Profiles of autism characteristics in thirteen genetic syndromes: a machine learning approach

Abstract

Abstract Background Phenotypic studies have identified distinct patterns of autistic characteristics in genetic syndromes associated with intellectual disability (ID), leading to diagnostic uncertainty and compromised access to autism-related support. Previous research has tended to include small samples and diverse measures, which limits the generalisability of findings. In this study, we generated detailed profiles of autistic characteristics in a large sample of > 1500 individuals with rare genetic syndromes. Methods Profiles of autistic characteristics based on the Social Communication Questionnaire (SCQ) scores were generated for thirteen genetic syndrome groups (Angelman n = 154, Cri du Chat n = 75, Cornelia de Lange n = 199, fragile X n = 297, Prader–Willi n = 278, Lowe n = 89, Smith–Magenis n = 54, Down n = 135, Sotos n = 40, Rubinstein–Taybi n = 102, 1p36 deletion n = 41, tuberous sclerosis complex n = 83 and Phelan–McDermid n = 35 syndromes). It was hypothesised that each syndrome group would evidence a degree of specificity in autistic characteristics. To test this hypothesis, a classification algorithm via support vector machine (SVM) learning was applied to scores from over 1500 individuals diagnosed with one of the thirteen genetic syndromes and autistic individuals who did not have a known genetic syndrome (ASD; n = 254). Self-help skills were included as an additional predictor. Results Genetic syndromes were associated with different but overlapping autism-related profiles, indicated by the substantial accuracy of the entire, multiclass SVM model (55% correctly classified individuals). Syndrome groups such as Angelman, fragile X, Prader–Willi, Rubinstein–Taybi and Cornelia de Lange showed greater phenotypic specificity than groups such as Cri du Chat, Lowe, Smith–Magenis, tuberous sclerosis complex, Sotos and Phelan-McDermid. The inclusion of the ASD reference group and self-help skills did not change the model accuracy. Limitations The key limitations of our study include a cross-sectional design, reliance on a screening tool which focuses primarily on social communication skills and imbalanced sample size across syndrome groups. Conclusions These findings replicate and extend previous work, demonstrating syndrome-specific profiles of autistic characteristics in people with genetic syndromes compared to autistic individuals without a genetic syndrome. This work calls for greater precision of assessment of autistic characteristics in individuals with genetic syndromes associated with ID.

Keywords

• Autism
• Genetic syndromes
• SVM
• Machine learning
• Behavioural phenotype