📣 Help Shape the Future of UKRI's Gateway to Research (GtR)

We're improving UKRI's Gateway to Research and are seeking your input! If you would be interested in being interviewed about the improvements we're making and to have your say about how we can make GtR more user-friendly, impactful, and effective for the Research and Innovation community, please email gateway@ukri.org.

Deep Learning Models for Fetal Monitoring and Decision Support in Labour

Lead Research Organisation: University of Portsmouth
Department Name: Sch of Computing

Abstract

Oxygen reaches the baby in the mother womb via the placenta and umbilical cord. During labour, the contractions squeeze the placenta and the cord, reducing the supply of oxygen to the baby. Most healthy babies cope well, but a small percentage are at risk of suffocation and brain injury. This causes each year in the UK about 100 healthy babies to die and more than 1,000 to sustain brain injury. Globally, each year, events during childbirth are estimated to account for 920,000 neonatal deaths; 1.1 million stillbirths during labour; and more than a million babies per year develop different important sequelae, from cerebral palsy to mental retardation, learning difficulties and other disabilities.
A baby which is at higher risk can be delivered urgently by emergency Caesarean section or instrumental vaginal extraction. To monitor the baby during labour, midwives and doctors in the developed countries use the cardiotocogram (CTG), which continuously displays the womb's contractions and the baby's heart rate. But our understanding of how to read the complex CTG graphs is limited and the patterns are difficult to interpret by eye, so some babies end up injured while at the same time many unnecessary emergency interventions are performed. Nearly 50% of the NHS litigation bill relates to maternity claims (in 2000-2010 these amounted to £3.1bn) and the majority of these are related to shortcomings in labour management and CTG interpretation.
With currently available computing power and routinely collected clinical data, methods using intelligent computer-based analysis can establish the relation of the CTG patterns and other clinical factors during labour to the baby's health at birth. In our pioneering work, we have already derived from the data a first prototype of a basic automated CTG analysis model, demonstrating proof-of-concept of using routinely collected maternal-fetal clinical data from pregnancy and childbirth (using retrospectively the data of more than 22,000 births at term). More recently, using the data from over 35,000 births at term, we conducted the first ever work on the application of deep learning methods for the analysis of CTG data, demonstrating their capability to learn from the raw CTG data and supersede all prior automated algorithms.
The main goal of this proposal is, based on an updated and unique Oxford dataset of over 100,000 births, to develop innovative deep learning models for personalised continuous fetal health risk assessment during labour. In particular, we propose to: (1) develop automated models for continuous fetal heart rate analysis based on the CTG data and clinical risk factors, incorporating convolutional neural networks and long short term memory networks into multimodal and stacked models; (2) develop a software App for tablets, capable of real-time CTG analysis and risk assessment of fetal compromise; (3) validate and demonstrate the capability, accuracy, and efficiency of the new models running on the App, by conducting simulations with real-time data at the Oxford NHS Trust as well as with retrospective data.
The output of this proposal will constitute a crucial building block towards the team's overarching goal to deliver at the bedside an individualised data-based tool for clinical decision support, preventing brain injury of the baby during labour. The main output will be a decision-support tool ready for prospective clinical tests, which holds clear potential benefits for the individuals, society, clinicians, and the NHS. It will address unresolved challenges in this clinical field, where improvements are painfully needed and are long overdue. This project is timely and represents excellent value for money, given the existing database and substantial prior work, the enormity of NHS litigation claims and the costs of unnecessary operative deliveries.

