Predictive modeling of adaptive behavior trajectories in autism: insights from a clinical cohort study

Research aimed at understanding how baseline clinical and demographic characteristics influence outcomes over time is critically important to inform individualized therapeutic programs for children with neurodevelopmental differences. This study characterizes adaptive behavior trajectories in children receiving medical and behavioral therapy within a network of care centers with a shared data-gathering mechanism for intake and longitudinal assessments. We then take the further step of utilizing intake data to develop machine-learning models which predict differences in those trajectories. Specifically, we evaluated data from 1225 autistic children, aged 20-90 months, using latent class growth mixture modeling (LCGMM) with scores on the Vineland Adaptive Behavior Scales, 3 rd Edition, as the primary outcome measure. The LCGMM analysis revealed two distinct clusters of adaptive behavior trajectories. The “Less Impairment/Improving Trajectory” group (≥66% of the sample) exhibited greater developmental change in adaptive behavior, while the “Higher Impairment/Stable Trajectory” group (≤33% of the sample) showed little change over time relative to age-matched normative data. For a subset of 729 children, we used machine learning algorithms to forecast adaptive behavior trajectories using clinical and sociodemographic data collected at the initial assessment, comparing elastic net GLM, support vector machine, and random forest. The best-performing random forest model predicted adaptive behavior trajectory with an accuracy rate of 77%. The strongest predictors in our model were socioeconomic status, history of developmental regression, child temperament, paternal age at the time of the child’s birth, baseline autism symptom severity, parent concerns about development, presence of ADHD symptoms, and parent concerns about mood. Notably, the inclusion of cumulative hours of applied behavioral analysis and developmental therapies in the machine learning models did not yield significant changes in performance metrics, indicating that increased therapy hours did not predict greater improvement. These findings extend our understanding of adaptive behavior development in autistic children and underscore the value of gathering comprehensive patient information at intake to tailor clinical care.