Early developmental trajectory phenotypes for risk stratification of autism spectrum disorder in very preterm infants: a machine learning approach.

The model was trained and tested in one regional cohort of very preterm children in southern Taiwan, so it may not apply to other health systems or populations. Of children the model predicted to have ASD, 24.7% had ASD, so many children predicted as having ASD did not have ASD. The model identified 64.2% of children with ASD, so it did not identify all children later diagnosed with ASD. Infants with congenital syndromes, brain malformation, blindness, hearing impairment, or severe cerebral palsy were excluded, so it says little about those children. ASD diagnosis was made at age 5, later than the commonly recommended 18 to 24 month screening window, so the model's timing may not match routine early screening. 153 children were lost to follow-up and had lower family socioeconomic status and maternal education, so the analysed group may not represent all survivors. The prediction performance was modest, with the best classifier correctly classifying 71.8% of children.

The study was funded by the Ministry of Science and Technology, Taiwan and National Cheng Kung University Hospital.

Do not read the model as a way to diagnose autism. Of children it predicted to have autism, 24.7% had autism, while of children it predicted not to have autism, 93.6% did not have autism.