Prelinguistic Predictors of 24-Month Expressive Language for Infants at High-Risk for ASD

Background: Language development for typically developing children occurs in the context of rich social interactions. Precursors of expressive language comprise a repertoire of nonverbal communicative behaviors, including gestures and facial expressions (Bates et al., 1987). Most children with autism spectrum disorder (ASD) experience difficulty with the acquisition and use of language, which has a significant impact on overall level of functioning. Moreover, the majority of high-risk infant siblings of children with ASD also exhibit language atypicalities, presumably indicative of the broader autism phenotype (e.g., Messinger et al., 2013). Identification of prelinguistic predictors of 24-month expressive language outcome could improve early detection of language delays and inform early intervention practices for high-risk siblings.

Objectives: The goals of this study are to: 1) identify prelinguistic predictors of 24-month expressive language outcome for high-risk and low-risk infants, and 2) compare linear regression and random forest methods for identification of predictor variables.

Methods: Participants include 39 high-risk and 56 low-risk infants seen for developmental and social-communication assessments at 12- and 24-months as part of a large, federally-funded longitudinal study examining risk and resilience in the first two years of life. The Communication and Symbolic Behavior Scales (CSBS-DP; Wetherby & Prizant, 2002) is a standardized assessment tool that examines communicative, social-affective, and symbolic abilities, resulting in seven clusters that make up three composites: Social (Emotion/Eye Gaze, Communication, Gestures), Speech (Sounds, Words), and Symbolic (Understanding, Object Use). The CSBS and Mullen Scales of Early Learning were administered at 12- and 24-months for each participant. Univariable and multivariable linear regressions were conducted using the 24-month Mullen Expressive Language T-score as the outcome. Possible predictor variables included: risk status, 12-month Mullen Expressive (EL) and Receptive (RL) Language subdomains, and all CSBS clusters. These results were compared to a random forest approach (a machine learning tool for regression) using the same predictor and outcome variables.

Results: High-risk infants scored significantly lower than low-risk infants on Mullen expressive language at 24-months (p<.001). The four most significant predictors resulting from the univariable regression were risk status (R²=.19,p<.001), CSBS Gestures (R²=.14,p<.001), CSBS Communication (R²=.11,p<.001), and CSBS Speech (R²=.11,p<.001). The multivariable regression revealed risk status, CSBS Speech, and CSBS Understanding as most predictive of expressive language (R²=.31, RMSE=11.36). Using random forests, the four most significant predictors were risk status, CSBS Speech, Mullen EL, and Mullen RL, (RMSE=10.96). A Spearman correlation indicated moderate agreement between the regression and random forest approach with regard to significant predictors (r_s=.53).

Conclusions: Results suggest several prelinguistic factors that significantly predict 24-month expressive language in high-risk and low-risk infants. There was moderate agreement between the linear regression and random forest approaches. Risk status was by far the most predictive of expressive language at 2-years. Additional significant predictors included gesture use, receptive language, very early speech (i.e., first words), and the overall frequency and function of communication. These results have significant implications for the development of very early intervention strategies that target all forms of communication, including gestures, first words, and receptive language, to promote language development in high-risk infants.