Read Time: 7 minutes, 742 words

Using artificial intelligence to diagnose autism in the womb

After measuring certain characteristics of unborn babies and following them through the first years of life, researchers input this data into an algorithm that could be used to predict autism diagnosis in other unborn babies.


shadow
Image Credit: Pregnancy Ultrasound. Canva.com (Copyrighted)

Autism Spectrum Disorder (ASD) is one of the most common developmental disorders diagnosed today. It affects an estimated 75 million people worldwide, and the number of cases has been gradually increasing since studies began in the 1960s. The symptoms generally include communication deficits, sensory processing differences, and repetitive actions or behaviors. It also exists as a spectrum divided into 3 levels, each based on the severity of these symptoms.

Because of how important early intervention therapies can be for children with autism, scientists have been turning their attention to finding ways of identifying ASD as early as possible. Currently, a reliable diagnosis can be made as soon as 2 years old, but is oftentimes delayed until the child affected is much older.

Right now, there is no tool we can use to give an easy diagnosis — the way you would use X-rays to see broken bones or blood tests to diagnose diabetes. Autism can only be diagnosed based on behavior, making it more subjective. Children with less severe symptoms often go unnoticed by parents and physicians, and more severe cases can look similar to other developmental disorders.

Recently, however, French scientists at the University of Limoges have developed a computer program that uses fetal characteristics to predict if a child will be diagnosed with autism as soon as one day after birth. Having this early diagnosis would allow families with autistic children to begin communication, social, and sensory therapies that could prove crucial to the child’s development.

In order to create the program, the scientists followed a large group of expectant mothers at the University of Limoges hospital through their pregnancy and until their children were around 6 years old. During the pregnancies, several measurements were taken of the babies using ultrasounds. These measurements included the size of the baby’s head and stomach, the length of their thighs, and when the baby began to turn upside down in preparation for birth. Measurements were also taken during and after labor, such as the baby’s heart rate and temperature. Lastly, the scientists documented medical histories of the mothers, specifically if she had any history of autoimmune disease.

Once all the data was collected, the scientists then kept track of the children’s health for several years and documented which children were later diagnosed with autism. They were then able to enter this data into a program, to “train” the computer. An example of this type of program that most people have experienced is Spotify’s playlist generating algorithm. After a while of listening and saving certain songs, Spotify begins to make connections between all the music you listen to and predict what other songs you might also like.

In this study, the program can begin to independently understand associations between different fetal characteristics and outcomes. With these newly found connections, the program would theoretically be able to predict the likelihood of an autism diagnosis in a new set of infants. However, this particular program would not be able to determine the severity, or level of autism that the child would have.

In order to test its accuracy, the scientists gave the program the information already collected from the babies and their mothers. They then compared the prediction made by the computer to the actual autism status of the children. The program presented a 77% positive predictive value, which means that if the program predicted that the child had autism, there is a 77% chance that this would turn out to be true.

The results from the program are some of the most promising we’ve seen from diagnostic tools for young children. The M-CHAT, which is a behavior questionnaire given to most parents in the U.S. seeking autism diagnoses, only has a positive predictive value of around 48%. Being able to drastically shorten the observation time needed to reach a more reliable diagnosis would allow for early treatments that are most effective while the child is still very young.

The scientists that designed the program explain that it is by no means ready to be used to diagnose autism today. However, it proves that designing computer programs capable of doing so in the future is entirely possible. Previously, attempts at predicting autism in very young children have involved genetic testing and brain scans, both of which haven’t proven to be effective enough by themselves to be used for the general public. The researchers say that future tests will likely be some combination of genetic testing, brain scans, and help from programs like this one.

Study Information

Original study: Machine learning analysis of pregnancy data enables early identification of a subpopulation of newborns with ASD

Study was published on: March 25, 2021

Study author(s): Hugues Caly, Hamed Rabiei, Perrine Coste-Mazeau, Sebastien Hantz, Sophie Alain, Jean-Luc Eyraud, Thierry Chianea, Catherine Caly, David Makowski, Nouchine Hadjikhani, Eric Lemonnier & Yehezkel Ben-Ari

The study was done at: University of Limoges (FRANCE)

The study was funded by: The Neurochlore Company and BABiomedical

Raw data availability: Contact BABiomedical company (they own the code)

Featured image credit: Pregnancy Ultrasound. Canva.com (Copyrighted)

This summary was edited by: Osama Alian