Planned Impact

Monitoring reliably and consistently the health of babies during childbirth remains a massive unmet clinical need worldwide. In the UK alone, over 1200 otherwise healthy term babies sustain permanent brain injury during labour. Recent reports have found that in the UK, problems with monitoring the fetal health during labour were a factor in over half of the infants with potentially preventable brain injury at term (Royal College of Obstetricians and Gynaecologists. Each Baby Counts: 2018 Progress Report. London: RCOG, 2018).
Nearly 50% of the NHS litigation bill relates to maternity claims, and between the years 2000 and 2010 these amounted to £3.1bn (The NHS Litigation Authority. Ten years of maternity claims: an analysis of NHS Litigation Authority data, 2012). The majority of these claims related to shortcomings in labour management and cardiotocgraphy (CTG) interpretation. In health economic terms, "more than 3,000 quality-adjusted life-years are lost annually to cerebral palsy from obstetric complications resulting in fetal hypoxia, at estimated costs of £62.9m" (Annual report of the Chief Medical Officer: the health of the 51%: Women, Dept. of Health, London, 2015). Reducing the rate of stillbirths, neonatal deaths and brain injuries that are caused during or soon after birth continues to be a priority for the NHS and it is one of the mandate objectives from central government (The Government's mandate to NHS England 2016-2017, Dept. of Health, 2017).
At the same time, the UK government has pledged to "ensure the benefits of NHS data and innovation are fully harnessed for patients in this country", and with this goal, has recently pledged £250 million for the NHS to invest in Artificial Intelligence (https://www.gov.uk/government/news/health-secretary-announces-250-million-investment-in-artificial-intelligence).
Our proposal fits right into the heart of the above national agendas, and is based on decades of experience of the team with machine learning and working with the CTG archives. Moreover, this proposal fits perfectly within the EPSRC Healthcare Technologies strategy, addressing one of the Grand Challenges: Optimising Treatment through effective diagnosis, patient-specific prediction and evidence-based intervention. In addition, this proposal also fits within the Cross-cutting research capabilities: it is focussed on innovative intelligent technologies that can have a transformative impact on diagnosis and monitoring in healthcare.
This collaborative EPSRC proposal is a crucial building block within our larger programme at the Oxford Centre for Fetal Monitoring Technologies (please see Case for Support - Fig.1 and Pathways to Impact part of this proposal). Our research project is designed to deliver novel advanced deep learning methods for CTG analysis and novel technical capability to interact with the clinicians through an App, supporting the wider project in preparation for a large i4i NIHR Product Development Award. Our overarching goal is to deliver to the bedside, for the first time, a disruptive data-driven individualised decision support tool for fetal health risk assessment during childbirth.
Finally, our proposal is highly innovative without an equivalent world-wide to date, both owing to the unique size and content of the Oxford dataset and to the applicants' expertise and prior work. The timing is perfect with current developments in deep learning methods, existing data archive, strong preliminary results, team's expertise, established collaborations and partnerships, and documented success with attracting funding from the NIHR. Our research programme holds clear potential benefits for the individuals, society, clinicians, and the NHS. Given the existing database, substantial prior work, enormity of NHS litigation claims, costs of unnecessary operative deliveries, and the quality-adjusted life-years lost to intrapartum hypoxic brain injury, this project represents excellent value for money.
 
Description Cardiotocography (CTG) is a widely used technique to monitor fetal heart rate (FHR) during labour and assess the health of the baby. A new deep learning (DL) models and framework to enhance the early detection of the babies at risk, leveraging both raw FHR signals and standard CTG features. Unlike traditional methods focusing on abnormal CTG traces (but not birth outcomes), this approach, backed by a substantial cohort of data records, demonstrates the potential of DL in predicting actual adverse outcomes. The DL framework uniquely combines a convolutional mechanism with a self-attention network, enhanced by a gating mechanism for more accurate feature extraction. Trained on a dataset of over 37,000 births, including 1,291 with abnormal outcomes, the model was evaluated and tested on a holdout set of 6,459 births and the open-access CTU-CHB CTG dataset of 552 births. The proposed DL model, demonstrates superior diagnostic accuracy, outperforming and surpassing state-of-the-art baseline performance and clinical benchmarks.
Exploitation Route The promising results of this investigation underscore its potential utility in assisting clinicians with labour clinical management, offering a powerful tool for improving neonatal care. However, it is crucial to acknowledge that one of the primary challenges encountered was accurately calculating the genuine false positive rate. Interventions informed by CTG readings could have potentially altered the natural outcomes of births, thus complicating the task of establishing a reliable clinical benchmark. Additionally, the framework raises explainability questions: it remains unclear why integrating raw FHR signals with standard CTG features results in superior performance compared to using standard CTG features alone. To enhance the model's predictive capacity and clinical relevance, future research directions should incorporate additional clinical risk factors, such as maternal age, gestational age, maternal fever, and the presence of meconium. Integrating these factors would provide a richer, more holistic context for the model's predictions, potentially uncovering nuanced relationships between various risk factors and adverse birth outcomes. Moreover, expanding the dataset to encompass a wider range of birth outcome categories, including different levels of fetal acidemia and degrees of neonatal compromise, could refine the framework ability to distinguish between various adverse outcomes.
Such advancements should pave the way for a more precise and clinically useful predictive tool. By enabling deep learning models to identify subtle indicators across a broad spectrum of birth scenarios, these efforts can significantly enhance the quality of neonatal care, offering a vital resource for healthcare professionals in managing labour and delivery processes more effectively.
Sectors Digital/Communication/Information Technologies (including Software)

