Use of machine learning to identify at birth babies at risk of developing Autism Spectrum Disorders

Autism Spectrum Disorders (ASDs) are associated with socialization and communication problems as well as restricted interests and stereotypic behavior. The diagnosis is usually obtained around 3-5 years old, yet we know that early psycho-educative methods enable to attenuate the severity of ASDs in particular if they are utilized early on. Since ASDs are “born” in the womb, we have tested the hypothesis that analysis of maternity data might enable to identify babies at risk of ASDs already at birth. 

We collected pregnancy data from the first trimester to birth of 65 children born in one maternity and diagnosed with ASD 2-5 years later. This was compared to 189 babies born in the same maternity and conditions, and who did not have ASD. Due to the large number of pregnancy parameters routinely available in French maternities (circa 116), we developed a supervised Machine Learning algorithm that will enable to analyze and compare files of babies with and without ASD, but also to rank the most impacting parameters that underlie the prognostic. 

Our results show that it is possible to identify 96% of babies that will not have an ASD diagnosis and 41% of babies who will, with a precision of 77%. Impacting parameters are very diverse, many of which have currently no understood link with ASD. Our aims now are to enlarge this trial in a few maternities to optimize the analysis in order to both increase the detection rate and also the reliability of impacting parameters so as to better understand the mechanisms leading to ASD.

An important note of caution: this is a prognostic and by no means a diagnostic, and its reliability is conditioned by the inclusion of larger populations with several hundreds of babies. Also, no single parameter allows to reach any conclusion, a large number of parameters is needed stressing the need for larger studies. Nevertheless, this is a first important step that if confirmed might facilitate an early use of psycho-educative techniques in order to ameliorate the well-being of children with ASD.

Yehezkel Ben-Ari CEO Neurochlore

References :

  • Caly H, Rabiei H, Coste-Mazeu P, Hantz S, Alain S, Eyraud J-L, Chianea T, Caly C, Makowski D, Hadjikhani N, Lemonnier E, Ben-Ari Y. Machine learning analysis of pregnancy data enables early identification of a subpopulation of newborns with ASD. 2021. Scientific Reports in press.

Découvrez tous les sites de Yehezkel Ben-Ari/Discover all Yehezkel Ben-Ari websites

Le Blog de Ben-Ari
B&A Biomedical
B&A Oncomedical