Automated Fetal and Neonatal Movement Assessment for Very Early Health Assessment
Lead Research Organisation:
Imperial College London
Department Name: Computing
Abstract
The time during pregnancy and the first weeks and months after birth are critical for sensorimotor development, and a distinct change in fetal or neonatal movements can be an indicator of neurological or motor-system compromise. As is commonly known, fetuses and neonates move a lot spontaneously. The quantity and quality of these movements evolve significantly during early development, and clear deviations in this trajectory are thought to be an indicator of adverse neurodevelopment.
In current clinical practice, fetal or neonatal movements are not systematically quantified, leading to under-diagnosis of conditions for which reduced or abnormal movements are the key characteristic. Fetal movements have been shown to be abnormal in the case of severe brain abnormalities (e.g., anencephaly) but there are also indications that neurological conditions such as autism and cerebral palsy (CP) result in abnormal movement signatures in very early life.
This project addresses the clear need for an objective and automated means to quantitatively assess fetal and neonatal movement patterns, to facilitate earlier diagnosis and more effective treatments of life-changing conditions affecting babies and children.
We will develop methods to track fetal and neonatal movements, use machine learning to elucidate links between specific movement patterns or characteristics of diseases, use computational modelling to provide understanding of movement signatures that are specific to particular illneses, and validate the efficacy of assessing movement patterns as a diagnostic tool.
We will validate our movement assessment algorithms and apply them to diagnose brain development; primarily CP leading to follow-up research on autism and stroke/seizure. One in 200 children in the UK suffer from CP caused by pre- or perinatal brain damage, but a formal diagnosis (and therefore appropriate therapy at the most critical early stage of life) is not possible before 24 months. One reason for this is the lack of appropriate reproducible movement assessment techniques. Diagnosis of other common neurological conditions affecting neonates such as strokes/seizures is currently only possible with continuous electroencephalogram (EEG) monitoring and/or Magnetic Resonance Imaging (MRI) of the brain and spine, but such technologies are not always available in less well equipped neonatal wards, or in developing countries. Furthermore, accurate interpretation of EEG and MRI requires specialist expertise that is not always widely available. Our approach will democratise this expertise and make it widely available, which will improve quality of care significantly.
In current clinical practice, fetal or neonatal movements are not systematically quantified, leading to under-diagnosis of conditions for which reduced or abnormal movements are the key characteristic. Fetal movements have been shown to be abnormal in the case of severe brain abnormalities (e.g., anencephaly) but there are also indications that neurological conditions such as autism and cerebral palsy (CP) result in abnormal movement signatures in very early life.
This project addresses the clear need for an objective and automated means to quantitatively assess fetal and neonatal movement patterns, to facilitate earlier diagnosis and more effective treatments of life-changing conditions affecting babies and children.
We will develop methods to track fetal and neonatal movements, use machine learning to elucidate links between specific movement patterns or characteristics of diseases, use computational modelling to provide understanding of movement signatures that are specific to particular illneses, and validate the efficacy of assessing movement patterns as a diagnostic tool.
We will validate our movement assessment algorithms and apply them to diagnose brain development; primarily CP leading to follow-up research on autism and stroke/seizure. One in 200 children in the UK suffer from CP caused by pre- or perinatal brain damage, but a formal diagnosis (and therefore appropriate therapy at the most critical early stage of life) is not possible before 24 months. One reason for this is the lack of appropriate reproducible movement assessment techniques. Diagnosis of other common neurological conditions affecting neonates such as strokes/seizures is currently only possible with continuous electroencephalogram (EEG) monitoring and/or Magnetic Resonance Imaging (MRI) of the brain and spine, but such technologies are not always available in less well equipped neonatal wards, or in developing countries. Furthermore, accurate interpretation of EEG and MRI requires specialist expertise that is not always widely available. Our approach will democratise this expertise and make it widely available, which will improve quality of care significantly.
Planned Impact
Who might benefit from this research?
To ensure maximal impact we will engage with
- healthcare professionals who might be impacted by this research which includes radiologists, public health professionals, general physicians, paediatricians, gynaecologists, physiotherapists, frontline health workers.
- Computer Scientists working in Machine Learning, Computer Vision, 3D reconstruction, high-performance computing and decision support systems.
- pregnant women and mothers and fathers with their infants
- healthcare technology companies all around the world.
- healthcare professionals interested in healthcare policy, public health and introduction of new healthcare interventions.
- biomedical engineers and clinicians who need standard models for comparison during diagnostics and to assess healthy movements.
How might they benefit from this research?
- the ultimate goal of the technology is to support front line health workers in identification of high risk babies;
- the developed models will serve as a baseline representation for healthy movements;
- Pregnant women and infants may benefit from quicker referral, access to antenatal care closer to home, and comfort (if not referred) that the pregnancy is not of concern. The system may also lead to earlier identification of problems before they become critical;
- Healthcare policy makers and professionals working in public health will be interested in understanding how this variant of motion modelling technology might have impact if introduced into clinical practice and home care.