Healthcare
 
Description Decision-support for individualised risk assessment of fetal health during labour: preventing fetal brain damage and death by utilising large, routinely collected datasets of cardiotocography and clinical risk factors
Amount £1,118,105 (GBP)
Funding ID NIHR202117 
Organisation National Institute for Health and Care Research 
Sector Public
Country United Kingdom
Start 04/2021 
End 04/2024
 
Title Deep Learning Models for Fetal Monitoring and Decision Support in Labour 
Description Cardiotocography (CTG) is an essential and widely adopted technique in obstetrics for monitoring fetal well-being during labor by simultaneously recording fetal heart rate (FHR) and uterine contractions. The goal of CTG is to identify fetal distress early and to make decisions on interventions that could mitigate risks to the fetus during childbirth. Despite its widespread use, the efficacy and utility of CTG in reducing neonatal and maternal morbidity have been subjects of ongoing debate. One of the core challenges with CTG lies in the interpretation of its outputs. The complex patterns of fetal heart rates and uterine contractions can be difficult to analyze accurately, leading to variability in interpretations among clinicians. This variability can result in both false positives-leading to unnecessary surgical interventions such as caesarean sections or operative deliveries-and false negatives, which may miss critical instances of fetal distress, potentially leading to adverse neonatal outcomes. Given these challenges, there is an emerging interest in the application of machine learning, particularly deep learning techniques, to enhance the interpretation of CTG data. Deep learning, a subset of artificial intelligence, is renowned for its capability to learn and make inferences from large datasets, identifying patterns and relationships that may not be immediately apparent to human observers. By training models on extensive datasets of CTG recordings, these techniques aim to reduce the subjectivity in CTG analysis, offering a more standardized and potentially accurate assessment of fetal well-being. This study introduces a deep learning (DL) framework to enhance the early detection of the babies at risk, leveraging both raw FHR signals and standard CTG features. Unlike traditional methods focusing on abnormal CTG traces (but not birth outcomes), this approach, backed by a substantial cohort, demonstrates the potential of DL in predicting actual adverse outcomes. The DL model uniquely combines a convolutional mechanism with a self-attention network, enhanced by a gating mechanism for more accurate feature extraction. Trained on a dataset of over 37,000 births, including 1,291 with abnormal outcomes, the model was evaluated tested on a holdout set of 6,459 births and the open-access CTU-CHB CTG dataset of 552 births. The proposed DL model, demonstrates superior diagnostic accuracy, outperforming method surpasses state-of-the-art baseline performance and clinical benchmarks. It achieved a sensitivity of 49.08% (95% CI, 46.01-53.36%) at a 15% false positive rate (FPR), compared to the clinical benchmark sensitivity of 37.70% (33.10-42.30%) and a previous model's 32.60% (28.20-37.30%) at a similar FPR. 
Type Of Material Computer model/algorithm 
Year Produced 2024 
Provided To Others? Yes  
Impact This study introduces a deep learning model designed to predict adverse birth outcomes by analysing FHR traces and widely recognized CTG features, up to two hours before delivery. The developed model marks a significant departure from previous research, which primarily concentrated on identifying abnormal CTG traces. Instead, our work leverages a large cohort of records to explore the full potential of deep learning in forecasting actual adverse birth outcomes. When evaluated against a hold-out testing set and an additional validation dataset, the model exhibited exceptional performance, significantly surpassing clinical benchmarks in the early detection of adverse outcomes. The promising results of this investigation underscore its potential utility in assisting clinicians in labour management, offering a powerful tool for improving neonatal care. However, it is crucial to acknowledge the study's limitations in order to fully appreciate its implications and areas for further research. One of the primary challenges encountered was accurately calculating the genuine false positive rate. Interventions informed by CTG readings could have potentially altered the natural outcomes of births, thus complicating the task of establishing a reliable clinical benchmark. Additionally, the model's lack of explainability raises questions; it remains to be further analysed why integrating raw FHR signals with standard CTG features results in superior performance compared to using standard CTG features alone. Another limitation pertains to the dataset's origin (from a single hospital), which may restrict the generalizability of the model's predictions. To ensure the model's broad applicability and robustness, future validation efforts should include data from a diverse array of populations and healthcare settings. To enhance the model predictive capacity and clinical relevance, future research directions should incorporate additional clinical risk factors, such as maternal age, gestational age, maternal fever, the presence of meconium, and others. Integrating these factors would provide a richer, more holistic context for the model inference capacity, potentially uncovering nuanced relationships between various risk factors and adverse birth outcomes. Moreover, expanding the dataset to encompass a wider range of birth outcome categories, including different levels of fetal acidemia and degrees of neonatal compromise, could refine the model ability to distinguish and predict various adverse outcomes. Such advancements would pave the way for a more precise and clinically useful predictive tool. By enabling deep learning models to identify subtle indicators across a broad spectrum of birth scenarios, these efforts could significantly enhance the quality of neonatal care, offering a vital resource for healthcare professionals in managing labour and delivery processes more effectively. 
 
