Machine-learning algorithms can help doctors diagnose children affected by fetal alcohol spectrum disorder, according to fresh research.
Fetal alcohol spectrum disorder (FASD) describes a range of conditions that are caused when a woman consumes excessive amounts of booze during her pregnancy. Such benders can impact the development of a fetus, causing the child to be born with impaired visual or hearing, learning difficulties, or health issues affecting the bones, kidneys or heart.
Diagnosing issues stemming from FASD is tricky. It’s not straightforward like a simple blood test or screening. It’s a time consuming process that requires seeing a range of health specialists, including a physician, psychologist, facial dysmorphologist and occupational therapist to look for telltale symptoms.
Researchers from the University of Southern California (USC) and Duke University in the US, and Queen’s University, Canada, have developed a tool that could make diagnosis easier.
First, a camera tracks the eye movements of children being tested for FASD as they watch short video clips. Children with FASD have abnormal eye behaviour as they tend to find it difficult to grasp visual concepts and focus. Convolutional neural networks analyse the eye movements to predict how likely the patient has FASD.
“Our method provides a comprehensive profile of distinct measures in domains including sensorimotor and visuospatial control, visual perception, attention, inhibition, working memory, academic functions, and brain structure,” the researchers wrote in a paper describing the tech, published in Frontiers in Neurology this month.
So this is how that terrifying killer AI will end us... by pushing us down hospital wait lists?READ MORE
The system was trained by analyzing data taken from eye movements of children diagnosed with FASD, coupled with psychometric test scores and brain scans of children affected by the disorder. 207 participants aged 5 to 18 were recruited for the project, 116 were healthy and 91 were affected by FASD. Data taken from 46 participants was used to test the performance of the system. The performance varied for different people, but the best result was 78.26 per cent accuracy.
So, sadly, not great, but certainly useful in clinical circumstances – and, don't forget, a first or early attempt at using neural networks for diagnosis.
The researchers hope that this could eventually help doctors perform tests quickly to help them with diagnosis and reach people that might not have as much access to healthcare.
“The new screening procedure only involves a camera and a computer screen, and can be applied to very young children,” said Chen Zhang, first author on the paper and postdoc at USC.
“It takes only 10 to 20 minutes and the cost should be affordable in most cases. The machine learning pipeline behind this gives out objective and consistent estimations in minutes.” ®