To ensure maximal impact we will engage with
- healthcare professionals who might be impacted by this research which includes radiologists, public health professionals, general physicians, paediatricians, gynaecologists, physiotherapists, frontline health workers.
- Computer Scientists working in Machine Learning, Computer Vision, 3D reconstruction, high-performance computing and decision support systems.
- pregnant women and mothers and fathers with their infants
- healthcare technology companies all around the world.
- healthcare professionals interested in healthcare policy, public health and introduction of new healthcare interventions.
- biomedical engineers and clinicians who need standard models for comparison during diagnostics and to assess healthy movements.
How might they benefit from this research?
- the ultimate goal of the technology is to support front line health workers in identification of high risk babies;
- the developed models will serve as a baseline representation for healthy movements;
- Pregnant women and infants may benefit from quicker referral, access to antenatal care closer to home, and comfort (if not referred) that the pregnancy is not of concern. The system may also lead to earlier identification of problems before they become critical;
- Healthcare policy makers and professionals working in public health will be interested in understanding how this variant of motion modelling technology might have impact if introduced into clinical practice and home care.
Publications
Tan J
(2022)
Detecting Outliers with Foreign Patch Interpolation
in Machine Learning for Biomedical Imaging
Tan J
(2020)
Detecting Outliers with Foreign Patch Interpolation
Schultes MT
(2022)
Barriers and Facilitators for Conducting Implementation Science in German-Speaking Countries: Findings from the Promote ImpSci Interview Study.
in Global implementation research and applications
Schmidtke L
(2021)
Unsupervised Human Pose Estimation through Transforming Shape Templates
Meng Q
(2019)
Weakly Supervised Estimation of Shadow Confidence Maps in Fetal Ultrasound Imaging
in IEEE Transactions on Medical Imaging
Matthew J
(2022)
Exploring a new paradigm for the fetal anomaly ultrasound scan: Artificial intelligence in real time.
in Prenatal diagnosis
Liu T
(2022)
Video Summarization Through Reinforcement Learning With a 3D Spatio-Temporal U-Net.
in IEEE transactions on image processing : a publication of the IEEE Signal Processing Society
Khatib N
(2023)
BORS/BJR TRAVELLING FELLOWSHIP ABSTRACT: MOTION CAPTURE OF NEONATAL INFANT KICKING MOVEMENTS CAN PROVIDE AN EARLY PREDICTION OF CEREBRAL PALSY
in Orthopaedic Proceedings
Description | We have introduced a groundbreaking approach for estimating the 2D and 3D positions of infants using standard video recordings from readily available devices, such as smartphones. This technique utilizes a learnable template matching problem, empowered by deep feature extraction, to transform a basic model into new data sets without the need for manually labeled examples. Typically, manual annotations pose a significant challenge in training machine learning models for pose estimation due to their time-consuming nature. Our innovative method significantly reduces the reliance on extensive human input, facilitating the rapid analysis of infant movements using nothing more than a simple camera placed beside the cot. In the broader scope of computer vision, human pose estimation plays a pivotal role in various applications, including augmented reality, video production, surveillance, and motion tracking. Specifically, in medical research, tracking movement can serve as a critical indicator of neurological disorders in infants. Despite the existence of numerous techniques, their practical application has been hindered by the demands for large, accurately annotated datasets and the challenge of adapting these methods to diverse human forms, such as those of children and infants. Our study introduces an innovative unsupervised technique to learn pose estimators applicable to both adults and infants. This method relies on deep learning to match templates to human body parts, represented by 2D Gaussian distributions, and incorporates a connectivity prior to ensure the model accurately reflects human anatomy. Our findings, validated across different datasets for both adults and infants, highlight the method's capability to efficiently learn pose estimation with minimal manual intervention, opening new avenues for high-throughput infant movement analysis without specialized equipment or setup. |
Exploitation Route | These methods can be used by any application that aim for pose tracking of anything that can be modelled with a skeleton. Applications in support for handicapped patients, wild-life monitoring, sports, etc. are imaginable. The code and models will be available on-line after potential protection of the invention. |
Sectors | Creative Economy Education Environment Healthcare Leisure Activities including Sports Recreation and Tourism Manufacturing including Industrial Biotechology |
Description | King's College London |
Organisation | King's College London |
Country | United Kingdom |
Sector | Academic/University |
PI Contribution | algorithm development and link to external tracking |
Collaborator Contribution | additional data collection |
Impact | none yet but we plan joint publications |
Start Year | 2019 |
Description | TU Munich |
Organisation | Imperial College London |
Country | United Kingdom |
Sector | Academic/University |
PI Contribution | We are working on geneal image registration algorithms |
Collaborator Contribution | A visiting PhD research from TUM spent some time with our team at Imperial |
Impact | Work in progress. |
Start Year | 2020 |
Description | UKER Erlangen |
Organisation | Friedrich-Alexander University Erlangen-Nuremberg |
Country | Germany |
Sector | Academic/University |
PI Contribution | new potential testing clinic in the children clinic at UKER Erlangen with Prof. Regina Trollmann |
Collaborator Contribution | none yet, we are currently working out how we can continue this project together |
Impact | none yet |
Start Year | 2022 |