James Cook University scientists believe they have made a breakthrough in the science of keeping babies alive by applying analytics and AI to newborns' vital statistics.
As part of her PhD work, JCU engineering lecturer Stephanie Baker led a pilot study that used a hybrid neural network to accurately predict how much risk individual premature babies face.
The study hopes to reduce alarming figures that show that complications resulting from premature birth are the leading cause of death in children under five. Meanwhile, over 50 per cent of neonatal deaths occur in pre-term infants.
In Australia, around eight per cent of babies are born prematurely (before 37 weeks gestation), each year. Most babies born prematurely are born between 32 and 36 weeks gestation, and almost all of these babies grow up to be healthy children.
However, some babies die as a result of being born too early because their organs are too immature to function properly outside the womb.
The issue has been that it is difficult to predict which pregnancies will end prematurely. Until now.
Baker, based in Queensland, says: “Pre-term birth rates are increasing almost everywhere. In neonatal intensive care units, assessment of mortality risks assists in making difficult decisions regarding which treatments should be used, and if and when the treatments are working effectively.”
But there are several limitations to this system. Generating the score requires complex manual measurements, extensive laboratory results, and the listing of maternal characteristics and existing conditions,” Baker says.
The alternative was measuring variables that don’t change – such as birth weight – that prevents re-calculation of the infant’s risk on an ongoing basis and doesn’t show their response to treatment.
JCU team’s research, published in the journey Computers in Biology and Medicine, had developed the Neonatal Artificial Intelligence Mortality Score (NAIMS), a hybrid neural network that relies on simple demographics and trends in heart and respiratory rate to determine mortality risk. The technique is fast, with no need for invasive procedures or knowledge of medical histories.
Baker explains that the research uses an open-access medical database called MIMIC-III, which contains comprehensive records of medical events and measurements from patent admission to discharge.
The team wanted to make a tool that could be used with little time cost for healthcare workers, so used inputs describing variation in heart and respiratory rates, birthweight and gestational age – all of which are routinely measured in neonatal ICUs.
The second stage is the neural network itself. For this work, the team combined two different types of neural networks.
“For this work, we combined two different types of neural networks. The first type is called a convolutional neural network, which are excellent at identifying the most important outputs. The second type is called a long-short memory network, which are good at identifying relationships between inputs. Creating a hybrid network allowed us to get the benefits of both network types,” she says.
The final stage of the output, which is a value between 0 – 1, indicates the mortality risk. “We were interested in how vital signs could be used to estimate mortality risk, and there were around 180 babies in the database that had information about vital signs available – so these babies formed our patient cohort,” Baker says.
“Using data generated over a 12-hour period, NAIMS showed strong performance in predicting an infant’s risk of mortality, within 3, 7 or 14 days. This is the first work I’m aware of that uses only easy-to-record demographics and respiratory rate and heart rate data to produce an accurate prediction of immediate mortality risk,” Baker says.
NAIMS has proved accurate when tested against hospital mortality records of pre-term babies and had the added advantage over existing schemes of being able to perform a risk assessment based on any 12 hours of data during the patient’s stay.
The next step, according to Baker, is to partner with local hospitals to gather more data and undertake further testing. “We also aim to conduct research into the prediction of other outcomes in neo-natal intensive care, such as the onset of sepsis and patient length of stay,” she says.