Early developmental trajectory phenotypes for risk stratification of autism spectrum disorder in very preterm infants: a machine learning approach
BSID-III trajectories from 6–24 months spot very preterm infants who are very unlikely to develop autism, letting you focus early-intervention resources on the rest.
01Research in Context
What this study did
The team tracked 441 very preterm babies with the BSID-III every three months from six to 24 months corrected age.
They fed the scores into a machine-learning model to see if early paths of cognitive, language, and fine-motor growth could flag later autism diagnosis.
Doctors confirmed ASD at 36 months with the ADOS-2.
What they found
Kids who later got an ASD label had flatter or falling curves on receptive language, expressive language, and fine motor.
The model’s “low-risk” rule was right 93.6 % of the time, so you can trust a negative flag.
Overall accuracy stayed modest, so some high-risk alerts will be false alarms.
How this fits with other research
Yaari et al. (2018) saw the same flat lines with the Mullen, but they only warned of general delay; Chen’s work sharpens the warning to ASD risk.
Velikos et al. (2015) first showed BSID-III dips at 12 months in preterm infants; Chen adds machine learning and longer follow-up to turn that snapshot into a forecast.
Seiverling et al. (2018) also used latent classes in toddlers with language delay and linked a low verbal track to later ASD; Chen proves the pattern starts even earlier in preterm babies.
Why it matters
You now have a clear, low-cost screen: run the BSID-III on schedule, plug the scores into the free model, and you can tell neonatal follow-up teams which families need earlier ABA referral and which can breathe easier.
Want CEUs on This Topic?
The ABA Clubhouse has 60+ free CEUs — live every Wednesday. Ethics, supervision & clinical topics.
Join Free →Enter the last two BSID-III scores of any preterm baby into the open-source trajectory calculator; if the child lands in the low-risk class, schedule routine monitoring instead of automatic ASD referral.
02At a glance
03Original abstract
Very preterm infants are at elevated risk for autism spectrum disorder (ASD), though early identification is challenging due to overlapping neurodevelopmental disorders. While the Bayley Scales of Infant and Toddler Development (BSID) is widely used for follow-up, it remains unclear whether domain-specific developmental trajectories—such as cognition, receptive and expressive communication, and fine and gross motor function assessed by the BSID, Third Edition (BSID-III)—can support the development of a prediction model for ASD risk by preschool age in this population. This population-based multicenter cohort study included infants born < 32 weeks’ gestation in 2011–2018. Neurodevelopment was assessed at 6, 12, and 24 months using domain-specific BSID-III scaled scores. ASD diagnosis was determined at age 5 years using the Autism Diagnostic Observation Schedule and the Autism Diagnostic Interview–Revised. Infants with congenital anomalies or severe sensorimotor impairments were excluded. Developmental trajectories were analyzed using locally estimated scatterplot smoothing. Six machine learning algorithms were used to evaluate ASD prediction based on neonatal risks and longitudinal domain-specific scaled score data. Of 583 very-preterm infants, 75 (12.9%) were diagnosed with ASD at preschool age. Infants later diagnosed with ASD exhibited persistently lower cognitive scores across the first two years of life (p < 0.05) and significantly slower development in receptive and expressive communication and fine motor skills during the second year (p < 0.0001 by 24 months) than infants without ASD. Gross motor trajectories did not differ significantly between groups. Integrating neurodevelopmental trajectories up to 24 months with neonatal risk factors improved prediction performance. The Support Vector Machine model yielded 71.8% accuracy (Area Under the Curve 0.69), with sensitivity of 64.2%, specificity of 72.9%, positive predictive value of 24.7%, and negative predictive value of 93.6%. Although the model shows promise in identifying infants at low likelihood of ASD, its overall predictive performance remains modest. The model was developed in a single regional cohort, potentially limiting generalizability. Preterm infants later diagnosed with ASD exhibit distinct, domain-specific developmental trajectories. The model’s high negative predictive value suggests that developmental trajectory phenotypes may support early risk stratification by identifying infants at low likelihood of ASD. The online version contains supplementary material available at 10.1186/s13229-025-00692-y.
Molecular Autism, 2025 · doi:10.1186/s13229-025-00692-y