Title Multimodal Deep Learning Models for predicting adverse birth outcomes 
Description Cardiotocography (CTG) is widely used to monitor fetal heart rate (FHR) during labour to assess the wellbeing of the baby. Interpretation of CTG patterns in clinical practice involves assessing visually features of the signal such as the baseline heart rate, variability, accelerations and deceleration patterns by a specialist clinician which is subjective and prone to errors. Computer-based methods have been developed to detect abnormal CTG patterns automatically by mimicking clinical guidelines, but these are still subject to limitations. Recently, data-driven approaches using deep learning (DL) methods have shown promising performance in the classification of CTG for detecting fetal acidemia around the time of birth. However, to allow adequate time for clinical decision and intervention, abnormal CTGs should be detected as early in labour as possible. We designed and developed new DL models, based on Convolutional Neural Networks (CNN) and Long Short-Time Memory (LSTM) neural network architectures. The ability of CNN to capture and extract important spectral features from the datasets is enhanced by adding LSTM, which are very good in acquiring temporal characteristics from time series data. The proposed DL models utilise CTGs from more than 51,000 births at term to classify those with and without severe compromise at birth, from the first hour and longer time traces of the FHR records, in predicting severe compromises and acidemia classes. We developed, simulated, and experimented three different CNN and LSTM based neural network architectures: 1D-CNN; 1D-CNN-LSTM sequential; and 1D-CNN-LSTM parallel. We also proposed a multi-modal architecture, where the 1D-CNN-LSTM and a 2D-CNN (that analyses time-frequency representation of the FHR) are connected in parallel. Results from the models' comparison are estimated using partial area under the curve (PAUC) between 0-10% false-positive rate and sensitivity at 95% specificity. These metrics defined superior performance of the 1D-CNN-LSTM parallel architecture and efficient usability of all three models for early prediction of the risk of severe compromise and acidemia outcomes. Next steps of the models development include adding the contraction signals in conjunction with some of the FHR macro features (such us: baseline mean, number of accelerations and decelerations; phase-rectified signal averaging (PRSA); short term and long term variability (STV/LTV); signal stability index; prolonged deceleration; etc.) as inputs to the models, which is expected to further improve their inference capability. 
Type Of Material Computer model/algorithm 
Year Produced 2023 
Provided To Others? Yes  
Impact Results from applying these models on early labour CTG data are clinically very important and encouraging as many of the infants born with severe compromise develop complications only later in the labour. Using the ability of these models for early prediction of the risk of compromise will help clinicians to take informed decision of whether intervene or avoid unnecessary interventions. 
 
