortcomings of visual CTG, computerised CTG was introduced to aid in decision making for abnormal FHR by standardising interpretations and reducing 'grey-zone' features. This will allow a quicker response to compromised fetal. A randomised control trial had shown that computerised CTG improved the quality of interpretations while minimising decision-making time. However, a meta-analysis of six studies showed no significant improvement in fetal well-being between visual and computerised CTG. A recent randomised control trial, the INFANT trial, investigated the ability of computerised CTG as a support decision tool to reduce poor new-borns outcomes demonstrated no significant difference between visual and computerised CTG and failure of computerised CTG to detect abnormal FHR accurately. From here, there is no evidence that computerised CTG improves outcomes. Hence, researchers had explored implementing machine learning (ML) for categorising FHR. ML is a robust tool in artificial intelligence that has received increasing attention within medical research due to its excellent decision-making outcomes. ML can learn and identify patterns from previous data to gain information and experience to make predictions on new data. While ML techniques showed promising results, but studies have so far been limited. For example, studies have used restricted numbers of 'cases' and 'controls' from specific clinical trial conditions, have been dependent on arbitrarily selected heart rate features, and/or used surrogate outcomes such as expert classification of normal/abnormal CTG. Rigorous development and validation of prediction models using large-scale real-world data and clinically significant neonatal outcomes are required before methods could be translated to clinical practice. Therefore, this project aims to develop and validate machine learning algorithms to predict fetal hypoxia using CTG analysis and clinical risk factors obtained from electronic health records.This collaborative project with the medical software company Clevermed offers new opportunities to analyse EFM alongside detailed clinical data. Clevermed produces the BadgerNet Platform used for neonatal and maternity records, including storage of EFM traces. The system is used in 26 Maternity units in the UK and Australia and New Zealand, capturing over 150,000 pregnancies and newborns per annum.