Integration of Longitudinal Cross-Domain Measures of Symptoms, Developmental Level and Adaptive Functioning to Predict Autism at 3 Years in High-Risk Infants
Objectives: To characterise early development of ASD, we integrated information across multiple behavioural and developmental domains and multiple time-points. Additionally, we used a machine learning approach to improve individual classification of ASD among high-risk siblings (HR) at an early age from different combinations of measures.
Methods: We used data of the Mullen Scales of Early Learning (MSEL), Vineland Adaptive Behavior Scales (VABS) and Autism Observation Scale for Infants (AOSI) between 8 and 36 months in a cohort of 71 low-risk controls (LR) and 161 HR siblings. Clinical outcome of HR siblings at 36 months (HR-Typical, HR-Atypical, HR-ASD) was established by expert clinical researchers. First, we examined whether LR and HR clinical outcome groups showed differences in cross-sectional measures at 8 and 14 months, and in developmental trajectories of behavioural and developmental measures between 8 and 36 months. Second, different combinations of the same measures at 8 and 14 months were integrated into a Least-Square Support Vector Machine (LS-SVM) classifier to discriminate between HR-ASD and both HR-Typical and HR-Atypical.
Results: We observed clear but small size group effects for Mullen and Vineland scores at 8 and 14 months, and larger group effects at 24 and 36 months (Figure 1). Group differentiation was found from 8 months, except for MSEL visual reception, receptive and expressive language scores, and VABS motor and social scores, showing increasing differentiation of groups over time. Overall, LR and HR-Typical showed higher developmental level and functioning, and lower ASD symptoms than HR-Atypical and HR-ASD. Individual classification of ASD clinical outcome was possible with moderate accuracy using VABS daily living scores at 14 months (Area Under the Curve, AUC: 71.3%; 95% Confidence Interval, CI: [55.6, 85.1]). The integration of measures from different domains did not significantly improve classification (Figure 2).
Conclusions: This study extends previous high-risk studies on early markers for ASD by integrating information from multiple measures and multiple time-points; testing models for the individual classification of ASD clinical outcome; and focusing on prediction at a younger age. Our results provide further evidence of the high inter-individual and intra-individual heterogeneity of ASD, which makes it difficult to predict the later development of the disorder from clinical manifestations at an early age. Further investigation is needed to understand the interplay of different domains in the first years of life leading to an ASD outcome, and the combination of measures from different domains can be extended to include more biological data to examine whether it would further improve predictive power.