NEONATAL SOCIETY ABSTRACTS
Postnatal Gestational Age Assessment Using Computer Vision and Deep Machine Learning – The Gestation Study
Presented at the Neonatal Society 2017 Summer Meeting (programme).
Henry C1, Ward C1, Torres-Torres M2, Valstar M2, Sharkey D1
1 Early Life Research, Academic Child Health, School of Medicine, University of Nottingham, Nottingham
2 School of Computer Science, University of Nottingham, Nottingham
Background: Global estimates of preterm birth are lacking in 125 of 184 countries covering ~71% of the 15 million preterm births annually (1,2). Many of these occur in low-middle income countries (LMICs) making it challenging to deliver targeted interventions to improve individual and population-based outcomes. Antenatal scans in LMICs are not readily available and postnatal clinical assessment of gestational age (GA) using the Ballard Score is unreliable (2). We hypothesised machine learning of foot, face or ear images could be used to estimate postnatal GA.
Methods: Still images of the face, foot and ear were collected from newborn infants with a known GA based on maternal first trimester ultrasound. Participants were stratified according to gestational age. We used deep machine learning techniques, combining convolutional neural networks (CNNs) and fully convolutional networks (FCNs), to train an algorithm to estimate GA. Blinded Ballard assessments were also performed. Receiver operator characteristic (ROC) curves along with positive and negative predictive values (PPV and NPV) were calculated. Ethical approval was given and the study funded by the Bill and Melinda Gates Foundation.
Results: Participants (n=93) images were split into a training data set (n=50) and algorithm testing set (n=43). 61% of babies were preterm with an equal mix of genders. When using foot and face images, along with birth weight, in the model the algorithm had a PPV of 0.89 (95% CI 0.72-0.96) and NPV of 0.91 (95% CI 0.73-0.98) for predicting prematurity with the AUC ROC 0.96 (95% CI 0.91-1.0, P<0.001). The root mean square error (RMSE) was calculated to estimate the accuracy of GA assessment which improved with doubling of the size of the dataset. This gave a RMSE of 1.15 weeks (8 days) with a standard deviation of 0.89 weeks (6 days) when using the ear, face and weight in the model. Blinded clinical Ballard score assessment of the same participants gave a RMSE of 25 days (SD 16 days).
Conclusion: Using small datasets and deep learning algorithms it is possible to capture images of the face, foot and ear of a newborn to accurately determine GA. This quick and simple method could be incorporated into a simple App to link with a central server in LMICs to provide 1) point of care advice for the individual baby 2) detailed epidemiological maps on births and prematurity allowing targeted population-based interventions. We are now developing a clinical trial in India to explore the feasibility of this approach further with an accuracy of ±1 week.
Corresponding author: firstname.lastname@example.org
1. Blencowe et al, Lancet 2012. 379(9832):p2162-72
2. Lee et al, Pediatrics 2016. 138(1): pii: e20153303