Description 5th Signal Processing and Monitoring (SPaM) in Labour International Workshop, Cagliari, Italy 
Form Of Engagement Activity A talk or presentation
Part Of Official Scheme? No
Geographic Reach International
Primary Audience Professional Practitioners
Results and Impact Prof. Jordanov and Prof. Georgieva gave a talk and presentation to the "5th SPaM in Labour International Workshop".
Presented some of the results from the project to a wide international audience of professionals, scientist, and business/industry participants. They took a part in a panel debate on the topics of interest and paths for further collaboration and preparation of research bids for funding were also discussed.
Year(s) Of Engagement Activity 2024
URL https://sites.unica.it/spam2024/
 
Description British Council UK-Vietnam research conference "Impacts of Climate Change and the Environment on Health", Hanoi, Vietnam 
Form Of Engagement Activity A talk or presentation
Part Of Official Scheme? No
Geographic Reach International
Primary Audience Professional Practitioners
Results and Impact Prof, Jordanov gave a talk/presentation " AI in Health: Deep Learning Models for Fetal Monitoring and Decision Support in Labour", British Council UK-Vietnam research conference "Impacts of Climate Change and the Environment on Health", Vietnam National University, Hanoi, (November 2023).
Year(s) Of Engagement Activity 2023
URL https://www.britishcouncil.vn/en/events/research-conference-impacts-climate-change-and-environment-h...
 
Description DECision-support for Intrapartum Data-driven Evaluation (DECIDE) 
Form Of Engagement Activity Participation in an activity, workshop or similar
Part Of Official Scheme? No
Geographic Reach International
Primary Audience Professional Practitioners
Results and Impact The Workshop was organised by Oxford Labour Monitoring Group and project team members at Worchester College, Oxford University, on 5th of October 2022. The presented and discussed topics included: Clinical usability and feasibility; Clinical and public users' requirements & needs: final report and actionable plan; Progress on early health economics model; Updates from ongoing national maternity initiatives and how we fit; Parent, Public & Patient Involvement; NHS adoption strategy and barriers; Kaunas study to collect prospective CTG data (International participation from Lithuania); Trustworthy and Ethical AI; also External data from Cambridge and London.
Year(s) Of Engagement Activity 2022
 
Description Engagement on twitter 
Form Of Engagement Activity Engagement focused website, blog or social media channel
Part Of Official Scheme? No
Geographic Reach International
Primary Audience Professional Practitioners
Results and Impact A twitter account with regular updates about our research or relevant events - @oxfordfetal
Year(s) Of Engagement Activity 2020,2021
URL https://twitter.com/oxfordfetal
 
Description First Summer School on Neuroinformatics, Neural Networks, and Neurocomputers 
Form Of Engagement Activity A talk or presentation
Part Of Official Scheme? No
Geographic Reach International
Primary Audience Postgraduate students
Results and Impact Gave a talk and presentation to the First Summer School on Neuroinformatics, Neural
Networks, and Neurocomputers, SS-N3BG 2023 "Deep Neural Networks and their applications, Case Study: Deep Learning Models for Fetal Monitoring in Labour (EPSRC/EP-V002511-1 project)".
Presented some results from the project to a wider international audience of postgraduate students, researches, and university academics.
Year(s) Of Engagement Activity 2023
URL https://www.knowledgeengineering.ai/summer-school
 
