Super Responders: Predicting Expressive Language Gains Among Minimally Verbal Children with Autism Spectrum Disorder

Background: Research in autism spectrum disorders (ASD) has focused on developing efficacious interventions addressing core deficits of ASD. However, the heterogeneous nature of ASD complicates the development of such interventions. Due to this heterogeneity, it is unlikely that there is a one size fits all intervention. It is important to understand for whom an intervention works. Advancements in data analytics, in particular machine learning, provide new methods to identify subgroups among a given population, and can potentially help to identify for whom intervention works best. Of particular interest are minimally verbal individuals. A targeted social communication intervention known as JASPER (Joint Attention, Symbolic Play, Engagement, and Regulation) has shown promise for improving language outcomes among minimally verbal children with ASD and may provide the context to examine the question of for whom an intervention benefits.

Objectives: This study aims to develop a model predicting expressive language gains among minimally verbal, preschool aged children with ASD that received JASPER. This study also explores the utility of classification and regression tree (CART) analysis to identify relationships between child characteristics and gains in expressive language.

Methods: Secondary data analysis was conducted on a sample of 99 minimally verbal, preschool age children with ASD, collected from participants across five previous intervention studies. Expressive language gains were calculated using expressive language age equivalents from the Mullen Scales for Early Learning. Predictors for the analyses were taken from demographic and behavioral assessments completed prior to intervention. Using expressive language age equivalent change scores, 47% (n = 47) of the sample were identified as “super responders,” children that exceeded expressive language gains expected through maturation. To predict responder status, all initial predictors were used to generate conditional inference forests, a machine learning process for variable selection.

Results: Conditional inference forest results identified three variables to be fitted into the final model; play diversity, requesting gestures, and fine motor age equivalent. A final conditional inference tree was created, with play diversity being the only significant predictor of responder status. Participants with higher entry play diversity scores (above 23) predicted super response while lower scores (23 or below) predicted slow response. Overall model accuracy was 67%, with a specificity of 55% and sensitivity of 78%. For comparison, stepwise logistic regression was run, and play diversity significantly predicted responder status (χ² (1) = 10.686, p = .001). Receiver operating characteristic curves were generated to compare model performance, and comparison of area under the curves for the two models showed no statistical difference (p = .82).

Conclusions: Overall accuracy of the conditional inference tree was moderate, and performed similarly to the more traditional logistic regression analysis. However, the conditional inference tree provides a cutoff point that may provide clinical utility over the regression results. Both models identify play diversity as an important predictor of expressive language gains from JASPER, a play based social communication intervention. Additionally, the model appears to be more sensitive to identifying slow responders. The role of play diversity and expressive language gains in JASPER is discussed.