Description PPI Panel meeting 
Form Of Engagement Activity A talk or presentation
Part Of Official Scheme? No
Geographic Reach Local
Primary Audience Public/other audiences
Results and Impact A project PPI Panel meeting was organised and led by the team's PPI co-leads. This involved several local members of the public with interest in maternity but no involvement professionally in the field.
Year(s) Of Engagement Activity 2022
 
Description Participant/presenter at the Pint of Science festival May 2023 
Form Of Engagement Activity A talk or presentation
Part Of Official Scheme? No
Geographic Reach Local
Primary Audience Public/other audiences
Results and Impact Prof Georgieva was one of the presenters at Oxford Pint of Science - delivering a talk and Q&A in informal setting targeting the general audience
https://pintofscience.co.uk/
Year(s) Of Engagement Activity 2023
URL https://pintofscience.co.uk/
 
Description Second Summer School on "Neuroinformatics, Neural Networks, and Neurocomputers", Sofia, Bulgaria 
Form Of Engagement Activity A talk or presentation
Part Of Official Scheme? No
Geographic Reach International
Primary Audience Postgraduate students
Results and Impact Gave a presentation and a talk to the Second Summer School "Neuroinformatics, Neural
Networks, and Neurocomputers", SS-N3BG 2024 "Deep Neural Networks for Predicting the Risk of Adverse Birth Outcome".
Presented some of the current results from the project to a wider international audience of postgraduate students, researches, and university academics.
Year(s) Of Engagement Activity 2024
URL https://www.knowledgeengineering.ai/summerschool2024
 
Description Webinar talk/presentation "Deep Learning Models for Fetal Monitoring and Decision Support in Labour" to N3-BG seminar 
Form Of Engagement Activity A talk or presentation
Part Of Official Scheme? No
Geographic Reach International
Primary Audience Other audiences
Results and Impact A online talk/presentation "Deep Learning Models for Fetal Monitoring and Decision Support in Labour" given to N3-BG seminar, which includes professionals/academics in AI field worldwide.
Year(s) Of Engagement Activity 2023
URL https://www.knowledgeengineering.ai/seminars
 
Description Webinar, presentation given by Dr Daniel Asfaw 
Form Of Engagement Activity A talk or presentation
Part Of Official Scheme? No
Geographic Reach National
Primary Audience Professional Practitioners
Results and Impact "Here at the PRU-MNHC we really look forward to our annual stakeholder day for our public involvement partners. The team enjoys the chats over lunch as much as the focussed research discussions, so it was something of a wrench to have to shift to an online meeting this year as well as an interesting online learning curve."
Year(s) Of Engagement Activity 2024
URL https://www.npeu.ox.ac.uk/pru-mnhc/news-and-blog/2021-september-public-involvement-meetings
 
Description Workshop "Rising Public Engagement and Awareness of the Project" 
Form Of Engagement Activity Participation in an activity, workshop or similar
Part Of Official Scheme? No
Geographic Reach National
Primary Audience Professional Practitioners
Results and Impact The workshop was organised by the EPSRC project team and was held in the Future Technology Centre (FTC), University of Portsmouth on 13th June 2023, hosted and chaired by the PI Dr. Ivan Jordanov.
The presented and discussed topics included:
• Current progress and next stages of project development;
• Developed Deep Learning Models for Fetal Monitoring and Risk Assessment;
• Data-driven fetal monitoring at Oxford (Risk assessment at th e onset of and during labour);
• Predicting Adverse Birth Outcomes During Labour Using Deep Learning;
• Parent, Patient and Public Involvement and Engagement (PPPIE);
• One size fits all? The role of social determinants of health on AI-driven CTG intrapartum decision-support tools in term babies;
• Fetal Monitoring Midwives' Perspective.

The workshop concluded with an open discussion about project achievements and the next stages of development.
Year(s) Of Engagement Activity 2023
URL http://www.port.ac.uk/FTC